[ { "path": ".gitignore", "content": "*.o\ncompareTopKAlgorithms\n" }, { "path": "LICENSE", "content": "MIT License\n\nCopyright (c) 2016 Anil Shanbhag\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n" }, { "path": "Makefile", "content": "OS_SIZE = $(shell uname -m | sed -e \"s/i.86/32/\" -e \"s/x86_64/64/\")\n\nCUDA_PATH ?= /usr/local/cuda-10.0\nCUDA_INC_PATH ?= $(CUDA_PATH)/include\nCUDA_BIN_PATH ?= $(CUDA_PATH)/bin\n\nifeq ($(OS_SIZE),32)\n CUDA_LIB_PATH ?= $(CUDA_PATH)/lib\nelse\n CUDA_LIB_PATH ?= $(CUDA_PATH)/lib64\nendif\n\nNVCC ?= $(CUDA_BIN_PATH)/nvcc\n\n# Make sure to choose the right SM for your device\nGENCODE_SM70 := -gencode arch=compute_70,code=sm_70\nGENCODE_SM50 := -gencode arch=compute_50,code=sm_50\nGENCODE_SM35 := -gencode arch=compute_35,code=sm_35\nGENCODE_FLAGS := $(GENCODE_SM50)\n\nLDFLAGS := -lcudart -lrt -lcurand -lm\nCFLAGS := -O3 -lineinfo -Xptxas=\"-dlcm=ca -v\"\nCUB_DIR = ./external/cub/\nSOURCE_DIR = ./src/\nTEST_DIR = ./test/\nINCLUDES := -I$(CUB_DIR) -I$(SOURCE_DIR) -I$(TEST_DIR)\n\n#obj/%.o: src/%.cu $(DEPS)\n#\t$(NVCC) $(CFLAGS) -I. $(INCLUDES) -g $(GENCODE_FLAGS) $< -o $@\n\ncompareTopKAlgorithms: test/compareTopKAlgorithms.cu src/bitonicTopK.cuh src/radixSelectTopK.cuh src/sortTopK.cuh\n\t$(NVCC) $(CFLAGS) $(INCLUDES) test/compareTopKAlgorithms.cu $(LDFLAGS) -o compareTopKAlgorithms\n\nclean:\n\trm -rf *.o compareTopKAlgorithms\n" }, { "path": "README.md", "content": "GPU-TopK\n========\n\nGPU-TopK implements efficient top-k runtimes for GPUs. The specific problem solved is given a array of entries (key-only or key-value), find the top-k entries based on value of key. \n\nThe package implements the following routines:\n\n* Bitonic Top-K: reduction algorithm based on bitonic sort\n* Radix Select Top-K: reduction of radix sort to compute top-k\n* Sort Top-K: sorts the entire array and selects the top-k entries\n\nFor full details of the algorithms, see our [paper](http://anilshanbhag.in/static/papers/gputopk_sigmod18.pdf)\n\n```\n@inproceedings{shanbhag2018efficient,\n title={Efficient Top-K query processing on massively parallel hardware},\n author={Shanbhag, Anil and Pirk, Holger and Madden, Samuel},\n booktitle={Proceedings of the 2018 International Conference on Management of Data},\n pages={1557--1570},\n year={2018},\n organization={ACM}\n}\n```\n\nUsage\n----\n\nThe individual implementations can used directly as standalone header files. For example, to use `RadixSelectTopK`, \n\n```\n#include \"radixSelectTopK.cuh\"\n...\nfloat* d_keys_in; // device pointer to the array\nuint num_items; // number of entries in the array\nuint k; // \nfloat* d_keys_out;// device pointer to the result array (needs to be of size atleast k)\nCachingDeviceAllocator& g_allocator // Cub memory allocator \nradixSelectTopK(d_keys_in, num_items, k, d_keys_out, g_allocator);\n```\n\nWe have implemented a testutil called `compareTopKAlgorithms` that can be used to benchmark the different algorithms. The testutil lets you test performance of the algorithms across standard data types and certain pre-defined distributions. To run the testutil:\n\n```\n# Edit Makefile to select the right Gencode for your GPU\n# For example: for V100 GPU set GENCODE_FLAGS to use GENCODE_SM70\n\nmake compareTopKAlgorithms\n./compareTopKAlgorithms\n```\n\nHere is an example tracelog:\n```\n$ ./compareTopKAlgorithms\nPlease enter the type of value you want to test:\n1-float\n2-double\n3-uint\n1\nPlease enter distribution type: 0\nPlease enter K: 32\nPlease enter number of tests to run per K: 3\nPlease enter start power (dataset size starts at 2^start)(max val: 29): 29\nPlease enter stop power (dataset size stops at 2^stop)(max val: 29): 29\nNOW STARTING A NEW K\n\nThe distribution is: UNIFORM FLOATS\nRunning test 1 of 3 for size: 536870912 and k: 32\nTESTING: 0 Sort\nTESTING: 2 Bitonic TopK\nTESTING: 1 Radix Select\nRunning test 2 of 3 for size: 536870912 and k: 32\nTESTING: 2 Bitonic TopK\nTESTING: 0 Sort\nTESTING: 1 Radix Select\nRunning test 3 of 3 for size: 536870912 and k: 32\nTESTING: 0 Sort\nTESTING: 1 Radix Select\nTESTING: 2 Bitonic TopK\n\n\nSort averaged: 219.273071 ms\nRadix Select averaged: 132.391724 ms\nBitonic TopK averaged: 134.959854 ms\nSort minimum: 215.583801 ms\nRadix Select minimum: 63.751999 ms\nBitonic TopK minimum: 28.718592 ms\nSort won 0 times\nRadix Select won 1 times\nBitonic TopK won 2 times\n```\n\nFor benchmarking, it is advisable to run the suite more than once inorder to have GPU warmed up. To see the full set of distributions implemented, check test/generateProblems.cuh. \n\nKnown Issues\n-----------\n1. Currently works for key-only, we will add key-value soon\n2. Works for data set size upto 2^29. This is due to inherent limitations of maximum array size on GPUs.\n3. Tested to work well on Nvidia Maxwell architecture and upwards (may not work on K80 GPU - you are welcome to submit a patch). \n4. BitonicTopK works only for K<=256, if you are testing K > 256, make sure to comment BitonicTopK\n" }, { "path": "external/cub/.cproject", "content": "\r\n\r\n\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\t\r\n\t\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n\t\r\n\t\t\r\n\t\r\n\t\r\n\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n\t\r\n\t\r\n\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n\t\r\n\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\t\r\n\t\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n\r\n" }, { "path": "external/cub/.project", "content": "\r\n\r\n\tGIT_CUB\r\n\t\r\n\t\r\n\t\r\n\t\r\n\t\t\r\n\t\t\torg.eclipse.cdt.managedbuilder.core.genmakebuilder\r\n\t\t\tclean,full,incremental,\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\torg.eclipse.cdt.managedbuilder.core.ScannerConfigBuilder\r\n\t\t\tfull,incremental,\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n\t\r\n\t\torg.eclipse.cdt.core.cnature\r\n\t\torg.eclipse.cdt.managedbuilder.core.managedBuildNature\r\n\t\torg.eclipse.cdt.managedbuilder.core.ScannerConfigNature\r\n\t\torg.eclipse.cdt.core.ccnature\r\n\t\r\n\r\n" }, { "path": "external/cub/.settings/.gitignore", "content": "/language.settings.xml\n" }, { "path": "external/cub/.settings/org.eclipse.cdt.codan.core.prefs", "content": "eclipse.preferences.version=1\r\norg.eclipse.cdt.codan.checkers.errnoreturn=Warning\r\norg.eclipse.cdt.codan.checkers.errnoreturn.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},implicit\\=>false}\r\norg.eclipse.cdt.codan.checkers.errreturnvalue=Error\r\norg.eclipse.cdt.codan.checkers.errreturnvalue.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.checkers.nocommentinside=-Error\r\norg.eclipse.cdt.codan.checkers.nocommentinside.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.checkers.nolinecomment=-Error\r\norg.eclipse.cdt.codan.checkers.nolinecomment.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.checkers.noreturn=Error\r\norg.eclipse.cdt.codan.checkers.noreturn.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},implicit\\=>false}\r\norg.eclipse.cdt.codan.internal.checkers.AbstractClassCreation=Error\r\norg.eclipse.cdt.codan.internal.checkers.AbstractClassCreation.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.AmbiguousProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.AmbiguousProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.AssignmentInConditionProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.AssignmentInConditionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.AssignmentToItselfProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.AssignmentToItselfProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.CaseBreakProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.CaseBreakProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},no_break_comment\\=>\"no break\",last_case_param\\=>true,empty_case_param\\=>false}\r\norg.eclipse.cdt.codan.internal.checkers.CatchByReference=Warning\r\norg.eclipse.cdt.codan.internal.checkers.CatchByReference.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},unknown\\=>false,exceptions\\=>()}\r\norg.eclipse.cdt.codan.internal.checkers.CircularReferenceProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.CircularReferenceProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.ClassMembersInitialization=Warning\r\norg.eclipse.cdt.codan.internal.checkers.ClassMembersInitialization.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},skip\\=>true}\r\norg.eclipse.cdt.codan.internal.checkers.FieldResolutionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.FieldResolutionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.FunctionResolutionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.FunctionResolutionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.InvalidArguments=Error\r\norg.eclipse.cdt.codan.internal.checkers.InvalidArguments.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.InvalidTemplateArgumentsProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.InvalidTemplateArgumentsProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.LabelStatementNotFoundProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.LabelStatementNotFoundProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.MemberDeclarationNotFoundProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.MemberDeclarationNotFoundProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.MethodResolutionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.MethodResolutionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.NamingConventionFunctionChecker=-Info\r\norg.eclipse.cdt.codan.internal.checkers.NamingConventionFunctionChecker.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},pattern\\=>\"^[a-z]\",macro\\=>true,exceptions\\=>()}\r\norg.eclipse.cdt.codan.internal.checkers.NonVirtualDestructorProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.NonVirtualDestructorProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.OverloadProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.OverloadProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.RedeclarationProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.RedeclarationProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.RedefinitionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.RedefinitionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.ReturnStyleProblem=-Warning\r\norg.eclipse.cdt.codan.internal.checkers.ReturnStyleProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.ScanfFormatStringSecurityProblem=-Warning\r\norg.eclipse.cdt.codan.internal.checkers.ScanfFormatStringSecurityProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.StatementHasNoEffectProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.StatementHasNoEffectProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},macro\\=>true,exceptions\\=>()}\r\norg.eclipse.cdt.codan.internal.checkers.SuggestedParenthesisProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.SuggestedParenthesisProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},paramNot\\=>false}\r\norg.eclipse.cdt.codan.internal.checkers.SuspiciousSemicolonProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.SuspiciousSemicolonProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},else\\=>false,afterelse\\=>false}\r\norg.eclipse.cdt.codan.internal.checkers.TypeResolutionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.TypeResolutionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\norg.eclipse.cdt.codan.internal.checkers.UnusedFunctionDeclarationProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.UnusedFunctionDeclarationProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},macro\\=>true}\r\norg.eclipse.cdt.codan.internal.checkers.UnusedStaticFunctionProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.UnusedStaticFunctionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},macro\\=>true}\r\norg.eclipse.cdt.codan.internal.checkers.UnusedVariableDeclarationProblem=Warning\r\norg.eclipse.cdt.codan.internal.checkers.UnusedVariableDeclarationProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true},macro\\=>true,exceptions\\=>(\"@(\\#)\",\"$Id\")}\r\norg.eclipse.cdt.codan.internal.checkers.VariableResolutionProblem=Error\r\norg.eclipse.cdt.codan.internal.checkers.VariableResolutionProblem.params={launchModes\\=>{RUN_ON_FULL_BUILD\\=>true,RUN_ON_INC_BUILD\\=>true,RUN_ON_FILE_OPEN\\=>false,RUN_ON_FILE_SAVE\\=>false,RUN_AS_YOU_TYPE\\=>true,RUN_ON_DEMAND\\=>true}}\r\nuseParentScope=false\r\n" }, { "path": "external/cub/.settings/org.eclipse.cdt.core.prefs", "content": "eclipse.preferences.version=1\r\nindexer/indexAllFiles=true\r\nindexer/indexAllHeaderVersions=false\r\nindexer/indexAllVersionsSpecificHeaders=\r\nindexer/indexOnOpen=false\r\nindexer/indexUnusedHeadersWithAlternateLang=false\r\nindexer/indexUnusedHeadersWithDefaultLang=true\r\nindexer/indexerId=org.eclipse.cdt.core.fastIndexer\r\nindexer/skipFilesLargerThanMB=8\r\nindexer/skipImplicitReferences=false\r\nindexer/skipIncludedFilesLargerThanMB=16\r\nindexer/skipMacroReferences=false\r\nindexer/skipReferences=false\r\nindexer/skipTypeReferences=false\r\nindexer/useHeuristicIncludeResolution=true\r\norg.eclipse.cdt.core.formatter.alignment_for_arguments_in_method_invocation=16\r\norg.eclipse.cdt.core.formatter.alignment_for_assignment=16\r\norg.eclipse.cdt.core.formatter.alignment_for_base_clause_in_type_declaration=48\r\norg.eclipse.cdt.core.formatter.alignment_for_binary_expression=16\r\norg.eclipse.cdt.core.formatter.alignment_for_compact_if=0\r\norg.eclipse.cdt.core.formatter.alignment_for_conditional_expression=48\r\norg.eclipse.cdt.core.formatter.alignment_for_conditional_expression_chain=18\r\norg.eclipse.cdt.core.formatter.alignment_for_constructor_initializer_list=0\r\norg.eclipse.cdt.core.formatter.alignment_for_declarator_list=16\r\norg.eclipse.cdt.core.formatter.alignment_for_enumerator_list=48\r\norg.eclipse.cdt.core.formatter.alignment_for_expression_list=0\r\norg.eclipse.cdt.core.formatter.alignment_for_expressions_in_array_initializer=16\r\norg.eclipse.cdt.core.formatter.alignment_for_member_access=0\r\norg.eclipse.cdt.core.formatter.alignment_for_overloaded_left_shift_chain=16\r\norg.eclipse.cdt.core.formatter.alignment_for_parameters_in_method_declaration=48\r\norg.eclipse.cdt.core.formatter.alignment_for_throws_clause_in_method_declaration=48\r\norg.eclipse.cdt.core.formatter.brace_position_for_array_initializer=next_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_block=next_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_block_in_case=end_of_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_method_declaration=next_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_namespace_declaration=end_of_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_switch=end_of_line\r\norg.eclipse.cdt.core.formatter.brace_position_for_type_declaration=next_line\r\norg.eclipse.cdt.core.formatter.comment.min_distance_between_code_and_line_comment=1\r\norg.eclipse.cdt.core.formatter.comment.never_indent_line_comments_on_first_column=true\r\norg.eclipse.cdt.core.formatter.comment.preserve_white_space_between_code_and_line_comments=true\r\norg.eclipse.cdt.core.formatter.compact_else_if=true\r\norg.eclipse.cdt.core.formatter.continuation_indentation=1\r\norg.eclipse.cdt.core.formatter.continuation_indentation_for_array_initializer=1\r\norg.eclipse.cdt.core.formatter.format_guardian_clause_on_one_line=false\r\norg.eclipse.cdt.core.formatter.indent_access_specifier_compare_to_type_header=false\r\norg.eclipse.cdt.core.formatter.indent_access_specifier_extra_spaces=0\r\norg.eclipse.cdt.core.formatter.indent_body_declarations_compare_to_access_specifier=true\r\norg.eclipse.cdt.core.formatter.indent_body_declarations_compare_to_namespace_header=false\r\norg.eclipse.cdt.core.formatter.indent_breaks_compare_to_cases=true\r\norg.eclipse.cdt.core.formatter.indent_declaration_compare_to_template_header=false\r\norg.eclipse.cdt.core.formatter.indent_empty_lines=false\r\norg.eclipse.cdt.core.formatter.indent_statements_compare_to_block=true\r\norg.eclipse.cdt.core.formatter.indent_statements_compare_to_body=true\r\norg.eclipse.cdt.core.formatter.indent_switchstatements_compare_to_cases=true\r\norg.eclipse.cdt.core.formatter.indent_switchstatements_compare_to_switch=false\r\norg.eclipse.cdt.core.formatter.indentation.size=4\r\norg.eclipse.cdt.core.formatter.insert_new_line_after_opening_brace_in_array_initializer=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_after_template_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_at_end_of_file_if_missing=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_catch_in_try_statement=insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_closing_brace_in_array_initializer=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_colon_in_constructor_initializer_list=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_else_in_if_statement=insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_identifier_in_function_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_before_while_in_do_statement=do not insert\r\norg.eclipse.cdt.core.formatter.insert_new_line_in_empty_block=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_assignment_operator=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_binary_operator=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_closing_angle_bracket_in_template_arguments=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_closing_angle_bracket_in_template_parameters=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_closing_brace_in_block=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_closing_paren_in_cast=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_colon_in_base_clause=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_colon_in_case=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_colon_in_conditional=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_colon_in_labeled_statement=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_array_initializer=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_base_types=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_declarator_list=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_enum_declarations=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_expression_list=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_method_declaration_parameters=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_method_declaration_throws=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_method_invocation_arguments=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_template_arguments=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_comma_in_template_parameters=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_angle_bracket_in_template_arguments=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_angle_bracket_in_template_parameters=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_brace_in_array_initializer=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_bracket=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_cast=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_catch=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_exception_specification=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_for=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_if=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_method_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_method_invocation=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_parenthesized_expression=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_switch=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_opening_paren_in_while=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_postfix_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_prefix_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_question_in_conditional=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_semicolon_in_for=insert\r\norg.eclipse.cdt.core.formatter.insert_space_after_unary_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_assignment_operator=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_binary_operator=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_angle_bracket_in_template_arguments=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_angle_bracket_in_template_parameters=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_brace_in_array_initializer=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_bracket=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_cast=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_catch=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_exception_specification=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_for=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_if=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_method_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_method_invocation=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_parenthesized_expression=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_switch=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_closing_paren_in_while=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_colon_in_base_clause=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_colon_in_case=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_colon_in_conditional=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_colon_in_default=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_colon_in_labeled_statement=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_array_initializer=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_base_types=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_declarator_list=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_enum_declarations=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_expression_list=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_method_declaration_parameters=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_method_declaration_throws=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_method_invocation_arguments=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_template_arguments=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_comma_in_template_parameters=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_angle_bracket_in_template_arguments=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_angle_bracket_in_template_parameters=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_array_initializer=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_block=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_method_declaration=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_namespace_declaration=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_switch=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_brace_in_type_declaration=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_bracket=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_catch=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_exception_specification=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_for=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_if=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_method_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_method_invocation=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_parenthesized_expression=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_switch=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_opening_paren_in_while=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_postfix_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_prefix_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_question_in_conditional=insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_semicolon=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_semicolon_in_for=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_before_unary_operator=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_between_empty_braces_in_array_initializer=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_between_empty_brackets=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_between_empty_parens_in_exception_specification=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_between_empty_parens_in_method_declaration=do not insert\r\norg.eclipse.cdt.core.formatter.insert_space_between_empty_parens_in_method_invocation=do not insert\r\norg.eclipse.cdt.core.formatter.join_wrapped_lines=true\r\norg.eclipse.cdt.core.formatter.keep_else_statement_on_same_line=false\r\norg.eclipse.cdt.core.formatter.keep_empty_array_initializer_on_one_line=false\r\norg.eclipse.cdt.core.formatter.keep_imple_if_on_one_line=true\r\norg.eclipse.cdt.core.formatter.keep_then_statement_on_same_line=false\r\norg.eclipse.cdt.core.formatter.lineSplit=80\r\norg.eclipse.cdt.core.formatter.number_of_empty_lines_to_preserve=1\r\norg.eclipse.cdt.core.formatter.put_empty_statement_on_new_line=true\r\norg.eclipse.cdt.core.formatter.tabulation.char=space\r\norg.eclipse.cdt.core.formatter.tabulation.size=4\r\norg.eclipse.cdt.core.formatter.use_tabs_only_for_leading_indentations=false\r\n" }, { "path": "external/cub/.settings/org.eclipse.cdt.ui.prefs", "content": "eclipse.preferences.version=1\r\nformatter_profile=_B40C\r\nformatter_settings_version=1\r\n" }, { "path": "external/cub/.settings/org.eclipse.core.runtime.prefs", "content": "content-types/enabled=true\r\ncontent-types/org.eclipse.cdt.core.cxxHeader/file-extensions=cuh\r\ncontent-types/org.eclipse.cdt.core.cxxSource/file-extensions=cu\r\neclipse.preferences.version=1\r\n" }, { "path": "external/cub/CHANGE_LOG.TXT", "content": "1.7.4 09/20/2017\n - Bug fixes: \n - Issue #114: Can't pair non-trivially-constructible values in radix sort\n - Issue #115: WarpReduce segmented reduction broken in CUDA 9 for logical warp sizes < 32 \n \t\t \n//-----------------------------------------------------------------------------\n\n1.7.3 08/28/2017\n - Bug fixes: \n - Issue #110: DeviceHistogram null-pointer exception bug for iterator inputs\n \t\t \n//-----------------------------------------------------------------------------\n\n1.7.2 08/26/2017\n - Bug fixes: \n - Issue #104: Device-wide reduction is now \"run-to-run\" deterministic for \n pseudo-associative reduction operators (like floating point addition)\n \t\t \n//-----------------------------------------------------------------------------\n\n1.7.1 08/18/2017\n - Updated Volta radix sorting tuning policies \n - Bug fixes: \n - Issue #104 (uint64_t warp-reduce broken for cub 1.7.0 on cuda 8 and older)\n - Issue #103 (Can't mix Thrust 9.0 and CUB)\n - Issue #102 (CUB pulls in windows.h which defines min/max macros that conflict with std::min/std::max)\n - Issue #99 (Radix sorting crashes NVCC on Windows 10 for SM52)\n - Issue #98 (cuda-memcheck: --tool initcheck failed with lineOfSight)\n - Issue #94 (Git clone size)\n - Issue #93 (accept iterators for segment offsets)\n - Issue #87 (CUB uses anonymous unions which is not valid C++)\n - Issue #44 (Check for C++ 11 should be changed that Visual Studio 2013 is also recognized as C++ 11 capable)\n \t\t \n//-----------------------------------------------------------------------------\n\n1.7.0 06/07/2017\n - Compatible with CUDA9 and SM7.x (Volta) independent thread scheduling \n - API change: remove cub::WarpAll() and cub::WarpAny(). These functions served to \n emulate __all and __any functionality for SM1.x devices, which did not have those \n operations. However, the SM1.x devices are now deprecated in CUDA, and the \n interfaces of the these two functions are now lacking the lane-mask needed \n for collectives to run on Volta SMs having independent thread scheduling. \n - Bug fixes: \n - Issue #86 Incorrect results with ReduceByKey\n \t\t \n//-----------------------------------------------------------------------------\n\n1.6.4 12/06/2016\n - Updated sm_5x, sm_6x tuning policies for radix sorting (3.5B and 3.4B \n 32b keys/s on TitanX and GTX 1080, respectively)\n - Bug fixes: \n - Restore fence work-around for scan (reduce-by-key, etc.) hangs \n in CUDA 8.5\n - Issue 65: DeviceSegmentedRadixSort should allow inputs to have \n pointer-to-const type \n - Mollify Clang device-side warnings\n - Remove out-dated VC project files\n \t\t \n//-----------------------------------------------------------------------------\n\n1.6.3 11/20/2016\n - API change: BlockLoad and BlockStore are now templated by the local\n data type, instead of the Iterator type. This allows for output iterators\n having \\p void as their \\p value_type (e.g., discard iterators).\n - Updated GP100 tuning policies for radix sorting (6.2B 32b keys/s)\n - Bug fixes: \n - Issue #74: Warpreduce executes reduction operator for out-of-bounds items\n - Issue #72 (cub:InequalityWrapper::operator() should be non-const)\n - Issue #71 (KeyVairPair won't work if Key has non-trivial ctor)\n\t\t- Issue #70 1.5.3 breaks BlockScan API. Retroactively reversioned\n\t\t from v1.5.3 -> v1.6 to appropriately indicate API change.\n\t\t- Issue #69 cub::BlockStore::Store doesn't compile if OutputIteratorT::value_type != T \n - Issue #68 (cub::TilePrefixCallbackOp::WarpReduce doesn't permit ptx \n arch specialization)\n\t\t- Improved support for Win32 platforms (warnings, alignment, etc)\n\t\t \n//-----------------------------------------------------------------------------\n\n1.6.2 (was 1.5.5) 10/25/2016\n - Updated Pascal tuning policies for radix sorting\n - Bug fixes: \n - Fix for arm64 compilation of caching allocator\n\n//-----------------------------------------------------------------------------\n\n1.6.1 (was 1.5.4) 10/14/2016\n - Bug fixes: \n - Fix for radix sorting bug introduced by scan refactorization\n\n//-----------------------------------------------------------------------------\n\n1.6.0 (was 1.5.3) 10/11/2016\n - API change: Device/block/warp-wide exclusive scans have been revised to now \n accept an \"initial value\" (instead of an \"identity value\") for seeding the \n computation with an arbitrary prefix. \n - API change: Device-wide reductions and scans can now have input sequence types that are \n different from output sequence types (as long as they are coercible)\n value\") for seeding the computation with an arbitrary prefix\n - Reduce repository size (move doxygen binary to doc repository)\n - Minor reductions in block-scan instruction count\n - Bug fixes: \n - Issue #55: warning in cub/device/dispatch/dispatch_reduce_by_key.cuh \n - Issue #59: cub::DeviceScan::ExclusiveSum can't prefix sum of float into double\n - Issue #58: Infinite loop in cub::CachingDeviceAllocator::NearestPowerOf\n - Issue #47: Caching allocator needs to clean up cuda error upon successful retry \n - Issue #46: Very high amount of needed memory from the cub::DeviceHistogram::HistogramEven routine\n - Issue #45: Caching Device Allocator fails with debug output enabled\n - Fix for generic-type reduce-by-key warpscan (sm3.x and newer)\n\n//-----------------------------------------------------------------------------\n\n1.5.2 03/21/2016\n\t- Improved medium-size scan performance for sm5x (Maxwell)\n - Refactored caching allocator for device memory\n \t\t- Spends less time locked\n\t\t- Failure to allocate a block from the runtime will retry once after\n\t\t freeing cached allocations\n\t\t- Now respects max-bin (issue where blocks in excess of max-bin were\n\t\t still being retained in free cache)\n\t\t- Uses C++11 mutex when available\n - Bug fixes: \n - Fix for generic-type reduce-by-key warpscan (sm3.x and newer)\n \n//-----------------------------------------------------------------------------\n\n1.5.1 12/28/2015\n - Bug fixes: \n - Fix for incorrect DeviceRadixSort output for some small problems on \n Maxwell SM52 architectures\n - Fix for macro redefinition warnings when compiling with Thrust sort\n \n//-----------------------------------------------------------------------------\n\n1.5.0 12/14/2015\n - New Features:\n - Added new segmented device-wide operations for device-wide sort and \n reduction primitives.\n - Bug fixes: \n - Fix for Git Issue 36 (Compilation error with GCC 4.8.4 nvcc 7.0.27) and\n Forums thread (ThreadLoad generates compiler errors when loading from \n pointer-to-const)\n - Fix for Git Issue 29 (DeviceRadixSort::SortKeys yields compiler \n errors)\n - Fix for Git Issue 26 (CUDA error: misaligned address after \n cub::DeviceRadixSort::SortKeys())\n - Fix for incorrect/crash on 0-length problems, e.g., Git Issue 25 (Floating \n point exception (core dumped) during cub::DeviceRadixSort::SortKeys)\n - Fix for CUDA 7.5 issues on SM 5.2 with SHFL-based warp-scan and warp-reduction \n on non-primitive data types (e.g., user-defined structs)\n - Fix for small radix sorting problems where 0 temporary bytes were \n required and users code was invoking malloc(0) on some systems where\n that returns NULL. (Impl assumed was asking for size again and was not \n running the sort.)\n \n//-----------------------------------------------------------------------------\n\n1.4.1 04/13/2015\n - Bug fixes: \n - Fixes for CUDA 7.0 issues with SHFL-based warp-scan and warp-reduction \n on non-primitive data types (e.g., user-defined structs)\n - Fixes for minor CUDA 7.0 performance regressions in cub::DeviceScan,\n DeviceReduceByKey\n - Fixes to allow cub::DeviceRadixSort and cub::BlockRadixSort on bool types\n - Remove requirement for callers to define the CUB_CDP macro \n when invoking CUB device-wide rountines using CUDA dynamic parallelism\n - Fix for headers not being included in the proper order (or missing includes)\n for some block-wide functions\n \n//-----------------------------------------------------------------------------\n\n1.4.0 03/18/2015\n - New Features:\n\t\t- Support and performance tuning for new Maxwell GPU architectures\n - Updated cub::DeviceHistogram implementation that provides the same \n \"histogram-even\" and \"histogram-range\" functionality as IPP/NPP.\n Provides extremely fast and, perhaps more importantly, very \n uniform performance response across diverse real-world datasets, \n including pathological (homogeneous) sample distributions (resilience)\n - New cub::DeviceSpmv methods for multiplying sparse matrices by \n dense vectors, load-balanced using a merge-based parallel decomposition.\n - New cub::DeviceRadixSort sorting entry-points that always return\n the sorted output into the specified buffer (as opposed to the \n cub::DoubleBuffer in which it could end up in either buffer)\n - New cub::DeviceRunLengthEncode::NonTrivialRuns for finding the starting \n offsets and lengths of all non-trivial runs (i.e., length > 1) of keys in \n a given sequence. (Useful for top-down partitioning algorithms like \n MSD sorting of very-large keys.)\n \n//-----------------------------------------------------------------------------\n\n1.3.2 07/28/2014\n - Bug fixes: \n - Fix for cub::DeviceReduce where reductions of small problems \n (small enough to only dispatch a single thread block) would run in \n the default stream (stream zero) regardless of whether an alternate\n stream was specified. \n \n//-----------------------------------------------------------------------------\n\n1.3.1 05/23/2014\n - Bug fixes: \n - Workaround for a benign WAW race warning reported by cuda-memcheck\n in BlockScan specialized for BLOCK_SCAN_WARP_SCANS algorithm.\n - Fix for bug in DeviceRadixSort where the algorithm may sort more \n key bits than the caller specified (up to the nearest radix digit).\n - Fix for ~3% DeviceRadixSort performance regression on Kepler and \n Fermi that was introduced in v1.3.0. \n\n//-----------------------------------------------------------------------------\n\n1.3.0 05/12/2014\n - New features:\n - CUB's collective (block-wide, warp-wide) primitives underwent a minor \n interface refactoring:\n - To provide the appropriate support for multidimensional thread blocks,\n The interfaces for collective classes are now template-parameterized \n by X, Y, and Z block dimensions (with BLOCK_DIM_Y and BLOCK_DIM_Z being \n optional, and BLOCK_DIM_X replacing BLOCK_THREADS). Furthermore, the \n constructors that accept remapped linear thread-identifiers have been \n removed: all primitives now assume a row-major thread-ranking for \n multidimensional thread blocks. \n - To allow the host program (compiled by the host-pass) to \n accurately determine the device-specific storage requirements for \n a given collective (compiled for each device-pass), the interfaces \n for collective classes are now (optionally) template-parameterized \n by the desired PTX compute capability. This is useful when \n aliasing collective storage to shared memory that has been \n allocated dynamically by the host at the kernel call site. \n - Most CUB programs having typical 1D usage should not require any \n changes to accomodate these updates.\n - Added new \"combination\" WarpScan methods for efficiently computing \n both inclusive and exclusive prefix scans (and sums).\n - Bug fixes: \n - Fixed bug in cub::WarpScan (which affected cub::BlockScan and \n cub::DeviceScan) where incorrect results (e.g., NAN) would often be \n returned when parameterized for floating-point types (fp32, fp64).\n - Workaround-fix for ptxas error when compiling with with -G flag on Linux \n (for debug instrumentation) \n - Misc. workaround-fixes for certain scan scenarios (using custom \n scan operators) where code compiled for SM1x is run on newer \n GPUs of higher compute-capability: the compiler could not tell\n which memory space was being used collective operations and was \n mistakenly using global ops instead of shared ops. \n\n//-----------------------------------------------------------------------------\n\n1.2.3 04/01/2014\n - Bug fixes: \n - Fixed access violation bug in DeviceReduce::ReduceByKey for non-primitive value types\n - Fixed code-snippet bug in ArgIndexInputIteratorT documentation \n\n//-----------------------------------------------------------------------------\n\n1.2.2 03/03/2014\n - New features:\n - Added MS VC++ project solutions for device-wide and block-wide examples \n - Performance:\n - Added a third algorithmic variant of cub::BlockReduce for improved performance\n when using commutative operators (e.g., numeric addition)\n - Bug fixes: \n - Fixed bug where inclusion of Thrust headers in a certain order prevented CUB device-wide primitives from working properly\n\n//-----------------------------------------------------------------------------\n\n1.2.0 02/25/2014\n - New features:\n - Added device-wide reduce-by-key (DeviceReduce::ReduceByKey, DeviceReduce::RunLengthEncode) \n - Performance\n - Improved DeviceScan, DeviceSelect, DevicePartition performance\n - Documentation and testing:\n - Compatible with CUDA 6.0\n - Added performance-portability plots for many device-wide primitives to doc \n - Update doc and tests to reflect iterator (in)compatibilities with CUDA 5.0 (and older) and Thrust 1.6 (and older).\n - Bug fixes \n - Revised the operation of temporary tile status bookkeeping for DeviceScan (and similar) to be safe for current code run on future platforms (now uses proper fences) \n - Fixed DeviceScan bug where Win32 alignment disagreements between host and device regarding user-defined data types would corrupt tile status\n - Fixed BlockScan bug where certain exclusive scans on custom data types for the BLOCK_SCAN_WARP_SCANS variant would return incorrect results for the first thread in the block\n - Added workaround for TexRefInputIteratorTto work with CUDA 6.0\n \n//-----------------------------------------------------------------------------\n\n1.1.1 12/11/2013\n - New features:\n - Added TexObjInputIteratorT, TexRefInputIteratorT, CacheModifiedInputIteratorT, and CacheModifiedOutputIterator types for loading & storing arbitrary types through the cache hierarchy. Compatible with Thrust API. \n - Added descending sorting to DeviceRadixSort and BlockRadixSort\n - Added min, max, arg-min, and arg-max to DeviceReduce\n - Added DeviceSelect (select-unique, select-if, and select-flagged)\n - Added DevicePartition (partition-if, partition-flagged)\n - Added generic cub::ShuffleUp(), cub::ShuffleDown(), and cub::ShuffleIndex() for warp-wide communication of arbitrary data types (SM3x+)\n - Added cub::MaxSmOccupancy() for accurately determining SM occupancy for any given kernel function pointer\n - Performance\n - Improved DeviceScan and DeviceRadixSort performance for older architectures (SM10-SM30)\n - Interface changes:\n - Refactored block-wide I/O (BlockLoad and BlockStore), removing cache-modifiers from their interfaces. The CacheModifiedInputIteratorTand CacheModifiedOutputIterator should now be used with BlockLoad and BlockStore to effect that behavior.\n - Rename device-wide \"stream_synchronous\" param to \"debug_synchronous\" to avoid confusion about usage\n - Documentation and testing:\n - Added simple examples of device-wide methods\n - Improved doxygen documentation and example snippets\n - Improved test coverege to include up to 21,000 kernel variants and 851,000 unit tests (per architecture, per platform)\n - Bug fixes \n - Fixed misc DeviceScan, BlockScan, DeviceReduce, and BlockReduce bugs when operating on non-primitive types for older architectures SM10-SM13\n - Fixed DeviceScan / WarpReduction bug: SHFL-based segmented reduction producting incorrect results for multi-word types (size > 4B) on Linux \n - Fixed BlockScan bug: For warpscan-based scans, not all threads in the first warp were entering the prefix callback functor\n - Fixed DeviceRadixSort bug: race condition with key-value pairs for pre-SM35 architectures\n - Fixed DeviceRadixSort bug: incorrect bitfield-extract behavior with long keys on 64bit Linux\n - Fixed BlockDiscontinuity bug: complation error in for types other than int32/uint32\n - CDP (device-callable) versions of device-wide methods now report the same temporary storage allocation size requirement as their host-callable counterparts\n \n\n//-----------------------------------------------------------------------------\n\n1.0.2 08/23/2013\n - Corrections to code snippet examples for BlockLoad, BlockStore, and BlockDiscontinuity\n - Cleaned up unnecessary/missing header includes. You can now safely #inlude a specific .cuh (instead of cub.cuh)\n - Bug/compilation fixes for BlockHistogram \n\n//-----------------------------------------------------------------------------\n\n1.0.1 08/08/2013\n - New collective interface idiom (specialize::construct::invoke).\n - Added best-in-class DeviceRadixSort. Implements short-circuiting for homogenous digit passes.\n - Added best-in-class DeviceScan. Implements single-pass \"adaptive-lookback\" strategy.\n - Significantly improved documentation (with example code snippets) \n - More extensive regression test suit for aggressively testing collective variants\n - Allow non-trially-constructed types (previously unions had prevented aliasing temporary storage of those types)\n - Improved support for Kepler SHFL (collective ops now use SHFL for types larger than 32b)\n - Better code generation for 64-bit addressing within BlockLoad/BlockStore\n - DeviceHistogram now supports histograms of arbitrary bins\n - Misc. fixes\n - Workarounds for SM10 codegen issues in uncommonly-used WarpScan/Reduce specializations\n - Updates to accommodate CUDA 5.5 dynamic parallelism \n\n\n//-----------------------------------------------------------------------------\n\n0.9.4 05/07/2013\n\n - Fixed compilation errors for SM10-SM13\n - Fixed compilation errors for some WarpScan entrypoints on SM30+\n - Added block-wide histogram (BlockHistogram256)\n - Added device-wide histogram (DeviceHistogram256)\n - Added new BlockScan algorithm variant BLOCK_SCAN_RAKING_MEMOIZE, which \n trades more register consumption for less shared memory I/O)\n - Updates to BlockRadixRank to use BlockScan (which improves performance\n on Kepler due to SHFL instruction)\n - Allow types other than C++ primitives to be used in WarpScan::*Sum methods \n if they only have operator + overloaded. (Previously they also required \n to support assignment from int(0).) \n - Update BlockReduce's BLOCK_REDUCE_WARP_REDUCTIONS algorithm to work even \n when block size is not an even multiple of warp size\n - Added work management utility descriptors (GridQueue, GridEvenShare)\n - Refactoring of DeviceAllocator interface and CachingDeviceAllocator \n implementation \n - Misc. documentation updates and corrections. \n \n//-----------------------------------------------------------------------------\n\n0.9.2 04/04/2013\n\n - Added WarpReduce. WarpReduce uses the SHFL instruction when applicable. \n BlockReduce now uses this WarpReduce instead of implementing its own.\n - Misc. fixes for 64-bit Linux compilation warnings and errors.\n - Misc. documentation updates and corrections. \n\n//-----------------------------------------------------------------------------\n\n0.9.1 03/09/2013\n\n - Fix for ambiguity in BlockScan::Reduce() between generic reduction and \n summation. Summation entrypoints are now called ::Sum(), similar to the \n convention in BlockScan.\n - Small edits to mainpage documentation and download tracking\n \n//-----------------------------------------------------------------------------\n\n0.9.0 03/07/2013 \n\n - Intial \"preview\" release. CUB is the first durable, high-performance library \n of cooperative block-level, warp-level, and thread-level primitives for CUDA \n kernel programming. More primitives and examples coming soon!\n " }, { "path": "external/cub/LICENSE.TXT", "content": "Copyright (c) 2010-2011, Duane Merrill. All rights reserved.\r\nCopyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\r\n\r\nRedistribution and use in source and binary forms, with or without\r\nmodification, are permitted provided that the following conditions are met:\r\n * Redistributions of source code must retain the above copyright\r\n notice, this list of conditions and the following disclaimer.\r\n * Redistributions in binary form must reproduce the above copyright\r\n notice, this list of conditions and the following disclaimer in the\r\n documentation and/or other materials provided with the distribution.\r\n * Neither the name of the NVIDIA CORPORATION nor the\r\n names of its contributors may be used to endorse or promote products\r\n derived from this software without specific prior written permission.\r\n\r\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\r\nANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\r\nWARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\r\nDISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\r\nDIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\r\n(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\r\nLOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\r\nON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\r\n(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\r\nSOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE." }, { "path": "external/cub/README.md", "content": "
\n

About CUB

\n\nCurrent release: v1.7.4 (09/20/2017)\n\nWe recommend the [CUB Project Website](http://nvlabs.github.com/cub) and the [cub-users discussion forum](http://groups.google.com/group/cub-users) for further information and examples.\n\nCUB provides state-of-the-art, reusable software components for every layer \nof the CUDA programming model:\n- [Device-wide primitives] (https://nvlabs.github.com/cub/group___device_module.html) \n - Sort, prefix scan, reduction, histogram, etc. \n - Compatible with CUDA dynamic parallelism\n- [Block-wide \"collective\" primitives] (https://nvlabs.github.com/cub/group___block_module.html)\n - I/O, sort, prefix scan, reduction, histogram, etc. \n - Compatible with arbitrary thread block sizes and types \n- [Warp-wide \"collective\" primitives] (https://nvlabs.github.com/cub/group___warp_module.html)\n - Warp-wide prefix scan, reduction, etc.\n - Safe and architecture-specific\n- [Thread and resource utilities](https://nvlabs.github.com/cub/group___thread_module.html)\n - PTX intrinsics, device reflection, texture-caching iterators, caching memory allocators, etc. \n\n![Orientation of collective primitives within the CUDA software stack](http://nvlabs.github.com/cub/cub_overview.png)\n\n
\n

A Simple Example

\n\n```C++\n#include \n \n// Block-sorting CUDA kernel\n__global__ void BlockSortKernel(int *d_in, int *d_out)\n{\n using namespace cub;\n\n // Specialize BlockRadixSort, BlockLoad, and BlockStore for 128 threads \n // owning 16 integer items each\n typedef BlockRadixSort BlockRadixSort;\n typedef BlockLoad BlockLoad;\n typedef BlockStore BlockStore;\n \n // Allocate shared memory\n __shared__ union {\n typename BlockRadixSort::TempStorage sort;\n typename BlockLoad::TempStorage load; \n typename BlockStore::TempStorage store; \n } temp_storage; \n\n int block_offset = blockIdx.x * (128 * 16);\t // OffsetT for this block's ment\n\n // Obtain a segment of 2048 consecutive keys that are blocked across threads\n int thread_keys[16];\n BlockLoad(temp_storage.load).Load(d_in + block_offset, thread_keys);\n __syncthreads();\n\n // Collectively sort the keys\n BlockRadixSort(temp_storage.sort).Sort(thread_keys);\n __syncthreads();\n\n // Store the sorted segment \n BlockStore(temp_storage.store).Store(d_out + block_offset, thread_keys);\n}\n```\n\nEach thread block uses cub::BlockRadixSort to collectively sort \nits own input segment. The class is specialized by the \ndata type being sorted, by the number of threads per block, by the number of \nkeys per thread, and implicitly by the targeted compilation architecture. \n\nThe cub::BlockLoad and cub::BlockStore classes are similarly specialized. \nFurthermore, to provide coalesced accesses to device memory, these primitives are \nconfigured to access memory using a striped access pattern (where consecutive threads \nsimultaneously access consecutive items) and then transpose the keys into \na [blocked arrangement](index.html#sec4sec3) of elements across threads. \n\nOnce specialized, these classes expose opaque \\p TempStorage member types. \nThe thread block uses these storage types to statically allocate the union of \nshared memory needed by the thread block. (Alternatively these storage types \ncould be aliased to global memory allocations).\n\n
\n

Stable Releases

\n\nCUB releases are labeled using version identifiers having three fields: \n*epoch.feature.update*. The *epoch* field corresponds to support for\na major change in the CUDA programming model. The *feature* field \ncorresponds to a stable set of features, functionality, and interface. The\n*update* field corresponds to a bug-fix or performance update for that\nfeature set. At the moment, we do not publicly provide non-stable releases \nsuch as development snapshots, beta releases or rolling releases. (Feel free\nto contact us if you would like such things.) See the \n[CUB Project Website](http://nvlabs.github.com/cub) for more information.\n\n
\n

Contributors

\n\nCUB is developed as an open-source project by [NVIDIA Research](http://research.nvidia.com). The primary contributor is [Duane Merrill](http://github.com/dumerrill).\n\n
\n

Open Source License

\n\nCUB is available under the \"New BSD\" open-source license:\n\n```\nCopyright (c) 2010-2011, Duane Merrill. All rights reserved.\nCopyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n\nRedistribution and use in source and binary forms, with or without\nmodification, are permitted provided that the following conditions are met:\n * Redistributions of source code must retain the above copyright\n notice, this list of conditions and the following disclaimer.\n * Redistributions in binary form must reproduce the above copyright\n notice, this list of conditions and the following disclaimer in the\n documentation and/or other materials provided with the distribution.\n * Neither the name of the NVIDIA CORPORATION nor the\n names of its contributors may be used to endorse or promote products\n derived from this software without specific prior written permission.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\nANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\nWARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\nDISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\nDIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\nLOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\nON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\nSOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n```\n" }, { "path": "external/cub/common.mk", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n\n\n#-------------------------------------------------------------------------------\n# Commandline Options\n#-------------------------------------------------------------------------------\n\n# [sm=] Compute-capability to compile for, e.g., \"sm=200,300,350\" (SM20 by default).\n \nCOMMA = ,\nifdef sm\n\tSM_ARCH = $(subst $(COMMA),-,$(sm))\nelse \n SM_ARCH = 200\nendif\n\nifeq (700, $(findstring 700, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_70,code=\\\"sm_70,compute_70\\\" \n SM_DEF \t\t+= -DSM700\n TEST_ARCH \t= 700\nendif\nifeq (620, $(findstring 620, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_62,code=\\\"sm_62,compute_62\\\" \n SM_DEF \t\t+= -DSM620\n TEST_ARCH \t= 620\nendif\nifeq (610, $(findstring 610, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_61,code=\\\"sm_61,compute_61\\\" \n SM_DEF \t\t+= -DSM610\n TEST_ARCH \t= 610\nendif\nifeq (600, $(findstring 600, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_60,code=\\\"sm_60,compute_60\\\" \n SM_DEF \t\t+= -DSM600\n TEST_ARCH \t= 600\nendif\nifeq (520, $(findstring 520, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_52,code=\\\"sm_52,compute_52\\\" \n SM_DEF \t\t+= -DSM520\n TEST_ARCH \t= 520\nendif\nifeq (370, $(findstring 370, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_37,code=\\\"sm_37,compute_37\\\" \n SM_DEF \t\t+= -DSM370\n TEST_ARCH \t= 370\nendif\nifeq (350, $(findstring 350, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_35,code=\\\"sm_35,compute_35\\\" \n SM_DEF \t\t+= -DSM350\n TEST_ARCH \t= 350\nendif\nifeq (300, $(findstring 300, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_30,code=\\\"sm_30,compute_30\\\"\n SM_DEF \t\t+= -DSM300\n TEST_ARCH \t= 300\nendif\nifeq (210, $(findstring 210, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_20,code=\\\"sm_21,compute_20\\\"\n SM_DEF \t\t+= -DSM210\n TEST_ARCH \t= 210\nendif\nifeq (200, $(findstring 200, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_20,code=\\\"sm_20,compute_20\\\"\n SM_DEF \t\t+= -DSM200\n TEST_ARCH \t= 200\nendif\nifeq (130, $(findstring 130, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_13,code=\\\"sm_13,compute_13\\\" \n SM_DEF \t\t+= -DSM130\n TEST_ARCH \t= 130\nendif\nifeq (120, $(findstring 120, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_12,code=\\\"sm_12,compute_12\\\" \n SM_DEF \t\t+= -DSM120\n TEST_ARCH \t= 120\nendif\nifeq (110, $(findstring 110, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_11,code=\\\"sm_11,compute_11\\\" \n SM_DEF \t\t+= -DSM110\n TEST_ARCH \t= 110\nendif\nifeq (100, $(findstring 100, $(SM_ARCH)))\n SM_TARGETS \t+= -gencode=arch=compute_10,code=\\\"sm_10,compute_10\\\" \n SM_DEF \t\t+= -DSM100\n TEST_ARCH \t= 100\nendif\n\n\n# [cdp=<0|1>] CDP enable option (default: no)\nifeq ($(cdp), 1)\n\tDEFINES += -DCUB_CDP\n\tCDP_SUFFIX = cdp\n NVCCFLAGS += -rdc=true -lcudadevrt\nelse\n\tCDP_SUFFIX = nocdp\nendif\n\n\n# [force32=<0|1>] Device addressing mode option (64-bit device pointers by default) \nifeq ($(force32), 1)\n\tCPU_ARCH = -m32\n\tCPU_ARCH_SUFFIX = i386\nelse\n\tCPU_ARCH = -m64\n\tCPU_ARCH_SUFFIX = x86_64\n NPPI = -lnppist\nendif\n\n\n# [abi=<0|1>] CUDA ABI option (enabled by default) \nifneq ($(abi), 0)\n\tABI_SUFFIX = abi\nelse \n\tNVCCFLAGS += -Xptxas -abi=no\n\tABI_SUFFIX = noabi\nendif\n\n\n# [open64=<0|1>] Middle-end compiler option (nvvm by default)\nifeq ($(open64), 1)\n\tNVCCFLAGS += -open64\n\tPTX_SUFFIX = open64\nelse \n\tPTX_SUFFIX = nvvm\nendif\n\n\n# [verbose=<0|1>] Verbose toolchain output from nvcc option\nifeq ($(verbose), 1)\n\tNVCCFLAGS += -v\nendif\n\n\n# [keep=<0|1>] Keep intermediate compilation artifacts option\nifeq ($(keep), 1)\n\tNVCCFLAGS += -keep\nendif\n\n# [debug=<0|1>] Generate debug mode code\nifeq ($(debug), 1)\n\tNVCCFLAGS += -G\nendif\n\n\n#-------------------------------------------------------------------------------\n# Compiler and compilation platform\n#-------------------------------------------------------------------------------\n\nCUB_DIR = $(dir $(lastword $(MAKEFILE_LIST)))\n\nNVCC = \"$(shell which nvcc)\"\nifdef nvccver\n NVCC_VERSION = $(nvccver)\nelse\n NVCC_VERSION = $(strip $(shell nvcc --version | grep release | sed 's/.*release //' | sed 's/,.*//'))\nendif\n\n# detect OS\nOSUPPER = $(shell uname -s 2>/dev/null | tr [:lower:] [:upper:])\n\n# Default flags: verbose kernel properties (regs, smem, cmem, etc.); runtimes for compilation phases \nNVCCFLAGS += $(SM_DEF) -Xptxas -v -Xcudafe -\\# \n\nifeq (WIN_NT, $(findstring WIN_NT, $(OSUPPER)))\n # For MSVC\n # Enable more warnings and treat as errors\n NVCCFLAGS += -Xcompiler /W3 -Xcompiler /WX\n # Disable excess x86 floating point precision that can lead to results being labeled incorrectly\n NVCCFLAGS += -Xcompiler /fp:strict\n # Help the compiler/linker work with huge numbers of kernels on Windows\n\tNVCCFLAGS += -Xcompiler /bigobj -Xcompiler /Zm500\n\tCC = cl\n\t\n\t# Multithreaded runtime\n\tNVCCFLAGS += -Xcompiler /MT\n\t\nifneq ($(force32), 1)\n\tCUDART_CYG = \"$(shell dirname $(NVCC))/../lib/Win32/cudart.lib\"\nelse\n\tCUDART_CYG = \"$(shell dirname $(NVCC))/../lib/x64/cudart.lib\"\nendif\n\tCUDART = \"$(shell cygpath -w $(CUDART_CYG))\"\nelse\n # For g++\n # Disable excess x86 floating point precision that can lead to results being labeled incorrectly\n NVCCFLAGS += -Xcompiler -ffloat-store\n CC = g++\nifneq ($(force32), 1)\n CUDART = \"$(shell dirname $(NVCC))/../lib/libcudart_static.a\"\nelse\n CUDART = \"$(shell dirname $(NVCC))/../lib64/libcudart_static.a\"\nendif\nendif\n\n# Suffix to append to each binary\nBIN_SUFFIX = sm$(SM_ARCH)_$(PTX_SUFFIX)_$(NVCC_VERSION)_$(ABI_SUFFIX)_$(CDP_SUFFIX)_$(CPU_ARCH_SUFFIX)\n\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nrwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nCUB_DEPS = \t$(call rwildcard, $(CUB_DIR),*.cuh) \\\n\t\t\t$(CUB_DIR)common.mk\n\t\t\n" }, { "path": "external/cub/cub/agent/agent_histogram.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentHistogram implements a stateful abstraction of CUDA thread blocks for participating in device-wide histogram .\n */\n\n#pragma once\n\n#include \n\n#include \"../util_type.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../grid/grid_queue.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy\n ******************************************************************************/\n\n/**\n *\n */\nenum BlockHistogramMemoryPreference\n{\n GMEM,\n SMEM,\n BLEND\n};\n\n\n/**\n * Parameterizable tuning policy type for AgentHistogram\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _PIXELS_PER_THREAD, ///< Pixels per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n bool _RLE_COMPRESS, ///< Whether to perform localized RLE to compress samples before histogramming\n BlockHistogramMemoryPreference _MEM_PREFERENCE, ///< Whether to prefer privatized shared-memory bins (versus privatized global-memory bins)\n bool _WORK_STEALING> ///< Whether to dequeue tiles from a global work queue\nstruct AgentHistogramPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n PIXELS_PER_THREAD = _PIXELS_PER_THREAD, ///< Pixels per thread (per tile of input)\n IS_RLE_COMPRESS = _RLE_COMPRESS, ///< Whether to perform localized RLE to compress samples before histogramming\n MEM_PREFERENCE = _MEM_PREFERENCE, ///< Whether to prefer privatized shared-memory bins (versus privatized global-memory bins)\n IS_WORK_STEALING = _WORK_STEALING, ///< Whether to dequeue tiles from a global work queue\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentHistogram implements a stateful abstraction of CUDA thread blocks for participating in device-wide histogram .\n */\ntemplate <\n typename AgentHistogramPolicyT, ///< Parameterized AgentHistogramPolicy tuning policy type\n int PRIVATIZED_SMEM_BINS, ///< Number of privatized shared-memory histogram bins of any channel. Zero indicates privatized counters to be maintained in device-accessible memory.\n int NUM_CHANNELS, ///< Number of channels interleaved in the input data. Supports up to four channels.\n int NUM_ACTIVE_CHANNELS, ///< Number of channels actively being histogrammed\n typename SampleIteratorT, ///< Random-access input iterator type for reading samples\n typename CounterT, ///< Integer type for counting sample occurrences per histogram bin\n typename PrivatizedDecodeOpT, ///< The transform operator type for determining privatized counter indices from samples, one for each channel\n typename OutputDecodeOpT, ///< The transform operator type for determining output bin-ids from privatized counter indices, one for each channel\n typename OffsetT, ///< Signed integer type for global offsets\n int PTX_ARCH = CUB_PTX_ARCH> ///< PTX compute capability\nstruct AgentHistogram\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// The sample type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n /// The pixel type of SampleT\n typedef typename CubVector::Type PixelT;\n\n /// The quad type of SampleT\n typedef typename CubVector::Type QuadT;\n\n /// Constants\n enum\n {\n BLOCK_THREADS = AgentHistogramPolicyT::BLOCK_THREADS,\n\n PIXELS_PER_THREAD = AgentHistogramPolicyT::PIXELS_PER_THREAD,\n SAMPLES_PER_THREAD = PIXELS_PER_THREAD * NUM_CHANNELS,\n QUADS_PER_THREAD = SAMPLES_PER_THREAD / 4,\n\n TILE_PIXELS = PIXELS_PER_THREAD * BLOCK_THREADS,\n TILE_SAMPLES = SAMPLES_PER_THREAD * BLOCK_THREADS,\n\n IS_RLE_COMPRESS = AgentHistogramPolicyT::IS_RLE_COMPRESS,\n\n MEM_PREFERENCE = (PRIVATIZED_SMEM_BINS > 0) ?\n AgentHistogramPolicyT::MEM_PREFERENCE :\n GMEM,\n\n IS_WORK_STEALING = AgentHistogramPolicyT::IS_WORK_STEALING,\n };\n\n /// Cache load modifier for reading input elements\n static const CacheLoadModifier LOAD_MODIFIER = AgentHistogramPolicyT::LOAD_MODIFIER;\n\n\n /// Input iterator wrapper type (for applying cache modifier)\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n SampleIteratorT>::Type // Directly use the supplied input iterator type\n WrappedSampleIteratorT;\n\n /// Pixel input iterator type (for applying cache modifier)\n typedef CacheModifiedInputIterator\n WrappedPixelIteratorT;\n\n /// Qaud input iterator type (for applying cache modifier)\n typedef CacheModifiedInputIterator\n WrappedQuadIteratorT;\n\n /// Parameterized BlockLoad type for samples\n typedef BlockLoad<\n SampleT,\n BLOCK_THREADS,\n SAMPLES_PER_THREAD,\n AgentHistogramPolicyT::LOAD_ALGORITHM>\n BlockLoadSampleT;\n\n /// Parameterized BlockLoad type for pixels\n typedef BlockLoad<\n PixelT,\n BLOCK_THREADS,\n PIXELS_PER_THREAD,\n AgentHistogramPolicyT::LOAD_ALGORITHM>\n BlockLoadPixelT;\n\n /// Parameterized BlockLoad type for quads\n typedef BlockLoad<\n QuadT,\n BLOCK_THREADS,\n QUADS_PER_THREAD,\n AgentHistogramPolicyT::LOAD_ALGORITHM>\n BlockLoadQuadT;\n\n /// Shared memory type required by this thread block\n struct _TempStorage\n {\n CounterT histograms[NUM_ACTIVE_CHANNELS][PRIVATIZED_SMEM_BINS + 1]; // Smem needed for block-privatized smem histogram (with 1 word of padding)\n\n int tile_idx;\n\n // Aliasable storage layout\n union Aliasable\n {\n typename BlockLoadSampleT::TempStorage sample_load; // Smem needed for loading a tile of samples\n typename BlockLoadPixelT::TempStorage pixel_load; // Smem needed for loading a tile of pixels\n typename BlockLoadQuadT::TempStorage quad_load; // Smem needed for loading a tile of quads\n\n } aliasable;\n };\n\n\n /// Temporary storage type (unionable)\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n /// Reference to temp_storage\n _TempStorage &temp_storage;\n\n /// Sample input iterator (with cache modifier applied, if possible)\n WrappedSampleIteratorT d_wrapped_samples;\n\n /// Native pointer for input samples (possibly NULL if unavailable)\n SampleT* d_native_samples;\n\n /// The number of output bins for each channel\n int (&num_output_bins)[NUM_ACTIVE_CHANNELS];\n\n /// The number of privatized bins for each channel\n int (&num_privatized_bins)[NUM_ACTIVE_CHANNELS];\n\n /// Reference to gmem privatized histograms for each channel\n CounterT* d_privatized_histograms[NUM_ACTIVE_CHANNELS];\n\n /// Reference to final output histograms (gmem)\n CounterT* (&d_output_histograms)[NUM_ACTIVE_CHANNELS];\n\n /// The transform operator for determining output bin-ids from privatized counter indices, one for each channel\n OutputDecodeOpT (&output_decode_op)[NUM_ACTIVE_CHANNELS];\n\n /// The transform operator for determining privatized counter indices from samples, one for each channel\n PrivatizedDecodeOpT (&privatized_decode_op)[NUM_ACTIVE_CHANNELS];\n\n /// Whether to prefer privatized smem counters vs privatized global counters\n bool prefer_smem;\n\n\n //---------------------------------------------------------------------\n // Initialize privatized bin counters\n //---------------------------------------------------------------------\n\n // Initialize privatized bin counters\n __device__ __forceinline__ void InitBinCounters(CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS])\n {\n // Initialize histogram bin counts to zeros\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n for (int privatized_bin = threadIdx.x; privatized_bin < num_privatized_bins[CHANNEL]; privatized_bin += BLOCK_THREADS)\n {\n privatized_histograms[CHANNEL][privatized_bin] = 0;\n }\n }\n\n // Barrier to make sure all threads are done updating counters\n CTA_SYNC();\n }\n\n\n // Initialize privatized bin counters. Specialized for privatized shared-memory counters\n __device__ __forceinline__ void InitSmemBinCounters()\n {\n CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];\n\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];\n\n InitBinCounters(privatized_histograms);\n }\n\n\n // Initialize privatized bin counters. Specialized for privatized global-memory counters\n __device__ __forceinline__ void InitGmemBinCounters()\n {\n InitBinCounters(d_privatized_histograms);\n }\n\n\n //---------------------------------------------------------------------\n // Update final output histograms\n //---------------------------------------------------------------------\n\n // Update final output histograms from privatized histograms\n __device__ __forceinline__ void StoreOutput(CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS])\n {\n // Barrier to make sure all threads are done updating counters\n CTA_SYNC();\n\n // Apply privatized bin counts to output bin counts\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n int channel_bins = num_privatized_bins[CHANNEL];\n for (int privatized_bin = threadIdx.x; \n privatized_bin < channel_bins; \n privatized_bin += BLOCK_THREADS)\n {\n int output_bin = -1;\n CounterT count = privatized_histograms[CHANNEL][privatized_bin];\n bool is_valid = count > 0;\n\n output_decode_op[CHANNEL].template BinSelect((SampleT) privatized_bin, output_bin, is_valid);\n\n if (output_bin >= 0)\n {\n atomicAdd(&d_output_histograms[CHANNEL][output_bin], count);\n }\n\n }\n }\n }\n\n\n // Update final output histograms from privatized histograms. Specialized for privatized shared-memory counters\n __device__ __forceinline__ void StoreSmemOutput()\n {\n CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];\n\n StoreOutput(privatized_histograms);\n }\n\n\n // Update final output histograms from privatized histograms. Specialized for privatized global-memory counters\n __device__ __forceinline__ void StoreGmemOutput()\n {\n StoreOutput(d_privatized_histograms);\n }\n\n\n //---------------------------------------------------------------------\n // Tile accumulation\n //---------------------------------------------------------------------\n\n // Accumulate pixels. Specialized for RLE compression.\n __device__ __forceinline__ void AccumulatePixels(\n SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],\n bool is_valid[PIXELS_PER_THREAD],\n CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS],\n Int2Type is_rle_compress)\n {\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n // Bin pixels\n int bins[PIXELS_PER_THREAD];\n\n #pragma unroll\n for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)\n {\n bins[PIXEL] = -1;\n privatized_decode_op[CHANNEL].template BinSelect(samples[PIXEL][CHANNEL], bins[PIXEL], is_valid[PIXEL]);\n }\n\n CounterT accumulator = 1;\n\n #pragma unroll\n for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD - 1; ++PIXEL)\n {\n if (bins[PIXEL] != bins[PIXEL + 1])\n {\n if (bins[PIXEL] >= 0)\n atomicAdd(privatized_histograms[CHANNEL] + bins[PIXEL], accumulator);\n\n accumulator = 0;\n }\n accumulator++;\n }\n\n // Last pixel\n if (bins[PIXELS_PER_THREAD - 1] >= 0)\n atomicAdd(privatized_histograms[CHANNEL] + bins[PIXELS_PER_THREAD - 1], accumulator);\n }\n }\n\n\n // Accumulate pixels. Specialized for individual accumulation of each pixel.\n __device__ __forceinline__ void AccumulatePixels(\n SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],\n bool is_valid[PIXELS_PER_THREAD],\n CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS],\n Int2Type is_rle_compress)\n {\n #pragma unroll\n for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)\n {\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n int bin = -1;\n privatized_decode_op[CHANNEL].template BinSelect(samples[PIXEL][CHANNEL], bin, is_valid[PIXEL]);\n if (bin >= 0)\n atomicAdd(privatized_histograms[CHANNEL] + bin, 1);\n }\n }\n }\n\n\n /**\n * Accumulate pixel, specialized for smem privatized histogram\n */\n __device__ __forceinline__ void AccumulateSmemPixels(\n SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],\n bool is_valid[PIXELS_PER_THREAD])\n {\n CounterT* privatized_histograms[NUM_ACTIVE_CHANNELS];\n\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n privatized_histograms[CHANNEL] = temp_storage.histograms[CHANNEL];\n\n AccumulatePixels(samples, is_valid, privatized_histograms, Int2Type());\n }\n\n\n /**\n * Accumulate pixel, specialized for gmem privatized histogram\n */\n __device__ __forceinline__ void AccumulateGmemPixels(\n SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS],\n bool is_valid[PIXELS_PER_THREAD])\n {\n AccumulatePixels(samples, is_valid, d_privatized_histograms, Int2Type());\n }\n\n\n\n //---------------------------------------------------------------------\n // Tile loading\n //---------------------------------------------------------------------\n\n // Load full, aligned tile using pixel iterator (multi-channel)\n template \n __device__ __forceinline__ void LoadFullAlignedTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type<_NUM_ACTIVE_CHANNELS> num_active_channels)\n {\n typedef PixelT AliasedPixels[PIXELS_PER_THREAD];\n\n WrappedPixelIteratorT d_wrapped_pixels((PixelT*) (d_native_samples + block_offset));\n\n // Load using a wrapped pixel iterator\n BlockLoadPixelT(temp_storage.aliasable.pixel_load).Load(\n d_wrapped_pixels,\n reinterpret_cast(samples));\n }\n\n // Load full, aligned tile using quad iterator (single-channel)\n __device__ __forceinline__ void LoadFullAlignedTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type<1> num_active_channels)\n {\n typedef QuadT AliasedQuads[QUADS_PER_THREAD];\n\n WrappedQuadIteratorT d_wrapped_quads((QuadT*) (d_native_samples + block_offset));\n\n // Load using a wrapped quad iterator\n BlockLoadQuadT(temp_storage.aliasable.quad_load).Load(\n d_wrapped_quads,\n reinterpret_cast(samples));\n }\n\n // Load full, aligned tile\n __device__ __forceinline__ void LoadTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type is_full_tile,\n Int2Type is_aligned)\n {\n LoadFullAlignedTile(block_offset, valid_samples, samples, Int2Type());\n }\n\n // Load full, mis-aligned tile using sample iterator\n __device__ __forceinline__ void LoadTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type is_full_tile,\n Int2Type is_aligned)\n {\n typedef SampleT AliasedSamples[SAMPLES_PER_THREAD];\n\n // Load using sample iterator\n BlockLoadSampleT(temp_storage.aliasable.sample_load).Load(\n d_wrapped_samples + block_offset,\n reinterpret_cast(samples));\n }\n\n // Load partially-full, aligned tile using the pixel iterator\n __device__ __forceinline__ void LoadTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type is_full_tile,\n Int2Type is_aligned)\n {\n typedef PixelT AliasedPixels[PIXELS_PER_THREAD];\n\n WrappedPixelIteratorT d_wrapped_pixels((PixelT*) (d_native_samples + block_offset));\n\n int valid_pixels = valid_samples / NUM_CHANNELS;\n\n // Load using a wrapped pixel iterator\n BlockLoadPixelT(temp_storage.aliasable.pixel_load).Load(\n d_wrapped_pixels,\n reinterpret_cast(samples),\n valid_pixels);\n }\n\n // Load partially-full, mis-aligned tile using sample iterator\n __device__ __forceinline__ void LoadTile(\n OffsetT block_offset,\n int valid_samples,\n SampleT (&samples)[PIXELS_PER_THREAD][NUM_CHANNELS],\n Int2Type is_full_tile,\n Int2Type is_aligned)\n {\n typedef SampleT AliasedSamples[SAMPLES_PER_THREAD];\n\n BlockLoadSampleT(temp_storage.aliasable.sample_load).Load(\n d_wrapped_samples + block_offset,\n reinterpret_cast(samples),\n valid_samples);\n }\n\n\n //---------------------------------------------------------------------\n // Tile processing\n //---------------------------------------------------------------------\n\n // Consume a tile of data samples\n template <\n bool IS_ALIGNED, // Whether the tile offset is aligned (quad-aligned for single-channel, pixel-aligned for multi-channel)\n bool IS_FULL_TILE> // Whether the tile is full\n __device__ __forceinline__ void ConsumeTile(OffsetT block_offset, int valid_samples)\n {\n SampleT samples[PIXELS_PER_THREAD][NUM_CHANNELS];\n bool is_valid[PIXELS_PER_THREAD];\n\n // Load tile\n LoadTile(\n block_offset,\n valid_samples,\n samples,\n Int2Type(),\n Int2Type());\n\n // Set valid flags\n #pragma unroll\n for (int PIXEL = 0; PIXEL < PIXELS_PER_THREAD; ++PIXEL)\n is_valid[PIXEL] = IS_FULL_TILE || (((threadIdx.x * PIXELS_PER_THREAD + PIXEL) * NUM_CHANNELS) < valid_samples);\n\n // Accumulate samples\n#if CUB_PTX_ARCH >= 120\n if (prefer_smem)\n AccumulateSmemPixels(samples, is_valid);\n else\n AccumulateGmemPixels(samples, is_valid);\n#else\n AccumulateGmemPixels(samples, is_valid);\n#endif\n\n }\n\n\n // Consume row tiles. Specialized for work-stealing from queue\n template \n __device__ __forceinline__ void ConsumeTiles(\n OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< The number of rows in the region of interest\n OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest\n int tiles_per_row, ///< Number of image tiles per row\n GridQueue tile_queue,\n Int2Type is_work_stealing)\n {\n\n int num_tiles = num_rows * tiles_per_row;\n int tile_idx = (blockIdx.y * gridDim.x) + blockIdx.x;\n OffsetT num_even_share_tiles = gridDim.x * gridDim.y;\n\n while (tile_idx < num_tiles)\n {\n int row = tile_idx / tiles_per_row;\n int col = tile_idx - (row * tiles_per_row);\n OffsetT row_offset = row * row_stride_samples;\n OffsetT col_offset = (col * TILE_SAMPLES);\n OffsetT tile_offset = row_offset + col_offset;\n\n if (col == tiles_per_row - 1)\n {\n // Consume a partially-full tile at the end of the row\n OffsetT num_remaining = (num_row_pixels * NUM_CHANNELS) - col_offset;\n ConsumeTile(tile_offset, num_remaining);\n } \n else\n {\n // Consume full tile\n ConsumeTile(tile_offset, TILE_SAMPLES);\n }\n\n CTA_SYNC();\n\n // Get next tile\n if (threadIdx.x == 0)\n temp_storage.tile_idx = tile_queue.Drain(1) + num_even_share_tiles;\n\n CTA_SYNC();\n\n tile_idx = temp_storage.tile_idx;\n }\n }\n\n\n // Consume row tiles. Specialized for even-share (striped across thread blocks)\n template \n __device__ __forceinline__ void ConsumeTiles(\n OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< The number of rows in the region of interest\n OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest\n int tiles_per_row, ///< Number of image tiles per row\n GridQueue tile_queue,\n Int2Type is_work_stealing)\n {\n for (int row = blockIdx.y; row < num_rows; row += gridDim.y)\n {\n OffsetT row_begin = row * row_stride_samples;\n OffsetT row_end = row_begin + (num_row_pixels * NUM_CHANNELS);\n OffsetT tile_offset = row_begin + (blockIdx.x * TILE_SAMPLES);\n\n while (tile_offset < row_end)\n {\n OffsetT num_remaining = row_end - tile_offset;\n\n if (num_remaining < TILE_SAMPLES)\n {\n // Consume partial tile\n ConsumeTile(tile_offset, num_remaining);\n break;\n }\n\n // Consume full tile\n ConsumeTile(tile_offset, TILE_SAMPLES);\n tile_offset += gridDim.x * TILE_SAMPLES;\n }\n }\n }\n\n\n //---------------------------------------------------------------------\n // Parameter extraction\n //---------------------------------------------------------------------\n\n // Return a native pixel pointer (specialized for CacheModifiedInputIterator types)\n template <\n CacheLoadModifier _MODIFIER,\n typename _ValueT,\n typename _OffsetT>\n __device__ __forceinline__ SampleT* NativePointer(CacheModifiedInputIterator<_MODIFIER, _ValueT, _OffsetT> itr)\n {\n return itr.ptr;\n }\n\n // Return a native pixel pointer (specialized for other types)\n template \n __device__ __forceinline__ SampleT* NativePointer(IteratorT itr)\n {\n return NULL;\n }\n\n\n\n //---------------------------------------------------------------------\n // Interface\n //---------------------------------------------------------------------\n\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentHistogram(\n TempStorage &temp_storage, ///< Reference to temp_storage\n SampleIteratorT d_samples, ///< Input data to reduce\n int (&num_output_bins)[NUM_ACTIVE_CHANNELS], ///< The number bins per final output histogram\n int (&num_privatized_bins)[NUM_ACTIVE_CHANNELS], ///< The number bins per privatized histogram\n CounterT* (&d_output_histograms)[NUM_ACTIVE_CHANNELS], ///< Reference to final output histograms\n CounterT* (&d_privatized_histograms)[NUM_ACTIVE_CHANNELS], ///< Reference to privatized histograms\n OutputDecodeOpT (&output_decode_op)[NUM_ACTIVE_CHANNELS], ///< The transform operator for determining output bin-ids from privatized counter indices, one for each channel\n PrivatizedDecodeOpT (&privatized_decode_op)[NUM_ACTIVE_CHANNELS]) ///< The transform operator for determining privatized counter indices from samples, one for each channel\n :\n temp_storage(temp_storage.Alias()),\n d_wrapped_samples(d_samples),\n num_output_bins(num_output_bins),\n num_privatized_bins(num_privatized_bins),\n d_output_histograms(d_output_histograms),\n privatized_decode_op(privatized_decode_op),\n output_decode_op(output_decode_op),\n d_native_samples(NativePointer(d_wrapped_samples)),\n prefer_smem((MEM_PREFERENCE == SMEM) ?\n true : // prefer smem privatized histograms\n (MEM_PREFERENCE == GMEM) ?\n false : // prefer gmem privatized histograms\n blockIdx.x & 1) // prefer blended privatized histograms\n {\n int blockId = (blockIdx.y * gridDim.x) + blockIdx.x;\n\n // Initialize the locations of this block's privatized histograms\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n this->d_privatized_histograms[CHANNEL] = d_privatized_histograms[CHANNEL] + (blockId * num_privatized_bins[CHANNEL]);\n }\n\n\n /**\n * Consume image\n */\n __device__ __forceinline__ void ConsumeTiles(\n OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< The number of rows in the region of interest\n OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest\n int tiles_per_row, ///< Number of image tiles per row\n GridQueue tile_queue) ///< Queue descriptor for assigning tiles of work to thread blocks\n {\n // Check whether all row starting offsets are quad-aligned (in single-channel) or pixel-aligned (in multi-channel)\n int quad_mask = AlignBytes::ALIGN_BYTES - 1;\n int pixel_mask = AlignBytes::ALIGN_BYTES - 1;\n size_t row_bytes = sizeof(SampleT) * row_stride_samples;\n\n bool quad_aligned_rows = (NUM_CHANNELS == 1) && (SAMPLES_PER_THREAD % 4 == 0) && // Single channel\n ((size_t(d_native_samples) & quad_mask) == 0) && // ptr is quad-aligned\n ((num_rows == 1) || ((row_bytes & quad_mask) == 0)); // number of row-samples is a multiple of the alignment of the quad\n\n bool pixel_aligned_rows = (NUM_CHANNELS > 1) && // Multi channel\n ((size_t(d_native_samples) & pixel_mask) == 0) && // ptr is pixel-aligned\n ((row_bytes & pixel_mask) == 0); // number of row-samples is a multiple of the alignment of the pixel\n\n // Whether rows are aligned and can be vectorized\n if ((d_native_samples != NULL) && (quad_aligned_rows || pixel_aligned_rows))\n ConsumeTiles(num_row_pixels, num_rows, row_stride_samples, tiles_per_row, tile_queue, Int2Type());\n else\n ConsumeTiles(num_row_pixels, num_rows, row_stride_samples, tiles_per_row, tile_queue, Int2Type());\n }\n\n\n /**\n * Initialize privatized bin counters. Specialized for privatized shared-memory counters\n */\n __device__ __forceinline__ void InitBinCounters()\n {\n if (prefer_smem)\n InitSmemBinCounters();\n else\n InitGmemBinCounters();\n }\n\n\n /**\n * Store privatized histogram to device-accessible memory. Specialized for privatized shared-memory counters\n */\n __device__ __forceinline__ void StoreOutput()\n {\n if (prefer_smem)\n StoreSmemOutput();\n else\n StoreGmemOutput();\n }\n\n\n};\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_radix_sort_downsweep.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * AgentRadixSortDownsweep implements a stateful abstraction of CUDA thread blocks for participating in device-wide radix sort downsweep .\n */\n\n\n#pragma once\n\n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_radix_rank.cuh\"\n#include \"../block/block_exchange.cuh\"\n#include \"../util_type.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Radix ranking algorithm\n */\nenum RadixRankAlgorithm\n{\n RADIX_RANK_BASIC,\n RADIX_RANK_MEMOIZE,\n RADIX_RANK_MATCH\n};\n\n/**\n * Parameterizable tuning policy type for AgentRadixSortDownsweep\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading keys (and values)\n RadixRankAlgorithm _RANK_ALGORITHM, ///< The radix ranking algorithm to use\n BlockScanAlgorithm _SCAN_ALGORITHM, ///< The block scan algorithm to use\n int _RADIX_BITS> ///< The number of radix bits, i.e., log2(bins)\nstruct AgentRadixSortDownsweepPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n RADIX_BITS = _RADIX_BITS, ///< The number of radix bits, i.e., log2(bins)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading keys (and values)\n static const RadixRankAlgorithm RANK_ALGORITHM = _RANK_ALGORITHM; ///< The radix ranking algorithm to use\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n\n\n\n\n/**\n * \\brief AgentRadixSortDownsweep implements a stateful abstraction of CUDA thread blocks for participating in device-wide radix sort downsweep .\n */\ntemplate <\n typename AgentRadixSortDownsweepPolicy, ///< Parameterized AgentRadixSortDownsweepPolicy tuning policy type\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< KeyT type\n typename ValueT, ///< ValueT type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentRadixSortDownsweep\n{\n //---------------------------------------------------------------------\n // Type definitions and constants\n //---------------------------------------------------------------------\n\n // Appropriate unsigned-bits representation of KeyT\n typedef typename Traits::UnsignedBits UnsignedBits;\n\n static const UnsignedBits LOWEST_KEY = Traits::LOWEST_KEY;\n static const UnsignedBits MAX_KEY = Traits::MAX_KEY;\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = AgentRadixSortDownsweepPolicy::LOAD_ALGORITHM;\n static const CacheLoadModifier LOAD_MODIFIER = AgentRadixSortDownsweepPolicy::LOAD_MODIFIER;\n static const RadixRankAlgorithm RANK_ALGORITHM = AgentRadixSortDownsweepPolicy::RANK_ALGORITHM;\n static const BlockScanAlgorithm SCAN_ALGORITHM = AgentRadixSortDownsweepPolicy::SCAN_ALGORITHM;\n\n enum\n {\n BLOCK_THREADS = AgentRadixSortDownsweepPolicy::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentRadixSortDownsweepPolicy::ITEMS_PER_THREAD,\n RADIX_BITS = AgentRadixSortDownsweepPolicy::RADIX_BITS,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n\n RADIX_DIGITS = 1 << RADIX_BITS,\n KEYS_ONLY = Equals::VALUE,\n };\n\n // Input iterator wrapper type (for applying cache modifier)s\n typedef CacheModifiedInputIterator KeysItr;\n typedef CacheModifiedInputIterator ValuesItr;\n\n // Radix ranking type to use\n typedef typename If<(RANK_ALGORITHM == RADIX_RANK_BASIC),\n BlockRadixRank,\n typename If<(RANK_ALGORITHM == RADIX_RANK_MEMOIZE),\n BlockRadixRank,\n BlockRadixRankMatch\n >::Type\n >::Type BlockRadixRankT;\n\n enum\n {\n /// Number of bin-starting offsets tracked per thread\n BINS_TRACKED_PER_THREAD = BlockRadixRankT::BINS_TRACKED_PER_THREAD\n };\n\n // BlockLoad type (keys)\n typedef BlockLoad<\n UnsignedBits,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n LOAD_ALGORITHM> BlockLoadKeysT;\n\n // BlockLoad type (values)\n typedef BlockLoad<\n ValueT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n LOAD_ALGORITHM> BlockLoadValuesT;\n\n // Value exchange array type\n typedef ValueT ValueExchangeT[TILE_ITEMS];\n\n /**\n * Shared memory storage layout\n */\n union __align__(16) _TempStorage\n {\n typename BlockLoadKeysT::TempStorage load_keys;\n typename BlockLoadValuesT::TempStorage load_values;\n typename BlockRadixRankT::TempStorage radix_rank;\n\n struct\n {\n UnsignedBits exchange_keys[TILE_ITEMS];\n OffsetT relative_bin_offsets[RADIX_DIGITS];\n };\n\n Uninitialized exchange_values;\n\n OffsetT exclusive_digit_prefix[RADIX_DIGITS];\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n // Shared storage for this CTA\n _TempStorage &temp_storage;\n\n // Input and output device pointers\n KeysItr d_keys_in;\n ValuesItr d_values_in;\n UnsignedBits *d_keys_out;\n ValueT *d_values_out;\n\n // The global scatter base offset for each digit (valid in the first RADIX_DIGITS threads)\n OffsetT bin_offset[BINS_TRACKED_PER_THREAD];\n\n // The least-significant bit position of the current digit to extract\n int current_bit;\n\n // Number of bits in current digit\n int num_bits;\n\n // Whether to short-cirucit\n int short_circuit;\n\n //---------------------------------------------------------------------\n // Utility methods\n //---------------------------------------------------------------------\n\n\n /**\n * Scatter ranked keys through shared memory, then to device-accessible memory\n */\n template \n __device__ __forceinline__ void ScatterKeys(\n UnsignedBits (&twiddled_keys)[ITEMS_PER_THREAD],\n OffsetT (&relative_bin_offsets)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n OffsetT valid_items)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n temp_storage.exchange_keys[ranks[ITEM]] = twiddled_keys[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n UnsignedBits key = temp_storage.exchange_keys[threadIdx.x + (ITEM * BLOCK_THREADS)];\n UnsignedBits digit = BFE(key, current_bit, num_bits);\n relative_bin_offsets[ITEM] = temp_storage.relative_bin_offsets[digit];\n\n // Un-twiddle\n key = Traits::TwiddleOut(key);\n\n if (FULL_TILE || \n (static_cast(threadIdx.x + (ITEM * BLOCK_THREADS)) < valid_items))\n {\n d_keys_out[relative_bin_offsets[ITEM] + threadIdx.x + (ITEM * BLOCK_THREADS)] = key;\n }\n }\n }\n\n\n /**\n * Scatter ranked values through shared memory, then to device-accessible memory\n */\n template \n __device__ __forceinline__ void ScatterValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n OffsetT (&relative_bin_offsets)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n OffsetT valid_items)\n {\n CTA_SYNC();\n\n ValueExchangeT &exchange_values = temp_storage.exchange_values.Alias();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n exchange_values[ranks[ITEM]] = values[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n ValueT value = exchange_values[threadIdx.x + (ITEM * BLOCK_THREADS)];\n\n if (FULL_TILE || \n (static_cast(threadIdx.x + (ITEM * BLOCK_THREADS)) < valid_items))\n {\n d_values_out[relative_bin_offsets[ITEM] + threadIdx.x + (ITEM * BLOCK_THREADS)] = value;\n }\n }\n }\n\n /**\n * Load a tile of keys (specialized for full tile, any ranking algorithm)\n */\n template \n __device__ __forceinline__ void LoadKeys(\n UnsignedBits (&keys)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n UnsignedBits oob_item,\n Int2Type is_full_tile,\n Int2Type<_RANK_ALGORITHM> rank_algorithm)\n {\n BlockLoadKeysT(temp_storage.load_keys).Load(\n d_keys_in + block_offset, keys);\n\n CTA_SYNC();\n }\n\n\n /**\n * Load a tile of keys (specialized for partial tile, any ranking algorithm)\n */\n template \n __device__ __forceinline__ void LoadKeys(\n UnsignedBits (&keys)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n UnsignedBits oob_item,\n Int2Type is_full_tile,\n Int2Type<_RANK_ALGORITHM> rank_algorithm)\n {\n BlockLoadKeysT(temp_storage.load_keys).Load(\n d_keys_in + block_offset, keys, valid_items, oob_item);\n\n CTA_SYNC();\n }\n\n\n /**\n * Load a tile of keys (specialized for full tile, match ranking algorithm)\n */\n __device__ __forceinline__ void LoadKeys(\n UnsignedBits (&keys)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n UnsignedBits oob_item,\n Int2Type is_full_tile,\n Int2Type rank_algorithm)\n {\n LoadDirectWarpStriped(threadIdx.x, d_keys_in + block_offset, keys);\n }\n\n\n /**\n * Load a tile of keys (specialized for partial tile, match ranking algorithm)\n */\n __device__ __forceinline__ void LoadKeys(\n UnsignedBits (&keys)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n UnsignedBits oob_item,\n Int2Type is_full_tile,\n Int2Type rank_algorithm)\n {\n LoadDirectWarpStriped(threadIdx.x, d_keys_in + block_offset, keys, valid_items, oob_item);\n }\n\n\n /**\n * Load a tile of values (specialized for full tile, any ranking algorithm)\n */\n template \n __device__ __forceinline__ void LoadValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n Int2Type is_full_tile,\n Int2Type<_RANK_ALGORITHM> rank_algorithm)\n {\n BlockLoadValuesT(temp_storage.load_values).Load(\n d_values_in + block_offset, values);\n\n CTA_SYNC();\n }\n\n\n /**\n * Load a tile of values (specialized for partial tile, any ranking algorithm)\n */\n template \n __device__ __forceinline__ void LoadValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n Int2Type is_full_tile,\n Int2Type<_RANK_ALGORITHM> rank_algorithm)\n {\n BlockLoadValuesT(temp_storage.load_values).Load(\n d_values_in + block_offset, values, valid_items);\n\n CTA_SYNC();\n }\n\n\n /**\n * Load a tile of items (specialized for full tile, match ranking algorithm)\n */\n __device__ __forceinline__ void LoadValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n volatile OffsetT valid_items,\n Int2Type is_full_tile,\n Int2Type rank_algorithm)\n {\n LoadDirectWarpStriped(threadIdx.x, d_values_in + block_offset, values);\n }\n\n\n /**\n * Load a tile of items (specialized for partial tile, match ranking algorithm)\n */\n __device__ __forceinline__ void LoadValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n volatile OffsetT valid_items,\n Int2Type is_full_tile,\n Int2Type rank_algorithm)\n {\n LoadDirectWarpStriped(threadIdx.x, d_values_in + block_offset, values, valid_items);\n }\n\n\n /**\n * Truck along associated values\n */\n template \n __device__ __forceinline__ void GatherScatterValues(\n OffsetT (&relative_bin_offsets)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n OffsetT block_offset,\n OffsetT valid_items,\n Int2Type /*is_keys_only*/)\n {\n CTA_SYNC();\n\n ValueT values[ITEMS_PER_THREAD];\n\n LoadValues(\n values,\n block_offset,\n valid_items,\n Int2Type(),\n Int2Type());\n\n ScatterValues(\n values,\n relative_bin_offsets,\n ranks,\n valid_items);\n }\n\n\n /**\n * Truck along associated values (specialized for key-only sorting)\n */\n template \n __device__ __forceinline__ void GatherScatterValues(\n OffsetT (&/*relative_bin_offsets*/)[ITEMS_PER_THREAD],\n int (&/*ranks*/)[ITEMS_PER_THREAD],\n OffsetT /*block_offset*/,\n OffsetT /*valid_items*/,\n Int2Type /*is_keys_only*/)\n {}\n\n\n /**\n * Process tile\n */\n template \n __device__ __forceinline__ void ProcessTile(\n OffsetT block_offset,\n const OffsetT &valid_items = TILE_ITEMS)\n {\n UnsignedBits keys[ITEMS_PER_THREAD];\n int ranks[ITEMS_PER_THREAD];\n OffsetT relative_bin_offsets[ITEMS_PER_THREAD];\n\n // Assign default (min/max) value to all keys\n UnsignedBits default_key = (IS_DESCENDING) ? LOWEST_KEY : MAX_KEY;\n\n // Load tile of keys\n LoadKeys(\n keys,\n block_offset,\n valid_items, \n default_key,\n Int2Type(),\n Int2Type());\n\n // Twiddle key bits if necessary\n #pragma unroll\n for (int KEY = 0; KEY < ITEMS_PER_THREAD; KEY++)\n {\n keys[KEY] = Traits::TwiddleIn(keys[KEY]);\n }\n\n // Rank the twiddled keys\n int exclusive_digit_prefix[BINS_TRACKED_PER_THREAD];\n BlockRadixRankT(temp_storage.radix_rank).RankKeys(\n keys,\n ranks,\n current_bit,\n num_bits,\n exclusive_digit_prefix);\n\n CTA_SYNC();\n\n // Share exclusive digit prefix\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n // Store exclusive prefix\n temp_storage.exclusive_digit_prefix[bin_idx] =\n exclusive_digit_prefix[track];\n }\n }\n\n CTA_SYNC();\n\n // Get inclusive digit prefix\n int inclusive_digit_prefix[BINS_TRACKED_PER_THREAD];\n\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n if (IS_DESCENDING)\n {\n // Get inclusive digit prefix from exclusive prefix (higher bins come first)\n inclusive_digit_prefix[track] = (bin_idx == 0) ?\n (BLOCK_THREADS * ITEMS_PER_THREAD) :\n temp_storage.exclusive_digit_prefix[bin_idx - 1];\n }\n else\n {\n // Get inclusive digit prefix from exclusive prefix (lower bins come first)\n inclusive_digit_prefix[track] = (bin_idx == RADIX_DIGITS - 1) ?\n (BLOCK_THREADS * ITEMS_PER_THREAD) :\n temp_storage.exclusive_digit_prefix[bin_idx + 1];\n }\n }\n }\n\n CTA_SYNC();\n\n // Update global scatter base offsets for each digit\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n bin_offset[track] -= exclusive_digit_prefix[track];\n temp_storage.relative_bin_offsets[bin_idx] = bin_offset[track];\n bin_offset[track] += inclusive_digit_prefix[track];\n }\n }\n\n CTA_SYNC();\n\n // Scatter keys\n ScatterKeys(keys, relative_bin_offsets, ranks, valid_items);\n\n // Gather/scatter values\n GatherScatterValues(relative_bin_offsets , ranks, block_offset, valid_items, Int2Type());\n }\n\n //---------------------------------------------------------------------\n // Copy shortcut\n //---------------------------------------------------------------------\n\n /**\n * Copy tiles within the range of input\n */\n template <\n typename InputIteratorT,\n typename T>\n __device__ __forceinline__ void Copy(\n InputIteratorT d_in,\n T *d_out,\n OffsetT block_offset,\n OffsetT block_end)\n {\n // Simply copy the input\n while (block_offset + TILE_ITEMS <= block_end)\n {\n T items[ITEMS_PER_THREAD];\n\n LoadDirectStriped(threadIdx.x, d_in + block_offset, items);\n CTA_SYNC();\n StoreDirectStriped(threadIdx.x, d_out + block_offset, items);\n\n block_offset += TILE_ITEMS;\n }\n\n // Clean up last partial tile with guarded-I/O\n if (block_offset < block_end)\n {\n OffsetT valid_items = block_end - block_offset;\n\n T items[ITEMS_PER_THREAD];\n\n LoadDirectStriped(threadIdx.x, d_in + block_offset, items, valid_items);\n CTA_SYNC();\n StoreDirectStriped(threadIdx.x, d_out + block_offset, items, valid_items);\n }\n }\n\n\n /**\n * Copy tiles within the range of input (specialized for NullType)\n */\n template \n __device__ __forceinline__ void Copy(\n InputIteratorT /*d_in*/,\n NullType * /*d_out*/,\n OffsetT /*block_offset*/,\n OffsetT /*block_end*/)\n {}\n\n\n //---------------------------------------------------------------------\n // Interface\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentRadixSortDownsweep(\n TempStorage &temp_storage,\n OffsetT (&bin_offset)[BINS_TRACKED_PER_THREAD],\n OffsetT num_items,\n const KeyT *d_keys_in,\n KeyT *d_keys_out,\n const ValueT *d_values_in,\n ValueT *d_values_out,\n int current_bit,\n int num_bits)\n :\n temp_storage(temp_storage.Alias()),\n d_keys_in(reinterpret_cast(d_keys_in)),\n d_values_in(d_values_in),\n d_keys_out(reinterpret_cast(d_keys_out)),\n d_values_out(d_values_out),\n current_bit(current_bit),\n num_bits(num_bits),\n short_circuit(1)\n {\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n this->bin_offset[track] = bin_offset[track];\n\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n // Short circuit if the histogram has only bin counts of only zeros or problem-size\n short_circuit = short_circuit && ((bin_offset[track] == 0) || (bin_offset[track] == num_items));\n }\n }\n\n short_circuit = CTA_SYNC_AND(short_circuit);\n }\n\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentRadixSortDownsweep(\n TempStorage &temp_storage,\n OffsetT num_items,\n OffsetT *d_spine,\n const KeyT *d_keys_in,\n KeyT *d_keys_out,\n const ValueT *d_values_in,\n ValueT *d_values_out,\n int current_bit,\n int num_bits)\n :\n temp_storage(temp_storage.Alias()),\n d_keys_in(reinterpret_cast(d_keys_in)),\n d_values_in(d_values_in),\n d_keys_out(reinterpret_cast(d_keys_out)),\n d_values_out(d_values_out),\n current_bit(current_bit),\n num_bits(num_bits),\n short_circuit(1)\n {\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n // Load digit bin offsets (each of the first RADIX_DIGITS threads will load an offset for that digit)\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n if (IS_DESCENDING)\n bin_idx = RADIX_DIGITS - bin_idx - 1;\n\n // Short circuit if the first block's histogram has only bin counts of only zeros or problem-size\n OffsetT first_block_bin_offset = d_spine[gridDim.x * bin_idx];\n short_circuit = short_circuit && ((first_block_bin_offset == 0) || (first_block_bin_offset == num_items));\n\n // Load my block's bin offset for my bin\n bin_offset[track] = d_spine[(gridDim.x * bin_idx) + blockIdx.x];\n }\n }\n\n short_circuit = CTA_SYNC_AND(short_circuit);\n }\n\n\n /**\n * Distribute keys from a segment of input tiles.\n */\n __device__ __forceinline__ void ProcessRegion(\n OffsetT block_offset,\n OffsetT block_end)\n {\n if (short_circuit)\n {\n // Copy keys\n Copy(d_keys_in, d_keys_out, block_offset, block_end);\n\n // Copy values\n Copy(d_values_in, d_values_out, block_offset, block_end);\n }\n else\n {\n // Process full tiles of tile_items\n while (block_offset + TILE_ITEMS <= block_end)\n {\n ProcessTile(block_offset);\n block_offset += TILE_ITEMS;\n\n CTA_SYNC();\n }\n\n // Clean up last partial tile with guarded-I/O\n if (block_offset < block_end)\n {\n ProcessTile(block_offset, block_end - block_offset);\n }\n\n }\n }\n\n};\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_radix_sort_upsweep.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * AgentRadixSortUpsweep implements a stateful abstraction of CUDA thread blocks for participating in device-wide radix sort upsweep .\n */\n\n#pragma once\n\n#include \"../thread/thread_reduce.cuh\"\n#include \"../thread/thread_load.cuh\"\n#include \"../warp/warp_reduce.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../util_type.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentRadixSortUpsweep\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading keys\n int _RADIX_BITS> ///< The number of radix bits, i.e., log2(bins)\nstruct AgentRadixSortUpsweepPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n RADIX_BITS = _RADIX_BITS, ///< The number of radix bits, i.e., log2(bins)\n };\n\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading keys\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentRadixSortUpsweep implements a stateful abstraction of CUDA thread blocks for participating in device-wide radix sort upsweep .\n */\ntemplate <\n typename AgentRadixSortUpsweepPolicy, ///< Parameterized AgentRadixSortUpsweepPolicy tuning policy type\n typename KeyT, ///< KeyT type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentRadixSortUpsweep\n{\n\n //---------------------------------------------------------------------\n // Type definitions and constants\n //---------------------------------------------------------------------\n\n typedef typename Traits::UnsignedBits UnsignedBits;\n\n // Integer type for digit counters (to be packed into words of PackedCounters)\n typedef unsigned char DigitCounter;\n\n // Integer type for packing DigitCounters into columns of shared memory banks\n typedef unsigned int PackedCounter;\n\n static const CacheLoadModifier LOAD_MODIFIER = AgentRadixSortUpsweepPolicy::LOAD_MODIFIER;\n\n enum\n {\n RADIX_BITS = AgentRadixSortUpsweepPolicy::RADIX_BITS,\n BLOCK_THREADS = AgentRadixSortUpsweepPolicy::BLOCK_THREADS,\n KEYS_PER_THREAD = AgentRadixSortUpsweepPolicy::ITEMS_PER_THREAD,\n\n RADIX_DIGITS = 1 << RADIX_BITS,\n\n LOG_WARP_THREADS = CUB_PTX_LOG_WARP_THREADS,\n WARP_THREADS = 1 << LOG_WARP_THREADS,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n TILE_ITEMS = BLOCK_THREADS * KEYS_PER_THREAD,\n\n BYTES_PER_COUNTER = sizeof(DigitCounter),\n LOG_BYTES_PER_COUNTER = Log2::VALUE,\n\n PACKING_RATIO = sizeof(PackedCounter) / sizeof(DigitCounter),\n LOG_PACKING_RATIO = Log2::VALUE,\n\n LOG_COUNTER_LANES = CUB_MAX(0, RADIX_BITS - LOG_PACKING_RATIO),\n COUNTER_LANES = 1 << LOG_COUNTER_LANES,\n\n // To prevent counter overflow, we must periodically unpack and aggregate the\n // digit counters back into registers. Each counter lane is assigned to a\n // warp for aggregation.\n\n LANES_PER_WARP = CUB_MAX(1, (COUNTER_LANES + WARPS - 1) / WARPS),\n\n // Unroll tiles in batches without risk of counter overflow\n UNROLL_COUNT = CUB_MIN(64, 255 / KEYS_PER_THREAD),\n UNROLLED_ELEMENTS = UNROLL_COUNT * TILE_ITEMS,\n };\n\n\n // Input iterator wrapper type (for applying cache modifier)s\n typedef CacheModifiedInputIterator KeysItr;\n\n /**\n * Shared memory storage layout\n */\n union __align__(16) _TempStorage\n {\n DigitCounter thread_counters[COUNTER_LANES][BLOCK_THREADS][PACKING_RATIO];\n PackedCounter packed_thread_counters[COUNTER_LANES][BLOCK_THREADS];\n OffsetT block_counters[WARP_THREADS][RADIX_DIGITS];\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Thread fields (aggregate state bundle)\n //---------------------------------------------------------------------\n\n // Shared storage for this CTA\n _TempStorage &temp_storage;\n\n // Thread-local counters for periodically aggregating composite-counter lanes\n OffsetT local_counts[LANES_PER_WARP][PACKING_RATIO];\n\n // Input and output device pointers\n KeysItr d_keys_in;\n\n // The least-significant bit position of the current digit to extract\n int current_bit;\n\n // Number of bits in current digit\n int num_bits;\n\n\n\n //---------------------------------------------------------------------\n // Helper structure for templated iteration\n //---------------------------------------------------------------------\n\n // Iterate\n template \n struct Iterate\n {\n // BucketKeys\n static __device__ __forceinline__ void BucketKeys(\n AgentRadixSortUpsweep &cta,\n UnsignedBits keys[KEYS_PER_THREAD])\n {\n cta.Bucket(keys[COUNT]);\n\n // Next\n Iterate::BucketKeys(cta, keys);\n }\n };\n\n // Terminate\n template \n struct Iterate\n {\n // BucketKeys\n static __device__ __forceinline__ void BucketKeys(AgentRadixSortUpsweep &/*cta*/, UnsignedBits /*keys*/[KEYS_PER_THREAD]) {}\n };\n\n\n //---------------------------------------------------------------------\n // Utility methods\n //---------------------------------------------------------------------\n\n /**\n * Decode a key and increment corresponding smem digit counter\n */\n __device__ __forceinline__ void Bucket(UnsignedBits key)\n {\n // Perform transform op\n UnsignedBits converted_key = Traits::TwiddleIn(key);\n\n // Extract current digit bits\n UnsignedBits digit = BFE(converted_key, current_bit, num_bits);\n\n // Get sub-counter offset\n UnsignedBits sub_counter = digit & (PACKING_RATIO - 1);\n\n // Get row offset\n UnsignedBits row_offset = digit >> LOG_PACKING_RATIO;\n\n // Increment counter\n temp_storage.thread_counters[row_offset][threadIdx.x][sub_counter]++;\n }\n\n\n /**\n * Reset composite counters\n */\n __device__ __forceinline__ void ResetDigitCounters()\n {\n #pragma unroll\n for (int LANE = 0; LANE < COUNTER_LANES; LANE++)\n {\n temp_storage.packed_thread_counters[LANE][threadIdx.x] = 0;\n }\n }\n\n\n /**\n * Reset the unpacked counters in each thread\n */\n __device__ __forceinline__ void ResetUnpackedCounters()\n {\n #pragma unroll\n for (int LANE = 0; LANE < LANES_PER_WARP; LANE++)\n {\n #pragma unroll\n for (int UNPACKED_COUNTER = 0; UNPACKED_COUNTER < PACKING_RATIO; UNPACKED_COUNTER++)\n {\n local_counts[LANE][UNPACKED_COUNTER] = 0;\n }\n }\n }\n\n\n /**\n * Extracts and aggregates the digit counters for each counter lane\n * owned by this warp\n */\n __device__ __forceinline__ void UnpackDigitCounts()\n {\n unsigned int warp_id = threadIdx.x >> LOG_WARP_THREADS;\n unsigned int warp_tid = LaneId();\n\n #pragma unroll\n for (int LANE = 0; LANE < LANES_PER_WARP; LANE++)\n {\n const int counter_lane = (LANE * WARPS) + warp_id;\n if (counter_lane < COUNTER_LANES)\n {\n #pragma unroll\n for (int PACKED_COUNTER = 0; PACKED_COUNTER < BLOCK_THREADS; PACKED_COUNTER += WARP_THREADS)\n {\n #pragma unroll\n for (int UNPACKED_COUNTER = 0; UNPACKED_COUNTER < PACKING_RATIO; UNPACKED_COUNTER++)\n {\n OffsetT counter = temp_storage.thread_counters[counter_lane][warp_tid + PACKED_COUNTER][UNPACKED_COUNTER];\n local_counts[LANE][UNPACKED_COUNTER] += counter;\n }\n }\n }\n }\n }\n\n\n /**\n * Processes a single, full tile\n */\n __device__ __forceinline__ void ProcessFullTile(OffsetT block_offset)\n {\n // Tile of keys\n UnsignedBits keys[KEYS_PER_THREAD];\n\n LoadDirectStriped(threadIdx.x, d_keys_in + block_offset, keys);\n\n // Prevent hoisting\n CTA_SYNC();\n\n // Bucket tile of keys\n Iterate<0, KEYS_PER_THREAD>::BucketKeys(*this, keys);\n }\n\n\n /**\n * Processes a single load (may have some threads masked off)\n */\n __device__ __forceinline__ void ProcessPartialTile(\n OffsetT block_offset,\n const OffsetT &block_end)\n {\n // Process partial tile if necessary using single loads\n block_offset += threadIdx.x;\n while (block_offset < block_end)\n {\n // Load and bucket key\n UnsignedBits key = d_keys_in[block_offset];\n Bucket(key);\n block_offset += BLOCK_THREADS;\n }\n }\n\n\n //---------------------------------------------------------------------\n // Interface\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentRadixSortUpsweep(\n TempStorage &temp_storage,\n const KeyT *d_keys_in,\n int current_bit,\n int num_bits)\n :\n temp_storage(temp_storage.Alias()),\n d_keys_in(reinterpret_cast(d_keys_in)),\n current_bit(current_bit),\n num_bits(num_bits)\n {}\n\n\n /**\n * Compute radix digit histograms from a segment of input tiles.\n */\n __device__ __forceinline__ void ProcessRegion(\n OffsetT block_offset,\n const OffsetT &block_end)\n {\n // Reset digit counters in smem and unpacked counters in registers\n ResetDigitCounters();\n ResetUnpackedCounters();\n\n // Unroll batches of full tiles\n while (block_offset + UNROLLED_ELEMENTS <= block_end)\n {\n for (int i = 0; i < UNROLL_COUNT; ++i)\n {\n ProcessFullTile(block_offset);\n block_offset += TILE_ITEMS;\n }\n\n CTA_SYNC();\n\n // Aggregate back into local_count registers to prevent overflow\n UnpackDigitCounts();\n\n CTA_SYNC();\n\n // Reset composite counters in lanes\n ResetDigitCounters();\n }\n\n // Unroll single full tiles\n while (block_offset + TILE_ITEMS <= block_end)\n {\n ProcessFullTile(block_offset);\n block_offset += TILE_ITEMS;\n }\n\n // Process partial tile if necessary\n ProcessPartialTile(\n block_offset,\n block_end);\n\n CTA_SYNC();\n\n // Aggregate back into local_count registers\n UnpackDigitCounts();\n }\n\n\n /**\n * Extract counts (saving them to the external array)\n */\n template \n __device__ __forceinline__ void ExtractCounts(\n OffsetT *counters,\n int bin_stride = 1,\n int bin_offset = 0)\n {\n unsigned int warp_id = threadIdx.x >> LOG_WARP_THREADS;\n unsigned int warp_tid = LaneId();\n\n // Place unpacked digit counters in shared memory\n #pragma unroll\n for (int LANE = 0; LANE < LANES_PER_WARP; LANE++)\n {\n int counter_lane = (LANE * WARPS) + warp_id;\n if (counter_lane < COUNTER_LANES)\n {\n int digit_row = counter_lane << LOG_PACKING_RATIO;\n\n #pragma unroll\n for (int UNPACKED_COUNTER = 0; UNPACKED_COUNTER < PACKING_RATIO; UNPACKED_COUNTER++)\n {\n int bin_idx = digit_row + UNPACKED_COUNTER;\n\n temp_storage.block_counters[warp_tid][bin_idx] =\n local_counts[LANE][UNPACKED_COUNTER];\n }\n }\n }\n\n CTA_SYNC();\n\n // Rake-reduce bin_count reductions\n\n // Whole blocks\n #pragma unroll\n for (int BIN_BASE = RADIX_DIGITS % BLOCK_THREADS;\n (BIN_BASE + BLOCK_THREADS) <= RADIX_DIGITS;\n BIN_BASE += BLOCK_THREADS)\n {\n int bin_idx = BIN_BASE + threadIdx.x;\n\n OffsetT bin_count = 0;\n #pragma unroll\n for (int i = 0; i < WARP_THREADS; ++i)\n bin_count += temp_storage.block_counters[i][bin_idx];\n\n if (IS_DESCENDING)\n bin_idx = RADIX_DIGITS - bin_idx - 1;\n\n counters[(bin_stride * bin_idx) + bin_offset] = bin_count;\n }\n\n // Remainder\n if ((RADIX_DIGITS % BLOCK_THREADS != 0) && (threadIdx.x < RADIX_DIGITS))\n {\n int bin_idx = threadIdx.x;\n\n OffsetT bin_count = 0;\n #pragma unroll\n for (int i = 0; i < WARP_THREADS; ++i)\n bin_count += temp_storage.block_counters[i][bin_idx];\n\n if (IS_DESCENDING)\n bin_idx = RADIX_DIGITS - bin_idx - 1;\n\n counters[(bin_stride * bin_idx) + bin_offset] = bin_count;\n }\n }\n\n\n /**\n * Extract counts\n */\n template \n __device__ __forceinline__ void ExtractCounts(\n OffsetT (&bin_count)[BINS_TRACKED_PER_THREAD]) ///< [out] The exclusive prefix sum for the digits [(threadIdx.x * BINS_TRACKED_PER_THREAD) ... (threadIdx.x * BINS_TRACKED_PER_THREAD) + BINS_TRACKED_PER_THREAD - 1]\n {\n unsigned int warp_id = threadIdx.x >> LOG_WARP_THREADS;\n unsigned int warp_tid = LaneId();\n\n // Place unpacked digit counters in shared memory\n #pragma unroll\n for (int LANE = 0; LANE < LANES_PER_WARP; LANE++)\n {\n int counter_lane = (LANE * WARPS) + warp_id;\n if (counter_lane < COUNTER_LANES)\n {\n int digit_row = counter_lane << LOG_PACKING_RATIO;\n\n #pragma unroll\n for (int UNPACKED_COUNTER = 0; UNPACKED_COUNTER < PACKING_RATIO; UNPACKED_COUNTER++)\n {\n int bin_idx = digit_row + UNPACKED_COUNTER;\n\n temp_storage.block_counters[warp_tid][bin_idx] =\n local_counts[LANE][UNPACKED_COUNTER];\n }\n }\n }\n\n CTA_SYNC();\n\n // Rake-reduce bin_count reductions\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n bin_count[track] = 0;\n\n #pragma unroll\n for (int i = 0; i < WARP_THREADS; ++i)\n bin_count[track] += temp_storage.block_counters[i][bin_idx];\n }\n }\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_reduce.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentReduce implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduction .\n */\n\n#pragma once\n\n#include \n\n#include \"../block/block_load.cuh\"\n#include \"../block/block_reduce.cuh\"\n#include \"../grid/grid_mapping.cuh\"\n#include \"../grid/grid_even_share.cuh\"\n#include \"../util_type.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentReduce\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n int _VECTOR_LOAD_LENGTH, ///< Number of items per vectorized load\n BlockReduceAlgorithm _BLOCK_ALGORITHM, ///< Cooperative block-wide reduction algorithm to use\n CacheLoadModifier _LOAD_MODIFIER> ///< Cache load modifier for reading input elements\nstruct AgentReducePolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n VECTOR_LOAD_LENGTH = _VECTOR_LOAD_LENGTH, ///< Number of items per vectorized load\n };\n\n static const BlockReduceAlgorithm BLOCK_ALGORITHM = _BLOCK_ALGORITHM; ///< Cooperative block-wide reduction algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n};\n\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentReduce implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduction .\n *\n * Each thread reduces only the values it loads. If \\p FIRST_TILE, this\n * partial reduction is stored into \\p thread_aggregate. Otherwise it is\n * accumulated into \\p thread_aggregate.\n */\ntemplate <\n typename AgentReducePolicy, ///< Parameterized AgentReducePolicy tuning policy type\n typename InputIteratorT, ///< Random-access iterator type for input\n typename OutputIteratorT, ///< Random-access iterator type for output\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOp> ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\nstruct AgentReduce\n{\n\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n /// The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n /// Vector type of InputT for data movement\n typedef typename CubVector::Type VectorT;\n\n /// Input iterator wrapper type (for applying cache modifier)\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n InputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedInputIteratorT;\n\n /// Constants\n enum\n {\n BLOCK_THREADS = AgentReducePolicy::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentReducePolicy::ITEMS_PER_THREAD,\n VECTOR_LOAD_LENGTH = CUB_MIN(ITEMS_PER_THREAD, AgentReducePolicy::VECTOR_LOAD_LENGTH),\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n\n // Can vectorize according to the policy if the input iterator is a native pointer to a primitive type\n ATTEMPT_VECTORIZATION = (VECTOR_LOAD_LENGTH > 1) &&\n (ITEMS_PER_THREAD % VECTOR_LOAD_LENGTH == 0) &&\n (IsPointer::VALUE) && Traits::PRIMITIVE,\n\n };\n\n static const CacheLoadModifier LOAD_MODIFIER = AgentReducePolicy::LOAD_MODIFIER;\n static const BlockReduceAlgorithm BLOCK_ALGORITHM = AgentReducePolicy::BLOCK_ALGORITHM;\n\n /// Parameterized BlockReduce primitive\n typedef BlockReduce BlockReduceT;\n\n /// Shared memory type required by this thread block\n struct _TempStorage\n {\n typename BlockReduceT::TempStorage reduce;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n InputIteratorT d_in; ///< Input data to reduce\n WrappedInputIteratorT d_wrapped_in; ///< Wrapped input data to reduce\n ReductionOp reduction_op; ///< Binary reduction operator\n\n\n //---------------------------------------------------------------------\n // Utility\n //---------------------------------------------------------------------\n\n\n // Whether or not the input is aligned with the vector type (specialized for types we can vectorize)\n template \n static __device__ __forceinline__ bool IsAligned(\n Iterator d_in,\n Int2Type /*can_vectorize*/)\n {\n return (size_t(d_in) & (sizeof(VectorT) - 1)) == 0;\n }\n\n // Whether or not the input is aligned with the vector type (specialized for types we cannot vectorize)\n template \n static __device__ __forceinline__ bool IsAligned(\n Iterator /*d_in*/,\n Int2Type /*can_vectorize*/)\n {\n return false;\n }\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentReduce(\n TempStorage& temp_storage, ///< Reference to temp_storage\n InputIteratorT d_in, ///< Input data to reduce\n ReductionOp reduction_op) ///< Binary reduction operator\n :\n temp_storage(temp_storage.Alias()),\n d_in(d_in),\n d_wrapped_in(d_in),\n reduction_op(reduction_op)\n {}\n\n\n //---------------------------------------------------------------------\n // Tile consumption\n //---------------------------------------------------------------------\n\n /**\n * Consume a full tile of input (non-vectorized)\n */\n template \n __device__ __forceinline__ void ConsumeTile(\n OutputT &thread_aggregate,\n OffsetT block_offset, ///< The offset the tile to consume\n int /*valid_items*/, ///< The number of valid items in the tile\n Int2Type /*is_full_tile*/, ///< Whether or not this is a full tile\n Int2Type /*can_vectorize*/) ///< Whether or not we can vectorize loads\n {\n OutputT items[ITEMS_PER_THREAD];\n\n // Load items in striped fashion\n LoadDirectStriped(threadIdx.x, d_wrapped_in + block_offset, items);\n\n // Reduce items within each thread stripe\n thread_aggregate = (IS_FIRST_TILE) ?\n internal::ThreadReduce(items, reduction_op) :\n internal::ThreadReduce(items, reduction_op, thread_aggregate);\n }\n\n\n /**\n * Consume a full tile of input (vectorized)\n */\n template \n __device__ __forceinline__ void ConsumeTile(\n OutputT &thread_aggregate,\n OffsetT block_offset, ///< The offset the tile to consume\n int /*valid_items*/, ///< The number of valid items in the tile\n Int2Type /*is_full_tile*/, ///< Whether or not this is a full tile\n Int2Type /*can_vectorize*/) ///< Whether or not we can vectorize loads\n {\n // Alias items as an array of VectorT and load it in striped fashion\n enum { WORDS = ITEMS_PER_THREAD / VECTOR_LOAD_LENGTH };\n\n // Fabricate a vectorized input iterator\n InputT *d_in_unqualified = const_cast(d_in) + block_offset + (threadIdx.x * VECTOR_LOAD_LENGTH);\n CacheModifiedInputIterator d_vec_in(\n reinterpret_cast(d_in_unqualified));\n\n // Load items as vector items\n InputT input_items[ITEMS_PER_THREAD];\n VectorT *vec_items = reinterpret_cast(input_items);\n #pragma unroll\n for (int i = 0; i < WORDS; ++i)\n vec_items[i] = d_vec_in[BLOCK_THREADS * i];\n\n // Convert from input type to output type\n OutputT items[ITEMS_PER_THREAD];\n #pragma unroll\n for (int i = 0; i < ITEMS_PER_THREAD; ++i)\n items[i] = input_items[i];\n\n // Reduce items within each thread stripe\n thread_aggregate = (IS_FIRST_TILE) ?\n internal::ThreadReduce(items, reduction_op) :\n internal::ThreadReduce(items, reduction_op, thread_aggregate);\n }\n\n\n /**\n * Consume a partial tile of input\n */\n template \n __device__ __forceinline__ void ConsumeTile(\n OutputT &thread_aggregate,\n OffsetT block_offset, ///< The offset the tile to consume\n int valid_items, ///< The number of valid items in the tile\n Int2Type /*is_full_tile*/, ///< Whether or not this is a full tile\n Int2Type /*can_vectorize*/) ///< Whether or not we can vectorize loads\n {\n // Partial tile\n int thread_offset = threadIdx.x;\n\n // Read first item\n if ((IS_FIRST_TILE) && (thread_offset < valid_items))\n {\n thread_aggregate = d_wrapped_in[block_offset + thread_offset];\n thread_offset += BLOCK_THREADS;\n }\n\n // Continue reading items (block-striped)\n while (thread_offset < valid_items)\n {\n OutputT item = d_wrapped_in[block_offset + thread_offset];\n thread_aggregate = reduction_op(thread_aggregate, item);\n thread_offset += BLOCK_THREADS;\n }\n }\n\n\n //---------------------------------------------------------------\n // Consume a contiguous segment of tiles\n //---------------------------------------------------------------------\n\n /**\n * \\brief Reduce a contiguous segment of input tiles\n */\n template \n __device__ __forceinline__ OutputT ConsumeRange(\n GridEvenShare &even_share, ///< GridEvenShare descriptor\n Int2Type can_vectorize) ///< Whether or not we can vectorize loads\n {\n OutputT thread_aggregate;\n\n if (even_share.block_offset + TILE_ITEMS > even_share.block_end)\n {\n // First tile isn't full (not all threads have valid items)\n int valid_items = even_share.block_end - even_share.block_offset;\n ConsumeTile(thread_aggregate, even_share.block_offset, valid_items, Int2Type(), can_vectorize);\n return BlockReduceT(temp_storage.reduce).Reduce(thread_aggregate, reduction_op, valid_items);\n }\n\n // At least one full block\n ConsumeTile(thread_aggregate, even_share.block_offset, TILE_ITEMS, Int2Type(), can_vectorize);\n even_share.block_offset += even_share.block_stride;\n\n // Consume subsequent full tiles of input\n while (even_share.block_offset + TILE_ITEMS <= even_share.block_end)\n {\n ConsumeTile(thread_aggregate, even_share.block_offset, TILE_ITEMS, Int2Type(), can_vectorize);\n even_share.block_offset += even_share.block_stride;\n }\n\n // Consume a partially-full tile\n if (even_share.block_offset < even_share.block_end)\n {\n int valid_items = even_share.block_end - even_share.block_offset;\n ConsumeTile(thread_aggregate, even_share.block_offset, valid_items, Int2Type(), can_vectorize);\n }\n\n // Compute block-wide reduction (all threads have valid items)\n return BlockReduceT(temp_storage.reduce).Reduce(thread_aggregate, reduction_op);\n }\n\n\n /**\n * \\brief Reduce a contiguous segment of input tiles\n */\n __device__ __forceinline__ OutputT ConsumeRange(\n OffsetT block_offset, ///< [in] Threadblock begin offset (inclusive)\n OffsetT block_end) ///< [in] Threadblock end offset (exclusive)\n {\n GridEvenShare even_share;\n even_share.template BlockInit(block_offset, block_end);\n\n return (IsAligned(d_in + block_offset, Int2Type())) ?\n ConsumeRange(even_share, Int2Type()) :\n ConsumeRange(even_share, Int2Type());\n }\n\n\n /**\n * Reduce a contiguous segment of input tiles\n */\n __device__ __forceinline__ OutputT ConsumeTiles(\n GridEvenShare &even_share) ///< [in] GridEvenShare descriptor\n {\n // Initialize GRID_MAPPING_STRIP_MINE even-share descriptor for this thread block\n even_share.template BlockInit();\n\n return (IsAligned(d_in, Int2Type())) ?\n ConsumeRange(even_share, Int2Type()) :\n ConsumeRange(even_share, Int2Type());\n\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_reduce_by_key.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentReduceByKey implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduce-value-by-key.\n */\n\n#pragma once\n\n#include \n\n#include \"single_pass_scan_operators.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../block/block_discontinuity.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../iterator/constant_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentReduceByKey\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentReduceByKeyPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentReduceByKey implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduce-value-by-key\n */\ntemplate <\n typename AgentReduceByKeyPolicyT, ///< Parameterized AgentReduceByKeyPolicy tuning policy type\n typename KeysInputIteratorT, ///< Random-access input iterator type for keys\n typename UniqueOutputIteratorT, ///< Random-access output iterator type for keys\n typename ValuesInputIteratorT, ///< Random-access input iterator type for values\n typename AggregatesOutputIteratorT, ///< Random-access output iterator type for values\n typename NumRunsOutputIteratorT, ///< Output iterator type for recording number of items selected\n typename EqualityOpT, ///< KeyT equality operator type\n typename ReductionOpT, ///< ValueT reduction operator type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentReduceByKey\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // The input keys type\n typedef typename std::iterator_traits::value_type KeyInputT;\n\n // The output keys type\n typedef typename If<(Equals::value_type, void>::VALUE), // KeyOutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type KeyOutputT; // ... else the output iterator's value type\n\n // The input values type\n typedef typename std::iterator_traits::value_type ValueInputT;\n\n // The output values type\n typedef typename If<(Equals::value_type, void>::VALUE), // ValueOutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type ValueOutputT; // ... else the output iterator's value type\n\n // Tuple type for scanning (pairs accumulated segment-value with segment-index)\n typedef KeyValuePair OffsetValuePairT;\n\n // Tuple type for pairing keys and values\n typedef KeyValuePair KeyValuePairT;\n\n // Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n // Guarded inequality functor\n template \n struct GuardedInequalityWrapper\n {\n _EqualityOpT op; ///< Wrapped equality operator\n int num_remaining; ///< Items remaining\n\n /// Constructor\n __host__ __device__ __forceinline__\n GuardedInequalityWrapper(_EqualityOpT op, int num_remaining) : op(op), num_remaining(num_remaining) {}\n\n /// Boolean inequality operator, returns (a != b)\n template \n __host__ __device__ __forceinline__ bool operator()(const T &a, const T &b, int idx) const\n {\n if (idx < num_remaining)\n return !op(a, b); // In bounds\n\n // Return true if first out-of-bounds item, false otherwise\n return (idx == num_remaining);\n }\n };\n\n\n // Constants\n enum\n {\n BLOCK_THREADS = AgentReduceByKeyPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentReduceByKeyPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n TWO_PHASE_SCATTER = (ITEMS_PER_THREAD > 1),\n\n // Whether or not the scan operation has a zero-valued identity value (true if we're performing addition on a primitive type)\n HAS_IDENTITY_ZERO = (Equals::VALUE) && (Traits::PRIMITIVE),\n };\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for keys\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n KeysInputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedKeysInputIteratorT;\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for values\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n ValuesInputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedValuesInputIteratorT;\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for fixup values\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n AggregatesOutputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedFixupInputIteratorT;\n\n // Reduce-value-by-segment scan operator\n typedef ReduceBySegmentOp ReduceBySegmentOpT;\n\n // Parameterized BlockLoad type for keys\n typedef BlockLoad<\n KeyOutputT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n AgentReduceByKeyPolicyT::LOAD_ALGORITHM>\n BlockLoadKeysT;\n\n // Parameterized BlockLoad type for values\n typedef BlockLoad<\n ValueOutputT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n AgentReduceByKeyPolicyT::LOAD_ALGORITHM>\n BlockLoadValuesT;\n\n // Parameterized BlockDiscontinuity type for keys\n typedef BlockDiscontinuity<\n KeyOutputT,\n BLOCK_THREADS>\n BlockDiscontinuityKeys;\n\n // Parameterized BlockScan type\n typedef BlockScan<\n OffsetValuePairT,\n BLOCK_THREADS,\n AgentReduceByKeyPolicyT::SCAN_ALGORITHM>\n BlockScanT;\n\n // Callback type for obtaining tile prefix during block scan\n typedef TilePrefixCallbackOp<\n OffsetValuePairT,\n ReduceBySegmentOpT,\n ScanTileStateT>\n TilePrefixCallbackOpT;\n\n // Key and value exchange types\n typedef KeyOutputT KeyExchangeT[TILE_ITEMS + 1];\n typedef ValueOutputT ValueExchangeT[TILE_ITEMS + 1];\n\n // Shared memory type for this thread block\n union _TempStorage\n {\n struct\n {\n typename BlockScanT::TempStorage scan; // Smem needed for tile scanning\n typename TilePrefixCallbackOpT::TempStorage prefix; // Smem needed for cooperative prefix callback\n typename BlockDiscontinuityKeys::TempStorage discontinuity; // Smem needed for discontinuity detection\n };\n\n // Smem needed for loading keys\n typename BlockLoadKeysT::TempStorage load_keys;\n\n // Smem needed for loading values\n typename BlockLoadValuesT::TempStorage load_values;\n\n // Smem needed for compacting key value pairs(allows non POD items in this union)\n Uninitialized raw_exchange;\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n WrappedKeysInputIteratorT d_keys_in; ///< Input keys\n UniqueOutputIteratorT d_unique_out; ///< Unique output keys\n WrappedValuesInputIteratorT d_values_in; ///< Input values\n AggregatesOutputIteratorT d_aggregates_out; ///< Output value aggregates\n NumRunsOutputIteratorT d_num_runs_out; ///< Output pointer for total number of segments identified\n EqualityOpT equality_op; ///< KeyT equality operator\n ReductionOpT reduction_op; ///< Reduction operator\n ReduceBySegmentOpT scan_op; ///< Reduce-by-segment scan operator\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n // Constructor\n __device__ __forceinline__\n AgentReduceByKey(\n TempStorage& temp_storage, ///< Reference to temp_storage\n KeysInputIteratorT d_keys_in, ///< Input keys\n UniqueOutputIteratorT d_unique_out, ///< Unique output keys\n ValuesInputIteratorT d_values_in, ///< Input values\n AggregatesOutputIteratorT d_aggregates_out, ///< Output value aggregates\n NumRunsOutputIteratorT d_num_runs_out, ///< Output pointer for total number of segments identified\n EqualityOpT equality_op, ///< KeyT equality operator\n ReductionOpT reduction_op) ///< ValueT reduction operator\n :\n temp_storage(temp_storage.Alias()),\n d_keys_in(d_keys_in),\n d_unique_out(d_unique_out),\n d_values_in(d_values_in),\n d_aggregates_out(d_aggregates_out),\n d_num_runs_out(d_num_runs_out),\n equality_op(equality_op),\n reduction_op(reduction_op),\n scan_op(reduction_op)\n {}\n\n\n //---------------------------------------------------------------------\n // Scatter utility methods\n //---------------------------------------------------------------------\n\n /**\n * Directly scatter flagged items to output offsets\n */\n __device__ __forceinline__ void ScatterDirect(\n KeyValuePairT (&scatter_items)[ITEMS_PER_THREAD],\n OffsetT (&segment_flags)[ITEMS_PER_THREAD],\n OffsetT (&segment_indices)[ITEMS_PER_THREAD])\n {\n // Scatter flagged keys and values\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (segment_flags[ITEM])\n {\n d_unique_out[segment_indices[ITEM]] = scatter_items[ITEM].key;\n d_aggregates_out[segment_indices[ITEM]] = scatter_items[ITEM].value;\n }\n }\n }\n\n\n /**\n * 2-phase scatter flagged items to output offsets\n *\n * The exclusive scan causes each head flag to be paired with the previous\n * value aggregate: the scatter offsets must be decremented for value aggregates\n */\n __device__ __forceinline__ void ScatterTwoPhase(\n KeyValuePairT (&scatter_items)[ITEMS_PER_THREAD],\n OffsetT (&segment_flags)[ITEMS_PER_THREAD],\n OffsetT (&segment_indices)[ITEMS_PER_THREAD],\n OffsetT num_tile_segments,\n OffsetT num_tile_segments_prefix)\n {\n CTA_SYNC();\n\n // Compact and scatter pairs\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (segment_flags[ITEM])\n {\n temp_storage.raw_exchange.Alias()[segment_indices[ITEM] - num_tile_segments_prefix] = scatter_items[ITEM];\n }\n }\n\n CTA_SYNC();\n\n for (int item = threadIdx.x; item < num_tile_segments; item += BLOCK_THREADS)\n {\n KeyValuePairT pair = temp_storage.raw_exchange.Alias()[item];\n d_unique_out[num_tile_segments_prefix + item] = pair.key;\n d_aggregates_out[num_tile_segments_prefix + item] = pair.value;\n }\n }\n\n\n /**\n * Scatter flagged items\n */\n __device__ __forceinline__ void Scatter(\n KeyValuePairT (&scatter_items)[ITEMS_PER_THREAD],\n OffsetT (&segment_flags)[ITEMS_PER_THREAD],\n OffsetT (&segment_indices)[ITEMS_PER_THREAD],\n OffsetT num_tile_segments,\n OffsetT num_tile_segments_prefix)\n {\n // Do a one-phase scatter if (a) two-phase is disabled or (b) the average number of selected items per thread is less than one\n if (TWO_PHASE_SCATTER && (num_tile_segments > BLOCK_THREADS))\n {\n ScatterTwoPhase(\n scatter_items,\n segment_flags,\n segment_indices,\n num_tile_segments,\n num_tile_segments_prefix);\n }\n else\n {\n ScatterDirect(\n scatter_items,\n segment_flags,\n segment_indices);\n }\n }\n\n\n //---------------------------------------------------------------------\n // Cooperatively scan a device-wide sequence of tiles with other CTAs\n //---------------------------------------------------------------------\n\n /**\n * Process a tile of input (dynamic chained scan)\n */\n template ///< Whether the current tile is the last tile\n __device__ __forceinline__ void ConsumeTile(\n OffsetT num_remaining, ///< Number of global input items remaining (including this tile)\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n KeyOutputT keys[ITEMS_PER_THREAD]; // Tile keys\n KeyOutputT prev_keys[ITEMS_PER_THREAD]; // Tile keys shuffled up\n ValueOutputT values[ITEMS_PER_THREAD]; // Tile values\n OffsetT head_flags[ITEMS_PER_THREAD]; // Segment head flags\n OffsetT segment_indices[ITEMS_PER_THREAD]; // Segment indices\n OffsetValuePairT scan_items[ITEMS_PER_THREAD]; // Zipped values and segment flags|indices\n KeyValuePairT scatter_items[ITEMS_PER_THREAD]; // Zipped key value pairs for scattering\n\n // Load keys\n if (IS_LAST_TILE)\n BlockLoadKeysT(temp_storage.load_keys).Load(d_keys_in + tile_offset, keys, num_remaining);\n else\n BlockLoadKeysT(temp_storage.load_keys).Load(d_keys_in + tile_offset, keys);\n\n // Load tile predecessor key in first thread\n KeyOutputT tile_predecessor;\n if (threadIdx.x == 0)\n {\n tile_predecessor = (tile_idx == 0) ?\n keys[0] : // First tile gets repeat of first item (thus first item will not be flagged as a head)\n d_keys_in[tile_offset - 1]; // Subsequent tiles get last key from previous tile\n }\n\n CTA_SYNC();\n\n // Load values\n if (IS_LAST_TILE)\n BlockLoadValuesT(temp_storage.load_values).Load(d_values_in + tile_offset, values, num_remaining);\n else\n BlockLoadValuesT(temp_storage.load_values).Load(d_values_in + tile_offset, values);\n\n CTA_SYNC();\n\n // Initialize head-flags and shuffle up the previous keys\n if (IS_LAST_TILE)\n {\n // Use custom flag operator to additionally flag the first out-of-bounds item\n GuardedInequalityWrapper flag_op(equality_op, num_remaining);\n BlockDiscontinuityKeys(temp_storage.discontinuity).FlagHeads(\n head_flags, keys, prev_keys, flag_op, tile_predecessor);\n }\n else\n {\n InequalityWrapper flag_op(equality_op);\n BlockDiscontinuityKeys(temp_storage.discontinuity).FlagHeads(\n head_flags, keys, prev_keys, flag_op, tile_predecessor);\n }\n\n // Zip values and head flags\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n scan_items[ITEM].value = values[ITEM];\n scan_items[ITEM].key = head_flags[ITEM];\n }\n\n // Perform exclusive tile scan\n OffsetValuePairT block_aggregate; // Inclusive block-wide scan aggregate\n OffsetT num_segments_prefix; // Number of segments prior to this tile\n ValueOutputT total_aggregate; // The tile prefix folded with block_aggregate\n if (tile_idx == 0)\n {\n // Scan first tile\n BlockScanT(temp_storage.scan).ExclusiveScan(scan_items, scan_items, scan_op, block_aggregate);\n num_segments_prefix = 0;\n total_aggregate = block_aggregate.value;\n\n // Update tile status if there are successor tiles\n if ((!IS_LAST_TILE) && (threadIdx.x == 0))\n tile_state.SetInclusive(0, block_aggregate);\n }\n else\n {\n // Scan non-first tile\n TilePrefixCallbackOpT prefix_op(tile_state, temp_storage.prefix, scan_op, tile_idx);\n BlockScanT(temp_storage.scan).ExclusiveScan(scan_items, scan_items, scan_op, prefix_op);\n\n block_aggregate = prefix_op.GetBlockAggregate();\n num_segments_prefix = prefix_op.GetExclusivePrefix().key;\n total_aggregate = reduction_op(\n prefix_op.GetExclusivePrefix().value,\n block_aggregate.value);\n }\n\n // Rezip scatter items and segment indices\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n scatter_items[ITEM].key = prev_keys[ITEM];\n scatter_items[ITEM].value = scan_items[ITEM].value;\n segment_indices[ITEM] = scan_items[ITEM].key;\n }\n\n // At this point, each flagged segment head has:\n // - The key for the previous segment\n // - The reduced value from the previous segment\n // - The segment index for the reduced value\n\n // Scatter flagged keys and values\n OffsetT num_tile_segments = block_aggregate.key;\n Scatter(scatter_items, head_flags, segment_indices, num_tile_segments, num_segments_prefix);\n\n // Last thread in last tile will output final count (and last pair, if necessary)\n if ((IS_LAST_TILE) && (threadIdx.x == BLOCK_THREADS - 1))\n {\n OffsetT num_segments = num_segments_prefix + num_tile_segments;\n\n // If the last tile is a whole tile, output the final_value\n if (num_remaining == TILE_ITEMS)\n {\n d_unique_out[num_segments] = keys[ITEMS_PER_THREAD - 1];\n d_aggregates_out[num_segments] = total_aggregate;\n num_segments++;\n }\n\n // Output the total number of items selected\n *d_num_runs_out = num_segments;\n }\n }\n\n\n /**\n * Scan tiles of items as part of a dynamic chained scan\n */\n __device__ __forceinline__ void ConsumeRange(\n int num_items, ///< Total number of input items\n ScanTileStateT& tile_state, ///< Global tile state descriptor\n int start_tile) ///< The starting tile for the current grid\n {\n // Blocks are launched in increasing order, so just assign one tile per block\n int tile_idx = start_tile + blockIdx.x; // Current tile index\n OffsetT tile_offset = OffsetT(TILE_ITEMS) * tile_idx; // Global offset for the current tile\n OffsetT num_remaining = num_items - tile_offset; // Remaining items (including this tile)\n\n if (num_remaining > TILE_ITEMS)\n {\n // Not last tile\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state);\n }\n else if (num_remaining > 0)\n {\n // Last tile\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state);\n }\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_rle.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentRle implements a stateful abstraction of CUDA thread blocks for participating in device-wide run-length-encode.\n */\n\n#pragma once\n\n#include \n\n#include \"single_pass_scan_operators.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../block/block_exchange.cuh\"\n#include \"../block/block_discontinuity.cuh\"\n#include \"../grid/grid_queue.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../iterator/constant_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentRle\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n bool _STORE_WARP_TIME_SLICING, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any store-related data transpositions (versus each warp having its own storage)\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentRlePolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n STORE_WARP_TIME_SLICING = _STORE_WARP_TIME_SLICING, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any store-related data transpositions (versus each warp having its own storage)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentRle implements a stateful abstraction of CUDA thread blocks for participating in device-wide run-length-encode \n */\ntemplate <\n typename AgentRlePolicyT, ///< Parameterized AgentRlePolicyT tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type for data\n typename OffsetsOutputIteratorT, ///< Random-access output iterator type for offset values\n typename LengthsOutputIteratorT, ///< Random-access output iterator type for length values\n typename EqualityOpT, ///< T equality operator type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentRle\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// The input value type\n typedef typename std::iterator_traits::value_type T;\n\n /// The lengths output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // LengthT = (if output iterator's value type is void) ?\n OffsetT, // ... then the OffsetT type,\n typename std::iterator_traits::value_type>::Type LengthT; // ... else the output iterator's value type\n\n /// Tuple type for scanning (pairs run-length and run-index)\n typedef KeyValuePair LengthOffsetPair;\n\n /// Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n // Constants\n enum\n {\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n BLOCK_THREADS = AgentRlePolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentRlePolicyT::ITEMS_PER_THREAD,\n WARP_ITEMS = WARP_THREADS * ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n /// Whether or not to sync after loading data\n SYNC_AFTER_LOAD = (AgentRlePolicyT::LOAD_ALGORITHM != BLOCK_LOAD_DIRECT),\n\n /// Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any store-related data transpositions (versus each warp having its own storage)\n STORE_WARP_TIME_SLICING = AgentRlePolicyT::STORE_WARP_TIME_SLICING,\n ACTIVE_EXCHANGE_WARPS = (STORE_WARP_TIME_SLICING) ? 1 : WARPS,\n };\n\n\n /**\n * Special operator that signals all out-of-bounds items are not equal to everything else,\n * forcing both (1) the last item to be tail-flagged and (2) all oob items to be marked\n * trivial.\n */\n template \n struct OobInequalityOp\n {\n OffsetT num_remaining;\n EqualityOpT equality_op;\n\n __device__ __forceinline__ OobInequalityOp(\n OffsetT num_remaining,\n EqualityOpT equality_op)\n :\n num_remaining(num_remaining),\n equality_op(equality_op)\n {}\n\n template \n __host__ __device__ __forceinline__ bool operator()(T first, T second, Index idx)\n {\n if (!LAST_TILE || (idx < num_remaining))\n return !equality_op(first, second);\n else\n return true;\n }\n };\n\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for data\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedVLengthnputIterator\n InputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedInputIteratorT;\n\n // Parameterized BlockLoad type for data\n typedef BlockLoad<\n T,\n AgentRlePolicyT::BLOCK_THREADS,\n AgentRlePolicyT::ITEMS_PER_THREAD,\n AgentRlePolicyT::LOAD_ALGORITHM>\n BlockLoadT;\n\n // Parameterized BlockDiscontinuity type for data\n typedef BlockDiscontinuity BlockDiscontinuityT;\n\n // Parameterized WarpScan type\n typedef WarpScan WarpScanPairs;\n\n // Reduce-length-by-run scan operator\n typedef ReduceBySegmentOp ReduceBySegmentOpT;\n\n // Callback type for obtaining tile prefix during block scan\n typedef TilePrefixCallbackOp<\n LengthOffsetPair,\n ReduceBySegmentOpT,\n ScanTileStateT>\n TilePrefixCallbackOpT;\n\n // Warp exchange types\n typedef WarpExchange WarpExchangePairs;\n\n typedef typename If::Type WarpExchangePairsStorage;\n\n typedef WarpExchange WarpExchangeOffsets;\n typedef WarpExchange WarpExchangeLengths;\n\n typedef LengthOffsetPair WarpAggregates[WARPS];\n\n // Shared memory type for this thread block\n struct _TempStorage\n {\n // Aliasable storage layout\n union Aliasable\n {\n struct\n {\n typename BlockDiscontinuityT::TempStorage discontinuity; // Smem needed for discontinuity detection\n typename WarpScanPairs::TempStorage warp_scan[WARPS]; // Smem needed for warp-synchronous scans\n Uninitialized warp_aggregates; // Smem needed for sharing warp-wide aggregates\n typename TilePrefixCallbackOpT::TempStorage prefix; // Smem needed for cooperative prefix callback\n };\n\n // Smem needed for input loading\n typename BlockLoadT::TempStorage load;\n\n // Aliasable layout needed for two-phase scatter\n union ScatterAliasable\n {\n unsigned long long align;\n WarpExchangePairsStorage exchange_pairs[ACTIVE_EXCHANGE_WARPS];\n typename WarpExchangeOffsets::TempStorage exchange_offsets[ACTIVE_EXCHANGE_WARPS];\n typename WarpExchangeLengths::TempStorage exchange_lengths[ACTIVE_EXCHANGE_WARPS];\n\n } scatter_aliasable;\n\n } aliasable;\n\n OffsetT tile_idx; // Shared tile index\n LengthOffsetPair tile_inclusive; // Inclusive tile prefix\n LengthOffsetPair tile_exclusive; // Exclusive tile prefix\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n\n WrappedInputIteratorT d_in; ///< Pointer to input sequence of data items\n OffsetsOutputIteratorT d_offsets_out; ///< Input run offsets\n LengthsOutputIteratorT d_lengths_out; ///< Output run lengths\n\n EqualityOpT equality_op; ///< T equality operator\n ReduceBySegmentOpT scan_op; ///< Reduce-length-by-flag scan operator\n OffsetT num_items; ///< Total number of input items\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n // Constructor\n __device__ __forceinline__\n AgentRle(\n TempStorage &temp_storage, ///< [in] Reference to temp_storage\n InputIteratorT d_in, ///< [in] Pointer to input sequence of data items\n OffsetsOutputIteratorT d_offsets_out, ///< [out] Pointer to output sequence of run offsets\n LengthsOutputIteratorT d_lengths_out, ///< [out] Pointer to output sequence of run lengths\n EqualityOpT equality_op, ///< [in] T equality operator\n OffsetT num_items) ///< [in] Total number of input items\n :\n temp_storage(temp_storage.Alias()),\n d_in(d_in),\n d_offsets_out(d_offsets_out),\n d_lengths_out(d_lengths_out),\n equality_op(equality_op),\n scan_op(cub::Sum()),\n num_items(num_items)\n {}\n\n\n //---------------------------------------------------------------------\n // Utility methods for initializing the selections\n //---------------------------------------------------------------------\n\n template \n __device__ __forceinline__ void InitializeSelections(\n OffsetT tile_offset,\n OffsetT num_remaining,\n T (&items)[ITEMS_PER_THREAD],\n LengthOffsetPair (&lengths_and_num_runs)[ITEMS_PER_THREAD])\n {\n bool head_flags[ITEMS_PER_THREAD];\n bool tail_flags[ITEMS_PER_THREAD];\n\n OobInequalityOp inequality_op(num_remaining, equality_op);\n\n if (FIRST_TILE && LAST_TILE)\n {\n // First-and-last-tile always head-flags the first item and tail-flags the last item\n\n BlockDiscontinuityT(temp_storage.aliasable.discontinuity).FlagHeadsAndTails(\n head_flags, tail_flags, items, inequality_op);\n }\n else if (FIRST_TILE)\n {\n // First-tile always head-flags the first item\n\n // Get the first item from the next tile\n T tile_successor_item;\n if (threadIdx.x == BLOCK_THREADS - 1)\n tile_successor_item = d_in[tile_offset + TILE_ITEMS];\n\n BlockDiscontinuityT(temp_storage.aliasable.discontinuity).FlagHeadsAndTails(\n head_flags, tail_flags, tile_successor_item, items, inequality_op);\n }\n else if (LAST_TILE)\n {\n // Last-tile always flags the last item\n\n // Get the last item from the previous tile\n T tile_predecessor_item;\n if (threadIdx.x == 0)\n tile_predecessor_item = d_in[tile_offset - 1];\n\n BlockDiscontinuityT(temp_storage.aliasable.discontinuity).FlagHeadsAndTails(\n head_flags, tile_predecessor_item, tail_flags, items, inequality_op);\n }\n else\n {\n // Get the first item from the next tile\n T tile_successor_item;\n if (threadIdx.x == BLOCK_THREADS - 1)\n tile_successor_item = d_in[tile_offset + TILE_ITEMS];\n\n // Get the last item from the previous tile\n T tile_predecessor_item;\n if (threadIdx.x == 0)\n tile_predecessor_item = d_in[tile_offset - 1];\n\n BlockDiscontinuityT(temp_storage.aliasable.discontinuity).FlagHeadsAndTails(\n head_flags, tile_predecessor_item, tail_flags, tile_successor_item, items, inequality_op);\n }\n\n // Zip counts and runs\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n lengths_and_num_runs[ITEM].key = head_flags[ITEM] && (!tail_flags[ITEM]);\n lengths_and_num_runs[ITEM].value = ((!head_flags[ITEM]) || (!tail_flags[ITEM]));\n }\n }\n\n //---------------------------------------------------------------------\n // Scan utility methods\n //---------------------------------------------------------------------\n\n /**\n * Scan of allocations\n */\n __device__ __forceinline__ void WarpScanAllocations(\n LengthOffsetPair &tile_aggregate,\n LengthOffsetPair &warp_aggregate,\n LengthOffsetPair &warp_exclusive_in_tile,\n LengthOffsetPair &thread_exclusive_in_warp,\n LengthOffsetPair (&lengths_and_num_runs)[ITEMS_PER_THREAD])\n {\n // Perform warpscans\n unsigned int warp_id = ((WARPS == 1) ? 0 : threadIdx.x / WARP_THREADS);\n int lane_id = LaneId();\n\n LengthOffsetPair identity;\n identity.key = 0;\n identity.value = 0;\n\n LengthOffsetPair thread_inclusive;\n LengthOffsetPair thread_aggregate = internal::ThreadReduce(lengths_and_num_runs, scan_op);\n WarpScanPairs(temp_storage.aliasable.warp_scan[warp_id]).Scan(\n thread_aggregate,\n thread_inclusive,\n thread_exclusive_in_warp,\n identity,\n scan_op);\n\n // Last lane in each warp shares its warp-aggregate\n if (lane_id == WARP_THREADS - 1)\n temp_storage.aliasable.warp_aggregates.Alias()[warp_id] = thread_inclusive;\n\n CTA_SYNC();\n\n // Accumulate total selected and the warp-wide prefix\n warp_exclusive_in_tile = identity;\n warp_aggregate = temp_storage.aliasable.warp_aggregates.Alias()[warp_id];\n tile_aggregate = temp_storage.aliasable.warp_aggregates.Alias()[0];\n\n #pragma unroll\n for (int WARP = 1; WARP < WARPS; ++WARP)\n {\n if (warp_id == WARP)\n warp_exclusive_in_tile = tile_aggregate;\n\n tile_aggregate = scan_op(tile_aggregate, temp_storage.aliasable.warp_aggregates.Alias()[WARP]);\n }\n }\n\n\n //---------------------------------------------------------------------\n // Utility methods for scattering selections\n //---------------------------------------------------------------------\n\n /**\n * Two-phase scatter, specialized for warp time-slicing\n */\n template \n __device__ __forceinline__ void ScatterTwoPhase(\n OffsetT tile_num_runs_exclusive_in_global,\n OffsetT warp_num_runs_aggregate,\n OffsetT warp_num_runs_exclusive_in_tile,\n OffsetT (&thread_num_runs_exclusive_in_warp)[ITEMS_PER_THREAD],\n LengthOffsetPair (&lengths_and_offsets)[ITEMS_PER_THREAD],\n Int2Type is_warp_time_slice)\n {\n unsigned int warp_id = ((WARPS == 1) ? 0 : threadIdx.x / WARP_THREADS);\n int lane_id = LaneId();\n\n // Locally compact items within the warp (first warp)\n if (warp_id == 0)\n {\n WarpExchangePairs(temp_storage.aliasable.scatter_aliasable.exchange_pairs[0]).ScatterToStriped(\n lengths_and_offsets, thread_num_runs_exclusive_in_warp);\n }\n\n // Locally compact items within the warp (remaining warps)\n #pragma unroll\n for (int SLICE = 1; SLICE < WARPS; ++SLICE)\n {\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n WarpExchangePairs(temp_storage.aliasable.scatter_aliasable.exchange_pairs[0]).ScatterToStriped(\n lengths_and_offsets, thread_num_runs_exclusive_in_warp);\n }\n }\n\n // Global scatter\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if ((ITEM * WARP_THREADS) < warp_num_runs_aggregate - lane_id)\n {\n OffsetT item_offset =\n tile_num_runs_exclusive_in_global +\n warp_num_runs_exclusive_in_tile +\n (ITEM * WARP_THREADS) + lane_id;\n\n // Scatter offset\n d_offsets_out[item_offset] = lengths_and_offsets[ITEM].key;\n\n // Scatter length if not the first (global) length\n if ((!FIRST_TILE) || (ITEM != 0) || (threadIdx.x > 0))\n {\n d_lengths_out[item_offset - 1] = lengths_and_offsets[ITEM].value;\n }\n }\n }\n }\n\n\n /**\n * Two-phase scatter\n */\n template \n __device__ __forceinline__ void ScatterTwoPhase(\n OffsetT tile_num_runs_exclusive_in_global,\n OffsetT warp_num_runs_aggregate,\n OffsetT warp_num_runs_exclusive_in_tile,\n OffsetT (&thread_num_runs_exclusive_in_warp)[ITEMS_PER_THREAD],\n LengthOffsetPair (&lengths_and_offsets)[ITEMS_PER_THREAD],\n Int2Type is_warp_time_slice)\n {\n unsigned int warp_id = ((WARPS == 1) ? 0 : threadIdx.x / WARP_THREADS);\n int lane_id = LaneId();\n\n // Unzip\n OffsetT run_offsets[ITEMS_PER_THREAD];\n LengthT run_lengths[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n run_offsets[ITEM] = lengths_and_offsets[ITEM].key;\n run_lengths[ITEM] = lengths_and_offsets[ITEM].value;\n }\n\n WarpExchangeOffsets(temp_storage.aliasable.scatter_aliasable.exchange_offsets[warp_id]).ScatterToStriped(\n run_offsets, thread_num_runs_exclusive_in_warp);\n\n WARP_SYNC(0xffffffff);\n\n WarpExchangeLengths(temp_storage.aliasable.scatter_aliasable.exchange_lengths[warp_id]).ScatterToStriped(\n run_lengths, thread_num_runs_exclusive_in_warp);\n\n // Global scatter\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if ((ITEM * WARP_THREADS) + lane_id < warp_num_runs_aggregate)\n {\n OffsetT item_offset =\n tile_num_runs_exclusive_in_global +\n warp_num_runs_exclusive_in_tile +\n (ITEM * WARP_THREADS) + lane_id;\n\n // Scatter offset\n d_offsets_out[item_offset] = run_offsets[ITEM];\n\n // Scatter length if not the first (global) length\n if ((!FIRST_TILE) || (ITEM != 0) || (threadIdx.x > 0))\n {\n d_lengths_out[item_offset - 1] = run_lengths[ITEM];\n }\n }\n }\n }\n\n\n /**\n * Direct scatter\n */\n template \n __device__ __forceinline__ void ScatterDirect(\n OffsetT tile_num_runs_exclusive_in_global,\n OffsetT warp_num_runs_aggregate,\n OffsetT warp_num_runs_exclusive_in_tile,\n OffsetT (&thread_num_runs_exclusive_in_warp)[ITEMS_PER_THREAD],\n LengthOffsetPair (&lengths_and_offsets)[ITEMS_PER_THREAD])\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (thread_num_runs_exclusive_in_warp[ITEM] < warp_num_runs_aggregate)\n {\n OffsetT item_offset =\n tile_num_runs_exclusive_in_global +\n warp_num_runs_exclusive_in_tile +\n thread_num_runs_exclusive_in_warp[ITEM];\n\n // Scatter offset\n d_offsets_out[item_offset] = lengths_and_offsets[ITEM].key;\n\n // Scatter length if not the first (global) length\n if (item_offset >= 1)\n {\n d_lengths_out[item_offset - 1] = lengths_and_offsets[ITEM].value;\n }\n }\n }\n }\n\n\n /**\n * Scatter\n */\n template \n __device__ __forceinline__ void Scatter(\n OffsetT tile_num_runs_aggregate,\n OffsetT tile_num_runs_exclusive_in_global,\n OffsetT warp_num_runs_aggregate,\n OffsetT warp_num_runs_exclusive_in_tile,\n OffsetT (&thread_num_runs_exclusive_in_warp)[ITEMS_PER_THREAD],\n LengthOffsetPair (&lengths_and_offsets)[ITEMS_PER_THREAD])\n {\n if ((ITEMS_PER_THREAD == 1) || (tile_num_runs_aggregate < BLOCK_THREADS))\n {\n // Direct scatter if the warp has any items\n if (warp_num_runs_aggregate)\n {\n ScatterDirect(\n tile_num_runs_exclusive_in_global,\n warp_num_runs_aggregate,\n warp_num_runs_exclusive_in_tile,\n thread_num_runs_exclusive_in_warp,\n lengths_and_offsets);\n }\n }\n else\n {\n // Scatter two phase\n ScatterTwoPhase(\n tile_num_runs_exclusive_in_global,\n warp_num_runs_aggregate,\n warp_num_runs_exclusive_in_tile,\n thread_num_runs_exclusive_in_warp,\n lengths_and_offsets,\n Int2Type());\n }\n }\n\n\n\n //---------------------------------------------------------------------\n // Cooperatively scan a device-wide sequence of tiles with other CTAs\n //---------------------------------------------------------------------\n\n /**\n * Process a tile of input (dynamic chained scan)\n */\n template <\n bool LAST_TILE>\n __device__ __forceinline__ LengthOffsetPair ConsumeTile(\n OffsetT num_items, ///< Total number of global input items\n OffsetT num_remaining, ///< Number of global input items remaining (including this tile)\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT &tile_status) ///< Global list of tile status\n {\n if (tile_idx == 0)\n {\n // First tile\n\n // Load items\n T items[ITEMS_PER_THREAD];\n if (LAST_TILE)\n BlockLoadT(temp_storage.aliasable.load).Load(d_in + tile_offset, items, num_remaining, T());\n else\n BlockLoadT(temp_storage.aliasable.load).Load(d_in + tile_offset, items);\n\n if (SYNC_AFTER_LOAD)\n CTA_SYNC();\n\n // Set flags\n LengthOffsetPair lengths_and_num_runs[ITEMS_PER_THREAD];\n\n InitializeSelections(\n tile_offset,\n num_remaining,\n items,\n lengths_and_num_runs);\n\n // Exclusive scan of lengths and runs\n LengthOffsetPair tile_aggregate;\n LengthOffsetPair warp_aggregate;\n LengthOffsetPair warp_exclusive_in_tile;\n LengthOffsetPair thread_exclusive_in_warp;\n\n WarpScanAllocations(\n tile_aggregate,\n warp_aggregate,\n warp_exclusive_in_tile,\n thread_exclusive_in_warp,\n lengths_and_num_runs);\n\n // Update tile status if this is not the last tile\n if (!LAST_TILE && (threadIdx.x == 0))\n tile_status.SetInclusive(0, tile_aggregate);\n\n // Update thread_exclusive_in_warp to fold in warp run-length\n if (thread_exclusive_in_warp.key == 0)\n thread_exclusive_in_warp.value += warp_exclusive_in_tile.value;\n\n LengthOffsetPair lengths_and_offsets[ITEMS_PER_THREAD];\n OffsetT thread_num_runs_exclusive_in_warp[ITEMS_PER_THREAD];\n LengthOffsetPair lengths_and_num_runs2[ITEMS_PER_THREAD];\n\n // Downsweep scan through lengths_and_num_runs\n internal::ThreadScanExclusive(lengths_and_num_runs, lengths_and_num_runs2, scan_op, thread_exclusive_in_warp);\n\n // Zip\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n lengths_and_offsets[ITEM].value = lengths_and_num_runs2[ITEM].value;\n lengths_and_offsets[ITEM].key = tile_offset + (threadIdx.x * ITEMS_PER_THREAD) + ITEM;\n thread_num_runs_exclusive_in_warp[ITEM] = (lengths_and_num_runs[ITEM].key) ?\n lengths_and_num_runs2[ITEM].key : // keep\n WARP_THREADS * ITEMS_PER_THREAD; // discard\n }\n\n OffsetT tile_num_runs_aggregate = tile_aggregate.key;\n OffsetT tile_num_runs_exclusive_in_global = 0;\n OffsetT warp_num_runs_aggregate = warp_aggregate.key;\n OffsetT warp_num_runs_exclusive_in_tile = warp_exclusive_in_tile.key;\n\n // Scatter\n Scatter(\n tile_num_runs_aggregate,\n tile_num_runs_exclusive_in_global,\n warp_num_runs_aggregate,\n warp_num_runs_exclusive_in_tile,\n thread_num_runs_exclusive_in_warp,\n lengths_and_offsets);\n\n // Return running total (inclusive of this tile)\n return tile_aggregate;\n }\n else\n {\n // Not first tile\n\n // Load items\n T items[ITEMS_PER_THREAD];\n if (LAST_TILE)\n BlockLoadT(temp_storage.aliasable.load).Load(d_in + tile_offset, items, num_remaining, T());\n else\n BlockLoadT(temp_storage.aliasable.load).Load(d_in + tile_offset, items);\n\n if (SYNC_AFTER_LOAD)\n CTA_SYNC();\n\n // Set flags\n LengthOffsetPair lengths_and_num_runs[ITEMS_PER_THREAD];\n\n InitializeSelections(\n tile_offset,\n num_remaining,\n items,\n lengths_and_num_runs);\n\n // Exclusive scan of lengths and runs\n LengthOffsetPair tile_aggregate;\n LengthOffsetPair warp_aggregate;\n LengthOffsetPair warp_exclusive_in_tile;\n LengthOffsetPair thread_exclusive_in_warp;\n\n WarpScanAllocations(\n tile_aggregate,\n warp_aggregate,\n warp_exclusive_in_tile,\n thread_exclusive_in_warp,\n lengths_and_num_runs);\n\n // First warp computes tile prefix in lane 0\n TilePrefixCallbackOpT prefix_op(tile_status, temp_storage.aliasable.prefix, Sum(), tile_idx);\n unsigned int warp_id = ((WARPS == 1) ? 0 : threadIdx.x / WARP_THREADS);\n if (warp_id == 0)\n {\n prefix_op(tile_aggregate);\n if (threadIdx.x == 0)\n temp_storage.tile_exclusive = prefix_op.exclusive_prefix;\n }\n\n CTA_SYNC();\n\n LengthOffsetPair tile_exclusive_in_global = temp_storage.tile_exclusive;\n\n // Update thread_exclusive_in_warp to fold in warp and tile run-lengths\n LengthOffsetPair thread_exclusive = scan_op(tile_exclusive_in_global, warp_exclusive_in_tile);\n if (thread_exclusive_in_warp.key == 0)\n thread_exclusive_in_warp.value += thread_exclusive.value;\n\n // Downsweep scan through lengths_and_num_runs\n LengthOffsetPair lengths_and_num_runs2[ITEMS_PER_THREAD];\n LengthOffsetPair lengths_and_offsets[ITEMS_PER_THREAD];\n OffsetT thread_num_runs_exclusive_in_warp[ITEMS_PER_THREAD];\n\n internal::ThreadScanExclusive(lengths_and_num_runs, lengths_and_num_runs2, scan_op, thread_exclusive_in_warp);\n\n // Zip\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n lengths_and_offsets[ITEM].value = lengths_and_num_runs2[ITEM].value;\n lengths_and_offsets[ITEM].key = tile_offset + (threadIdx.x * ITEMS_PER_THREAD) + ITEM;\n thread_num_runs_exclusive_in_warp[ITEM] = (lengths_and_num_runs[ITEM].key) ?\n lengths_and_num_runs2[ITEM].key : // keep\n WARP_THREADS * ITEMS_PER_THREAD; // discard\n }\n\n OffsetT tile_num_runs_aggregate = tile_aggregate.key;\n OffsetT tile_num_runs_exclusive_in_global = tile_exclusive_in_global.key;\n OffsetT warp_num_runs_aggregate = warp_aggregate.key;\n OffsetT warp_num_runs_exclusive_in_tile = warp_exclusive_in_tile.key;\n\n // Scatter\n Scatter(\n tile_num_runs_aggregate,\n tile_num_runs_exclusive_in_global,\n warp_num_runs_aggregate,\n warp_num_runs_exclusive_in_tile,\n thread_num_runs_exclusive_in_warp,\n lengths_and_offsets);\n\n // Return running total (inclusive of this tile)\n return prefix_op.inclusive_prefix;\n }\n }\n\n\n /**\n * Scan tiles of items as part of a dynamic chained scan\n */\n template ///< Output iterator type for recording number of items selected\n __device__ __forceinline__ void ConsumeRange(\n int num_tiles, ///< Total number of input tiles\n ScanTileStateT& tile_status, ///< Global list of tile status\n NumRunsIteratorT d_num_runs_out) ///< Output pointer for total number of runs identified\n {\n // Blocks are launched in increasing order, so just assign one tile per block\n int tile_idx = (blockIdx.x * gridDim.y) + blockIdx.y; // Current tile index\n OffsetT tile_offset = tile_idx * TILE_ITEMS; // Global offset for the current tile\n OffsetT num_remaining = num_items - tile_offset; // Remaining items (including this tile)\n\n if (tile_idx < num_tiles - 1)\n {\n // Not the last tile (full)\n ConsumeTile(num_items, num_remaining, tile_idx, tile_offset, tile_status);\n }\n else if (num_remaining > 0)\n {\n // The last tile (possibly partially-full)\n LengthOffsetPair running_total = ConsumeTile(num_items, num_remaining, tile_idx, tile_offset, tile_status);\n\n if (threadIdx.x == 0)\n {\n // Output the total number of items selected\n *d_num_runs_out = running_total.key;\n\n // The inclusive prefix contains accumulated length reduction for the last run\n if (running_total.key > 0)\n d_lengths_out[running_total.key - 1] = running_total.value;\n }\n }\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_scan.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentScan implements a stateful abstraction of CUDA thread blocks for participating in device-wide prefix scan .\n */\n\n#pragma once\n\n#include \n\n#include \"single_pass_scan_operators.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../grid/grid_queue.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentScan\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n BlockStoreAlgorithm _STORE_ALGORITHM, ///< The BlockStore algorithm to use\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentScanPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockStoreAlgorithm STORE_ALGORITHM = _STORE_ALGORITHM; ///< The BlockStore algorithm to use\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentScan implements a stateful abstraction of CUDA thread blocks for participating in device-wide prefix scan .\n */\ntemplate <\n typename AgentScanPolicyT, ///< Parameterized AgentScanPolicyT tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type\n typename OutputIteratorT, ///< Random-access output iterator type\n typename ScanOpT, ///< Scan functor type\n typename InitValueT, ///< The init_value element for ScanOpT type (cub::NullType for inclusive scan)\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentScan\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Tile status descriptor interface type\n typedef ScanTileState ScanTileStateT;\n\n // Input iterator wrapper type (for applying cache modifier)\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n InputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedInputIteratorT;\n\n // Constants\n enum\n {\n IS_INCLUSIVE = Equals::VALUE, // Inclusive scan if no init_value type is provided\n BLOCK_THREADS = AgentScanPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentScanPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n // Parameterized BlockLoad type\n typedef BlockLoad<\n OutputT,\n AgentScanPolicyT::BLOCK_THREADS,\n AgentScanPolicyT::ITEMS_PER_THREAD,\n AgentScanPolicyT::LOAD_ALGORITHM>\n BlockLoadT;\n\n // Parameterized BlockStore type\n typedef BlockStore<\n OutputT,\n AgentScanPolicyT::BLOCK_THREADS,\n AgentScanPolicyT::ITEMS_PER_THREAD,\n AgentScanPolicyT::STORE_ALGORITHM>\n BlockStoreT;\n\n // Parameterized BlockScan type\n typedef BlockScan<\n OutputT,\n AgentScanPolicyT::BLOCK_THREADS,\n AgentScanPolicyT::SCAN_ALGORITHM>\n BlockScanT;\n\n // Callback type for obtaining tile prefix during block scan\n typedef TilePrefixCallbackOp<\n OutputT,\n ScanOpT,\n ScanTileStateT>\n TilePrefixCallbackOpT;\n\n // Stateful BlockScan prefix callback type for managing a running total while scanning consecutive tiles\n typedef BlockScanRunningPrefixOp<\n OutputT,\n ScanOpT>\n RunningPrefixCallbackOp;\n\n // Shared memory type for this thread block\n union _TempStorage\n {\n typename BlockLoadT::TempStorage load; // Smem needed for tile loading\n typename BlockStoreT::TempStorage store; // Smem needed for tile storing\n\n struct\n {\n typename TilePrefixCallbackOpT::TempStorage prefix; // Smem needed for cooperative prefix callback\n typename BlockScanT::TempStorage scan; // Smem needed for tile scanning\n };\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n WrappedInputIteratorT d_in; ///< Input data\n OutputIteratorT d_out; ///< Output data\n ScanOpT scan_op; ///< Binary scan operator\n InitValueT init_value; ///< The init_value element for ScanOpT\n\n\n //---------------------------------------------------------------------\n // Block scan utility methods\n //---------------------------------------------------------------------\n\n /**\n * Exclusive scan specialization (first tile)\n */\n __device__ __forceinline__\n void ScanTile(\n OutputT (&items)[ITEMS_PER_THREAD],\n OutputT init_value,\n ScanOpT scan_op,\n OutputT &block_aggregate,\n Int2Type /*is_inclusive*/)\n {\n BlockScanT(temp_storage.scan).ExclusiveScan(items, items, init_value, scan_op, block_aggregate);\n block_aggregate = scan_op(init_value, block_aggregate);\n }\n\n\n /**\n * Inclusive scan specialization (first tile)\n */\n __device__ __forceinline__\n void ScanTile(\n OutputT (&items)[ITEMS_PER_THREAD],\n InitValueT /*init_value*/,\n ScanOpT scan_op,\n OutputT &block_aggregate,\n Int2Type /*is_inclusive*/)\n {\n BlockScanT(temp_storage.scan).InclusiveScan(items, items, scan_op, block_aggregate);\n }\n\n\n /**\n * Exclusive scan specialization (subsequent tiles)\n */\n template \n __device__ __forceinline__\n void ScanTile(\n OutputT (&items)[ITEMS_PER_THREAD],\n ScanOpT scan_op,\n PrefixCallback &prefix_op,\n Int2Type /*is_inclusive*/)\n {\n BlockScanT(temp_storage.scan).ExclusiveScan(items, items, scan_op, prefix_op);\n }\n\n\n /**\n * Inclusive scan specialization (subsequent tiles)\n */\n template \n __device__ __forceinline__\n void ScanTile(\n OutputT (&items)[ITEMS_PER_THREAD],\n ScanOpT scan_op,\n PrefixCallback &prefix_op,\n Int2Type /*is_inclusive*/)\n {\n BlockScanT(temp_storage.scan).InclusiveScan(items, items, scan_op, prefix_op);\n }\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n // Constructor\n __device__ __forceinline__\n AgentScan(\n TempStorage& temp_storage, ///< Reference to temp_storage\n InputIteratorT d_in, ///< Input data\n OutputIteratorT d_out, ///< Output data\n ScanOpT scan_op, ///< Binary scan operator\n InitValueT init_value) ///< Initial value to seed the exclusive scan\n :\n temp_storage(temp_storage.Alias()),\n d_in(d_in),\n d_out(d_out),\n scan_op(scan_op),\n init_value(init_value)\n {}\n\n\n //---------------------------------------------------------------------\n // Cooperatively scan a device-wide sequence of tiles with other CTAs\n //---------------------------------------------------------------------\n\n /**\n * Process a tile of input (dynamic chained scan)\n */\n template ///< Whether the current tile is the last tile\n __device__ __forceinline__ void ConsumeTile(\n OffsetT num_remaining, ///< Number of global input items remaining (including this tile)\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n // Load items\n OutputT items[ITEMS_PER_THREAD];\n\n if (IS_LAST_TILE)\n BlockLoadT(temp_storage.load).Load(d_in + tile_offset, items, num_remaining);\n else\n BlockLoadT(temp_storage.load).Load(d_in + tile_offset, items);\n\n CTA_SYNC();\n\n // Perform tile scan\n if (tile_idx == 0)\n {\n // Scan first tile\n OutputT block_aggregate;\n ScanTile(items, init_value, scan_op, block_aggregate, Int2Type());\n if ((!IS_LAST_TILE) && (threadIdx.x == 0))\n tile_state.SetInclusive(0, block_aggregate);\n }\n else\n {\n // Scan non-first tile\n TilePrefixCallbackOpT prefix_op(tile_state, temp_storage.prefix, scan_op, tile_idx);\n ScanTile(items, scan_op, prefix_op, Int2Type());\n }\n\n CTA_SYNC();\n\n // Store items\n if (IS_LAST_TILE)\n BlockStoreT(temp_storage.store).Store(d_out + tile_offset, items, num_remaining);\n else\n BlockStoreT(temp_storage.store).Store(d_out + tile_offset, items);\n }\n\n\n /**\n * Scan tiles of items as part of a dynamic chained scan\n */\n __device__ __forceinline__ void ConsumeRange(\n int num_items, ///< Total number of input items\n ScanTileStateT& tile_state, ///< Global tile state descriptor\n int start_tile) ///< The starting tile for the current grid\n {\n // Blocks are launched in increasing order, so just assign one tile per block\n int tile_idx = start_tile + blockIdx.x; // Current tile index\n OffsetT tile_offset = OffsetT(TILE_ITEMS) * tile_idx; // Global offset for the current tile\n OffsetT num_remaining = num_items - tile_offset; // Remaining items (including this tile)\n\n if (num_remaining > TILE_ITEMS)\n {\n // Not last tile\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state);\n }\n else if (num_remaining > 0)\n {\n // Last tile\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state);\n }\n }\n\n\n //---------------------------------------------------------------------\n // Scan an sequence of consecutive tiles (independent of other thread blocks)\n //---------------------------------------------------------------------\n\n /**\n * Process a tile of input\n */\n template <\n bool IS_FIRST_TILE,\n bool IS_LAST_TILE>\n __device__ __forceinline__ void ConsumeTile(\n OffsetT tile_offset, ///< Tile offset\n RunningPrefixCallbackOp& prefix_op, ///< Running prefix operator\n int valid_items = TILE_ITEMS) ///< Number of valid items in the tile\n {\n // Load items\n OutputT items[ITEMS_PER_THREAD];\n\n if (IS_LAST_TILE)\n BlockLoadT(temp_storage.load).Load(d_in + tile_offset, items, valid_items);\n else\n BlockLoadT(temp_storage.load).Load(d_in + tile_offset, items);\n\n CTA_SYNC();\n\n // Block scan\n if (IS_FIRST_TILE)\n {\n OutputT block_aggregate;\n ScanTile(items, init_value, scan_op, block_aggregate, Int2Type());\n prefix_op.running_total = block_aggregate;\n }\n else\n {\n ScanTile(items, scan_op, prefix_op, Int2Type());\n }\n\n CTA_SYNC();\n\n // Store items\n if (IS_LAST_TILE)\n BlockStoreT(temp_storage.store).Store(d_out + tile_offset, items, valid_items);\n else\n BlockStoreT(temp_storage.store).Store(d_out + tile_offset, items);\n }\n\n\n /**\n * Scan a consecutive share of input tiles\n */\n __device__ __forceinline__ void ConsumeRange(\n OffsetT range_offset, ///< [in] Threadblock begin offset (inclusive)\n OffsetT range_end) ///< [in] Threadblock end offset (exclusive)\n {\n BlockScanRunningPrefixOp prefix_op(scan_op);\n\n if (range_offset + TILE_ITEMS <= range_end)\n {\n // Consume first tile of input (full)\n ConsumeTile(range_offset, prefix_op);\n range_offset += TILE_ITEMS;\n\n // Consume subsequent full tiles of input\n while (range_offset + TILE_ITEMS <= range_end)\n {\n ConsumeTile(range_offset, prefix_op);\n range_offset += TILE_ITEMS;\n }\n\n // Consume a partially-full tile\n if (range_offset < range_end)\n {\n int valid_items = range_end - range_offset;\n ConsumeTile(range_offset, prefix_op, valid_items);\n }\n }\n else\n {\n // Consume the first tile of input (partially-full)\n int valid_items = range_end - range_offset;\n ConsumeTile(range_offset, prefix_op, valid_items);\n }\n }\n\n\n /**\n * Scan a consecutive share of input tiles, seeded with the specified prefix value\n */\n __device__ __forceinline__ void ConsumeRange(\n OffsetT range_offset, ///< [in] Threadblock begin offset (inclusive)\n OffsetT range_end, ///< [in] Threadblock end offset (exclusive)\n OutputT prefix) ///< [in] The prefix to apply to the scan segment\n {\n BlockScanRunningPrefixOp prefix_op(prefix, scan_op);\n\n // Consume full tiles of input\n while (range_offset + TILE_ITEMS <= range_end)\n {\n ConsumeTile(range_offset, prefix_op);\n range_offset += TILE_ITEMS;\n }\n\n // Consume a partially-full tile\n if (range_offset < range_end)\n {\n int valid_items = range_end - range_offset;\n ConsumeTile(range_offset, prefix_op, valid_items);\n }\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_segment_fixup.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentSegmentFixup implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduce-value-by-key.\n */\n\n#pragma once\n\n#include \n\n#include \"single_pass_scan_operators.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../block/block_discontinuity.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../iterator/constant_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentSegmentFixup\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentSegmentFixupPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n/**\n * \\brief AgentSegmentFixup implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduce-value-by-key\n */\ntemplate <\n typename AgentSegmentFixupPolicyT, ///< Parameterized AgentSegmentFixupPolicy tuning policy type\n typename PairsInputIteratorT, ///< Random-access input iterator type for keys\n typename AggregatesOutputIteratorT, ///< Random-access output iterator type for values\n typename EqualityOpT, ///< KeyT equality operator type\n typename ReductionOpT, ///< ValueT reduction operator type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct AgentSegmentFixup\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // Data type of key-value input iterator\n typedef typename std::iterator_traits::value_type KeyValuePairT;\n\n // Value type\n typedef typename KeyValuePairT::Value ValueT;\n\n // Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n // Constants\n enum\n {\n BLOCK_THREADS = AgentSegmentFixupPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentSegmentFixupPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n\n // Whether or not do fixup using RLE + global atomics\n USE_ATOMIC_FIXUP = (CUB_PTX_ARCH >= 350) && \n (Equals::VALUE || \n Equals::VALUE ||\n Equals::VALUE ||\n Equals::VALUE),\n\n // Whether or not the scan operation has a zero-valued identity value (true if we're performing addition on a primitive type)\n HAS_IDENTITY_ZERO = (Equals::VALUE) && (Traits::PRIMITIVE),\n };\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for keys\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n PairsInputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedPairsInputIteratorT;\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for fixup values\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n AggregatesOutputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedFixupInputIteratorT;\n\n // Reduce-value-by-segment scan operator\n typedef ReduceByKeyOp ReduceBySegmentOpT;\n\n // Parameterized BlockLoad type for pairs\n typedef BlockLoad<\n KeyValuePairT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n AgentSegmentFixupPolicyT::LOAD_ALGORITHM>\n BlockLoadPairs;\n\n // Parameterized BlockScan type\n typedef BlockScan<\n KeyValuePairT,\n BLOCK_THREADS,\n AgentSegmentFixupPolicyT::SCAN_ALGORITHM>\n BlockScanT;\n\n // Callback type for obtaining tile prefix during block scan\n typedef TilePrefixCallbackOp<\n KeyValuePairT,\n ReduceBySegmentOpT,\n ScanTileStateT>\n TilePrefixCallbackOpT;\n\n // Shared memory type for this thread block\n union _TempStorage\n {\n struct\n {\n typename BlockScanT::TempStorage scan; // Smem needed for tile scanning\n typename TilePrefixCallbackOpT::TempStorage prefix; // Smem needed for cooperative prefix callback\n };\n\n // Smem needed for loading keys\n typename BlockLoadPairs::TempStorage load_pairs;\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n WrappedPairsInputIteratorT d_pairs_in; ///< Input keys\n AggregatesOutputIteratorT d_aggregates_out; ///< Output value aggregates\n WrappedFixupInputIteratorT d_fixup_in; ///< Fixup input values\n InequalityWrapper inequality_op; ///< KeyT inequality operator\n ReductionOpT reduction_op; ///< Reduction operator\n ReduceBySegmentOpT scan_op; ///< Reduce-by-segment scan operator\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n // Constructor\n __device__ __forceinline__\n AgentSegmentFixup(\n TempStorage& temp_storage, ///< Reference to temp_storage\n PairsInputIteratorT d_pairs_in, ///< Input keys\n AggregatesOutputIteratorT d_aggregates_out, ///< Output value aggregates\n EqualityOpT equality_op, ///< KeyT equality operator\n ReductionOpT reduction_op) ///< ValueT reduction operator\n :\n temp_storage(temp_storage.Alias()),\n d_pairs_in(d_pairs_in),\n d_aggregates_out(d_aggregates_out),\n d_fixup_in(d_aggregates_out),\n inequality_op(equality_op),\n reduction_op(reduction_op),\n scan_op(reduction_op)\n {}\n\n\n //---------------------------------------------------------------------\n // Cooperatively scan a device-wide sequence of tiles with other CTAs\n //---------------------------------------------------------------------\n\n\n /**\n * Process input tile. Specialized for atomic-fixup\n */\n template \n __device__ __forceinline__ void ConsumeTile(\n OffsetT num_remaining, ///< Number of global input items remaining (including this tile)\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state, ///< Global tile state descriptor\n Int2Type use_atomic_fixup) ///< Marker whether to use atomicAdd (instead of reduce-by-key)\n {\n KeyValuePairT pairs[ITEMS_PER_THREAD];\n\n // Load pairs\n KeyValuePairT oob_pair;\n oob_pair.key = -1;\n\n if (IS_LAST_TILE)\n BlockLoadPairs(temp_storage.load_pairs).Load(d_pairs_in + tile_offset, pairs, num_remaining, oob_pair);\n else\n BlockLoadPairs(temp_storage.load_pairs).Load(d_pairs_in + tile_offset, pairs);\n\n // RLE \n #pragma unroll\n for (int ITEM = 1; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n ValueT* d_scatter = d_aggregates_out + pairs[ITEM - 1].key;\n if (pairs[ITEM].key != pairs[ITEM - 1].key)\n atomicAdd(d_scatter, pairs[ITEM - 1].value);\n else\n pairs[ITEM].value = reduction_op(pairs[ITEM - 1].value, pairs[ITEM].value);\n }\n\n // Flush last item if valid\n ValueT* d_scatter = d_aggregates_out + pairs[ITEMS_PER_THREAD - 1].key;\n if ((!IS_LAST_TILE) || (pairs[ITEMS_PER_THREAD - 1].key >= 0))\n atomicAdd(d_scatter, pairs[ITEMS_PER_THREAD - 1].value);\n }\n\n\n /**\n * Process input tile. Specialized for reduce-by-key fixup\n */\n template \n __device__ __forceinline__ void ConsumeTile(\n OffsetT num_remaining, ///< Number of global input items remaining (including this tile)\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state, ///< Global tile state descriptor\n Int2Type use_atomic_fixup) ///< Marker whether to use atomicAdd (instead of reduce-by-key)\n {\n KeyValuePairT pairs[ITEMS_PER_THREAD];\n KeyValuePairT scatter_pairs[ITEMS_PER_THREAD];\n\n // Load pairs\n KeyValuePairT oob_pair;\n oob_pair.key = -1;\n\n if (IS_LAST_TILE)\n BlockLoadPairs(temp_storage.load_pairs).Load(d_pairs_in + tile_offset, pairs, num_remaining, oob_pair);\n else\n BlockLoadPairs(temp_storage.load_pairs).Load(d_pairs_in + tile_offset, pairs);\n\n CTA_SYNC();\n\n KeyValuePairT tile_aggregate;\n if (tile_idx == 0)\n {\n // Exclusive scan of values and segment_flags\n BlockScanT(temp_storage.scan).ExclusiveScan(pairs, scatter_pairs, scan_op, tile_aggregate);\n\n // Update tile status if this is not the last tile\n if (threadIdx.x == 0)\n {\n // Set first segment id to not trigger a flush (invalid from exclusive scan)\n scatter_pairs[0].key = pairs[0].key;\n\n if (!IS_LAST_TILE)\n tile_state.SetInclusive(0, tile_aggregate);\n\n }\n }\n else\n {\n // Exclusive scan of values and segment_flags\n TilePrefixCallbackOpT prefix_op(tile_state, temp_storage.prefix, scan_op, tile_idx);\n BlockScanT(temp_storage.scan).ExclusiveScan(pairs, scatter_pairs, scan_op, prefix_op);\n tile_aggregate = prefix_op.GetBlockAggregate();\n }\n\n // Scatter updated values\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (scatter_pairs[ITEM].key != pairs[ITEM].key)\n {\n // Update the value at the key location\n ValueT value = d_fixup_in[scatter_pairs[ITEM].key];\n value = reduction_op(value, scatter_pairs[ITEM].value);\n\n d_aggregates_out[scatter_pairs[ITEM].key] = value;\n }\n }\n\n // Finalize the last item\n if (IS_LAST_TILE)\n {\n // Last thread will output final count and last item, if necessary\n if (threadIdx.x == BLOCK_THREADS - 1)\n {\n // If the last tile is a whole tile, the inclusive prefix contains accumulated value reduction for the last segment\n if (num_remaining == TILE_ITEMS)\n {\n // Update the value at the key location\n OffsetT last_key = pairs[ITEMS_PER_THREAD - 1].key;\n d_aggregates_out[last_key] = reduction_op(tile_aggregate.value, d_fixup_in[last_key]);\n }\n }\n }\n }\n\n\n /**\n * Scan tiles of items as part of a dynamic chained scan\n */\n __device__ __forceinline__ void ConsumeRange(\n int num_items, ///< Total number of input items\n int num_tiles, ///< Total number of input tiles\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n // Blocks are launched in increasing order, so just assign one tile per block\n int tile_idx = (blockIdx.x * gridDim.y) + blockIdx.y; // Current tile index\n OffsetT tile_offset = tile_idx * TILE_ITEMS; // Global offset for the current tile\n OffsetT num_remaining = num_items - tile_offset; // Remaining items (including this tile)\n\n if (num_remaining > TILE_ITEMS)\n {\n // Not the last tile (full)\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state, Int2Type());\n }\n else if (num_remaining > 0)\n {\n // The last tile (possibly partially-full)\n ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state, Int2Type());\n }\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_select_if.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentSelectIf implements a stateful abstraction of CUDA thread blocks for participating in device-wide select.\n */\n\n#pragma once\n\n#include \n\n#include \"single_pass_scan_operators.cuh\"\n#include \"../block/block_load.cuh\"\n#include \"../block/block_store.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../block/block_exchange.cuh\"\n#include \"../block/block_discontinuity.cuh\"\n#include \"../grid/grid_queue.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentSelectIf\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentSelectIfPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\n\n/**\n * \\brief AgentSelectIf implements a stateful abstraction of CUDA thread blocks for participating in device-wide selection\n *\n * Performs functor-based selection if SelectOpT functor type != NullType\n * Otherwise performs flag-based selection if FlagsInputIterator's value type != NullType\n * Otherwise performs discontinuity selection (keep unique)\n */\ntemplate <\n typename AgentSelectIfPolicyT, ///< Parameterized AgentSelectIfPolicy tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type for selection items\n typename FlagsInputIteratorT, ///< Random-access input iterator type for selections (NullType* if a selection functor or discontinuity flagging is to be used for selection)\n typename SelectedOutputIteratorT, ///< Random-access input iterator type for selection_flags items\n typename SelectOpT, ///< Selection operator type (NullType if selections or discontinuity flagging is to be used for selection)\n typename EqualityOpT, ///< Equality operator type (NullType if selection functor or selections is to be used for selection)\n typename OffsetT, ///< Signed integer type for global offsets\n bool KEEP_REJECTS> ///< Whether or not we push rejected items to the back of the output\nstruct AgentSelectIf\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // The flag value type\n typedef typename std::iterator_traits::value_type FlagT;\n\n // Tile status descriptor interface type\n typedef ScanTileState ScanTileStateT;\n\n // Constants\n enum\n {\n USE_SELECT_OP,\n USE_SELECT_FLAGS,\n USE_DISCONTINUITY,\n\n BLOCK_THREADS = AgentSelectIfPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentSelectIfPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n TWO_PHASE_SCATTER = (ITEMS_PER_THREAD > 1),\n\n SELECT_METHOD = (!Equals::VALUE) ?\n USE_SELECT_OP :\n (!Equals::VALUE) ?\n USE_SELECT_FLAGS :\n USE_DISCONTINUITY\n };\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for items\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n InputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedInputIteratorT;\n\n // Cache-modified Input iterator wrapper type (for applying cache modifier) for values\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedValuesInputIterator\n FlagsInputIteratorT>::Type // Directly use the supplied input iterator type\n WrappedFlagsInputIteratorT;\n\n // Parameterized BlockLoad type for input data\n typedef BlockLoad<\n OutputT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n AgentSelectIfPolicyT::LOAD_ALGORITHM>\n BlockLoadT;\n\n // Parameterized BlockLoad type for flags\n typedef BlockLoad<\n FlagT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n AgentSelectIfPolicyT::LOAD_ALGORITHM>\n BlockLoadFlags;\n\n // Parameterized BlockDiscontinuity type for items\n typedef BlockDiscontinuity<\n OutputT,\n BLOCK_THREADS>\n BlockDiscontinuityT;\n\n // Parameterized BlockScan type\n typedef BlockScan<\n OffsetT,\n BLOCK_THREADS,\n AgentSelectIfPolicyT::SCAN_ALGORITHM>\n BlockScanT;\n\n // Callback type for obtaining tile prefix during block scan\n typedef TilePrefixCallbackOp<\n OffsetT,\n cub::Sum,\n ScanTileStateT>\n TilePrefixCallbackOpT;\n\n // Item exchange type\n typedef OutputT ItemExchangeT[TILE_ITEMS];\n\n // Shared memory type for this thread block\n union _TempStorage\n {\n struct\n {\n typename BlockScanT::TempStorage scan; // Smem needed for tile scanning\n typename TilePrefixCallbackOpT::TempStorage prefix; // Smem needed for cooperative prefix callback\n typename BlockDiscontinuityT::TempStorage discontinuity; // Smem needed for discontinuity detection\n };\n\n // Smem needed for loading items\n typename BlockLoadT::TempStorage load_items;\n\n // Smem needed for loading values\n typename BlockLoadFlags::TempStorage load_flags;\n\n // Smem needed for compacting items (allows non POD items in this union)\n Uninitialized raw_exchange;\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n _TempStorage& temp_storage; ///< Reference to temp_storage\n WrappedInputIteratorT d_in; ///< Input items\n SelectedOutputIteratorT d_selected_out; ///< Unique output items\n WrappedFlagsInputIteratorT d_flags_in; ///< Input selection flags (if applicable)\n InequalityWrapper inequality_op; ///< T inequality operator\n SelectOpT select_op; ///< Selection operator\n OffsetT num_items; ///< Total number of input items\n\n\n //---------------------------------------------------------------------\n // Constructor\n //---------------------------------------------------------------------\n\n // Constructor\n __device__ __forceinline__\n AgentSelectIf(\n TempStorage &temp_storage, ///< Reference to temp_storage\n InputIteratorT d_in, ///< Input data\n FlagsInputIteratorT d_flags_in, ///< Input selection flags (if applicable)\n SelectedOutputIteratorT d_selected_out, ///< Output data\n SelectOpT select_op, ///< Selection operator\n EqualityOpT equality_op, ///< Equality operator\n OffsetT num_items) ///< Total number of input items\n :\n temp_storage(temp_storage.Alias()),\n d_in(d_in),\n d_flags_in(d_flags_in),\n d_selected_out(d_selected_out),\n select_op(select_op),\n inequality_op(equality_op),\n num_items(num_items)\n {}\n\n\n //---------------------------------------------------------------------\n // Utility methods for initializing the selections\n //---------------------------------------------------------------------\n\n /**\n * Initialize selections (specialized for selection operator)\n */\n template \n __device__ __forceinline__ void InitializeSelections(\n OffsetT /*tile_offset*/,\n OffsetT num_tile_items,\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n Int2Type /*select_method*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n // Out-of-bounds items are selection_flags\n selection_flags[ITEM] = 1;\n\n if (!IS_LAST_TILE || (OffsetT(threadIdx.x * ITEMS_PER_THREAD) + ITEM < num_tile_items))\n selection_flags[ITEM] = select_op(items[ITEM]);\n }\n }\n\n\n /**\n * Initialize selections (specialized for valid flags)\n */\n template \n __device__ __forceinline__ void InitializeSelections(\n OffsetT tile_offset,\n OffsetT num_tile_items,\n OutputT (&/*items*/)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n Int2Type /*select_method*/)\n {\n CTA_SYNC();\n\n FlagT flags[ITEMS_PER_THREAD];\n\n if (IS_LAST_TILE)\n {\n // Out-of-bounds items are selection_flags\n BlockLoadFlags(temp_storage.load_flags).Load(d_flags_in + tile_offset, flags, num_tile_items, 1);\n }\n else\n {\n BlockLoadFlags(temp_storage.load_flags).Load(d_flags_in + tile_offset, flags);\n }\n\n // Convert flag type to selection_flags type\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n selection_flags[ITEM] = flags[ITEM];\n }\n }\n\n\n /**\n * Initialize selections (specialized for discontinuity detection)\n */\n template \n __device__ __forceinline__ void InitializeSelections(\n OffsetT tile_offset,\n OffsetT num_tile_items,\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n Int2Type /*select_method*/)\n {\n if (IS_FIRST_TILE)\n {\n CTA_SYNC();\n\n // Set head selection_flags. First tile sets the first flag for the first item\n BlockDiscontinuityT(temp_storage.discontinuity).FlagHeads(selection_flags, items, inequality_op);\n }\n else\n {\n OutputT tile_predecessor;\n if (threadIdx.x == 0)\n tile_predecessor = d_in[tile_offset - 1];\n\n CTA_SYNC();\n\n BlockDiscontinuityT(temp_storage.discontinuity).FlagHeads(selection_flags, items, inequality_op, tile_predecessor);\n }\n\n // Set selection flags for out-of-bounds items\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n // Set selection_flags for out-of-bounds items\n if ((IS_LAST_TILE) && (OffsetT(threadIdx.x * ITEMS_PER_THREAD) + ITEM >= num_tile_items))\n selection_flags[ITEM] = 1;\n }\n }\n\n\n //---------------------------------------------------------------------\n // Scatter utility methods\n //---------------------------------------------------------------------\n\n /**\n * Scatter flagged items to output offsets (specialized for direct scattering)\n */\n template \n __device__ __forceinline__ void ScatterDirect(\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n OffsetT (&selection_indices)[ITEMS_PER_THREAD],\n OffsetT num_selections)\n {\n // Scatter flagged items\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (selection_flags[ITEM])\n {\n if ((!IS_LAST_TILE) || selection_indices[ITEM] < num_selections)\n {\n d_selected_out[selection_indices[ITEM]] = items[ITEM];\n }\n }\n }\n }\n\n\n /**\n * Scatter flagged items to output offsets (specialized for two-phase scattering)\n */\n template \n __device__ __forceinline__ void ScatterTwoPhase(\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n OffsetT (&selection_indices)[ITEMS_PER_THREAD],\n int /*num_tile_items*/, ///< Number of valid items in this tile\n int num_tile_selections, ///< Number of selections in this tile\n OffsetT num_selections_prefix, ///< Total number of selections prior to this tile\n OffsetT /*num_rejected_prefix*/, ///< Total number of rejections prior to this tile\n Int2Type /*is_keep_rejects*/) ///< Marker type indicating whether to keep rejected items in the second partition\n {\n CTA_SYNC();\n\n // Compact and scatter items\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int local_scatter_offset = selection_indices[ITEM] - num_selections_prefix;\n if (selection_flags[ITEM])\n {\n temp_storage.raw_exchange.Alias()[local_scatter_offset] = items[ITEM];\n }\n }\n\n CTA_SYNC();\n\n for (int item = threadIdx.x; item < num_tile_selections; item += BLOCK_THREADS)\n {\n d_selected_out[num_selections_prefix + item] = temp_storage.raw_exchange.Alias()[item];\n }\n }\n\n\n /**\n * Scatter flagged items to output offsets (specialized for two-phase scattering)\n */\n template \n __device__ __forceinline__ void ScatterTwoPhase(\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n OffsetT (&selection_indices)[ITEMS_PER_THREAD],\n int num_tile_items, ///< Number of valid items in this tile\n int num_tile_selections, ///< Number of selections in this tile\n OffsetT num_selections_prefix, ///< Total number of selections prior to this tile\n OffsetT num_rejected_prefix, ///< Total number of rejections prior to this tile\n Int2Type /*is_keep_rejects*/) ///< Marker type indicating whether to keep rejected items in the second partition\n {\n CTA_SYNC();\n\n int tile_num_rejections = num_tile_items - num_tile_selections;\n\n // Scatter items to shared memory (rejections first)\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int item_idx = (threadIdx.x * ITEMS_PER_THREAD) + ITEM;\n int local_selection_idx = selection_indices[ITEM] - num_selections_prefix;\n int local_rejection_idx = item_idx - local_selection_idx;\n int local_scatter_offset = (selection_flags[ITEM]) ?\n tile_num_rejections + local_selection_idx :\n local_rejection_idx;\n\n temp_storage.raw_exchange.Alias()[local_scatter_offset] = items[ITEM];\n }\n\n CTA_SYNC();\n\n // Gather items from shared memory and scatter to global\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int item_idx = (ITEM * BLOCK_THREADS) + threadIdx.x;\n int rejection_idx = item_idx;\n int selection_idx = item_idx - tile_num_rejections;\n OffsetT scatter_offset = (item_idx < tile_num_rejections) ?\n num_items - num_rejected_prefix - rejection_idx - 1 :\n num_selections_prefix + selection_idx;\n\n OutputT item = temp_storage.raw_exchange.Alias()[item_idx];\n\n if (!IS_LAST_TILE || (item_idx < num_tile_items))\n {\n d_selected_out[scatter_offset] = item;\n }\n }\n }\n\n\n /**\n * Scatter flagged items\n */\n template \n __device__ __forceinline__ void Scatter(\n OutputT (&items)[ITEMS_PER_THREAD],\n OffsetT (&selection_flags)[ITEMS_PER_THREAD],\n OffsetT (&selection_indices)[ITEMS_PER_THREAD],\n int num_tile_items, ///< Number of valid items in this tile\n int num_tile_selections, ///< Number of selections in this tile\n OffsetT num_selections_prefix, ///< Total number of selections prior to this tile\n OffsetT num_rejected_prefix, ///< Total number of rejections prior to this tile\n OffsetT num_selections) ///< Total number of selections including this tile\n {\n // Do a two-phase scatter if (a) keeping both partitions or (b) two-phase is enabled and the average number of selection_flags items per thread is greater than one\n if (KEEP_REJECTS || (TWO_PHASE_SCATTER && (num_tile_selections > BLOCK_THREADS)))\n {\n ScatterTwoPhase(\n items,\n selection_flags,\n selection_indices,\n num_tile_items,\n num_tile_selections,\n num_selections_prefix,\n num_rejected_prefix,\n Int2Type());\n }\n else\n {\n ScatterDirect(\n items,\n selection_flags,\n selection_indices,\n num_selections);\n }\n }\n\n //---------------------------------------------------------------------\n // Cooperatively scan a device-wide sequence of tiles with other CTAs\n //---------------------------------------------------------------------\n\n\n /**\n * Process first tile of input (dynamic chained scan). Returns the running count of selections (including this tile)\n */\n template \n __device__ __forceinline__ OffsetT ConsumeFirstTile(\n int num_tile_items, ///< Number of input items comprising this tile\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n OutputT items[ITEMS_PER_THREAD];\n OffsetT selection_flags[ITEMS_PER_THREAD];\n OffsetT selection_indices[ITEMS_PER_THREAD];\n\n // Load items\n if (IS_LAST_TILE)\n BlockLoadT(temp_storage.load_items).Load(d_in + tile_offset, items, num_tile_items);\n else\n BlockLoadT(temp_storage.load_items).Load(d_in + tile_offset, items);\n\n // Initialize selection_flags\n InitializeSelections(\n tile_offset,\n num_tile_items,\n items,\n selection_flags,\n Int2Type());\n\n CTA_SYNC();\n\n // Exclusive scan of selection_flags\n OffsetT num_tile_selections;\n BlockScanT(temp_storage.scan).ExclusiveSum(selection_flags, selection_indices, num_tile_selections);\n\n if (threadIdx.x == 0)\n {\n // Update tile status if this is not the last tile\n if (!IS_LAST_TILE)\n tile_state.SetInclusive(0, num_tile_selections);\n }\n\n // Discount any out-of-bounds selections\n if (IS_LAST_TILE)\n num_tile_selections -= (TILE_ITEMS - num_tile_items);\n\n // Scatter flagged items\n Scatter(\n items,\n selection_flags,\n selection_indices,\n num_tile_items,\n num_tile_selections,\n 0,\n 0,\n num_tile_selections);\n\n return num_tile_selections;\n }\n\n\n /**\n * Process subsequent tile of input (dynamic chained scan). Returns the running count of selections (including this tile)\n */\n template \n __device__ __forceinline__ OffsetT ConsumeSubsequentTile(\n int num_tile_items, ///< Number of input items comprising this tile\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n OutputT items[ITEMS_PER_THREAD];\n OffsetT selection_flags[ITEMS_PER_THREAD];\n OffsetT selection_indices[ITEMS_PER_THREAD];\n\n // Load items\n if (IS_LAST_TILE)\n BlockLoadT(temp_storage.load_items).Load(d_in + tile_offset, items, num_tile_items);\n else\n BlockLoadT(temp_storage.load_items).Load(d_in + tile_offset, items);\n\n // Initialize selection_flags\n InitializeSelections(\n tile_offset,\n num_tile_items,\n items,\n selection_flags,\n Int2Type());\n\n CTA_SYNC();\n\n // Exclusive scan of values and selection_flags\n TilePrefixCallbackOpT prefix_op(tile_state, temp_storage.prefix, cub::Sum(), tile_idx);\n BlockScanT(temp_storage.scan).ExclusiveSum(selection_flags, selection_indices, prefix_op);\n\n OffsetT num_tile_selections = prefix_op.GetBlockAggregate();\n OffsetT num_selections = prefix_op.GetInclusivePrefix();\n OffsetT num_selections_prefix = prefix_op.GetExclusivePrefix();\n OffsetT num_rejected_prefix = (tile_idx * TILE_ITEMS) - num_selections_prefix;\n\n // Discount any out-of-bounds selections\n if (IS_LAST_TILE)\n {\n int num_discount = TILE_ITEMS - num_tile_items;\n num_selections -= num_discount;\n num_tile_selections -= num_discount;\n }\n\n // Scatter flagged items\n Scatter(\n items,\n selection_flags,\n selection_indices,\n num_tile_items,\n num_tile_selections,\n num_selections_prefix,\n num_rejected_prefix,\n num_selections);\n\n return num_selections;\n }\n\n\n /**\n * Process a tile of input\n */\n template \n __device__ __forceinline__ OffsetT ConsumeTile(\n int num_tile_items, ///< Number of input items comprising this tile\n int tile_idx, ///< Tile index\n OffsetT tile_offset, ///< Tile offset\n ScanTileStateT& tile_state) ///< Global tile state descriptor\n {\n OffsetT num_selections;\n if (tile_idx == 0)\n {\n num_selections = ConsumeFirstTile(num_tile_items, tile_offset, tile_state);\n }\n else\n {\n num_selections = ConsumeSubsequentTile(num_tile_items, tile_idx, tile_offset, tile_state);\n }\n\n return num_selections;\n }\n\n\n /**\n * Scan tiles of items as part of a dynamic chained scan\n */\n template ///< Output iterator type for recording number of items selection_flags\n __device__ __forceinline__ void ConsumeRange(\n int num_tiles, ///< Total number of input tiles\n ScanTileStateT& tile_state, ///< Global tile state descriptor\n NumSelectedIteratorT d_num_selected_out) ///< Output total number selection_flags\n {\n // Blocks are launched in increasing order, so just assign one tile per block\n int tile_idx = (blockIdx.x * gridDim.y) + blockIdx.y; // Current tile index\n OffsetT tile_offset = tile_idx * TILE_ITEMS; // Global offset for the current tile\n\n if (tile_idx < num_tiles - 1)\n {\n // Not the last tile (full)\n ConsumeTile(TILE_ITEMS, tile_idx, tile_offset, tile_state);\n }\n else\n {\n // The last tile (possibly partially-full)\n OffsetT num_remaining = num_items - tile_offset;\n OffsetT num_selections = ConsumeTile(num_remaining, tile_idx, tile_offset, tile_state);\n\n if (threadIdx.x == 0)\n {\n // Output the total number of items selection_flags\n *d_num_selected_out = num_selections;\n }\n }\n }\n\n};\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/agent_spmv_orig.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::AgentSpmv implements a stateful abstraction of CUDA thread blocks for participating in device-wide SpMV.\n */\n\n#pragma once\n\n#include \n\n#include \"../util_type.cuh\"\n#include \"../block/block_reduce.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../block/block_exchange.cuh\"\n#include \"../thread/thread_search.cuh\"\n#include \"../thread/thread_operators.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../iterator/counting_input_iterator.cuh\"\n#include \"../iterator/tex_ref_input_iterator.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Tuning policy\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for AgentSpmv\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n CacheLoadModifier _ROW_OFFSETS_SEARCH_LOAD_MODIFIER, ///< Cache load modifier for reading CSR row-offsets during search\n CacheLoadModifier _ROW_OFFSETS_LOAD_MODIFIER, ///< Cache load modifier for reading CSR row-offsets\n CacheLoadModifier _COLUMN_INDICES_LOAD_MODIFIER, ///< Cache load modifier for reading CSR column-indices\n CacheLoadModifier _VALUES_LOAD_MODIFIER, ///< Cache load modifier for reading CSR values\n CacheLoadModifier _VECTOR_VALUES_LOAD_MODIFIER, ///< Cache load modifier for reading vector values\n bool _DIRECT_LOAD_NONZEROS, ///< Whether to load nonzeros directly from global during sequential merging (vs. pre-staged through shared memory)\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct AgentSpmvPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n DIRECT_LOAD_NONZEROS = _DIRECT_LOAD_NONZEROS, ///< Whether to load nonzeros directly from global during sequential merging (pre-staged through shared memory)\n };\n\n static const CacheLoadModifier ROW_OFFSETS_SEARCH_LOAD_MODIFIER = _ROW_OFFSETS_SEARCH_LOAD_MODIFIER; ///< Cache load modifier for reading CSR row-offsets\n static const CacheLoadModifier ROW_OFFSETS_LOAD_MODIFIER = _ROW_OFFSETS_LOAD_MODIFIER; ///< Cache load modifier for reading CSR row-offsets\n static const CacheLoadModifier COLUMN_INDICES_LOAD_MODIFIER = _COLUMN_INDICES_LOAD_MODIFIER; ///< Cache load modifier for reading CSR column-indices\n static const CacheLoadModifier VALUES_LOAD_MODIFIER = _VALUES_LOAD_MODIFIER; ///< Cache load modifier for reading CSR values\n static const CacheLoadModifier VECTOR_VALUES_LOAD_MODIFIER = _VECTOR_VALUES_LOAD_MODIFIER; ///< Cache load modifier for reading vector values\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n\n};\n\n\n/******************************************************************************\n * Thread block abstractions\n ******************************************************************************/\n\ntemplate <\n typename ValueT, ///< Matrix and vector value type\n typename OffsetT> ///< Signed integer type for sequence offsets\nstruct SpmvParams\n{\n ValueT* d_values; ///< Pointer to the array of \\p num_nonzeros values of the corresponding nonzero elements of matrix A.\n OffsetT* d_row_end_offsets; ///< Pointer to the array of \\p m offsets demarcating the end of every row in \\p d_column_indices and \\p d_values\n OffsetT* d_column_indices; ///< Pointer to the array of \\p num_nonzeros column-indices of the corresponding nonzero elements of matrix A. (Indices are zero-valued.)\n ValueT* d_vector_x; ///< Pointer to the array of \\p num_cols values corresponding to the dense input vector x\n ValueT* d_vector_y; ///< Pointer to the array of \\p num_rows values corresponding to the dense output vector y\n int num_rows; ///< Number of rows of matrix A.\n int num_cols; ///< Number of columns of matrix A.\n int num_nonzeros; ///< Number of nonzero elements of matrix A.\n ValueT alpha; ///< Alpha multiplicand\n ValueT beta; ///< Beta addend-multiplicand\n\n TexRefInputIterator t_vector_x;\n};\n\n\n/**\n * \\brief AgentSpmv implements a stateful abstraction of CUDA thread blocks for participating in device-wide SpMV.\n */\ntemplate <\n typename AgentSpmvPolicyT, ///< Parameterized AgentSpmvPolicy tuning policy type\n typename ValueT, ///< Matrix and vector value type\n typename OffsetT, ///< Signed integer type for sequence offsets\n bool HAS_ALPHA, ///< Whether the input parameter \\p alpha is 1\n bool HAS_BETA, ///< Whether the input parameter \\p beta is 0\n int PTX_ARCH = CUB_PTX_ARCH> ///< PTX compute capability\nstruct AgentSpmv\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// Constants\n enum\n {\n BLOCK_THREADS = AgentSpmvPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = AgentSpmvPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n /// 2D merge path coordinate type\n typedef typename CubVector::Type CoordinateT;\n\n /// Input iterator wrapper types (for applying cache modifiers)\n\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::ROW_OFFSETS_SEARCH_LOAD_MODIFIER,\n OffsetT,\n OffsetT>\n RowOffsetsSearchIteratorT;\n\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::ROW_OFFSETS_LOAD_MODIFIER,\n OffsetT,\n OffsetT>\n RowOffsetsIteratorT;\n\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::COLUMN_INDICES_LOAD_MODIFIER,\n OffsetT,\n OffsetT>\n ColumnIndicesIteratorT;\n\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::VALUES_LOAD_MODIFIER,\n ValueT,\n OffsetT>\n ValueIteratorT;\n\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::VECTOR_VALUES_LOAD_MODIFIER,\n ValueT,\n OffsetT>\n VectorValueIteratorT;\n\n // Tuple type for scanning (pairs accumulated segment-value with segment-index)\n typedef KeyValuePair KeyValuePairT;\n\n // Reduce-value-by-segment scan operator\n typedef ReduceByKeyOp ReduceBySegmentOpT;\n\n // BlockReduce specialization\n typedef BlockReduce<\n ValueT,\n BLOCK_THREADS,\n BLOCK_REDUCE_WARP_REDUCTIONS>\n BlockReduceT;\n\n // BlockScan specialization\n typedef BlockScan<\n KeyValuePairT,\n BLOCK_THREADS,\n AgentSpmvPolicyT::SCAN_ALGORITHM>\n BlockScanT;\n\n // BlockScan specialization\n typedef BlockScan<\n ValueT,\n BLOCK_THREADS,\n AgentSpmvPolicyT::SCAN_ALGORITHM>\n BlockPrefixSumT;\n\n // BlockExchange specialization\n typedef BlockExchange<\n ValueT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD>\n BlockExchangeT;\n\n /// Merge item type (either a non-zero value or a row-end offset)\n union MergeItem\n {\n // Value type to pair with index type OffsetT (NullType if loading values directly during merge)\n typedef typename If::Type MergeValueT;\n\n OffsetT row_end_offset;\n MergeValueT nonzero;\n };\n\n /// Shared memory type required by this thread block\n struct _TempStorage\n {\n CoordinateT tile_coords[2];\n\n union Aliasable\n {\n // Smem needed for tile of merge items\n MergeItem merge_items[ITEMS_PER_THREAD + TILE_ITEMS + 1];\n\n // Smem needed for block exchange\n typename BlockExchangeT::TempStorage exchange;\n\n // Smem needed for block-wide reduction\n typename BlockReduceT::TempStorage reduce;\n\n // Smem needed for tile scanning\n typename BlockScanT::TempStorage scan;\n\n // Smem needed for tile prefix sum\n typename BlockPrefixSumT::TempStorage prefix_sum;\n\n } aliasable;\n };\n\n /// Temporary storage type (unionable)\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n\n _TempStorage& temp_storage; /// Reference to temp_storage\n\n SpmvParams& spmv_params;\n\n ValueIteratorT wd_values; ///< Wrapped pointer to the array of \\p num_nonzeros values of the corresponding nonzero elements of matrix A.\n RowOffsetsIteratorT wd_row_end_offsets; ///< Wrapped Pointer to the array of \\p m offsets demarcating the end of every row in \\p d_column_indices and \\p d_values\n ColumnIndicesIteratorT wd_column_indices; ///< Wrapped Pointer to the array of \\p num_nonzeros column-indices of the corresponding nonzero elements of matrix A. (Indices are zero-valued.)\n VectorValueIteratorT wd_vector_x; ///< Wrapped Pointer to the array of \\p num_cols values corresponding to the dense input vector x\n VectorValueIteratorT wd_vector_y; ///< Wrapped Pointer to the array of \\p num_cols values corresponding to the dense input vector x\n\n\n //---------------------------------------------------------------------\n // Interface\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__ AgentSpmv(\n TempStorage& temp_storage, ///< Reference to temp_storage\n SpmvParams& spmv_params) ///< SpMV input parameter bundle\n :\n temp_storage(temp_storage.Alias()),\n spmv_params(spmv_params),\n wd_values(spmv_params.d_values),\n wd_row_end_offsets(spmv_params.d_row_end_offsets),\n wd_column_indices(spmv_params.d_column_indices),\n wd_vector_x(spmv_params.d_vector_x),\n wd_vector_y(spmv_params.d_vector_y)\n {}\n\n\n\n\n /**\n * Consume a merge tile, specialized for direct-load of nonzeros\n */\n __device__ __forceinline__ KeyValuePairT ConsumeTile(\n int tile_idx,\n CoordinateT tile_start_coord,\n CoordinateT tile_end_coord,\n Int2Type is_direct_load) ///< Marker type indicating whether to load nonzeros directly during path-discovery or beforehand in batch\n {\n int tile_num_rows = tile_end_coord.x - tile_start_coord.x;\n int tile_num_nonzeros = tile_end_coord.y - tile_start_coord.y;\n OffsetT* s_tile_row_end_offsets = &temp_storage.aliasable.merge_items[0].row_end_offset;\n\n // Gather the row end-offsets for the merge tile into shared memory\n for (int item = threadIdx.x; item <= tile_num_rows; item += BLOCK_THREADS)\n {\n s_tile_row_end_offsets[item] = wd_row_end_offsets[tile_start_coord.x + item];\n }\n\n CTA_SYNC();\n\n // Search for the thread's starting coordinate within the merge tile\n CountingInputIterator tile_nonzero_indices(tile_start_coord.y);\n CoordinateT thread_start_coord;\n\n MergePathSearch(\n OffsetT(threadIdx.x * ITEMS_PER_THREAD), // Diagonal\n s_tile_row_end_offsets, // List A\n tile_nonzero_indices, // List B\n tile_num_rows,\n tile_num_nonzeros,\n thread_start_coord);\n\n CTA_SYNC(); // Perf-sync\n\n // Compute the thread's merge path segment\n CoordinateT thread_current_coord = thread_start_coord;\n KeyValuePairT scan_segment[ITEMS_PER_THREAD];\n\n ValueT running_total = 0.0;\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n OffsetT nonzero_idx = CUB_MIN(tile_nonzero_indices[thread_current_coord.y], spmv_params.num_nonzeros - 1);\n OffsetT column_idx = wd_column_indices[nonzero_idx];\n ValueT value = wd_values[nonzero_idx];\n\n ValueT vector_value = spmv_params.t_vector_x[column_idx];\n#if (CUB_PTX_ARCH >= 350)\n vector_value = wd_vector_x[column_idx];\n#endif\n ValueT nonzero = value * vector_value;\n\n OffsetT row_end_offset = s_tile_row_end_offsets[thread_current_coord.x];\n\n if (tile_nonzero_indices[thread_current_coord.y] < row_end_offset)\n {\n // Move down (accumulate)\n running_total += nonzero;\n scan_segment[ITEM].value = running_total;\n scan_segment[ITEM].key = tile_num_rows;\n ++thread_current_coord.y;\n }\n else\n {\n // Move right (reset)\n scan_segment[ITEM].value = running_total;\n scan_segment[ITEM].key = thread_current_coord.x;\n running_total = 0.0;\n ++thread_current_coord.x;\n }\n }\n\n CTA_SYNC();\n\n // Block-wide reduce-value-by-segment\n KeyValuePairT tile_carry;\n ReduceBySegmentOpT scan_op;\n KeyValuePairT scan_item;\n\n scan_item.value = running_total;\n scan_item.key = thread_current_coord.x;\n\n BlockScanT(temp_storage.aliasable.scan).ExclusiveScan(scan_item, scan_item, scan_op, tile_carry);\n\n if (tile_num_rows > 0)\n {\n if (threadIdx.x == 0)\n scan_item.key = -1;\n\n // Direct scatter\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (scan_segment[ITEM].key < tile_num_rows)\n {\n if (scan_item.key == scan_segment[ITEM].key)\n scan_segment[ITEM].value = scan_item.value + scan_segment[ITEM].value;\n\n if (HAS_ALPHA)\n {\n scan_segment[ITEM].value *= spmv_params.alpha;\n }\n\n if (HAS_BETA)\n {\n // Update the output vector element\n ValueT addend = spmv_params.beta * wd_vector_y[tile_start_coord.x + scan_segment[ITEM].key];\n scan_segment[ITEM].value += addend;\n }\n\n // Set the output vector element\n spmv_params.d_vector_y[tile_start_coord.x + scan_segment[ITEM].key] = scan_segment[ITEM].value;\n }\n }\n }\n\n // Return the tile's running carry-out\n return tile_carry;\n }\n\n\n\n /**\n * Consume a merge tile, specialized for indirect load of nonzeros\n */\n __device__ __forceinline__ KeyValuePairT ConsumeTile(\n int tile_idx,\n CoordinateT tile_start_coord,\n CoordinateT tile_end_coord,\n Int2Type is_direct_load) ///< Marker type indicating whether to load nonzeros directly during path-discovery or beforehand in batch\n {\n int tile_num_rows = tile_end_coord.x - tile_start_coord.x;\n int tile_num_nonzeros = tile_end_coord.y - tile_start_coord.y;\n\n#if (CUB_PTX_ARCH >= 520)\n\n/*\n OffsetT* s_tile_row_end_offsets = &temp_storage.merge_items[tile_num_nonzeros].row_end_offset;\n ValueT* s_tile_nonzeros = &temp_storage.merge_items[0].nonzero;\n\n OffsetT col_indices[ITEMS_PER_THREAD];\n ValueT mat_values[ITEMS_PER_THREAD];\n int nonzero_indices[ITEMS_PER_THREAD];\n\n // Gather the nonzeros for the merge tile into shared memory\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n nonzero_indices[ITEM] = threadIdx.x + (ITEM * BLOCK_THREADS);\n\n ValueIteratorT a = wd_values + tile_start_coord.y + nonzero_indices[ITEM];\n ColumnIndicesIteratorT ci = wd_column_indices + tile_start_coord.y + nonzero_indices[ITEM];\n\n col_indices[ITEM] = (nonzero_indices[ITEM] < tile_num_nonzeros) ? *ci : 0;\n mat_values[ITEM] = (nonzero_indices[ITEM] < tile_num_nonzeros) ? *a : 0.0;\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n VectorValueIteratorT x = wd_vector_x + col_indices[ITEM];\n mat_values[ITEM] *= *x;\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n ValueT *s = s_tile_nonzeros + nonzero_indices[ITEM];\n\n *s = mat_values[ITEM];\n }\n\n CTA_SYNC();\n\n*/\n\n OffsetT* s_tile_row_end_offsets = &temp_storage.merge_items[0].row_end_offset;\n ValueT* s_tile_nonzeros = &temp_storage.merge_items[tile_num_rows + ITEMS_PER_THREAD].nonzero;\n\n // Gather the nonzeros for the merge tile into shared memory\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int nonzero_idx = threadIdx.x + (ITEM * BLOCK_THREADS);\n\n ValueIteratorT a = wd_values + tile_start_coord.y + nonzero_idx;\n ColumnIndicesIteratorT ci = wd_column_indices + tile_start_coord.y + nonzero_idx;\n ValueT* s = s_tile_nonzeros + nonzero_idx;\n\n if (nonzero_idx < tile_num_nonzeros)\n {\n\n OffsetT column_idx = *ci;\n ValueT value = *a;\n\n ValueT vector_value = spmv_params.t_vector_x[column_idx];\n vector_value = wd_vector_x[column_idx];\n\n ValueT nonzero = value * vector_value;\n\n *s = nonzero;\n }\n }\n\n\n#else\n\n OffsetT* s_tile_row_end_offsets = &temp_storage.aliasable.merge_items[0].row_end_offset;\n ValueT* s_tile_nonzeros = &temp_storage.aliasable.merge_items[tile_num_rows + ITEMS_PER_THREAD].nonzero;\n\n // Gather the nonzeros for the merge tile into shared memory\n if (tile_num_nonzeros > 0)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int nonzero_idx = threadIdx.x + (ITEM * BLOCK_THREADS);\n nonzero_idx = CUB_MIN(nonzero_idx, tile_num_nonzeros - 1);\n\n OffsetT column_idx = wd_column_indices[tile_start_coord.y + nonzero_idx];\n ValueT value = wd_values[tile_start_coord.y + nonzero_idx];\n\n ValueT vector_value = spmv_params.t_vector_x[column_idx];\n#if (CUB_PTX_ARCH >= 350)\n vector_value = wd_vector_x[column_idx];\n#endif\n ValueT nonzero = value * vector_value;\n\n s_tile_nonzeros[nonzero_idx] = nonzero;\n }\n }\n\n#endif\n\n // Gather the row end-offsets for the merge tile into shared memory\n #pragma unroll 1\n for (int item = threadIdx.x; item <= tile_num_rows; item += BLOCK_THREADS)\n {\n s_tile_row_end_offsets[item] = wd_row_end_offsets[tile_start_coord.x + item];\n }\n\n CTA_SYNC();\n\n // Search for the thread's starting coordinate within the merge tile\n CountingInputIterator tile_nonzero_indices(tile_start_coord.y);\n CoordinateT thread_start_coord;\n\n MergePathSearch(\n OffsetT(threadIdx.x * ITEMS_PER_THREAD), // Diagonal\n s_tile_row_end_offsets, // List A\n tile_nonzero_indices, // List B\n tile_num_rows,\n tile_num_nonzeros,\n thread_start_coord);\n\n CTA_SYNC(); // Perf-sync\n\n // Compute the thread's merge path segment\n CoordinateT thread_current_coord = thread_start_coord;\n KeyValuePairT scan_segment[ITEMS_PER_THREAD];\n ValueT running_total = 0.0;\n\n OffsetT row_end_offset = s_tile_row_end_offsets[thread_current_coord.x];\n ValueT nonzero = s_tile_nonzeros[thread_current_coord.y];\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (tile_nonzero_indices[thread_current_coord.y] < row_end_offset)\n {\n // Move down (accumulate)\n scan_segment[ITEM].value = nonzero;\n running_total += nonzero;\n ++thread_current_coord.y;\n nonzero = s_tile_nonzeros[thread_current_coord.y];\n }\n else\n {\n // Move right (reset)\n scan_segment[ITEM].value = 0.0;\n running_total = 0.0;\n ++thread_current_coord.x;\n row_end_offset = s_tile_row_end_offsets[thread_current_coord.x];\n }\n\n scan_segment[ITEM].key = thread_current_coord.x;\n }\n\n CTA_SYNC();\n\n // Block-wide reduce-value-by-segment\n KeyValuePairT tile_carry;\n ReduceBySegmentOpT scan_op;\n KeyValuePairT scan_item;\n\n scan_item.value = running_total;\n scan_item.key = thread_current_coord.x;\n\n BlockScanT(temp_storage.aliasable.scan).ExclusiveScan(scan_item, scan_item, scan_op, tile_carry);\n\n if (threadIdx.x == 0)\n {\n scan_item.key = thread_start_coord.x;\n scan_item.value = 0.0;\n }\n\n if (tile_num_rows > 0)\n {\n\n CTA_SYNC();\n\n // Scan downsweep and scatter\n ValueT* s_partials = &temp_storage.aliasable.merge_items[0].nonzero;\n\n if (scan_item.key != scan_segment[0].key)\n {\n s_partials[scan_item.key] = scan_item.value;\n }\n else\n {\n scan_segment[0].value += scan_item.value;\n }\n\n #pragma unroll\n for (int ITEM = 1; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (scan_segment[ITEM - 1].key != scan_segment[ITEM].key)\n {\n s_partials[scan_segment[ITEM - 1].key] = scan_segment[ITEM - 1].value;\n }\n else\n {\n scan_segment[ITEM].value += scan_segment[ITEM - 1].value;\n }\n }\n\n CTA_SYNC();\n\n #pragma unroll 1\n for (int item = threadIdx.x; item < tile_num_rows; item += BLOCK_THREADS)\n {\n spmv_params.d_vector_y[tile_start_coord.x + item] = s_partials[item];\n }\n }\n\n // Return the tile's running carry-out\n return tile_carry;\n }\n\n\n\n\n\n\n\n /**\n * Consume a merge tile, specialized for indirect load of nonzeros\n * /\n template \n __device__ __forceinline__ KeyValuePairT ConsumeTile1(\n int tile_idx,\n CoordinateT tile_start_coord,\n CoordinateT tile_end_coord,\n IsDirectLoadT is_direct_load) ///< Marker type indicating whether to load nonzeros directly during path-discovery or beforehand in batch\n {\n int tile_num_rows = tile_end_coord.x - tile_start_coord.x;\n int tile_num_nonzeros = tile_end_coord.y - tile_start_coord.y;\n\n OffsetT* s_tile_row_end_offsets = &temp_storage.merge_items[0].row_end_offset;\n\n int warp_idx = threadIdx.x / WARP_THREADS;\n int lane_idx = LaneId();\n\n // Gather the row end-offsets for the merge tile into shared memory\n #pragma unroll 1\n for (int item = threadIdx.x; item <= tile_num_rows; item += BLOCK_THREADS)\n {\n s_tile_row_end_offsets[item] = wd_row_end_offsets[tile_start_coord.x + item];\n }\n\n CTA_SYNC();\n\n // Search for warp start/end coords\n if (lane_idx == 0)\n {\n MergePathSearch(\n OffsetT(warp_idx * ITEMS_PER_WARP), // Diagonal\n s_tile_row_end_offsets, // List A\n CountingInputIterator(tile_start_coord.y), // List B\n tile_num_rows,\n tile_num_nonzeros,\n temp_storage.warp_coords[warp_idx]);\n\n CoordinateT last = {tile_num_rows, tile_num_nonzeros};\n temp_storage.warp_coords[WARPS] = last;\n }\n\n CTA_SYNC();\n\n CoordinateT warp_coord = temp_storage.warp_coords[warp_idx];\n CoordinateT warp_end_coord = temp_storage.warp_coords[warp_idx + 1];\n OffsetT warp_nonzero_idx = tile_start_coord.y + warp_coord.y;\n\n // Consume whole rows\n #pragma unroll 1\n for (; warp_coord.x < warp_end_coord.x; ++warp_coord.x)\n {\n ValueT row_total = 0.0;\n OffsetT row_end_offset = s_tile_row_end_offsets[warp_coord.x];\n\n #pragma unroll 1\n for (OffsetT nonzero_idx = warp_nonzero_idx + lane_idx;\n nonzero_idx < row_end_offset;\n nonzero_idx += WARP_THREADS)\n {\n OffsetT column_idx = wd_column_indices[nonzero_idx];\n ValueT value = wd_values[nonzero_idx];\n ValueT vector_value = wd_vector_x[column_idx];\n row_total += value * vector_value;\n }\n\n // Warp reduce\n row_total = WarpReduceT(temp_storage.warp_reduce[warp_idx]).Sum(row_total);\n\n // Output\n if (lane_idx == 0)\n {\n spmv_params.d_vector_y[tile_start_coord.x + warp_coord.x] = row_total;\n }\n\n warp_nonzero_idx = row_end_offset;\n }\n\n // Consume partial portion of thread's last row\n if (warp_nonzero_idx < tile_start_coord.y + warp_end_coord.y)\n {\n ValueT row_total = 0.0;\n for (OffsetT nonzero_idx = warp_nonzero_idx + lane_idx;\n nonzero_idx < tile_start_coord.y + warp_end_coord.y;\n nonzero_idx += WARP_THREADS)\n {\n\n OffsetT column_idx = wd_column_indices[nonzero_idx];\n ValueT value = wd_values[nonzero_idx];\n ValueT vector_value = wd_vector_x[column_idx];\n row_total += value * vector_value;\n }\n\n // Warp reduce\n row_total = WarpReduceT(temp_storage.warp_reduce[warp_idx]).Sum(row_total);\n\n // Output\n if (lane_idx == 0)\n {\n spmv_params.d_vector_y[tile_start_coord.x + warp_coord.x] = row_total;\n }\n }\n\n // Return the tile's running carry-out\n KeyValuePairT tile_carry(tile_num_rows, 0.0);\n return tile_carry;\n }\n*/\n\n\n\n\n\n\n\n /**\n * Consume a merge tile, specialized for indirect load of nonzeros\n * /\n __device__ __forceinline__ KeyValuePairT ConsumeTile2(\n int tile_idx,\n CoordinateT tile_start_coord,\n CoordinateT tile_end_coord,\n Int2Type is_direct_load) ///< Marker type indicating whether to load nonzeros directly during path-discovery or beforehand in batch\n {\n int tile_num_rows = tile_end_coord.x - tile_start_coord.x;\n int tile_num_nonzeros = tile_end_coord.y - tile_start_coord.y;\n\n ValueT* s_tile_nonzeros = &temp_storage.merge_items[0].nonzero;\n\n ValueT nonzeros[ITEMS_PER_THREAD];\n\n // Gather the nonzeros for the merge tile into shared memory\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int nonzero_idx = threadIdx.x + (ITEM * BLOCK_THREADS);\n nonzero_idx = CUB_MIN(nonzero_idx, tile_num_nonzeros - 1);\n\n OffsetT column_idx = wd_column_indices[tile_start_coord.y + nonzero_idx];\n ValueT value = wd_values[tile_start_coord.y + nonzero_idx];\n\n ValueT vector_value = spmv_params.t_vector_x[column_idx];\n#if (CUB_PTX_ARCH >= 350)\n vector_value = wd_vector_x[column_idx];\n#endif\n\n nonzeros[ITEM] = value * vector_value;\n }\n\n // Exchange striped->blocked\n BlockExchangeT(temp_storage.exchange).StripedToBlocked(nonzeros);\n\n CTA_SYNC();\n\n // Compute an inclusive prefix sum\n BlockPrefixSumT(temp_storage.prefix_sum).InclusiveSum(nonzeros, nonzeros);\n\n CTA_SYNC();\n\n if (threadIdx.x == 0)\n s_tile_nonzeros[0] = 0.0;\n\n // Scatter back to smem\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int item_idx = (threadIdx.x * ITEMS_PER_THREAD) + ITEM + 1;\n s_tile_nonzeros[item_idx] = nonzeros[ITEM];\n }\n\n CTA_SYNC();\n\n // Gather the row end-offsets for the merge tile into shared memory\n #pragma unroll 1\n for (int item = threadIdx.x; item < tile_num_rows; item += BLOCK_THREADS)\n {\n OffsetT start = CUB_MAX(wd_row_end_offsets[tile_start_coord.x + item - 1], tile_start_coord.y);\n OffsetT end = wd_row_end_offsets[tile_start_coord.x + item];\n\n start -= tile_start_coord.y;\n end -= tile_start_coord.y;\n\n ValueT row_partial = s_tile_nonzeros[end] - s_tile_nonzeros[start];\n\n spmv_params.d_vector_y[tile_start_coord.x + item] = row_partial;\n }\n\n // Get the tile's carry-out\n KeyValuePairT tile_carry;\n if (threadIdx.x == 0)\n {\n tile_carry.key = tile_num_rows;\n\n OffsetT start = CUB_MAX(wd_row_end_offsets[tile_end_coord.x - 1], tile_start_coord.y);\n start -= tile_start_coord.y;\n OffsetT end = tile_num_nonzeros;\n\n tile_carry.value = s_tile_nonzeros[end] - s_tile_nonzeros[start];\n }\n\n // Return the tile's running carry-out\n return tile_carry;\n }\n*/\n\n\n /**\n * Consume input tile\n */\n __device__ __forceinline__ void ConsumeTile(\n CoordinateT* d_tile_coordinates, ///< [in] Pointer to the temporary array of tile starting coordinates\n KeyValuePairT* d_tile_carry_pairs, ///< [out] Pointer to the temporary array carry-out dot product row-ids, one per block\n int num_merge_tiles) ///< [in] Number of merge tiles\n {\n int tile_idx = (blockIdx.x * gridDim.y) + blockIdx.y; // Current tile index\n\n if (tile_idx >= num_merge_tiles)\n return;\n\n // Read our starting coordinates\n if (threadIdx.x < 2)\n {\n if (d_tile_coordinates == NULL)\n {\n // Search our starting coordinates\n OffsetT diagonal = (tile_idx + threadIdx.x) * TILE_ITEMS;\n CoordinateT tile_coord;\n CountingInputIterator nonzero_indices(0);\n\n // Search the merge path\n MergePathSearch(\n diagonal,\n RowOffsetsSearchIteratorT(spmv_params.d_row_end_offsets),\n nonzero_indices,\n spmv_params.num_rows,\n spmv_params.num_nonzeros,\n tile_coord);\n\n temp_storage.tile_coords[threadIdx.x] = tile_coord;\n }\n else\n {\n temp_storage.tile_coords[threadIdx.x] = d_tile_coordinates[tile_idx + threadIdx.x];\n }\n }\n\n CTA_SYNC();\n\n CoordinateT tile_start_coord = temp_storage.tile_coords[0];\n CoordinateT tile_end_coord = temp_storage.tile_coords[1];\n\n // Consume multi-segment tile\n KeyValuePairT tile_carry = ConsumeTile(\n tile_idx,\n tile_start_coord,\n tile_end_coord,\n Int2Type());\n\n // Output the tile's carry-out\n if (threadIdx.x == 0)\n {\n if (HAS_ALPHA)\n tile_carry.value *= spmv_params.alpha;\n\n tile_carry.key += tile_start_coord.x;\n d_tile_carry_pairs[tile_idx] = tile_carry;\n }\n }\n\n\n};\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/agent/single_pass_scan_operators.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Callback operator types for supplying BlockScan prefixes\n */\n\n#pragma once\n\n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../warp/warp_reduce.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Prefix functor type for maintaining a running prefix while scanning a\n * region independent of other thread blocks\n ******************************************************************************/\n\n/**\n * Stateful callback operator type for supplying BlockScan prefixes.\n * Maintains a running prefix that can be applied to consecutive\n * BlockScan operations.\n */\ntemplate <\n typename T, ///< BlockScan value type\n typename ScanOpT> ///< Wrapped scan operator type\nstruct BlockScanRunningPrefixOp\n{\n ScanOpT op; ///< Wrapped scan operator\n T running_total; ///< Running block-wide prefix\n\n /// Constructor\n __device__ __forceinline__ BlockScanRunningPrefixOp(ScanOpT op)\n :\n op(op)\n {}\n\n /// Constructor\n __device__ __forceinline__ BlockScanRunningPrefixOp(\n T starting_prefix,\n ScanOpT op)\n :\n op(op),\n running_total(starting_prefix)\n {}\n\n /**\n * Prefix callback operator. Returns the block-wide running_total in thread-0.\n */\n __device__ __forceinline__ T operator()(\n const T &block_aggregate) ///< The aggregate sum of the BlockScan inputs\n {\n T retval = running_total;\n running_total = op(running_total, block_aggregate);\n return retval;\n }\n};\n\n\n/******************************************************************************\n * Generic tile status interface types for block-cooperative scans\n ******************************************************************************/\n\n/**\n * Enumerations of tile status\n */\nenum ScanTileStatus\n{\n SCAN_TILE_OOB, // Out-of-bounds (e.g., padding)\n SCAN_TILE_INVALID = 99, // Not yet processed\n SCAN_TILE_PARTIAL, // Tile aggregate is available\n SCAN_TILE_INCLUSIVE, // Inclusive tile prefix is available\n};\n\n\n/**\n * Tile status interface.\n */\ntemplate <\n typename T,\n bool SINGLE_WORD = Traits::PRIMITIVE>\nstruct ScanTileState;\n\n\n/**\n * Tile status interface specialized for scan status and value types\n * that can be combined into one machine word that can be\n * read/written coherently in a single access.\n */\ntemplate \nstruct ScanTileState\n{\n // Status word type\n typedef typename If<(sizeof(T) == 8),\n long long,\n typename If<(sizeof(T) == 4),\n int,\n typename If<(sizeof(T) == 2),\n short,\n char>::Type>::Type>::Type StatusWord;\n\n\n // Unit word type\n typedef typename If<(sizeof(T) == 8),\n longlong2,\n typename If<(sizeof(T) == 4),\n int2,\n typename If<(sizeof(T) == 2),\n int,\n uchar2>::Type>::Type>::Type TxnWord;\n\n\n // Device word type\n struct TileDescriptor\n {\n StatusWord status;\n T value;\n };\n\n\n // Constants\n enum\n {\n TILE_STATUS_PADDING = CUB_PTX_WARP_THREADS,\n };\n\n\n // Device storage\n TxnWord *d_tile_descriptors;\n\n /// Constructor\n __host__ __device__ __forceinline__\n ScanTileState()\n :\n d_tile_descriptors(NULL)\n {}\n\n\n /// Initializer\n __host__ __device__ __forceinline__\n cudaError_t Init(\n int /*num_tiles*/, ///< [in] Number of tiles\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t /*temp_storage_bytes*/) ///< [in] Size in bytes of \\t d_temp_storage allocation\n {\n d_tile_descriptors = reinterpret_cast(d_temp_storage);\n return cudaSuccess;\n }\n\n\n /**\n * Compute device memory needed for tile status\n */\n __host__ __device__ __forceinline__\n static cudaError_t AllocationSize(\n int num_tiles, ///< [in] Number of tiles\n size_t &temp_storage_bytes) ///< [out] Size in bytes of \\t d_temp_storage allocation\n {\n temp_storage_bytes = (num_tiles + TILE_STATUS_PADDING) * sizeof(TileDescriptor); // bytes needed for tile status descriptors\n return cudaSuccess;\n }\n\n\n /**\n * Initialize (from device)\n */\n __device__ __forceinline__ void InitializeStatus(int num_tiles)\n {\n int tile_idx = (blockIdx.x * blockDim.x) + threadIdx.x;\n\n TxnWord val = TxnWord();\n TileDescriptor *descriptor = reinterpret_cast(&val);\n\n if (tile_idx < num_tiles)\n {\n // Not-yet-set\n descriptor->status = StatusWord(SCAN_TILE_INVALID);\n d_tile_descriptors[TILE_STATUS_PADDING + tile_idx] = val;\n }\n\n if ((blockIdx.x == 0) && (threadIdx.x < TILE_STATUS_PADDING))\n {\n // Padding\n descriptor->status = StatusWord(SCAN_TILE_OOB);\n d_tile_descriptors[threadIdx.x] = val;\n }\n }\n\n\n /**\n * Update the specified tile's inclusive value and corresponding status\n */\n __device__ __forceinline__ void SetInclusive(int tile_idx, T tile_inclusive)\n {\n TileDescriptor tile_descriptor;\n tile_descriptor.status = SCAN_TILE_INCLUSIVE;\n tile_descriptor.value = tile_inclusive;\n\n TxnWord alias;\n *reinterpret_cast(&alias) = tile_descriptor;\n ThreadStore(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx, alias);\n }\n\n\n /**\n * Update the specified tile's partial value and corresponding status\n */\n __device__ __forceinline__ void SetPartial(int tile_idx, T tile_partial)\n {\n TileDescriptor tile_descriptor;\n tile_descriptor.status = SCAN_TILE_PARTIAL;\n tile_descriptor.value = tile_partial;\n\n TxnWord alias;\n *reinterpret_cast(&alias) = tile_descriptor;\n ThreadStore(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx, alias);\n }\n\n /**\n * Wait for the corresponding tile to become non-invalid\n */\n __device__ __forceinline__ void WaitForValid(\n int tile_idx,\n StatusWord &status,\n T &value)\n {\n TileDescriptor tile_descriptor;\n do\n {\n __threadfence_block(); // prevent hoisting loads from loop\n TxnWord alias = ThreadLoad(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx);\n tile_descriptor = reinterpret_cast(alias);\n\n } while (WARP_ANY((tile_descriptor.status == SCAN_TILE_INVALID), 0xffffffff));\n\n status = tile_descriptor.status;\n value = tile_descriptor.value;\n }\n\n};\n\n\n\n/**\n * Tile status interface specialized for scan status and value types that\n * cannot be combined into one machine word.\n */\ntemplate \nstruct ScanTileState\n{\n // Status word type\n typedef char StatusWord;\n\n // Constants\n enum\n {\n TILE_STATUS_PADDING = CUB_PTX_WARP_THREADS,\n };\n\n // Device storage\n StatusWord *d_tile_status;\n T *d_tile_partial;\n T *d_tile_inclusive;\n\n /// Constructor\n __host__ __device__ __forceinline__\n ScanTileState()\n :\n d_tile_status(NULL),\n d_tile_partial(NULL),\n d_tile_inclusive(NULL)\n {}\n\n\n /// Initializer\n __host__ __device__ __forceinline__\n cudaError_t Init(\n int num_tiles, ///< [in] Number of tiles\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t temp_storage_bytes) ///< [in] Size in bytes of \\t d_temp_storage allocation\n {\n cudaError_t error = cudaSuccess;\n do\n {\n void* allocations[3];\n size_t allocation_sizes[3];\n\n allocation_sizes[0] = (num_tiles + TILE_STATUS_PADDING) * sizeof(StatusWord); // bytes needed for tile status descriptors\n allocation_sizes[1] = (num_tiles + TILE_STATUS_PADDING) * sizeof(Uninitialized); // bytes needed for partials\n allocation_sizes[2] = (num_tiles + TILE_STATUS_PADDING) * sizeof(Uninitialized); // bytes needed for inclusives\n\n // Compute allocation pointers into the single storage blob\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n\n // Alias the offsets\n d_tile_status = reinterpret_cast(allocations[0]);\n d_tile_partial = reinterpret_cast(allocations[1]);\n d_tile_inclusive = reinterpret_cast(allocations[2]);\n }\n while (0);\n\n return error;\n }\n\n\n /**\n * Compute device memory needed for tile status\n */\n __host__ __device__ __forceinline__\n static cudaError_t AllocationSize(\n int num_tiles, ///< [in] Number of tiles\n size_t &temp_storage_bytes) ///< [out] Size in bytes of \\t d_temp_storage allocation\n {\n // Specify storage allocation requirements\n size_t allocation_sizes[3];\n allocation_sizes[0] = (num_tiles + TILE_STATUS_PADDING) * sizeof(StatusWord); // bytes needed for tile status descriptors\n allocation_sizes[1] = (num_tiles + TILE_STATUS_PADDING) * sizeof(Uninitialized); // bytes needed for partials\n allocation_sizes[2] = (num_tiles + TILE_STATUS_PADDING) * sizeof(Uninitialized); // bytes needed for inclusives\n\n // Set the necessary size of the blob\n void* allocations[3];\n return CubDebug(AliasTemporaries(NULL, temp_storage_bytes, allocations, allocation_sizes));\n }\n\n\n /**\n * Initialize (from device)\n */\n __device__ __forceinline__ void InitializeStatus(int num_tiles)\n {\n int tile_idx = (blockIdx.x * blockDim.x) + threadIdx.x;\n if (tile_idx < num_tiles)\n {\n // Not-yet-set\n d_tile_status[TILE_STATUS_PADDING + tile_idx] = StatusWord(SCAN_TILE_INVALID);\n }\n\n if ((blockIdx.x == 0) && (threadIdx.x < TILE_STATUS_PADDING))\n {\n // Padding\n d_tile_status[threadIdx.x] = StatusWord(SCAN_TILE_OOB);\n }\n }\n\n\n /**\n * Update the specified tile's inclusive value and corresponding status\n */\n __device__ __forceinline__ void SetInclusive(int tile_idx, T tile_inclusive)\n {\n // Update tile inclusive value\n ThreadStore(d_tile_inclusive + TILE_STATUS_PADDING + tile_idx, tile_inclusive);\n\n // Fence\n __threadfence();\n\n // Update tile status\n ThreadStore(d_tile_status + TILE_STATUS_PADDING + tile_idx, StatusWord(SCAN_TILE_INCLUSIVE));\n }\n\n\n /**\n * Update the specified tile's partial value and corresponding status\n */\n __device__ __forceinline__ void SetPartial(int tile_idx, T tile_partial)\n {\n // Update tile partial value\n ThreadStore(d_tile_partial + TILE_STATUS_PADDING + tile_idx, tile_partial);\n\n // Fence\n __threadfence();\n\n // Update tile status\n ThreadStore(d_tile_status + TILE_STATUS_PADDING + tile_idx, StatusWord(SCAN_TILE_PARTIAL));\n }\n\n /**\n * Wait for the corresponding tile to become non-invalid\n */\n __device__ __forceinline__ void WaitForValid(\n int tile_idx,\n StatusWord &status,\n T &value)\n {\n do {\n status = ThreadLoad(d_tile_status + TILE_STATUS_PADDING + tile_idx);\n\n __threadfence(); // prevent hoisting loads from loop or loads below above this one\n\n } while (status == SCAN_TILE_INVALID);\n\n if (status == StatusWord(SCAN_TILE_PARTIAL)) \n value = ThreadLoad(d_tile_partial + TILE_STATUS_PADDING + tile_idx);\n else\n value = ThreadLoad(d_tile_inclusive + TILE_STATUS_PADDING + tile_idx);\n }\n};\n\n\n/******************************************************************************\n * ReduceByKey tile status interface types for block-cooperative scans\n ******************************************************************************/\n\n/**\n * Tile status interface for reduction by key.\n *\n */\ntemplate <\n typename ValueT,\n typename KeyT,\n bool SINGLE_WORD = (Traits::PRIMITIVE) && (sizeof(ValueT) + sizeof(KeyT) < 16)>\nstruct ReduceByKeyScanTileState;\n\n\n/**\n * Tile status interface for reduction by key, specialized for scan status and value types that\n * cannot be combined into one machine word.\n */\ntemplate <\n typename ValueT,\n typename KeyT>\nstruct ReduceByKeyScanTileState :\n ScanTileState >\n{\n typedef ScanTileState > SuperClass;\n\n /// Constructor\n __host__ __device__ __forceinline__\n ReduceByKeyScanTileState() : SuperClass() {}\n};\n\n\n/**\n * Tile status interface for reduction by key, specialized for scan status and value types that\n * can be combined into one machine word that can be read/written coherently in a single access.\n */\ntemplate <\n typename ValueT,\n typename KeyT>\nstruct ReduceByKeyScanTileState\n{\n typedef KeyValuePairKeyValuePairT;\n\n // Constants\n enum\n {\n PAIR_SIZE = sizeof(ValueT) + sizeof(KeyT),\n TXN_WORD_SIZE = 1 << Log2::VALUE,\n STATUS_WORD_SIZE = TXN_WORD_SIZE - PAIR_SIZE,\n\n TILE_STATUS_PADDING = CUB_PTX_WARP_THREADS,\n };\n\n // Status word type\n typedef typename If<(STATUS_WORD_SIZE == 8),\n long long,\n typename If<(STATUS_WORD_SIZE == 4),\n int,\n typename If<(STATUS_WORD_SIZE == 2),\n short,\n char>::Type>::Type>::Type StatusWord;\n\n // Status word type\n typedef typename If<(TXN_WORD_SIZE == 16),\n longlong2,\n typename If<(TXN_WORD_SIZE == 8),\n long long,\n int>::Type>::Type TxnWord;\n\n // Device word type (for when sizeof(ValueT) == sizeof(KeyT))\n struct TileDescriptorBigStatus\n {\n KeyT key;\n ValueT value;\n StatusWord status;\n };\n\n // Device word type (for when sizeof(ValueT) != sizeof(KeyT))\n struct TileDescriptorLittleStatus\n {\n ValueT value;\n StatusWord status;\n KeyT key;\n };\n\n // Device word type\n typedef typename If<\n (sizeof(ValueT) == sizeof(KeyT)),\n TileDescriptorBigStatus,\n TileDescriptorLittleStatus>::Type\n TileDescriptor;\n\n\n // Device storage\n TxnWord *d_tile_descriptors;\n\n\n /// Constructor\n __host__ __device__ __forceinline__\n ReduceByKeyScanTileState()\n :\n d_tile_descriptors(NULL)\n {}\n\n\n /// Initializer\n __host__ __device__ __forceinline__\n cudaError_t Init(\n int /*num_tiles*/, ///< [in] Number of tiles\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t /*temp_storage_bytes*/) ///< [in] Size in bytes of \\t d_temp_storage allocation\n {\n d_tile_descriptors = reinterpret_cast(d_temp_storage);\n return cudaSuccess;\n }\n\n\n /**\n * Compute device memory needed for tile status\n */\n __host__ __device__ __forceinline__\n static cudaError_t AllocationSize(\n int num_tiles, ///< [in] Number of tiles\n size_t &temp_storage_bytes) ///< [out] Size in bytes of \\t d_temp_storage allocation\n {\n temp_storage_bytes = (num_tiles + TILE_STATUS_PADDING) * sizeof(TileDescriptor); // bytes needed for tile status descriptors\n return cudaSuccess;\n }\n\n\n /**\n * Initialize (from device)\n */\n __device__ __forceinline__ void InitializeStatus(int num_tiles)\n {\n int tile_idx = (blockIdx.x * blockDim.x) + threadIdx.x;\n TxnWord val = TxnWord();\n TileDescriptor *descriptor = reinterpret_cast(&val);\n\n if (tile_idx < num_tiles)\n {\n // Not-yet-set\n descriptor->status = StatusWord(SCAN_TILE_INVALID);\n d_tile_descriptors[TILE_STATUS_PADDING + tile_idx] = val;\n }\n\n if ((blockIdx.x == 0) && (threadIdx.x < TILE_STATUS_PADDING))\n {\n // Padding\n descriptor->status = StatusWord(SCAN_TILE_OOB);\n d_tile_descriptors[threadIdx.x] = val;\n }\n }\n\n\n /**\n * Update the specified tile's inclusive value and corresponding status\n */\n __device__ __forceinline__ void SetInclusive(int tile_idx, KeyValuePairT tile_inclusive)\n {\n TileDescriptor tile_descriptor;\n tile_descriptor.status = SCAN_TILE_INCLUSIVE;\n tile_descriptor.value = tile_inclusive.value;\n tile_descriptor.key = tile_inclusive.key;\n\n TxnWord alias;\n *reinterpret_cast(&alias) = tile_descriptor;\n ThreadStore(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx, alias);\n }\n\n\n /**\n * Update the specified tile's partial value and corresponding status\n */\n __device__ __forceinline__ void SetPartial(int tile_idx, KeyValuePairT tile_partial)\n {\n TileDescriptor tile_descriptor;\n tile_descriptor.status = SCAN_TILE_PARTIAL;\n tile_descriptor.value = tile_partial.value;\n tile_descriptor.key = tile_partial.key;\n\n TxnWord alias;\n *reinterpret_cast(&alias) = tile_descriptor;\n ThreadStore(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx, alias);\n }\n\n /**\n * Wait for the corresponding tile to become non-invalid\n */\n __device__ __forceinline__ void WaitForValid(\n int tile_idx,\n StatusWord &status,\n KeyValuePairT &value)\n {\n// TxnWord alias = ThreadLoad(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx);\n// TileDescriptor tile_descriptor = reinterpret_cast(alias);\n//\n// while (tile_descriptor.status == SCAN_TILE_INVALID)\n// {\n// __threadfence_block(); // prevent hoisting loads from loop\n//\n// alias = ThreadLoad(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx);\n// tile_descriptor = reinterpret_cast(alias);\n// }\n//\n// status = tile_descriptor.status;\n// value.value = tile_descriptor.value;\n// value.key = tile_descriptor.key;\n\n TileDescriptor tile_descriptor;\n do\n {\n __threadfence_block(); // prevent hoisting loads from loop\n TxnWord alias = ThreadLoad(d_tile_descriptors + TILE_STATUS_PADDING + tile_idx);\n tile_descriptor = reinterpret_cast(alias);\n\n } while (WARP_ANY((tile_descriptor.status == SCAN_TILE_INVALID), 0xffffffff));\n\n status = tile_descriptor.status;\n value.value = tile_descriptor.value;\n value.key = tile_descriptor.key;\n }\n\n};\n\n\n/******************************************************************************\n * Prefix call-back operator for coupling local block scan within a\n * block-cooperative scan\n ******************************************************************************/\n\n/**\n * Stateful block-scan prefix functor. Provides the the running prefix for\n * the current tile by using the call-back warp to wait on on\n * aggregates/prefixes from predecessor tiles to become available.\n */\ntemplate <\n typename T,\n typename ScanOpT,\n typename ScanTileStateT,\n int PTX_ARCH = CUB_PTX_ARCH>\nstruct TilePrefixCallbackOp\n{\n // Parameterized warp reduce\n typedef WarpReduce WarpReduceT;\n\n // Temporary storage type\n struct _TempStorage\n {\n typename WarpReduceT::TempStorage warp_reduce;\n T exclusive_prefix;\n T inclusive_prefix;\n T block_aggregate;\n };\n\n // Alias wrapper allowing temporary storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n // Type of status word\n typedef typename ScanTileStateT::StatusWord StatusWord;\n\n // Fields\n _TempStorage& temp_storage; ///< Reference to a warp-reduction instance\n ScanTileStateT& tile_status; ///< Interface to tile status\n ScanOpT scan_op; ///< Binary scan operator\n int tile_idx; ///< The current tile index\n T exclusive_prefix; ///< Exclusive prefix for the tile\n T inclusive_prefix; ///< Inclusive prefix for the tile\n\n // Constructor\n __device__ __forceinline__\n TilePrefixCallbackOp(\n ScanTileStateT &tile_status,\n TempStorage &temp_storage,\n ScanOpT scan_op,\n int tile_idx)\n :\n temp_storage(temp_storage.Alias()),\n tile_status(tile_status),\n scan_op(scan_op),\n tile_idx(tile_idx) {}\n\n\n // Block until all predecessors within the warp-wide window have non-invalid status\n __device__ __forceinline__\n void ProcessWindow(\n int predecessor_idx, ///< Preceding tile index to inspect\n StatusWord &predecessor_status, ///< [out] Preceding tile status\n T &window_aggregate) ///< [out] Relevant partial reduction from this window of preceding tiles\n {\n T value;\n tile_status.WaitForValid(predecessor_idx, predecessor_status, value);\n\n // Perform a segmented reduction to get the prefix for the current window.\n // Use the swizzled scan operator because we are now scanning *down* towards thread0.\n\n int tail_flag = (predecessor_status == StatusWord(SCAN_TILE_INCLUSIVE));\n window_aggregate = WarpReduceT(temp_storage.warp_reduce).TailSegmentedReduce(\n value,\n tail_flag,\n SwizzleScanOp(scan_op));\n }\n\n\n // BlockScan prefix callback functor (called by the first warp)\n __device__ __forceinline__\n T operator()(T block_aggregate)\n {\n\n // Update our status with our tile-aggregate\n if (threadIdx.x == 0)\n {\n temp_storage.block_aggregate = block_aggregate;\n tile_status.SetPartial(tile_idx, block_aggregate);\n }\n\n int predecessor_idx = tile_idx - threadIdx.x - 1;\n StatusWord predecessor_status;\n T window_aggregate;\n\n // Wait for the warp-wide window of predecessor tiles to become valid\n ProcessWindow(predecessor_idx, predecessor_status, window_aggregate);\n\n // The exclusive tile prefix starts out as the current window aggregate\n exclusive_prefix = window_aggregate;\n\n // Keep sliding the window back until we come across a tile whose inclusive prefix is known\n while (WARP_ALL((predecessor_status != StatusWord(SCAN_TILE_INCLUSIVE)), 0xffffffff))\n {\n predecessor_idx -= CUB_PTX_WARP_THREADS;\n\n // Update exclusive tile prefix with the window prefix\n ProcessWindow(predecessor_idx, predecessor_status, window_aggregate);\n exclusive_prefix = scan_op(window_aggregate, exclusive_prefix);\n }\n\n // Compute the inclusive tile prefix and update the status for this tile\n if (threadIdx.x == 0)\n {\n inclusive_prefix = scan_op(exclusive_prefix, block_aggregate);\n tile_status.SetInclusive(tile_idx, inclusive_prefix);\n\n temp_storage.exclusive_prefix = exclusive_prefix;\n temp_storage.inclusive_prefix = inclusive_prefix;\n }\n\n // Return exclusive_prefix\n return exclusive_prefix;\n }\n\n // Get the exclusive prefix stored in temporary storage\n __device__ __forceinline__\n T GetExclusivePrefix()\n {\n return temp_storage.exclusive_prefix;\n }\n\n // Get the inclusive prefix stored in temporary storage\n __device__ __forceinline__\n T GetInclusivePrefix()\n {\n return temp_storage.inclusive_prefix;\n }\n\n // Get the block aggregate stored in temporary storage\n __device__ __forceinline__\n T GetBlockAggregate()\n {\n return temp_storage.block_aggregate;\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_adjacent_difference.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockDiscontinuity class provides [collective](index.html#sec0) methods for flagging discontinuities within an ordered set of items partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../util_type.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockAdjacentDifference\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n\n /// Shared memory storage layout type (last element from each thread's input)\n struct _TempStorage\n {\n T first_items[BLOCK_THREADS];\n T last_items[BLOCK_THREADS];\n };\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /// Specialization for when FlagOp has third index param\n template ::HAS_PARAM>\n struct ApplyOp\n {\n // Apply flag operator\n static __device__ __forceinline__ T FlagT(FlagOp flag_op, const T &a, const T &b, int idx)\n {\n return flag_op(b, a, idx);\n }\n };\n\n /// Specialization for when FlagOp does not have a third index param\n template \n struct ApplyOp\n {\n // Apply flag operator\n static __device__ __forceinline__ T FlagT(FlagOp flag_op, const T &a, const T &b, int /*idx*/)\n {\n return flag_op(b, a);\n }\n };\n\n /// Templated unrolling of item comparison (inductive case)\n template \n struct Iterate\n {\n // Head flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagHeads(\n int linear_tid,\n FlagT (&flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n preds[ITERATION] = input[ITERATION - 1];\n\n flags[ITERATION] = ApplyOp::FlagT(\n flag_op,\n preds[ITERATION],\n input[ITERATION],\n (linear_tid * ITEMS_PER_THREAD) + ITERATION);\n\n Iterate::FlagHeads(linear_tid, flags, input, preds, flag_op);\n }\n\n // Tail flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagTails(\n int linear_tid,\n FlagT (&flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n flags[ITERATION] = ApplyOp::FlagT(\n flag_op,\n input[ITERATION],\n input[ITERATION + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITERATION + 1);\n\n Iterate::FlagTails(linear_tid, flags, input, flag_op);\n }\n\n };\n\n /// Templated unrolling of item comparison (termination case)\n template \n struct Iterate\n {\n // Head flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagHeads(\n int /*linear_tid*/,\n FlagT (&/*flags*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&/*input*/)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&/*preds*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp /*flag_op*/) ///< [in] Binary boolean flag predicate\n {}\n\n // Tail flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagTails(\n int /*linear_tid*/,\n FlagT (&/*flags*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&/*input*/)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp /*flag_op*/) ///< [in] Binary boolean flag predicate\n {}\n };\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\npublic:\n\n /// \\smemstorage{BlockDiscontinuity}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockAdjacentDifference()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockAdjacentDifference(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Head flag operations\n *********************************************************************/\n //@{\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share last item\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n if (linear_tid == 0)\n {\n // Set flag for first thread-item (preds[0] is undefined)\n head_flags[0] = 1;\n }\n else\n {\n preds[0] = temp_storage.last_items[linear_tid - 1];\n head_flags[0] = ApplyOp::FlagT(flag_op, preds[0], input[0], linear_tid * ITEMS_PER_THREAD);\n }\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n }\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_predecessor_item) ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n {\n // Share last item\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(flag_op, preds[0], input[0], linear_tid * ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n }\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n T preds[ITEMS_PER_THREAD];\n FlagHeads(head_flags, input, preds, flag_op);\n }\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_predecessor_item) ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n {\n T preds[ITEMS_PER_THREAD];\n FlagHeads(head_flags, input, preds, flag_op, tile_predecessor_item);\n }\n\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagTails(\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first item\n temp_storage.first_items[linear_tid] = input[0];\n\n CTA_SYNC();\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagTails(\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_successor_item) ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n {\n // Share first item\n temp_storage.first_items[linear_tid] = input[0];\n\n CTA_SYNC();\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = temp_storage.last_items[linear_tid - 1];\n if (linear_tid == 0)\n {\n head_flags[0] = 1;\n }\n else\n {\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n }\n\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T tile_successor_item, ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n if (linear_tid == 0)\n {\n head_flags[0] = 1;\n }\n else\n {\n preds[0] = temp_storage.last_items[linear_tid - 1];\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n }\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T tile_predecessor_item, ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T tile_predecessor_item, ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T tile_successor_item, ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/block/block_discontinuity.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockDiscontinuity class provides [collective](index.html#sec0) methods for flagging discontinuities within an ordered set of items partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../util_type.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief The BlockDiscontinuity class provides [collective](index.html#sec0) methods for flagging discontinuities within an ordered set of items partitioned across a CUDA thread block. ![](discont_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam T The data type to be flagged.\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - A set of \"head flags\" (or \"tail flags\") is often used to indicate corresponding items\n * that differ from their predecessors (or successors). For example, head flags are convenient\n * for demarcating disjoint data segments as part of a segmented scan or reduction.\n * - \\blocked\n *\n * \\par Performance Considerations\n * - \\granularity\n *\n * \\par A Simple Example\n * \\blockcollective{BlockDiscontinuity}\n * \\par\n * The code snippet below illustrates the head flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute head flags for discontinuities in the segment\n * int head_flags[4];\n * BlockDiscontinuity(temp_storage).FlagHeads(head_flags, thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], [3,4,4,4], ... }.\n * The corresponding output \\p head_flags in those threads will be\n * { [1,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n *\n * \\par Performance Considerations\n * - Incurs zero bank conflicts for most types\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockDiscontinuity\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n\n /// Shared memory storage layout type (last element from each thread's input)\n struct _TempStorage\n {\n T first_items[BLOCK_THREADS];\n T last_items[BLOCK_THREADS];\n };\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /// Specialization for when FlagOp has third index param\n template ::HAS_PARAM>\n struct ApplyOp\n {\n // Apply flag operator\n static __device__ __forceinline__ bool FlagT(FlagOp flag_op, const T &a, const T &b, int idx)\n {\n return flag_op(a, b, idx);\n }\n };\n\n /// Specialization for when FlagOp does not have a third index param\n template \n struct ApplyOp\n {\n // Apply flag operator\n static __device__ __forceinline__ bool FlagT(FlagOp flag_op, const T &a, const T &b, int /*idx*/)\n {\n return flag_op(a, b);\n }\n };\n\n /// Templated unrolling of item comparison (inductive case)\n template \n struct Iterate\n {\n // Head flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagHeads(\n int linear_tid,\n FlagT (&flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n preds[ITERATION] = input[ITERATION - 1];\n\n flags[ITERATION] = ApplyOp::FlagT(\n flag_op,\n preds[ITERATION],\n input[ITERATION],\n (linear_tid * ITEMS_PER_THREAD) + ITERATION);\n\n Iterate::FlagHeads(linear_tid, flags, input, preds, flag_op);\n }\n\n // Tail flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagTails(\n int linear_tid,\n FlagT (&flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n flags[ITERATION] = ApplyOp::FlagT(\n flag_op,\n input[ITERATION],\n input[ITERATION + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITERATION + 1);\n\n Iterate::FlagTails(linear_tid, flags, input, flag_op);\n }\n\n };\n\n /// Templated unrolling of item comparison (termination case)\n template \n struct Iterate\n {\n // Head flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagHeads(\n int /*linear_tid*/,\n FlagT (&/*flags*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&/*input*/)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&/*preds*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp /*flag_op*/) ///< [in] Binary boolean flag predicate\n {}\n\n // Tail flags\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n static __device__ __forceinline__ void FlagTails(\n int /*linear_tid*/,\n FlagT (&/*flags*/)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&/*input*/)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp /*flag_op*/) ///< [in] Binary boolean flag predicate\n {}\n };\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\npublic:\n\n /// \\smemstorage{BlockDiscontinuity}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockDiscontinuity()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockDiscontinuity(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Head flag operations\n *********************************************************************/\n //@{\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share last item\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n if (linear_tid == 0)\n {\n // Set flag for first thread-item (preds[0] is undefined)\n head_flags[0] = 1;\n }\n else\n {\n preds[0] = temp_storage.last_items[linear_tid - 1];\n head_flags[0] = ApplyOp::FlagT(flag_op, preds[0], input[0], linear_tid * ITEMS_PER_THREAD);\n }\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n }\n\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&preds)[ITEMS_PER_THREAD], ///< [out] Calling thread's predecessor items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_predecessor_item) ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n {\n // Share last item\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(flag_op, preds[0], input[0], linear_tid * ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n }\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n /**\n * \\brief Sets head flags indicating discontinuities between items partitioned across the thread block, for which the first item has no reference and is always flagged.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is always flagged.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute head flags for discontinuities in the segment\n * int head_flags[4];\n * BlockDiscontinuity(temp_storage).FlagHeads(head_flags, thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], [3,4,4,4], ... }.\n * The corresponding output \\p head_flags in those threads will be\n * { [1,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n T preds[ITEMS_PER_THREAD];\n FlagHeads(head_flags, input, preds, flag_op);\n }\n\n\n /**\n * \\brief Sets head flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is compared\n * against \\p tile_predecessor_item.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Have thread0 obtain the predecessor item for the entire tile\n * int tile_predecessor_item;\n * if (threadIdx.x == 0) tile_predecessor_item == ...\n *\n * // Collectively compute head flags for discontinuities in the segment\n * int head_flags[4];\n * BlockDiscontinuity(temp_storage).FlagHeads(\n * head_flags, thread_data, cub::Inequality(), tile_predecessor_item);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], [3,4,4,4], ... },\n * and that \\p tile_predecessor_item is \\p 0. The corresponding output \\p head_flags in those threads will be\n * { [0,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeads(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_predecessor_item) ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n {\n T preds[ITEMS_PER_THREAD];\n FlagHeads(head_flags, input, preds, flag_op, tile_predecessor_item);\n }\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Tail flag operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Sets tail flags indicating discontinuities between items partitioned across the thread block, for which the last item has no reference and is always flagged.\n *\n * \\par\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is always flagged.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute tail flags for discontinuities in the segment\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(tail_flags, thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] }.\n * The corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,1] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagTails(\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first item\n temp_storage.first_items[linear_tid] = input[0];\n\n CTA_SYNC();\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n /**\n * \\brief Sets tail flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is compared\n * against \\p tile_successor_item.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Have thread127 obtain the successor item for the entire tile\n * int tile_successor_item;\n * if (threadIdx.x == 127) tile_successor_item == ...\n *\n * // Collectively compute tail flags for discontinuities in the segment\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(\n * tail_flags, thread_data, cub::Inequality(), tile_successor_item);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] }\n * and that \\p tile_successor_item is \\p 125. The corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,0] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagTails(\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op, ///< [in] Binary boolean flag predicate\n T tile_successor_item) ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n {\n // Share first item\n temp_storage.first_items[linear_tid] = input[0];\n\n CTA_SYNC();\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Head & tail flag operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Sets both head and tail flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is always flagged.\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is always flagged.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head- and tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute head and flags for discontinuities in the segment\n * int head_flags[4];\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(\n * head_flags, tail_flags, thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] }\n * and that the tile_successor_item is \\p 125. The corresponding output \\p head_flags\n * in those threads will be { [1,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n * and the corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,1] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = temp_storage.last_items[linear_tid - 1];\n if (linear_tid == 0)\n {\n head_flags[0] = 1;\n }\n else\n {\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n }\n\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n /**\n * \\brief Sets both head and tail flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is always flagged.\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is compared\n * against \\p tile_predecessor_item.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head- and tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Have thread127 obtain the successor item for the entire tile\n * int tile_successor_item;\n * if (threadIdx.x == 127) tile_successor_item == ...\n *\n * // Collectively compute head and flags for discontinuities in the segment\n * int head_flags[4];\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(\n * head_flags, tail_flags, tile_successor_item, thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] }\n * and that the tile_successor_item is \\p 125. The corresponding output \\p head_flags\n * in those threads will be { [1,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n * and the corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,0] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T tile_successor_item, ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n if (linear_tid == 0)\n {\n head_flags[0] = 1;\n }\n else\n {\n preds[0] = temp_storage.last_items[linear_tid - 1];\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n }\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n /**\n * \\brief Sets both head and tail flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is compared\n * against \\p tile_predecessor_item.\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is always flagged.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head- and tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Have thread0 obtain the predecessor item for the entire tile\n * int tile_predecessor_item;\n * if (threadIdx.x == 0) tile_predecessor_item == ...\n *\n * // Have thread127 obtain the successor item for the entire tile\n * int tile_successor_item;\n * if (threadIdx.x == 127) tile_successor_item == ...\n *\n * // Collectively compute head and flags for discontinuities in the segment\n * int head_flags[4];\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(\n * head_flags, tile_predecessor_item, tail_flags, tile_successor_item,\n * thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] },\n * that the \\p tile_predecessor_item is \\p 0, and that the\n * \\p tile_successor_item is \\p 125. The corresponding output \\p head_flags\n * in those threads will be { [0,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n * and the corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,1] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T tile_predecessor_item, ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n\n // Set flag for last thread-item\n tail_flags[ITEMS_PER_THREAD - 1] = (linear_tid == BLOCK_THREADS - 1) ?\n 1 : // Last thread\n ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n temp_storage.first_items[linear_tid + 1],\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n /**\n * \\brief Sets both head and tail flags indicating discontinuities between items partitioned across the thread block.\n *\n * \\par\n * - The flag head_flagsi is set for item\n * inputi when\n * flag_op(previous-item, inputi)\n * returns \\p true (where previous-item is either the preceding item\n * in the same thread or the last item in the previous thread).\n * - For thread0, item input0 is compared\n * against \\p tile_predecessor_item.\n * - The flag tail_flagsi is set for item\n * inputi when\n * flag_op(inputi, next-item)\n * returns \\p true (where next-item is either the next item\n * in the same thread or the first item in the next thread).\n * - For threadBLOCK_THREADS-1, item\n * inputITEMS_PER_THREAD-1 is compared\n * against \\p tile_successor_item.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the head- and tail-flagging of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockDiscontinuity for a 1D block of 128 threads on type int\n * typedef cub::BlockDiscontinuity BlockDiscontinuity;\n *\n * // Allocate shared memory for BlockDiscontinuity\n * __shared__ typename BlockDiscontinuity::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Have thread0 obtain the predecessor item for the entire tile\n * int tile_predecessor_item;\n * if (threadIdx.x == 0) tile_predecessor_item == ...\n *\n * // Have thread127 obtain the successor item for the entire tile\n * int tile_successor_item;\n * if (threadIdx.x == 127) tile_successor_item == ...\n *\n * // Collectively compute head and flags for discontinuities in the segment\n * int head_flags[4];\n * int tail_flags[4];\n * BlockDiscontinuity(temp_storage).FlagTails(\n * head_flags, tile_predecessor_item, tail_flags, tile_successor_item,\n * thread_data, cub::Inequality());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,0,1,1], [1,1,1,1], [2,3,3,3], ..., [124,125,125,125] },\n * that the \\p tile_predecessor_item is \\p 0, and that the\n * \\p tile_successor_item is \\p 125. The corresponding output \\p head_flags\n * in those threads will be { [0,0,1,0], [0,0,0,0], [1,1,0,0], [0,1,0,0], ... }.\n * and the corresponding output \\p tail_flags in those threads will be\n * { [0,1,0,0], [0,0,0,1], [1,0,0,...], ..., [1,0,0,0] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam FlagT [inferred] The flag type (must be an integer type)\n * \\tparam FlagOp [inferred] Binary predicate functor type having member T operator()(const T &a, const T &b) or member T operator()(const T &a, const T &b, unsigned int b_index), and returning \\p true if a discontinuity exists between \\p a and \\p b, otherwise \\p false. \\p b_index is the rank of b in the aggregate tile of data.\n */\n template <\n int ITEMS_PER_THREAD,\n typename FlagT,\n typename FlagOp>\n __device__ __forceinline__ void FlagHeadsAndTails(\n FlagT (&head_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity head_flags\n T tile_predecessor_item, ///< [in] [thread0 only] Item with which to compare the first tile item (input0 from thread0).\n FlagT (&tail_flags)[ITEMS_PER_THREAD], ///< [out] Calling thread's discontinuity tail_flags\n T tile_successor_item, ///< [in] [threadBLOCK_THREADS-1 only] Item with which to compare the last tile item (inputITEMS_PER_THREAD-1 from threadBLOCK_THREADS-1).\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n FlagOp flag_op) ///< [in] Binary boolean flag predicate\n {\n // Share first and last items\n temp_storage.first_items[linear_tid] = input[0];\n temp_storage.last_items[linear_tid] = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n T preds[ITEMS_PER_THREAD];\n\n // Set flag for first thread-item\n preds[0] = (linear_tid == 0) ?\n tile_predecessor_item : // First thread\n temp_storage.last_items[linear_tid - 1];\n\n head_flags[0] = ApplyOp::FlagT(\n flag_op,\n preds[0],\n input[0],\n linear_tid * ITEMS_PER_THREAD);\n\n // Set flag for last thread-item\n T successor_item = (linear_tid == BLOCK_THREADS - 1) ?\n tile_successor_item : // Last thread\n temp_storage.first_items[linear_tid + 1];\n\n tail_flags[ITEMS_PER_THREAD - 1] = ApplyOp::FlagT(\n flag_op,\n input[ITEMS_PER_THREAD - 1],\n successor_item,\n (linear_tid * ITEMS_PER_THREAD) + ITEMS_PER_THREAD);\n\n // Set head_flags for remaining items\n Iterate<1, ITEMS_PER_THREAD>::FlagHeads(linear_tid, head_flags, input, preds, flag_op);\n\n // Set tail_flags for remaining items\n Iterate<0, ITEMS_PER_THREAD - 1>::FlagTails(linear_tid, tail_flags, input, flag_op);\n }\n\n\n\n\n //@} end member group\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/block/block_exchange.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockExchange class provides [collective](index.html#sec0) methods for rearranging data partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../util_ptx.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_macro.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief The BlockExchange class provides [collective](index.html#sec0) methods for rearranging data partitioned across a CUDA thread block. ![](transpose_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam T The data type to be exchanged.\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ITEMS_PER_THREAD The number of items partitioned onto each thread.\n * \\tparam WARP_TIME_SLICING [optional] When \\p true, only use enough shared memory for a single warp's worth of tile data, time-slicing the block-wide exchange over multiple synchronized rounds. Yields a smaller memory footprint at the expense of decreased parallelism. (Default: false)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - It is commonplace for blocks of threads to rearrange data items between\n * threads. For example, the device-accessible memory subsystem prefers access patterns\n * where data items are \"striped\" across threads (where consecutive threads access consecutive items),\n * yet most block-wide operations prefer a \"blocked\" partitioning of items across threads\n * (where consecutive items belong to a single thread).\n * - BlockExchange supports the following types of data exchanges:\n * - Transposing between [blocked](index.html#sec5sec3) and [striped](index.html#sec5sec3) arrangements\n * - Transposing between [blocked](index.html#sec5sec3) and [warp-striped](index.html#sec5sec3) arrangements\n * - Scattering ranked items to a [blocked arrangement](index.html#sec5sec3)\n * - Scattering ranked items to a [striped arrangement](index.html#sec5sec3)\n * - \\rowmajor\n *\n * \\par A Simple Example\n * \\blockcollective{BlockExchange}\n * \\par\n * The code snippet below illustrates the conversion from a \"blocked\" to a \"striped\" arrangement\n * of 512 integer items partitioned across 128 threads where each thread owns 4 items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockExchange for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockExchange BlockExchange;\n *\n * // Allocate shared memory for BlockExchange\n * __shared__ typename BlockExchange::TempStorage temp_storage;\n *\n * // Load a tile of data striped across threads\n * int thread_data[4];\n * cub::LoadDirectStriped<128>(threadIdx.x, d_data, thread_data);\n *\n * // Collectively exchange data into a blocked arrangement across threads\n * BlockExchange(temp_storage).StripedToBlocked(thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of striped input \\p thread_data across the block of threads is\n * { [0,128,256,384], [1,129,257,385], ..., [127,255,383,511] }.\n * The corresponding output \\p thread_data in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n *\n * \\par Performance Considerations\n * - Proper device-specific padding ensures zero bank conflicts for most types.\n *\n */\ntemplate <\n typename InputT,\n int BLOCK_DIM_X,\n int ITEMS_PER_THREAD,\n bool WARP_TIME_SLICING = false,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockExchange\n{\nprivate:\n\n /******************************************************************************\n * Constants\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n LOG_WARP_THREADS = CUB_LOG_WARP_THREADS(PTX_ARCH),\n WARP_THREADS = 1 << LOG_WARP_THREADS,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n LOG_SMEM_BANKS = CUB_LOG_SMEM_BANKS(PTX_ARCH),\n SMEM_BANKS = 1 << LOG_SMEM_BANKS,\n\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n\n TIME_SLICES = (WARP_TIME_SLICING) ? WARPS : 1,\n\n TIME_SLICED_THREADS = (WARP_TIME_SLICING) ? CUB_MIN(BLOCK_THREADS, WARP_THREADS) : BLOCK_THREADS,\n TIME_SLICED_ITEMS = TIME_SLICED_THREADS * ITEMS_PER_THREAD,\n\n WARP_TIME_SLICED_THREADS = CUB_MIN(BLOCK_THREADS, WARP_THREADS),\n WARP_TIME_SLICED_ITEMS = WARP_TIME_SLICED_THREADS * ITEMS_PER_THREAD,\n\n // Insert padding to avoid bank conflicts during raking when items per thread is a power of two and > 4 (otherwise we can typically use 128b loads)\n INSERT_PADDING = (ITEMS_PER_THREAD > 4) && (PowerOfTwo::VALUE),\n PADDING_ITEMS = (INSERT_PADDING) ? (TIME_SLICED_ITEMS >> LOG_SMEM_BANKS) : 0,\n };\n\n /******************************************************************************\n * Type definitions\n ******************************************************************************/\n\n /// Shared memory storage layout type\n struct __align__(16) _TempStorage\n {\n InputT buff[TIME_SLICED_ITEMS + PADDING_ITEMS];\n };\n\npublic:\n\n /// \\smemstorage{BlockExchange}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\nprivate:\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n unsigned int lane_id;\n unsigned int warp_id;\n unsigned int warp_offset;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /**\n * Transposes data items from blocked arrangement to striped arrangement. Specialized for no timeslicing.\n */\n template \n __device__ __forceinline__ void BlockedToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (linear_tid * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = int(ITEM * BLOCK_THREADS) + linear_tid;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n /**\n * Transposes data items from blocked arrangement to striped arrangement. Specialized for warp-timeslicing.\n */\n template \n __device__ __forceinline__ void BlockedToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n InputT temp_items[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int SLICE = 0; SLICE < TIME_SLICES; SLICE++)\n {\n const int SLICE_OFFSET = SLICE * TIME_SLICED_ITEMS;\n const int SLICE_OOB = SLICE_OFFSET + TIME_SLICED_ITEMS;\n\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (lane_id * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n // Read a strip of items\n const int STRIP_OFFSET = ITEM * BLOCK_THREADS;\n const int STRIP_OOB = STRIP_OFFSET + BLOCK_THREADS;\n\n if ((SLICE_OFFSET < STRIP_OOB) && (SLICE_OOB > STRIP_OFFSET))\n {\n int item_offset = STRIP_OFFSET + linear_tid - SLICE_OFFSET;\n if ((item_offset >= 0) && (item_offset < TIME_SLICED_ITEMS))\n {\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n }\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n output_items[ITEM] = temp_items[ITEM];\n }\n }\n\n\n /**\n * Transposes data items from blocked arrangement to warp-striped arrangement. Specialized for no timeslicing\n */\n template \n __device__ __forceinline__ void BlockedToWarpStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = warp_offset + ITEM + (lane_id * ITEMS_PER_THREAD);\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = warp_offset + (ITEM * WARP_TIME_SLICED_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n /**\n * Transposes data items from blocked arrangement to warp-striped arrangement. Specialized for warp-timeslicing\n */\n template \n __device__ __forceinline__ void BlockedToWarpStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n if (warp_id == 0)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ITEM + (lane_id * ITEMS_PER_THREAD);\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (ITEM * WARP_TIME_SLICED_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n #pragma unroll\n for (unsigned int SLICE = 1; SLICE < TIME_SLICES; ++SLICE)\n {\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ITEM + (lane_id * ITEMS_PER_THREAD);\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (ITEM * WARP_TIME_SLICED_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n }\n\n\n /**\n * Transposes data items from striped arrangement to blocked arrangement. Specialized for no timeslicing.\n */\n template \n __device__ __forceinline__ void StripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = int(ITEM * BLOCK_THREADS) + linear_tid;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n // No timeslicing\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (linear_tid * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n /**\n * Transposes data items from striped arrangement to blocked arrangement. Specialized for warp-timeslicing.\n */\n template \n __device__ __forceinline__ void StripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n // Warp time-slicing\n InputT temp_items[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int SLICE = 0; SLICE < TIME_SLICES; SLICE++)\n {\n const int SLICE_OFFSET = SLICE * TIME_SLICED_ITEMS;\n const int SLICE_OOB = SLICE_OFFSET + TIME_SLICED_ITEMS;\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n // Write a strip of items\n const int STRIP_OFFSET = ITEM * BLOCK_THREADS;\n const int STRIP_OOB = STRIP_OFFSET + BLOCK_THREADS;\n\n if ((SLICE_OFFSET < STRIP_OOB) && (SLICE_OOB > STRIP_OFFSET))\n {\n int item_offset = STRIP_OFFSET + linear_tid - SLICE_OFFSET;\n if ((item_offset >= 0) && (item_offset < TIME_SLICED_ITEMS))\n {\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n }\n }\n\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (lane_id * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n output_items[ITEM] = temp_items[ITEM];\n }\n }\n\n\n /**\n * Transposes data items from warp-striped arrangement to blocked arrangement. Specialized for no timeslicing\n */\n template \n __device__ __forceinline__ void WarpStripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = warp_offset + (ITEM * WARP_TIME_SLICED_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = warp_offset + ITEM + (lane_id * ITEMS_PER_THREAD);\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n /**\n * Transposes data items from warp-striped arrangement to blocked arrangement. Specialized for warp-timeslicing\n */\n template \n __device__ __forceinline__ void WarpStripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (unsigned int SLICE = 0; SLICE < TIME_SLICES; ++SLICE)\n {\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (ITEM * WARP_TIME_SLICED_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ITEM + (lane_id * ITEMS_PER_THREAD);\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n }\n\n\n /**\n * Exchanges data items annotated by rank into blocked arrangement. Specialized for no timeslicing.\n */\n template \n __device__ __forceinline__ void ScatterToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM];\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (linear_tid * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n /**\n * Exchanges data items annotated by rank into blocked arrangement. Specialized for warp-timeslicing.\n */\n template \n __device__ __forceinline__ void ScatterToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n Int2Type /*time_slicing*/)\n {\n InputT temp_items[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int SLICE = 0; SLICE < TIME_SLICES; SLICE++)\n {\n CTA_SYNC();\n\n const int SLICE_OFFSET = TIME_SLICED_ITEMS * SLICE;\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM] - SLICE_OFFSET;\n if ((item_offset >= 0) && (item_offset < WARP_TIME_SLICED_ITEMS))\n {\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n }\n\n CTA_SYNC();\n\n if (warp_id == SLICE)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (lane_id * ITEMS_PER_THREAD) + ITEM;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n temp_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n output_items[ITEM] = temp_items[ITEM];\n }\n }\n\n\n /**\n * Exchanges data items annotated by rank into striped arrangement. Specialized for no timeslicing.\n */\n template \n __device__ __forceinline__ void ScatterToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n Int2Type /*time_slicing*/)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM];\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = int(ITEM * BLOCK_THREADS) + linear_tid;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n /**\n * Exchanges data items annotated by rank into striped arrangement. Specialized for warp-timeslicing.\n */\n template \n __device__ __forceinline__ void ScatterToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between blocked and striped arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items to exchange, converting between blocked and striped arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n Int2Type /*time_slicing*/)\n {\n InputT temp_items[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int SLICE = 0; SLICE < TIME_SLICES; SLICE++)\n {\n const int SLICE_OFFSET = SLICE * TIME_SLICED_ITEMS;\n const int SLICE_OOB = SLICE_OFFSET + TIME_SLICED_ITEMS;\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM] - SLICE_OFFSET;\n if ((item_offset >= 0) && (item_offset < WARP_TIME_SLICED_ITEMS))\n {\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n // Read a strip of items\n const int STRIP_OFFSET = ITEM * BLOCK_THREADS;\n const int STRIP_OOB = STRIP_OFFSET + BLOCK_THREADS;\n\n if ((SLICE_OFFSET < STRIP_OOB) && (SLICE_OOB > STRIP_OFFSET))\n {\n int item_offset = STRIP_OFFSET + linear_tid - SLICE_OFFSET;\n if ((item_offset >= 0) && (item_offset < TIME_SLICED_ITEMS))\n {\n if (INSERT_PADDING) item_offset += item_offset >> LOG_SMEM_BANKS;\n temp_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n }\n }\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n output_items[ITEM] = temp_items[ITEM];\n }\n }\n\n\npublic:\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockExchange()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n warp_id((WARPS == 1) ? 0 : linear_tid / WARP_THREADS),\n lane_id(LaneId()),\n warp_offset(warp_id * WARP_TIME_SLICED_ITEMS)\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockExchange(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n lane_id(LaneId()),\n warp_id((WARPS == 1) ? 0 : linear_tid / WARP_THREADS),\n warp_offset(warp_id * WARP_TIME_SLICED_ITEMS)\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Structured exchanges\n *********************************************************************/\n //@{\n\n /**\n * \\brief Transposes data items from striped arrangement to blocked arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the conversion from a \"striped\" to a \"blocked\" arrangement\n * of 512 integer items partitioned across 128 threads where each thread owns 4 items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockExchange for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockExchange BlockExchange;\n *\n * // Allocate shared memory for BlockExchange\n * __shared__ typename BlockExchange::TempStorage temp_storage;\n *\n * // Load a tile of ordered data into a striped arrangement across block threads\n * int thread_data[4];\n * cub::LoadDirectStriped<128>(threadIdx.x, d_data, thread_data);\n *\n * // Collectively exchange data into a blocked arrangement across threads\n * BlockExchange(temp_storage).StripedToBlocked(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of striped input \\p thread_data across the block of threads is\n * { [0,128,256,384], [1,129,257,385], ..., [127,255,383,511] } after loading from device-accessible memory.\n * The corresponding output \\p thread_data in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n *\n */\n template \n __device__ __forceinline__ void StripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD]) ///< [out] Items from exchange, converting between striped and blocked arrangements.\n {\n StripedToBlocked(input_items, output_items, Int2Type());\n }\n\n\n /**\n * \\brief Transposes data items from blocked arrangement to striped arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the conversion from a \"blocked\" to a \"striped\" arrangement\n * of 512 integer items partitioned across 128 threads where each thread owns 4 items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockExchange for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockExchange BlockExchange;\n *\n * // Allocate shared memory for BlockExchange\n * __shared__ typename BlockExchange::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively exchange data into a striped arrangement across threads\n * BlockExchange(temp_storage).BlockedToStriped(thread_data, thread_data);\n *\n * // Store data striped across block threads into an ordered tile\n * cub::StoreDirectStriped(threadIdx.x, d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of blocked input \\p thread_data across the block of threads is\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n * The corresponding output \\p thread_data in those threads will be\n * { [0,128,256,384], [1,129,257,385], ..., [127,255,383,511] } in\n * preparation for storing to device-accessible memory.\n *\n */\n template \n __device__ __forceinline__ void BlockedToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD]) ///< [out] Items from exchange, converting between striped and blocked arrangements.\n {\n BlockedToStriped(input_items, output_items, Int2Type());\n }\n\n\n\n /**\n * \\brief Transposes data items from warp-striped arrangement to blocked arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the conversion from a \"warp-striped\" to a \"blocked\" arrangement\n * of 512 integer items partitioned across 128 threads where each thread owns 4 items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockExchange for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockExchange BlockExchange;\n *\n * // Allocate shared memory for BlockExchange\n * __shared__ typename BlockExchange::TempStorage temp_storage;\n *\n * // Load a tile of ordered data into a warp-striped arrangement across warp threads\n * int thread_data[4];\n * cub::LoadSWarptriped(threadIdx.x, d_data, thread_data);\n *\n * // Collectively exchange data into a blocked arrangement across threads\n * BlockExchange(temp_storage).WarpStripedToBlocked(thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of warp-striped input \\p thread_data across the block of threads is\n * { [0,32,64,96], [1,33,65,97], [2,34,66,98], ..., [415,447,479,511] }\n * after loading from device-accessible memory. (The first 128 items are striped across\n * the first warp of 32 threads, the second 128 items are striped across the second warp, etc.)\n * The corresponding output \\p thread_data in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n *\n */\n template \n __device__ __forceinline__ void WarpStripedToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD]) ///< [out] Items from exchange, converting between striped and blocked arrangements.\n {\n WarpStripedToBlocked(input_items, output_items, Int2Type());\n }\n\n\n\n /**\n * \\brief Transposes data items from blocked arrangement to warp-striped arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the conversion from a \"blocked\" to a \"warp-striped\" arrangement\n * of 512 integer items partitioned across 128 threads where each thread owns 4 items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockExchange for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockExchange BlockExchange;\n *\n * // Allocate shared memory for BlockExchange\n * __shared__ typename BlockExchange::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively exchange data into a warp-striped arrangement across threads\n * BlockExchange(temp_storage).BlockedToWarpStriped(thread_data, thread_data);\n *\n * // Store data striped across warp threads into an ordered tile\n * cub::StoreDirectStriped(threadIdx.x, d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of blocked input \\p thread_data across the block of threads is\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n * The corresponding output \\p thread_data in those threads will be\n * { [0,32,64,96], [1,33,65,97], [2,34,66,98], ..., [415,447,479,511] }\n * in preparation for storing to device-accessible memory. (The first 128 items are striped across\n * the first warp of 32 threads, the second 128 items are striped across the second warp, etc.)\n *\n */\n template \n __device__ __forceinline__ void BlockedToWarpStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD]) ///< [out] Items from exchange, converting between striped and blocked arrangements.\n {\n BlockedToWarpStriped(input_items, output_items, Int2Type());\n }\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Scatter exchanges\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Exchanges data items annotated by rank into blocked arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\tparam OffsetT [inferred] Signed integer type for local offsets\n */\n template \n __device__ __forceinline__ void ScatterToBlocked(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items from exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n ScatterToBlocked(input_items, output_items, ranks, Int2Type());\n }\n\n\n\n /**\n * \\brief Exchanges data items annotated by rank into striped arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\tparam OffsetT [inferred] Signed integer type for local offsets\n */\n template \n __device__ __forceinline__ void ScatterToStriped(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items from exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n ScatterToStriped(input_items, output_items, ranks, Int2Type());\n }\n\n\n\n /**\n * \\brief Exchanges data items annotated by rank into striped arrangement. Items with rank -1 are not exchanged.\n *\n * \\par\n * - \\smemreuse\n *\n * \\tparam OffsetT [inferred] Signed integer type for local offsets\n */\n template \n __device__ __forceinline__ void ScatterToStripedGuarded(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items from exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM];\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n if (ranks[ITEM] >= 0)\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = int(ITEM * BLOCK_THREADS) + linear_tid;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n\n\n /**\n * \\brief Exchanges valid data items annotated by rank into striped arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\tparam OffsetT [inferred] Signed integer type for local offsets\n * \\tparam ValidFlag [inferred] FlagT type denoting which items are valid\n */\n template \n __device__ __forceinline__ void ScatterToStripedFlagged(\n InputT input_items[ITEMS_PER_THREAD], ///< [in] Items to exchange, converting between striped and blocked arrangements.\n OutputT output_items[ITEMS_PER_THREAD], ///< [out] Items from exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n ValidFlag is_valid[ITEMS_PER_THREAD]) ///< [in] Corresponding flag denoting item validity\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = ranks[ITEM];\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n if (is_valid[ITEM])\n temp_storage.buff[item_offset] = input_items[ITEM];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = int(ITEM * BLOCK_THREADS) + linear_tid;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n output_items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n\n //@} end member group\n\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n __device__ __forceinline__ void StripedToBlocked(\n InputT items[ITEMS_PER_THREAD]) ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n {\n StripedToBlocked(items, items);\n }\n\n __device__ __forceinline__ void BlockedToStriped(\n InputT items[ITEMS_PER_THREAD]) ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n {\n BlockedToStriped(items, items);\n }\n\n __device__ __forceinline__ void WarpStripedToBlocked(\n InputT items[ITEMS_PER_THREAD]) ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n {\n WarpStripedToBlocked(items, items);\n }\n\n __device__ __forceinline__ void BlockedToWarpStriped(\n InputT items[ITEMS_PER_THREAD]) ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n {\n BlockedToWarpStriped(items, items);\n }\n\n template \n __device__ __forceinline__ void ScatterToBlocked(\n InputT items[ITEMS_PER_THREAD], ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n ScatterToBlocked(items, items, ranks);\n }\n\n template \n __device__ __forceinline__ void ScatterToStriped(\n InputT items[ITEMS_PER_THREAD], ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n ScatterToStriped(items, items, ranks);\n }\n\n template \n __device__ __forceinline__ void ScatterToStripedGuarded(\n InputT items[ITEMS_PER_THREAD], ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n ScatterToStripedGuarded(items, items, ranks);\n }\n\n template \n __device__ __forceinline__ void ScatterToStripedFlagged(\n InputT items[ITEMS_PER_THREAD], ///< [in-out] Items to exchange, converting between striped and blocked arrangements.\n OffsetT ranks[ITEMS_PER_THREAD], ///< [in] Corresponding scatter ranks\n ValidFlag is_valid[ITEMS_PER_THREAD]) ///< [in] Corresponding flag denoting item validity\n {\n ScatterToStriped(items, items, ranks, is_valid);\n }\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n};\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\ntemplate <\n typename T,\n int ITEMS_PER_THREAD,\n int LOGICAL_WARP_THREADS = CUB_PTX_WARP_THREADS,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass WarpExchange\n{\nprivate:\n\n /******************************************************************************\n * Constants\n ******************************************************************************/\n\n /// Constants\n enum\n {\n // Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n WARP_ITEMS = (ITEMS_PER_THREAD * LOGICAL_WARP_THREADS) + 1,\n\n LOG_SMEM_BANKS = CUB_LOG_SMEM_BANKS(PTX_ARCH),\n SMEM_BANKS = 1 << LOG_SMEM_BANKS,\n\n // Insert padding if the number of items per thread is a power of two and > 4 (otherwise we can typically use 128b loads)\n INSERT_PADDING = (ITEMS_PER_THREAD > 4) && (PowerOfTwo::VALUE),\n PADDING_ITEMS = (INSERT_PADDING) ? (WARP_ITEMS >> LOG_SMEM_BANKS) : 0,\n };\n\n /******************************************************************************\n * Type definitions\n ******************************************************************************/\n\n /// Shared memory storage layout type\n struct _TempStorage\n {\n T buff[WARP_ITEMS + PADDING_ITEMS];\n };\n\npublic:\n\n /// \\smemstorage{WarpExchange}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\nprivate:\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n _TempStorage &temp_storage;\n int lane_id;\n\npublic:\n\n /******************************************************************************\n * Construction\n ******************************************************************************/\n\n /// Constructor\n __device__ __forceinline__ WarpExchange(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n lane_id(IS_ARCH_WARP ?\n LaneId() :\n LaneId() % LOGICAL_WARP_THREADS)\n {}\n\n\n /******************************************************************************\n * Interface\n ******************************************************************************/\n\n /**\n * \\brief Exchanges valid data items annotated by rank into striped arrangement.\n *\n * \\par\n * - \\smemreuse\n *\n * \\tparam OffsetT [inferred] Signed integer type for local offsets\n */\n template \n __device__ __forceinline__ void ScatterToStriped(\n T items[ITEMS_PER_THREAD], ///< [in-out] Items to exchange\n OffsetT ranks[ITEMS_PER_THREAD]) ///< [in] Corresponding scatter ranks\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (INSERT_PADDING) ranks[ITEM] = SHR_ADD(ranks[ITEM], LOG_SMEM_BANKS, ranks[ITEM]);\n temp_storage.buff[ranks[ITEM]] = items[ITEM];\n }\n\n WARP_SYNC(0xffffffff);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n int item_offset = (ITEM * LOGICAL_WARP_THREADS) + lane_id;\n if (INSERT_PADDING) item_offset = SHR_ADD(item_offset, LOG_SMEM_BANKS, item_offset);\n items[ITEM] = temp_storage.buff[item_offset];\n }\n }\n\n};\n\n\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_histogram.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockHistogram class provides [collective](index.html#sec0) methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"specializations/block_histogram_sort.cuh\"\n#include \"specializations/block_histogram_atomic.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Algorithmic variants\n ******************************************************************************/\n\n/**\n * \\brief BlockHistogramAlgorithm enumerates alternative algorithms for the parallel construction of block-wide histograms.\n */\nenum BlockHistogramAlgorithm\n{\n\n /**\n * \\par Overview\n * Sorting followed by differentiation. Execution is comprised of two phases:\n * -# Sort the data using efficient radix sort\n * -# Look for \"runs\" of same-valued keys by detecting discontinuities; the run-lengths are histogram bin counts.\n *\n * \\par Performance Considerations\n * Delivers consistent throughput regardless of sample bin distribution.\n */\n BLOCK_HISTO_SORT,\n\n\n /**\n * \\par Overview\n * Use atomic addition to update byte counts directly\n *\n * \\par Performance Considerations\n * Performance is strongly tied to the hardware implementation of atomic\n * addition, and may be significantly degraded for non uniformly-random\n * input distributions where many concurrent updates are likely to be\n * made to the same bin counter.\n */\n BLOCK_HISTO_ATOMIC,\n};\n\n\n\n/******************************************************************************\n * Block histogram\n ******************************************************************************/\n\n\n/**\n * \\brief The BlockHistogram class provides [collective](index.html#sec0) methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block. ![](histogram_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam T The sample type being histogrammed (must be castable to an integer bin identifier)\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ITEMS_PER_THREAD The number of items per thread\n * \\tparam BINS The number bins within the histogram\n * \\tparam ALGORITHM [optional] cub::BlockHistogramAlgorithm enumerator specifying the underlying algorithm to use (default: cub::BLOCK_HISTO_SORT)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - A histogram\n * counts the number of observations that fall into each of the disjoint categories (known as bins).\n * - BlockHistogram can be optionally specialized to use different algorithms:\n * -# cub::BLOCK_HISTO_SORT. Sorting followed by differentiation. [More...](\\ref cub::BlockHistogramAlgorithm)\n * -# cub::BLOCK_HISTO_ATOMIC. Use atomic addition to update byte counts directly. [More...](\\ref cub::BlockHistogramAlgorithm)\n *\n * \\par Performance Considerations\n * - \\granularity\n *\n * \\par A Simple Example\n * \\blockcollective{BlockHistogram}\n * \\par\n * The code snippet below illustrates a 256-bin histogram of 512 integer samples that\n * are partitioned across 128 threads where each thread owns 4 samples.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize a 256-bin BlockHistogram type for a 1D block of 128 threads having 4 character samples each\n * typedef cub::BlockHistogram BlockHistogram;\n *\n * // Allocate shared memory for BlockHistogram\n * __shared__ typename BlockHistogram::TempStorage temp_storage;\n *\n * // Allocate shared memory for block-wide histogram bin counts\n * __shared__ unsigned int smem_histogram[256];\n *\n * // Obtain input samples per thread\n * unsigned char data[4];\n * ...\n *\n * // Compute the block-wide histogram\n * BlockHistogram(temp_storage).Histogram(data, smem_histogram);\n *\n * \\endcode\n *\n * \\par Performance and Usage Considerations\n * - The histogram output can be constructed in shared or device-accessible memory\n * - See cub::BlockHistogramAlgorithm for performance details regarding algorithmic alternatives\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n int ITEMS_PER_THREAD,\n int BINS,\n BlockHistogramAlgorithm ALGORITHM = BLOCK_HISTO_SORT,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockHistogram\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n /**\n * Ensure the template parameterization meets the requirements of the\n * targeted device architecture. BLOCK_HISTO_ATOMIC can only be used\n * on version SM120 or later. Otherwise BLOCK_HISTO_SORT is used\n * regardless.\n */\n static const BlockHistogramAlgorithm SAFE_ALGORITHM =\n ((ALGORITHM == BLOCK_HISTO_ATOMIC) && (PTX_ARCH < 120)) ?\n BLOCK_HISTO_SORT :\n ALGORITHM;\n\n /// Internal specialization.\n typedef typename If<(SAFE_ALGORITHM == BLOCK_HISTO_SORT),\n BlockHistogramSort,\n BlockHistogramAtomic >::Type InternalBlockHistogram;\n\n /// Shared memory storage layout type for BlockHistogram\n typedef typename InternalBlockHistogram::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\npublic:\n\n /// \\smemstorage{BlockHistogram}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockHistogram()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockHistogram(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Histogram operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Initialize the shared histogram counters to zero.\n *\n * \\par Snippet\n * The code snippet below illustrates a the initialization and update of a\n * histogram of 512 integer samples that are partitioned across 128 threads\n * where each thread owns 4 samples.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize a 256-bin BlockHistogram type for a 1D block of 128 threads having 4 character samples each\n * typedef cub::BlockHistogram BlockHistogram;\n *\n * // Allocate shared memory for BlockHistogram\n * __shared__ typename BlockHistogram::TempStorage temp_storage;\n *\n * // Allocate shared memory for block-wide histogram bin counts\n * __shared__ unsigned int smem_histogram[256];\n *\n * // Obtain input samples per thread\n * unsigned char thread_samples[4];\n * ...\n *\n * // Initialize the block-wide histogram\n * BlockHistogram(temp_storage).InitHistogram(smem_histogram);\n *\n * // Update the block-wide histogram\n * BlockHistogram(temp_storage).Composite(thread_samples, smem_histogram);\n *\n * \\endcode\n *\n * \\tparam CounterT [inferred] Histogram counter type\n */\n template \n __device__ __forceinline__ void InitHistogram(CounterT histogram[BINS])\n {\n // Initialize histogram bin counts to zeros\n int histo_offset = 0;\n\n #pragma unroll\n for(; histo_offset + BLOCK_THREADS <= BINS; histo_offset += BLOCK_THREADS)\n {\n histogram[histo_offset + linear_tid] = 0;\n }\n // Finish up with guarded initialization if necessary\n if ((BINS % BLOCK_THREADS != 0) && (histo_offset + linear_tid < BINS))\n {\n histogram[histo_offset + linear_tid] = 0;\n }\n }\n\n\n /**\n * \\brief Constructs a block-wide histogram in shared/device-accessible memory. Each thread contributes an array of input elements.\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a 256-bin histogram of 512 integer samples that\n * are partitioned across 128 threads where each thread owns 4 samples.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize a 256-bin BlockHistogram type for a 1D block of 128 threads having 4 character samples each\n * typedef cub::BlockHistogram BlockHistogram;\n *\n * // Allocate shared memory for BlockHistogram\n * __shared__ typename BlockHistogram::TempStorage temp_storage;\n *\n * // Allocate shared memory for block-wide histogram bin counts\n * __shared__ unsigned int smem_histogram[256];\n *\n * // Obtain input samples per thread\n * unsigned char thread_samples[4];\n * ...\n *\n * // Compute the block-wide histogram\n * BlockHistogram(temp_storage).Histogram(thread_samples, smem_histogram);\n *\n * \\endcode\n *\n * \\tparam CounterT [inferred] Histogram counter type\n */\n template <\n typename CounterT >\n __device__ __forceinline__ void Histogram(\n T (&items)[ITEMS_PER_THREAD], ///< [in] Calling thread's input values to histogram\n CounterT histogram[BINS]) ///< [out] Reference to shared/device-accessible memory histogram\n {\n // Initialize histogram bin counts to zeros\n InitHistogram(histogram);\n\n CTA_SYNC();\n\n // Composite the histogram\n InternalBlockHistogram(temp_storage).Composite(items, histogram);\n }\n\n\n\n /**\n * \\brief Updates an existing block-wide histogram in shared/device-accessible memory. Each thread composites an array of input elements.\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a the initialization and update of a\n * histogram of 512 integer samples that are partitioned across 128 threads\n * where each thread owns 4 samples.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize a 256-bin BlockHistogram type for a 1D block of 128 threads having 4 character samples each\n * typedef cub::BlockHistogram BlockHistogram;\n *\n * // Allocate shared memory for BlockHistogram\n * __shared__ typename BlockHistogram::TempStorage temp_storage;\n *\n * // Allocate shared memory for block-wide histogram bin counts\n * __shared__ unsigned int smem_histogram[256];\n *\n * // Obtain input samples per thread\n * unsigned char thread_samples[4];\n * ...\n *\n * // Initialize the block-wide histogram\n * BlockHistogram(temp_storage).InitHistogram(smem_histogram);\n *\n * // Update the block-wide histogram\n * BlockHistogram(temp_storage).Composite(thread_samples, smem_histogram);\n *\n * \\endcode\n *\n * \\tparam CounterT [inferred] Histogram counter type\n */\n template <\n typename CounterT >\n __device__ __forceinline__ void Composite(\n T (&items)[ITEMS_PER_THREAD], ///< [in] Calling thread's input values to histogram\n CounterT histogram[BINS]) ///< [out] Reference to shared/device-accessible memory histogram\n {\n InternalBlockHistogram(temp_storage).Composite(items, histogram);\n }\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_load.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Operations for reading linear tiles of data into the CUDA thread block.\n */\n\n#pragma once\n\n#include \n\n#include \"block_exchange.cuh\"\n#include \"../iterator/cache_modified_input_iterator.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_macro.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIo\n * @{\n */\n\n\n/******************************************************************//**\n * \\name Blocked arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n\n/**\n * \\brief Load a linear segment of items into a blocked arrangement across the thread block.\n *\n * \\blocked\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectBlocked(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n InputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);\n\n // Load directly in thread-blocked order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n items[ITEM] = thread_itr[ITEM];\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a blocked arrangement across the thread block, guarded by range.\n *\n * \\blocked\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectBlocked(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n{\n InputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if ((linear_tid * ITEMS_PER_THREAD) + ITEM < valid_items)\n {\n items[ITEM] = thread_itr[ITEM];\n }\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a blocked arrangement across the thread block, guarded by range, with a fall-back assignment of out-of-bound elements..\n *\n * \\blocked\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n typename DefaultT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectBlocked(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n{\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n items[ITEM] = oob_default;\n\n LoadDirectBlocked(linear_tid, block_itr, items, valid_items);\n}\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n/**\n * Internal implementation for load vectorization\n */\ntemplate <\n CacheLoadModifier MODIFIER,\n typename T,\n int ITEMS_PER_THREAD>\n__device__ __forceinline__ void InternalLoadDirectBlockedVectorized(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n T *block_ptr, ///< [in] Input pointer for loading from\n T (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n // Biggest memory access word that T is a whole multiple of\n typedef typename UnitWord::DeviceWord DeviceWord;\n\n enum\n {\n TOTAL_WORDS = sizeof(items) / sizeof(DeviceWord),\n\n VECTOR_SIZE = (TOTAL_WORDS % 4 == 0) ?\n 4 :\n (TOTAL_WORDS % 2 == 0) ?\n 2 :\n 1,\n\n VECTORS_PER_THREAD = TOTAL_WORDS / VECTOR_SIZE,\n };\n\n // Vector type\n typedef typename CubVector::Type Vector;\n\n // Vector items\n Vector vec_items[VECTORS_PER_THREAD];\n\n // Aliased input ptr\n Vector* vec_ptr = reinterpret_cast(block_ptr) + (linear_tid * VECTORS_PER_THREAD);\n\n // Load directly in thread-blocked order\n #pragma unroll\n for (int ITEM = 0; ITEM < VECTORS_PER_THREAD; ITEM++)\n {\n vec_items[ITEM] = ThreadLoad(vec_ptr + ITEM);\n }\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n items[ITEM] = *(reinterpret_cast(vec_items) + ITEM);\n }\n}\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/**\n * \\brief Load a linear segment of items into a blocked arrangement across the thread block.\n *\n * \\blocked\n *\n * The input offset (\\p block_ptr + \\p block_offset) must be quad-item aligned\n *\n * The following conditions will prevent vectorization and loading will fall back to cub::BLOCK_LOAD_DIRECT:\n * - \\p ITEMS_PER_THREAD is odd\n * - The data type \\p T is not a built-in primitive or CUDA vector type (e.g., \\p short, \\p int2, \\p double, \\p float2, etc.)\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD>\n__device__ __forceinline__ void LoadDirectBlockedVectorized(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n T *block_ptr, ///< [in] Input pointer for loading from\n T (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n InternalLoadDirectBlockedVectorized(linear_tid, block_ptr, items);\n}\n\n\n//@} end member group\n/******************************************************************//**\n * \\name Striped arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n\n/**\n * \\brief Load a linear segment of items into a striped arrangement across the thread block.\n *\n * \\striped\n *\n * \\tparam BLOCK_THREADS The thread block size in threads\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n int BLOCK_THREADS,\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n InputIteratorT thread_itr = block_itr + linear_tid;\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n items[ITEM] = thread_itr[ITEM * BLOCK_THREADS];\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a striped arrangement across the thread block, guarded by range\n *\n * \\striped\n *\n * \\tparam BLOCK_THREADS The thread block size in threads\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n int BLOCK_THREADS,\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n{\n InputIteratorT thread_itr = block_itr + linear_tid;\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (linear_tid + (ITEM * BLOCK_THREADS) < valid_items)\n {\n items[ITEM] = thread_itr[ITEM * BLOCK_THREADS];\n }\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a striped arrangement across the thread block, guarded by range, with a fall-back assignment of out-of-bound elements.\n *\n * \\striped\n *\n * \\tparam BLOCK_THREADS The thread block size in threads\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n int BLOCK_THREADS,\n typename InputT,\n typename DefaultT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n{\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n items[ITEM] = oob_default;\n\n LoadDirectStriped(linear_tid, block_itr, items, valid_items);\n}\n\n\n\n//@} end member group\n/******************************************************************//**\n * \\name Warp-striped arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n\n/**\n * \\brief Load a linear segment of items into a warp-striped arrangement across the thread block.\n *\n * \\warpstriped\n *\n * \\par Usage Considerations\n * The number of threads in the thread block must be a multiple of the architecture's warp size.\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectWarpStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n int tid = linear_tid & (CUB_PTX_WARP_THREADS - 1);\n int wid = linear_tid >> CUB_PTX_LOG_WARP_THREADS;\n int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;\n\n InputIteratorT thread_itr = block_itr + warp_offset + tid ;\n\n // Load directly in warp-striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n items[ITEM] = thread_itr[(ITEM * CUB_PTX_WARP_THREADS)];\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a warp-striped arrangement across the thread block, guarded by range\n *\n * \\warpstriped\n *\n * \\par Usage Considerations\n * The number of threads in the thread block must be a multiple of the architecture's warp size.\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectWarpStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n{\n int tid = linear_tid & (CUB_PTX_WARP_THREADS - 1);\n int wid = linear_tid >> CUB_PTX_LOG_WARP_THREADS;\n int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;\n\n InputIteratorT thread_itr = block_itr + warp_offset + tid ;\n\n // Load directly in warp-striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (warp_offset + tid + (ITEM * CUB_PTX_WARP_THREADS) < valid_items)\n {\n items[ITEM] = thread_itr[(ITEM * CUB_PTX_WARP_THREADS)];\n }\n }\n}\n\n\n/**\n * \\brief Load a linear segment of items into a warp-striped arrangement across the thread block, guarded by range, with a fall-back assignment of out-of-bound elements.\n *\n * \\warpstriped\n *\n * \\par Usage Considerations\n * The number of threads in the thread block must be a multiple of the architecture's warp size.\n *\n * \\tparam T [inferred] The data type to load.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam InputIteratorT [inferred] The random-access iterator type for input \\iterator.\n */\ntemplate <\n typename InputT,\n typename DefaultT,\n int ITEMS_PER_THREAD,\n typename InputIteratorT>\n__device__ __forceinline__ void LoadDirectWarpStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n{\n // Load directly in warp-striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n items[ITEM] = oob_default;\n\n LoadDirectWarpStriped(linear_tid, block_itr, items, valid_items);\n}\n\n\n\n//@} end member group\n\n/** @} */ // end group UtilIo\n\n\n\n//-----------------------------------------------------------------------------\n// Generic BlockLoad abstraction\n//-----------------------------------------------------------------------------\n\n/**\n * \\brief cub::BlockLoadAlgorithm enumerates alternative algorithms for cub::BlockLoad to read a linear segment of data from memory into a blocked arrangement across a CUDA thread block.\n */\n\n/**\n * \\brief cub::BlockLoadAlgorithm enumerates alternative algorithms for cub::BlockLoad to read a linear segment of data from memory into a blocked arrangement across a CUDA thread block.\n */\nenum BlockLoadAlgorithm\n{\n /**\n * \\par Overview\n *\n * A [blocked arrangement](index.html#sec5sec3) of data is read\n * directly from memory.\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) decreases as the\n * access stride between threads increases (i.e., the number items per thread).\n */\n BLOCK_LOAD_DIRECT,\n\n /**\n * \\par Overview\n *\n * A [blocked arrangement](index.html#sec5sec3) of data is read\n * from memory using CUDA's built-in vectorized loads as a coalescing optimization.\n * For example, ld.global.v4.s32 instructions will be generated\n * when \\p T = \\p int and \\p ITEMS_PER_THREAD % 4 == 0.\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high until the the\n * access stride between threads (i.e., the number items per thread) exceeds the\n * maximum vector load width (typically 4 items or 64B, whichever is lower).\n * - The following conditions will prevent vectorization and loading will fall back to cub::BLOCK_LOAD_DIRECT:\n * - \\p ITEMS_PER_THREAD is odd\n * - The \\p InputIteratorTis not a simple pointer type\n * - The block input offset is not quadword-aligned\n * - The data type \\p T is not a built-in primitive or CUDA vector type (e.g., \\p short, \\p int2, \\p double, \\p float2, etc.)\n */\n BLOCK_LOAD_VECTORIZE,\n\n /**\n * \\par Overview\n *\n * A [striped arrangement](index.html#sec5sec3) of data is read\n * efficiently from memory and then locally transposed into a\n * [blocked arrangement](index.html#sec5sec3).\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items loaded per thread.\n * - The local reordering incurs slightly longer latencies and throughput than the\n * direct cub::BLOCK_LOAD_DIRECT and cub::BLOCK_LOAD_VECTORIZE alternatives.\n */\n BLOCK_LOAD_TRANSPOSE,\n\n\n /**\n * \\par Overview\n *\n * A [warp-striped arrangement](index.html#sec5sec3) of data is\n * read efficiently from memory and then locally transposed into a\n * [blocked arrangement](index.html#sec5sec3).\n *\n * \\par Usage Considerations\n * - BLOCK_THREADS must be a multiple of WARP_THREADS\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items loaded per thread.\n * - The local reordering incurs slightly larger latencies than the\n * direct cub::BLOCK_LOAD_DIRECT and cub::BLOCK_LOAD_VECTORIZE alternatives.\n * - Provisions more shared storage, but incurs smaller latencies than the\n * BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED alternative.\n */\n BLOCK_LOAD_WARP_TRANSPOSE,\n\n\n /**\n * \\par Overview\n *\n * Like \\p BLOCK_LOAD_WARP_TRANSPOSE, a [warp-striped arrangement](index.html#sec5sec3)\n * of data is read directly from memory and then is locally transposed into a\n * [blocked arrangement](index.html#sec5sec3). To reduce the shared memory\n * requirement, only one warp's worth of shared memory is provisioned and is\n * subsequently time-sliced among warps.\n *\n * \\par Usage Considerations\n * - BLOCK_THREADS must be a multiple of WARP_THREADS\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items loaded per thread.\n * - Provisions less shared memory temporary storage, but incurs larger\n * latencies than the BLOCK_LOAD_WARP_TRANSPOSE alternative.\n */\n BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED,\n};\n\n\n/**\n * \\brief The BlockLoad class provides [collective](index.html#sec0) data movement methods for loading a linear segment of items from memory into a [blocked arrangement](index.html#sec5sec3) across a CUDA thread block. ![](block_load_logo.png)\n * \\ingroup BlockModule\n * \\ingroup UtilIo\n *\n * \\tparam InputT The data type to read into (which must be convertible from the input iterator's value type).\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ITEMS_PER_THREAD The number of consecutive items partitioned onto each thread.\n * \\tparam ALGORITHM [optional] cub::BlockLoadAlgorithm tuning policy. default: cub::BLOCK_LOAD_DIRECT.\n * \\tparam WARP_TIME_SLICING [optional] Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage). (default: false)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - The BlockLoad class provides a single data movement abstraction that can be specialized\n * to implement different cub::BlockLoadAlgorithm strategies. This facilitates different\n * performance policies for different architectures, data types, granularity sizes, etc.\n * - BlockLoad can be optionally specialized by different data movement strategies:\n * -# cub::BLOCK_LOAD_DIRECT. A [blocked arrangement](index.html#sec5sec3)\n * of data is read directly from memory. [More...](\\ref cub::BlockLoadAlgorithm)\n * -# cub::BLOCK_LOAD_VECTORIZE. A [blocked arrangement](index.html#sec5sec3)\n * of data is read directly from memory using CUDA's built-in vectorized loads as a\n * coalescing optimization. [More...](\\ref cub::BlockLoadAlgorithm)\n * -# cub::BLOCK_LOAD_TRANSPOSE. A [striped arrangement](index.html#sec5sec3)\n * of data is read directly from memory and is then locally transposed into a\n * [blocked arrangement](index.html#sec5sec3). [More...](\\ref cub::BlockLoadAlgorithm)\n * -# cub::BLOCK_LOAD_WARP_TRANSPOSE. A [warp-striped arrangement](index.html#sec5sec3)\n * of data is read directly from memory and is then locally transposed into a\n * [blocked arrangement](index.html#sec5sec3). [More...](\\ref cub::BlockLoadAlgorithm)\n * -# cub::BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED,. A [warp-striped arrangement](index.html#sec5sec3)\n * of data is read directly from memory and is then locally transposed into a\n * [blocked arrangement](index.html#sec5sec3) one warp at a time. [More...](\\ref cub::BlockLoadAlgorithm)\n * - \\rowmajor\n *\n * \\par A Simple Example\n * \\blockcollective{BlockLoad}\n * \\par\n * The code snippet below illustrates the loading of a linear\n * segment of 512 integers into a \"blocked\" arrangement across 128 threads where each\n * thread owns 4 consecutive items. The load is specialized for \\p BLOCK_LOAD_WARP_TRANSPOSE,\n * meaning memory references are efficiently coalesced using a warp-striped access\n * pattern (after which items are locally reordered among threads).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockLoad for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockLoad BlockLoad;\n *\n * // Allocate shared memory for BlockLoad\n * __shared__ typename BlockLoad::TempStorage temp_storage;\n *\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage).Load(d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, 1, 2, 3, 4, 5, ....\n * The set of \\p thread_data across the block of threads in those threads will be\n * { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n *\n */\ntemplate <\n typename InputT,\n int BLOCK_DIM_X,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm ALGORITHM = BLOCK_LOAD_DIRECT,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockLoad\n{\nprivate:\n\n /******************************************************************************\n * Constants and typed definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n\n /******************************************************************************\n * Algorithmic variants\n ******************************************************************************/\n\n /// Load helper\n template \n struct LoadInternal;\n\n\n /**\n * BLOCK_LOAD_DIRECT specialization of load helper\n */\n template \n struct LoadInternal\n {\n /// Shared memory storage layout type\n typedef NullType TempStorage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ LoadInternal(\n TempStorage &/*temp_storage*/,\n int linear_tid)\n :\n linear_tid(linear_tid)\n {}\n\n /// Load a linear segment of items from memory\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n LoadDirectBlocked(linear_tid, block_itr, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n LoadDirectBlocked(linear_tid, block_itr, items, valid_items);\n }\n\n /// Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n LoadDirectBlocked(linear_tid, block_itr, items, valid_items, oob_default);\n }\n\n };\n\n\n /**\n * BLOCK_LOAD_VECTORIZE specialization of load helper\n */\n template \n struct LoadInternal\n {\n /// Shared memory storage layout type\n typedef NullType TempStorage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ LoadInternal(\n TempStorage &/*temp_storage*/,\n int linear_tid)\n :\n linear_tid(linear_tid)\n {}\n\n /// Load a linear segment of items from memory, specialized for native pointer types (attempts vectorization)\n template \n __device__ __forceinline__ void Load(\n InputT *block_ptr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n InternalLoadDirectBlockedVectorized(linear_tid, block_ptr, items);\n }\n\n /// Load a linear segment of items from memory, specialized for native pointer types (attempts vectorization)\n template \n __device__ __forceinline__ void Load(\n const InputT *block_ptr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n InternalLoadDirectBlockedVectorized(linear_tid, block_ptr, items);\n }\n\n /// Load a linear segment of items from memory, specialized for native pointer types (attempts vectorization)\n template <\n CacheLoadModifier MODIFIER,\n typename ValueType,\n typename OffsetT>\n __device__ __forceinline__ void Load(\n CacheModifiedInputIterator block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n InternalLoadDirectBlockedVectorized(linear_tid, block_itr.ptr, items);\n }\n\n /// Load a linear segment of items from memory, specialized for opaque input iterators (skips vectorization)\n template \n __device__ __forceinline__ void Load(\n _InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n LoadDirectBlocked(linear_tid, block_itr, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range (skips vectorization)\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n LoadDirectBlocked(linear_tid, block_itr, items, valid_items);\n }\n\n /// Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements (skips vectorization)\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n LoadDirectBlocked(linear_tid, block_itr, items, valid_items, oob_default);\n }\n\n };\n\n\n /**\n * BLOCK_LOAD_TRANSPOSE specialization of load helper\n */\n template \n struct LoadInternal\n {\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ LoadInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Load a linear segment of items from memory\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load{\n {\n LoadDirectStriped(linear_tid, block_itr, items);\n BlockExchange(temp_storage).StripedToBlocked(items, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n BlockExchange(temp_storage).StripedToBlocked(items, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectStriped(linear_tid, block_itr, items, temp_storage.valid_items, oob_default);\n BlockExchange(temp_storage).StripedToBlocked(items, items);\n }\n\n };\n\n\n /**\n * BLOCK_LOAD_WARP_TRANSPOSE specialization of load helper\n */\n template \n struct LoadInternal\n {\n enum\n {\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)\n };\n\n // Assert BLOCK_THREADS must be a multiple of WARP_THREADS\n CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), \"BLOCK_THREADS must be a multiple of WARP_THREADS\");\n\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ LoadInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Load a linear segment of items from memory\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load{\n {\n LoadDirectWarpStriped(linear_tid, block_itr, items);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n\n\n /// Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items, oob_default);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n };\n\n\n /**\n * BLOCK_LOAD_WARP_TRANSPOSE_TIMESLICED specialization of load helper\n */\n template \n struct LoadInternal\n {\n enum\n {\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)\n };\n\n // Assert BLOCK_THREADS must be a multiple of WARP_THREADS\n CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), \"BLOCK_THREADS must be a multiple of WARP_THREADS\");\n\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ LoadInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Load a linear segment of items from memory\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load{\n {\n LoadDirectWarpStriped(linear_tid, block_itr, items);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n\n /// Load a linear segment of items from memory, guarded by range\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n\n\n /// Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n LoadDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items, oob_default);\n BlockExchange(temp_storage).WarpStripedToBlocked(items, items);\n }\n };\n\n\n /******************************************************************************\n * Type definitions\n ******************************************************************************/\n\n /// Internal load implementation to use\n typedef LoadInternal InternalLoad;\n\n\n /// Shared memory storage layout type\n typedef typename InternalLoad::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\npublic:\n\n /// \\smemstorage{BlockLoad}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockLoad()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockLoad(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Data movement\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Load a linear segment of items from memory.\n *\n * \\par\n * - \\blocked\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the loading of a linear\n * segment of 512 integers into a \"blocked\" arrangement across 128 threads where each\n * thread owns 4 consecutive items. The load is specialized for \\p BLOCK_LOAD_WARP_TRANSPOSE,\n * meaning memory references are efficiently coalesced using a warp-striped access\n * pattern (after which items are locally reordered among threads).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockLoad for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockLoad BlockLoad;\n *\n * // Allocate shared memory for BlockLoad\n * __shared__ typename BlockLoad::TempStorage temp_storage;\n *\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage).Load(d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, 1, 2, 3, 4, 5, ....\n * The set of \\p thread_data across the block of threads in those threads will be\n * { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n *\n */\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n {\n InternalLoad(temp_storage, linear_tid).Load(block_itr, items);\n }\n\n\n /**\n * \\brief Load a linear segment of items from memory, guarded by range.\n *\n * \\par\n * - \\blocked\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the guarded loading of a linear\n * segment of 512 integers into a \"blocked\" arrangement across 128 threads where each\n * thread owns 4 consecutive items. The load is specialized for \\p BLOCK_LOAD_WARP_TRANSPOSE,\n * meaning memory references are efficiently coalesced using a warp-striped access\n * pattern (after which items are locally reordered among threads).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, int valid_items, ...)\n * {\n * // Specialize BlockLoad for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockLoad BlockLoad;\n *\n * // Allocate shared memory for BlockLoad\n * __shared__ typename BlockLoad::TempStorage temp_storage;\n *\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage).Load(d_data, thread_data, valid_items);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, 1, 2, 3, 4, 5, 6... and \\p valid_items is \\p 5.\n * The set of \\p thread_data across the block of threads in those threads will be\n * { [0,1,2,3], [4,?,?,?], ..., [?,?,?,?] }, with only the first two threads\n * being unmasked to load portions of valid data (and other items remaining unassigned).\n *\n */\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items) ///< [in] Number of valid items to load\n {\n InternalLoad(temp_storage, linear_tid).Load(block_itr, items, valid_items);\n }\n\n\n /**\n * \\brief Load a linear segment of items from memory, guarded by range, with a fall-back assignment of out-of-bound elements\n *\n * \\par\n * - \\blocked\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the guarded loading of a linear\n * segment of 512 integers into a \"blocked\" arrangement across 128 threads where each\n * thread owns 4 consecutive items. The load is specialized for \\p BLOCK_LOAD_WARP_TRANSPOSE,\n * meaning memory references are efficiently coalesced using a warp-striped access\n * pattern (after which items are locally reordered among threads).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, int valid_items, ...)\n * {\n * // Specialize BlockLoad for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockLoad BlockLoad;\n *\n * // Allocate shared memory for BlockLoad\n * __shared__ typename BlockLoad::TempStorage temp_storage;\n *\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage).Load(d_data, thread_data, valid_items, -1);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, 1, 2, 3, 4, 5, 6...,\n * \\p valid_items is \\p 5, and the out-of-bounds default is \\p -1.\n * The set of \\p thread_data across the block of threads in those threads will be\n * { [0,1,2,3], [4,-1,-1,-1], ..., [-1,-1,-1,-1] }, with only the first two threads\n * being unmasked to load portions of valid data (and other items are assigned \\p -1)\n *\n */\n template \n __device__ __forceinline__ void Load(\n InputIteratorT block_itr, ///< [in] The thread block's base input iterator for loading from\n InputT (&items)[ITEMS_PER_THREAD], ///< [out] Data to load\n int valid_items, ///< [in] Number of valid items to load\n DefaultT oob_default) ///< [in] Default value to assign out-of-bound items\n {\n InternalLoad(temp_storage, linear_tid).Load(block_itr, items, valid_items, oob_default);\n }\n\n\n //@} end member group\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_radix_rank.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockRadixRank provides operations for ranking unsigned integer types within a CUDA thread block\n */\n\n#pragma once\n\n#include \n\n#include \"../thread/thread_reduce.cuh\"\n#include \"../thread/thread_scan.cuh\"\n#include \"../block/block_scan.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief BlockRadixRank provides operations for ranking unsigned integer types within a CUDA thread block.\n * \\ingroup BlockModule\n *\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam RADIX_BITS The number of radix bits per digit place\n * \\tparam IS_DESCENDING Whether or not the sorted-order is high-to-low\n * \\tparam MEMOIZE_OUTER_SCAN [optional] Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure (default: true for architectures SM35 and newer, false otherwise). See BlockScanAlgorithm::BLOCK_SCAN_RAKING_MEMOIZE for more details.\n * \\tparam INNER_SCAN_ALGORITHM [optional] The cub::BlockScanAlgorithm algorithm to use (default: cub::BLOCK_SCAN_WARP_SCANS)\n * \\tparam SMEM_CONFIG [optional] Shared memory bank mode (default: \\p cudaSharedMemBankSizeFourByte)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * Blah...\n * - Keys must be in a form suitable for radix ranking (i.e., unsigned bits).\n * - \\blocked\n *\n * \\par Performance Considerations\n * - \\granularity\n *\n * \\par Examples\n * \\par\n * - Example 1: Simple radix rank of 32-bit integer keys\n * \\code\n * #include \n *\n * template \n * __global__ void ExampleKernel(...)\n * {\n *\n * \\endcode\n */\ntemplate <\n int BLOCK_DIM_X,\n int RADIX_BITS,\n bool IS_DESCENDING,\n bool MEMOIZE_OUTER_SCAN = (CUB_PTX_ARCH >= 350) ? true : false,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM = BLOCK_SCAN_WARP_SCANS,\n cudaSharedMemConfig SMEM_CONFIG = cudaSharedMemBankSizeFourByte,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockRadixRank\n{\nprivate:\n\n /******************************************************************************\n * Type definitions and constants\n ******************************************************************************/\n\n // Integer type for digit counters (to be packed into words of type PackedCounters)\n typedef unsigned short DigitCounter;\n\n // Integer type for packing DigitCounters into columns of shared memory banks\n typedef typename If<(SMEM_CONFIG == cudaSharedMemBankSizeEightByte),\n unsigned long long,\n unsigned int>::Type PackedCounter;\n\n enum\n {\n // The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n RADIX_DIGITS = 1 << RADIX_BITS,\n\n LOG_WARP_THREADS = CUB_LOG_WARP_THREADS(PTX_ARCH),\n WARP_THREADS = 1 << LOG_WARP_THREADS,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n BYTES_PER_COUNTER = sizeof(DigitCounter),\n LOG_BYTES_PER_COUNTER = Log2::VALUE,\n\n PACKING_RATIO = sizeof(PackedCounter) / sizeof(DigitCounter),\n LOG_PACKING_RATIO = Log2::VALUE,\n\n LOG_COUNTER_LANES = CUB_MAX((RADIX_BITS - LOG_PACKING_RATIO), 0), // Always at least one lane\n COUNTER_LANES = 1 << LOG_COUNTER_LANES,\n\n // The number of packed counters per thread (plus one for padding)\n PADDED_COUNTER_LANES = COUNTER_LANES + 1,\n RAKING_SEGMENT = PADDED_COUNTER_LANES,\n };\n\npublic:\n\n enum\n {\n /// Number of bin-starting offsets tracked per thread\n BINS_TRACKED_PER_THREAD = CUB_MAX(1, RADIX_DIGITS / BLOCK_THREADS),\n };\n\nprivate:\n\n\n /// BlockScan type\n typedef BlockScan<\n PackedCounter,\n BLOCK_DIM_X,\n INNER_SCAN_ALGORITHM,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n BlockScan;\n\n\n /// Shared memory storage layout type for BlockRadixRank\n struct __align__(16) _TempStorage\n {\n union Aliasable\n {\n DigitCounter digit_counters[PADDED_COUNTER_LANES][BLOCK_THREADS][PACKING_RATIO];\n PackedCounter raking_grid[BLOCK_THREADS][RAKING_SEGMENT];\n\n } aliasable;\n\n // Storage for scanning local ranks\n typename BlockScan::TempStorage block_scan;\n };\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n /// Copy of raking segment, promoted to registers\n PackedCounter cached_segment[RAKING_SEGMENT];\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /**\n * Internal storage allocator\n */\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /**\n * Performs upsweep raking reduction, returning the aggregate\n */\n __device__ __forceinline__ PackedCounter Upsweep()\n {\n PackedCounter *smem_raking_ptr = temp_storage.aliasable.raking_grid[linear_tid];\n PackedCounter *raking_ptr;\n\n if (MEMOIZE_OUTER_SCAN)\n {\n // Copy data into registers\n #pragma unroll\n for (int i = 0; i < RAKING_SEGMENT; i++)\n {\n cached_segment[i] = smem_raking_ptr[i];\n }\n raking_ptr = cached_segment;\n }\n else\n {\n raking_ptr = smem_raking_ptr;\n }\n\n return internal::ThreadReduce(raking_ptr, Sum());\n }\n\n\n /// Performs exclusive downsweep raking scan\n __device__ __forceinline__ void ExclusiveDownsweep(\n PackedCounter raking_partial)\n {\n PackedCounter *smem_raking_ptr = temp_storage.aliasable.raking_grid[linear_tid];\n\n PackedCounter *raking_ptr = (MEMOIZE_OUTER_SCAN) ?\n cached_segment :\n smem_raking_ptr;\n\n // Exclusive raking downsweep scan\n internal::ThreadScanExclusive(raking_ptr, raking_ptr, Sum(), raking_partial);\n\n if (MEMOIZE_OUTER_SCAN)\n {\n // Copy data back to smem\n #pragma unroll\n for (int i = 0; i < RAKING_SEGMENT; i++)\n {\n smem_raking_ptr[i] = cached_segment[i];\n }\n }\n }\n\n\n /**\n * Reset shared memory digit counters\n */\n __device__ __forceinline__ void ResetCounters()\n {\n // Reset shared memory digit counters\n #pragma unroll\n for (int LANE = 0; LANE < PADDED_COUNTER_LANES; LANE++)\n {\n *((PackedCounter*) temp_storage.aliasable.digit_counters[LANE][linear_tid]) = 0;\n }\n }\n\n\n /**\n * Block-scan prefix callback\n */\n struct PrefixCallBack\n {\n __device__ __forceinline__ PackedCounter operator()(PackedCounter block_aggregate)\n {\n PackedCounter block_prefix = 0;\n\n // Propagate totals in packed fields\n #pragma unroll\n for (int PACKED = 1; PACKED < PACKING_RATIO; PACKED++)\n {\n block_prefix += block_aggregate << (sizeof(DigitCounter) * 8 * PACKED);\n }\n\n return block_prefix;\n }\n };\n\n\n /**\n * Scan shared memory digit counters.\n */\n __device__ __forceinline__ void ScanCounters()\n {\n // Upsweep scan\n PackedCounter raking_partial = Upsweep();\n\n // Compute exclusive sum\n PackedCounter exclusive_partial;\n PrefixCallBack prefix_call_back;\n BlockScan(temp_storage.block_scan).ExclusiveSum(raking_partial, exclusive_partial, prefix_call_back);\n\n // Downsweep scan with exclusive partial\n ExclusiveDownsweep(exclusive_partial);\n }\n\npublic:\n\n /// \\smemstorage{BlockScan}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockRadixRank()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockRadixRank(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Raking\n *********************************************************************/\n //@{\n\n /**\n * \\brief Rank keys.\n */\n template <\n typename UnsignedBits,\n int KEYS_PER_THREAD>\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&keys)[KEYS_PER_THREAD], ///< [in] Keys for this tile\n int (&ranks)[KEYS_PER_THREAD], ///< [out] For each key, the local rank within the tile\n int current_bit, ///< [in] The least-significant bit position of the current digit to extract\n int num_bits) ///< [in] The number of bits in the current digit\n {\n DigitCounter thread_prefixes[KEYS_PER_THREAD]; // For each key, the count of previous keys in this tile having the same digit\n DigitCounter* digit_counters[KEYS_PER_THREAD]; // For each key, the byte-offset of its corresponding digit counter in smem\n\n // Reset shared memory digit counters\n ResetCounters();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < KEYS_PER_THREAD; ++ITEM)\n {\n // Get digit\n unsigned int digit = BFE(keys[ITEM], current_bit, num_bits);\n\n // Get sub-counter\n unsigned int sub_counter = digit >> LOG_COUNTER_LANES;\n\n // Get counter lane\n unsigned int counter_lane = digit & (COUNTER_LANES - 1);\n\n if (IS_DESCENDING)\n {\n sub_counter = PACKING_RATIO - 1 - sub_counter;\n counter_lane = COUNTER_LANES - 1 - counter_lane;\n }\n\n // Pointer to smem digit counter\n digit_counters[ITEM] = &temp_storage.aliasable.digit_counters[counter_lane][linear_tid][sub_counter];\n\n // Load thread-exclusive prefix\n thread_prefixes[ITEM] = *digit_counters[ITEM];\n\n // Store inclusive prefix\n *digit_counters[ITEM] = thread_prefixes[ITEM] + 1;\n }\n\n CTA_SYNC();\n\n // Scan shared memory counters\n ScanCounters();\n\n CTA_SYNC();\n\n // Extract the local ranks of each key\n for (int ITEM = 0; ITEM < KEYS_PER_THREAD; ++ITEM)\n {\n // Add in thread block exclusive prefix\n ranks[ITEM] = thread_prefixes[ITEM] + *digit_counters[ITEM];\n }\n }\n\n\n /**\n * \\brief Rank keys. For the lower \\p RADIX_DIGITS threads, digit counts for each digit are provided for the corresponding thread.\n */\n template <\n typename UnsignedBits,\n int KEYS_PER_THREAD>\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&keys)[KEYS_PER_THREAD], ///< [in] Keys for this tile\n int (&ranks)[KEYS_PER_THREAD], ///< [out] For each key, the local rank within the tile (out parameter)\n int current_bit, ///< [in] The least-significant bit position of the current digit to extract\n int num_bits, ///< [in] The number of bits in the current digit\n int (&exclusive_digit_prefix)[BINS_TRACKED_PER_THREAD]) ///< [out] The exclusive prefix sum for the digits [(threadIdx.x * BINS_TRACKED_PER_THREAD) ... (threadIdx.x * BINS_TRACKED_PER_THREAD) + BINS_TRACKED_PER_THREAD - 1]\n {\n // Rank keys\n RankKeys(keys, ranks, current_bit, num_bits);\n\n // Get the inclusive and exclusive digit totals corresponding to the calling thread.\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (linear_tid * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n if (IS_DESCENDING)\n bin_idx = RADIX_DIGITS - bin_idx - 1;\n\n // Obtain ex/inclusive digit counts. (Unfortunately these all reside in the\n // first counter column, resulting in unavoidable bank conflicts.)\n unsigned int counter_lane = (bin_idx & (COUNTER_LANES - 1));\n unsigned int sub_counter = bin_idx >> (LOG_COUNTER_LANES);\n\n exclusive_digit_prefix[track] = temp_storage.aliasable.digit_counters[counter_lane][0][sub_counter];\n }\n }\n }\n};\n\n\n\n\n\n/**\n * Radix-rank using match.any\n */\ntemplate <\n int BLOCK_DIM_X,\n int RADIX_BITS,\n bool IS_DESCENDING,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM = BLOCK_SCAN_WARP_SCANS,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockRadixRankMatch\n{\nprivate:\n\n /******************************************************************************\n * Type definitions and constants\n ******************************************************************************/\n\n typedef int32_t RankT;\n typedef int32_t DigitCounterT;\n\n enum\n {\n // The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n RADIX_DIGITS = 1 << RADIX_BITS,\n\n LOG_WARP_THREADS = CUB_LOG_WARP_THREADS(PTX_ARCH),\n WARP_THREADS = 1 << LOG_WARP_THREADS,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n PADDED_WARPS = ((WARPS & 0x1) == 0) ?\n WARPS + 1 :\n WARPS,\n\n COUNTERS = PADDED_WARPS * RADIX_DIGITS,\n RAKING_SEGMENT = (COUNTERS + BLOCK_THREADS - 1) / BLOCK_THREADS,\n PADDED_RAKING_SEGMENT = ((RAKING_SEGMENT & 0x1) == 0) ?\n RAKING_SEGMENT + 1 :\n RAKING_SEGMENT,\n };\n\npublic:\n\n enum\n {\n /// Number of bin-starting offsets tracked per thread\n BINS_TRACKED_PER_THREAD = CUB_MAX(1, RADIX_DIGITS / BLOCK_THREADS),\n };\n\nprivate:\n\n /// BlockScan type\n typedef BlockScan<\n DigitCounterT,\n BLOCK_THREADS,\n INNER_SCAN_ALGORITHM,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n BlockScanT;\n\n\n /// Shared memory storage layout type for BlockRadixRank\n struct __align__(16) _TempStorage\n {\n typename BlockScanT::TempStorage block_scan;\n\n union __align__(16) Aliasable\n {\n volatile DigitCounterT warp_digit_counters[RADIX_DIGITS][PADDED_WARPS];\n DigitCounterT raking_grid[BLOCK_THREADS][PADDED_RAKING_SEGMENT];\n\n } aliasable;\n };\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\n\npublic:\n\n /// \\smemstorage{BlockScan}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockRadixRankMatch(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Raking\n *********************************************************************/\n //@{\n\n /**\n * \\brief Rank keys.\n */\n template <\n typename UnsignedBits,\n int KEYS_PER_THREAD>\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&keys)[KEYS_PER_THREAD], ///< [in] Keys for this tile\n int (&ranks)[KEYS_PER_THREAD], ///< [out] For each key, the local rank within the tile\n int current_bit, ///< [in] The least-significant bit position of the current digit to extract\n int num_bits) ///< [in] The number of bits in the current digit\n {\n // Initialize shared digit counters\n\n #pragma unroll\n for (int ITEM = 0; ITEM < PADDED_RAKING_SEGMENT; ++ITEM)\n temp_storage.aliasable.raking_grid[linear_tid][ITEM] = 0;\n\n CTA_SYNC();\n\n // Each warp will strip-mine its section of input, one strip at a time\n\n volatile DigitCounterT *digit_counters[KEYS_PER_THREAD];\n uint32_t lane_id = LaneId();\n uint32_t warp_id = linear_tid >> LOG_WARP_THREADS;\n uint32_t lane_mask_lt = LaneMaskLt();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < KEYS_PER_THREAD; ++ITEM)\n {\n // My digit\n uint32_t digit = BFE(keys[ITEM], current_bit, num_bits);\n\n if (IS_DESCENDING)\n digit = RADIX_DIGITS - digit - 1;\n\n // Mask of peers who have same digit as me\n uint32_t peer_mask = MatchAny(digit);\n\n // Pointer to smem digit counter for this key\n digit_counters[ITEM] = &temp_storage.aliasable.warp_digit_counters[digit][warp_id];\n\n // Number of occurrences in previous strips\n DigitCounterT warp_digit_prefix = *digit_counters[ITEM];\n\n // Warp-sync\n WARP_SYNC(0xFFFFFFFF);\n\n // Number of peers having same digit as me\n int32_t digit_count = __popc(peer_mask);\n\n // Number of lower-ranked peers having same digit seen so far\n int32_t peer_digit_prefix = __popc(peer_mask & lane_mask_lt);\n\n if (peer_digit_prefix == 0)\n {\n // First thread for each digit updates the shared warp counter\n *digit_counters[ITEM] = DigitCounterT(warp_digit_prefix + digit_count);\n }\n\n // Warp-sync\n WARP_SYNC(0xFFFFFFFF);\n\n // Number of prior keys having same digit\n ranks[ITEM] = warp_digit_prefix + DigitCounterT(peer_digit_prefix);\n }\n\n CTA_SYNC();\n\n // Scan warp counters\n\n DigitCounterT scan_counters[PADDED_RAKING_SEGMENT];\n\n #pragma unroll\n for (int ITEM = 0; ITEM < PADDED_RAKING_SEGMENT; ++ITEM)\n scan_counters[ITEM] = temp_storage.aliasable.raking_grid[linear_tid][ITEM];\n\n BlockScanT(temp_storage.block_scan).ExclusiveSum(scan_counters, scan_counters);\n\n #pragma unroll\n for (int ITEM = 0; ITEM < PADDED_RAKING_SEGMENT; ++ITEM)\n temp_storage.aliasable.raking_grid[linear_tid][ITEM] = scan_counters[ITEM];\n\n CTA_SYNC();\n\n // Seed ranks with counter values from previous warps\n #pragma unroll\n for (int ITEM = 0; ITEM < KEYS_PER_THREAD; ++ITEM)\n ranks[ITEM] += *digit_counters[ITEM];\n }\n\n\n /**\n * \\brief Rank keys. For the lower \\p RADIX_DIGITS threads, digit counts for each digit are provided for the corresponding thread.\n */\n template <\n typename UnsignedBits,\n int KEYS_PER_THREAD>\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&keys)[KEYS_PER_THREAD], ///< [in] Keys for this tile\n int (&ranks)[KEYS_PER_THREAD], ///< [out] For each key, the local rank within the tile (out parameter)\n int current_bit, ///< [in] The least-significant bit position of the current digit to extract\n int num_bits, ///< [in] The number of bits in the current digit\n int (&exclusive_digit_prefix)[BINS_TRACKED_PER_THREAD]) ///< [out] The exclusive prefix sum for the digits [(threadIdx.x * BINS_TRACKED_PER_THREAD) ... (threadIdx.x * BINS_TRACKED_PER_THREAD) + BINS_TRACKED_PER_THREAD - 1]\n {\n RankKeys(keys, ranks, current_bit, num_bits);\n\n // Get exclusive count for each digit\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (linear_tid * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n {\n if (IS_DESCENDING)\n bin_idx = RADIX_DIGITS - bin_idx - 1;\n\n exclusive_digit_prefix[track] = temp_storage.aliasable.warp_digit_counters[bin_idx][0];\n }\n }\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/block/block_radix_sort.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockRadixSort class provides [collective](index.html#sec0) methods for radix sorting of items partitioned across a CUDA thread block.\n */\n\n\n#pragma once\n\n#include \"block_exchange.cuh\"\n#include \"block_radix_rank.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief The BlockRadixSort class provides [collective](index.html#sec0) methods for sorting items partitioned across a CUDA thread block using a radix sorting method. ![](sorting_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam KeyT KeyT type\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ITEMS_PER_THREAD The number of items per thread\n * \\tparam ValueT [optional] ValueT type (default: cub::NullType, which indicates a keys-only sort)\n * \\tparam RADIX_BITS [optional] The number of radix bits per digit place (default: 4 bits)\n * \\tparam MEMOIZE_OUTER_SCAN [optional] Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure (default: true for architectures SM35 and newer, false otherwise).\n * \\tparam INNER_SCAN_ALGORITHM [optional] The cub::BlockScanAlgorithm algorithm to use (default: cub::BLOCK_SCAN_WARP_SCANS)\n * \\tparam SMEM_CONFIG [optional] Shared memory bank mode (default: \\p cudaSharedMemBankSizeFourByte)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - The [radix sorting method](http://en.wikipedia.org/wiki/Radix_sort) arranges\n * items into ascending order. It relies upon a positional representation for\n * keys, i.e., each key is comprised of an ordered sequence of symbols (e.g., digits,\n * characters, etc.) specified from least-significant to most-significant. For a\n * given input sequence of keys and a set of rules specifying a total ordering\n * of the symbolic alphabet, the radix sorting method produces a lexicographic\n * ordering of those keys.\n * - BlockRadixSort can sort all of the built-in C++ numeric primitive types, e.g.:\n * unsigned char, \\p int, \\p double, etc. Within each key, the implementation treats fixed-length\n * bit-sequences of \\p RADIX_BITS as radix digit places. Although the direct radix sorting\n * method can only be applied to unsigned integral types, BlockRadixSort\n * is able to sort signed and floating-point types via simple bit-wise transformations\n * that ensure lexicographic key ordering.\n * - \\rowmajor\n *\n * \\par Performance Considerations\n * - \\granularity\n *\n * \\par A Simple Example\n * \\blockcollective{BlockRadixSort}\n * \\par\n * The code snippet below illustrates a sort of 512 integer keys that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * ...\n *\n * // Collectively sort the keys\n * BlockRadixSort(temp_storage).Sort(thread_keys);\n *\n * ...\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n *\n */\ntemplate <\n typename KeyT,\n int BLOCK_DIM_X,\n int ITEMS_PER_THREAD,\n typename ValueT = NullType,\n int RADIX_BITS = 4,\n bool MEMOIZE_OUTER_SCAN = (CUB_PTX_ARCH >= 350) ? true : false,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM = BLOCK_SCAN_WARP_SCANS,\n cudaSharedMemConfig SMEM_CONFIG = cudaSharedMemBankSizeFourByte,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockRadixSort\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n enum\n {\n // The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n // Whether or not there are values to be trucked along with keys\n KEYS_ONLY = Equals::VALUE,\n };\n\n // KeyT traits and unsigned bits type\n typedef Traits KeyTraits;\n typedef typename KeyTraits::UnsignedBits UnsignedBits;\n\n /// Ascending BlockRadixRank utility type\n typedef BlockRadixRank<\n BLOCK_DIM_X,\n RADIX_BITS,\n false,\n MEMOIZE_OUTER_SCAN,\n INNER_SCAN_ALGORITHM,\n SMEM_CONFIG,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n AscendingBlockRadixRank;\n\n /// Descending BlockRadixRank utility type\n typedef BlockRadixRank<\n BLOCK_DIM_X,\n RADIX_BITS,\n true,\n MEMOIZE_OUTER_SCAN,\n INNER_SCAN_ALGORITHM,\n SMEM_CONFIG,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n DescendingBlockRadixRank;\n\n /// BlockExchange utility type for keys\n typedef BlockExchange BlockExchangeKeys;\n\n /// BlockExchange utility type for values\n typedef BlockExchange BlockExchangeValues;\n\n /// Shared memory storage layout type\n union _TempStorage\n {\n typename AscendingBlockRadixRank::TempStorage asending_ranking_storage;\n typename DescendingBlockRadixRank::TempStorage descending_ranking_storage;\n typename BlockExchangeKeys::TempStorage exchange_keys;\n typename BlockExchangeValues::TempStorage exchange_values;\n };\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n /// Rank keys (specialized for ascending sort)\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&unsigned_keys)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n int begin_bit,\n int pass_bits,\n Int2Type /*is_descending*/)\n {\n AscendingBlockRadixRank(temp_storage.asending_ranking_storage).RankKeys(\n unsigned_keys,\n ranks,\n begin_bit,\n pass_bits);\n }\n\n /// Rank keys (specialized for descending sort)\n __device__ __forceinline__ void RankKeys(\n UnsignedBits (&unsigned_keys)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n int begin_bit,\n int pass_bits,\n Int2Type /*is_descending*/)\n {\n DescendingBlockRadixRank(temp_storage.descending_ranking_storage).RankKeys(\n unsigned_keys,\n ranks,\n begin_bit,\n pass_bits);\n }\n\n /// ExchangeValues (specialized for key-value sort, to-blocked arrangement)\n __device__ __forceinline__ void ExchangeValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n Int2Type /*is_keys_only*/,\n Int2Type /*is_blocked*/)\n {\n CTA_SYNC();\n\n // Exchange values through shared memory in blocked arrangement\n BlockExchangeValues(temp_storage.exchange_values).ScatterToBlocked(values, ranks);\n }\n\n /// ExchangeValues (specialized for key-value sort, to-striped arrangement)\n __device__ __forceinline__ void ExchangeValues(\n ValueT (&values)[ITEMS_PER_THREAD],\n int (&ranks)[ITEMS_PER_THREAD],\n Int2Type /*is_keys_only*/,\n Int2Type /*is_blocked*/)\n {\n CTA_SYNC();\n\n // Exchange values through shared memory in blocked arrangement\n BlockExchangeValues(temp_storage.exchange_values).ScatterToStriped(values, ranks);\n }\n\n /// ExchangeValues (specialized for keys-only sort)\n template \n __device__ __forceinline__ void ExchangeValues(\n ValueT (&/*values*/)[ITEMS_PER_THREAD],\n int (&/*ranks*/)[ITEMS_PER_THREAD],\n Int2Type /*is_keys_only*/,\n Int2Type /*is_blocked*/)\n {}\n\n /// Sort blocked arrangement\n template \n __device__ __forceinline__ void SortBlocked(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< Values to sort\n int begin_bit, ///< The beginning (least-significant) bit index needed for key comparison\n int end_bit, ///< The past-the-end (most-significant) bit index needed for key comparison\n Int2Type is_descending, ///< Tag whether is a descending-order sort\n Int2Type is_keys_only) ///< Tag whether is keys-only sort\n {\n UnsignedBits (&unsigned_keys)[ITEMS_PER_THREAD] =\n reinterpret_cast(keys);\n\n // Twiddle bits if necessary\n #pragma unroll\n for (int KEY = 0; KEY < ITEMS_PER_THREAD; KEY++)\n {\n unsigned_keys[KEY] = KeyTraits::TwiddleIn(unsigned_keys[KEY]);\n }\n\n // Radix sorting passes\n while (true)\n {\n int pass_bits = CUB_MIN(RADIX_BITS, end_bit - begin_bit);\n\n // Rank the blocked keys\n int ranks[ITEMS_PER_THREAD];\n RankKeys(unsigned_keys, ranks, begin_bit, pass_bits, is_descending);\n begin_bit += RADIX_BITS;\n\n CTA_SYNC();\n\n // Exchange keys through shared memory in blocked arrangement\n BlockExchangeKeys(temp_storage.exchange_keys).ScatterToBlocked(keys, ranks);\n\n // Exchange values through shared memory in blocked arrangement\n ExchangeValues(values, ranks, is_keys_only, Int2Type());\n\n // Quit if done\n if (begin_bit >= end_bit) break;\n\n CTA_SYNC();\n }\n\n // Untwiddle bits if necessary\n #pragma unroll\n for (int KEY = 0; KEY < ITEMS_PER_THREAD; KEY++)\n {\n unsigned_keys[KEY] = KeyTraits::TwiddleOut(unsigned_keys[KEY]);\n }\n }\n\npublic:\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n /// Sort blocked -> striped arrangement\n template \n __device__ __forceinline__ void SortBlockedToStriped(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< Values to sort\n int begin_bit, ///< The beginning (least-significant) bit index needed for key comparison\n int end_bit, ///< The past-the-end (most-significant) bit index needed for key comparison\n Int2Type is_descending, ///< Tag whether is a descending-order sort\n Int2Type is_keys_only) ///< Tag whether is keys-only sort\n {\n UnsignedBits (&unsigned_keys)[ITEMS_PER_THREAD] =\n reinterpret_cast(keys);\n\n // Twiddle bits if necessary\n #pragma unroll\n for (int KEY = 0; KEY < ITEMS_PER_THREAD; KEY++)\n {\n unsigned_keys[KEY] = KeyTraits::TwiddleIn(unsigned_keys[KEY]);\n }\n\n // Radix sorting passes\n while (true)\n {\n int pass_bits = CUB_MIN(RADIX_BITS, end_bit - begin_bit);\n\n // Rank the blocked keys\n int ranks[ITEMS_PER_THREAD];\n RankKeys(unsigned_keys, ranks, begin_bit, pass_bits, is_descending);\n begin_bit += RADIX_BITS;\n\n CTA_SYNC();\n\n // Check if this is the last pass\n if (begin_bit >= end_bit)\n {\n // Last pass exchanges keys through shared memory in striped arrangement\n BlockExchangeKeys(temp_storage.exchange_keys).ScatterToStriped(keys, ranks);\n\n // Last pass exchanges through shared memory in striped arrangement\n ExchangeValues(values, ranks, is_keys_only, Int2Type());\n\n // Quit\n break;\n }\n\n // Exchange keys through shared memory in blocked arrangement\n BlockExchangeKeys(temp_storage.exchange_keys).ScatterToBlocked(keys, ranks);\n\n // Exchange values through shared memory in blocked arrangement\n ExchangeValues(values, ranks, is_keys_only, Int2Type());\n\n CTA_SYNC();\n }\n\n // Untwiddle bits if necessary\n #pragma unroll\n for (int KEY = 0; KEY < ITEMS_PER_THREAD; KEY++)\n {\n unsigned_keys[KEY] = KeyTraits::TwiddleOut(unsigned_keys[KEY]);\n }\n }\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n /// \\smemstorage{BlockRadixSort}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockRadixSort()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockRadixSort(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Sorting (blocked arrangements)\n *********************************************************************/\n //@{\n\n /**\n * \\brief Performs an ascending block-wide radix sort over a [blocked arrangement](index.html#sec5sec3) of keys.\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * ...\n *\n * // Collectively sort the keys\n * BlockRadixSort(temp_storage).Sort(thread_keys);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }.\n * The corresponding output \\p thread_keys in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n */\n __device__ __forceinline__ void Sort(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n NullType values[ITEMS_PER_THREAD];\n\n SortBlocked(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n /**\n * \\brief Performs an ascending block-wide radix sort across a [blocked arrangement](index.html#sec5sec3) of keys and values.\n *\n * \\par\n * - BlockRadixSort can only accommodate one associated tile of values. To \"truck along\"\n * more than one tile of values, simply perform a key-value sort of the keys paired\n * with a temporary value array that enumerates the key indices. The reordered indices\n * can then be used as a gather-vector for exchanging other associated tile data through\n * shared memory.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys and values that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive pairs.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys and values each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * int thread_values[4];\n * ...\n *\n * // Collectively sort the keys and values among block threads\n * BlockRadixSort(temp_storage).Sort(thread_keys, thread_values);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [0,1,2,3], [4,5,6,7], [8,9,10,11], ..., [508,509,510,511] }.\n *\n */\n __device__ __forceinline__ void Sort(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< [in-out] Values to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n SortBlocked(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n /**\n * \\brief Performs a descending block-wide radix sort over a [blocked arrangement](index.html#sec5sec3) of keys.\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * ...\n *\n * // Collectively sort the keys\n * BlockRadixSort(temp_storage).Sort(thread_keys);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }.\n * The corresponding output \\p thread_keys in those threads will be\n * { [511,510,509,508], [11,10,9,8], [7,6,5,4], ..., [3,2,1,0] }.\n */\n __device__ __forceinline__ void SortDescending(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n NullType values[ITEMS_PER_THREAD];\n\n SortBlocked(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n /**\n * \\brief Performs a descending block-wide radix sort across a [blocked arrangement](index.html#sec5sec3) of keys and values.\n *\n * \\par\n * - BlockRadixSort can only accommodate one associated tile of values. To \"truck along\"\n * more than one tile of values, simply perform a key-value sort of the keys paired\n * with a temporary value array that enumerates the key indices. The reordered indices\n * can then be used as a gather-vector for exchanging other associated tile data through\n * shared memory.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys and values that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive pairs.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys and values each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * int thread_values[4];\n * ...\n *\n * // Collectively sort the keys and values among block threads\n * BlockRadixSort(temp_storage).Sort(thread_keys, thread_values);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [511,510,509,508], [11,10,9,8], [7,6,5,4], ..., [3,2,1,0] }.\n *\n */\n __device__ __forceinline__ void SortDescending(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< [in-out] Values to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n SortBlocked(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Sorting (blocked arrangement -> striped arrangement)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Performs an ascending radix sort across a [blocked arrangement](index.html#sec5sec3) of keys, leaving them in a [striped arrangement](index.html#sec5sec3).\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys that\n * are initially partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive keys. The final partitioning is striped.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * ...\n *\n * // Collectively sort the keys\n * BlockRadixSort(temp_storage).SortBlockedToStriped(thread_keys);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [0,128,256,384], [1,129,257,385], [2,130,258,386], ..., [127,255,383,511] }.\n *\n */\n __device__ __forceinline__ void SortBlockedToStriped(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n NullType values[ITEMS_PER_THREAD];\n\n SortBlockedToStriped(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n /**\n * \\brief Performs an ascending radix sort across a [blocked arrangement](index.html#sec5sec3) of keys and values, leaving them in a [striped arrangement](index.html#sec5sec3).\n *\n * \\par\n * - BlockRadixSort can only accommodate one associated tile of values. To \"truck along\"\n * more than one tile of values, simply perform a key-value sort of the keys paired\n * with a temporary value array that enumerates the key indices. The reordered indices\n * can then be used as a gather-vector for exchanging other associated tile data through\n * shared memory.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys and values that\n * are initially partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive pairs. The final partitioning is striped.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys and values each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * int thread_values[4];\n * ...\n *\n * // Collectively sort the keys and values among block threads\n * BlockRadixSort(temp_storage).SortBlockedToStriped(thread_keys, thread_values);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [0,128,256,384], [1,129,257,385], [2,130,258,386], ..., [127,255,383,511] }.\n *\n */\n __device__ __forceinline__ void SortBlockedToStriped(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< [in-out] Values to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n SortBlockedToStriped(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n /**\n * \\brief Performs a descending radix sort across a [blocked arrangement](index.html#sec5sec3) of keys, leaving them in a [striped arrangement](index.html#sec5sec3).\n *\n * \\par\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys that\n * are initially partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive keys. The final partitioning is striped.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * ...\n *\n * // Collectively sort the keys\n * BlockRadixSort(temp_storage).SortBlockedToStriped(thread_keys);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [511,383,255,127], [386,258,130,2], [385,257,128,1], ..., [384,256,128,0] }.\n *\n */\n __device__ __forceinline__ void SortDescendingBlockedToStriped(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n NullType values[ITEMS_PER_THREAD];\n\n SortBlockedToStriped(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n /**\n * \\brief Performs a descending radix sort across a [blocked arrangement](index.html#sec5sec3) of keys and values, leaving them in a [striped arrangement](index.html#sec5sec3).\n *\n * \\par\n * - BlockRadixSort can only accommodate one associated tile of values. To \"truck along\"\n * more than one tile of values, simply perform a key-value sort of the keys paired\n * with a temporary value array that enumerates the key indices. The reordered indices\n * can then be used as a gather-vector for exchanging other associated tile data through\n * shared memory.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sort of 512 integer keys and values that\n * are initially partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive pairs. The final partitioning is striped.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockRadixSort for a 1D block of 128 threads owning 4 integer keys and values each\n * typedef cub::BlockRadixSort BlockRadixSort;\n *\n * // Allocate shared memory for BlockRadixSort\n * __shared__ typename BlockRadixSort::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_keys[4];\n * int thread_values[4];\n * ...\n *\n * // Collectively sort the keys and values among block threads\n * BlockRadixSort(temp_storage).SortBlockedToStriped(thread_keys, thread_values);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_keys across the block of threads is\n * { [0,511,1,510], [2,509,3,508], [4,507,5,506], ..., [254,257,255,256] }. The\n * corresponding output \\p thread_keys in those threads will be\n * { [511,383,255,127], [386,258,130,2], [385,257,128,1], ..., [384,256,128,0] }.\n *\n */\n __device__ __forceinline__ void SortDescendingBlockedToStriped(\n KeyT (&keys)[ITEMS_PER_THREAD], ///< [in-out] Keys to sort\n ValueT (&values)[ITEMS_PER_THREAD], ///< [in-out] Values to sort\n int begin_bit = 0, ///< [in] [optional] The beginning (least-significant) bit index needed for key comparison\n int end_bit = sizeof(KeyT) * 8) ///< [in] [optional] The past-the-end (most-significant) bit index needed for key comparison\n {\n SortBlockedToStriped(keys, values, begin_bit, end_bit, Int2Type(), Int2Type());\n }\n\n\n //@} end member group\n\n};\n\n/**\n * \\example example_block_radix_sort.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_raking_layout.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockRakingLayout provides a conflict-free shared memory layout abstraction for warp-raking across thread block data.\n */\n\n\n#pragma once\n\n#include \"../util_macro.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief BlockRakingLayout provides a conflict-free shared memory layout abstraction for 1D raking across thread block data. ![](raking.png)\n * \\ingroup BlockModule\n *\n * \\par Overview\n * This type facilitates a shared memory usage pattern where a block of CUDA\n * threads places elements into shared memory and then reduces the active\n * parallelism to one \"raking\" warp of threads for serially aggregating consecutive\n * sequences of shared items. Padding is inserted to eliminate bank conflicts\n * (for most data types).\n *\n * \\tparam T The data type to be exchanged.\n * \\tparam BLOCK_THREADS The thread block size in threads.\n * \\tparam PTX_ARCH [optional] \\ptxversion\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int PTX_ARCH = CUB_PTX_ARCH>\nstruct BlockRakingLayout\n{\n //---------------------------------------------------------------------\n // Constants and type definitions\n //---------------------------------------------------------------------\n\n enum\n {\n /// The total number of elements that need to be cooperatively reduced\n SHARED_ELEMENTS = BLOCK_THREADS,\n\n /// Maximum number of warp-synchronous raking threads\n MAX_RAKING_THREADS = CUB_MIN(BLOCK_THREADS, CUB_WARP_THREADS(PTX_ARCH)),\n\n /// Number of raking elements per warp-synchronous raking thread (rounded up)\n SEGMENT_LENGTH = (SHARED_ELEMENTS + MAX_RAKING_THREADS - 1) / MAX_RAKING_THREADS,\n\n /// Never use a raking thread that will have no valid data (e.g., when BLOCK_THREADS is 62 and SEGMENT_LENGTH is 2, we should only use 31 raking threads)\n RAKING_THREADS = (SHARED_ELEMENTS + SEGMENT_LENGTH - 1) / SEGMENT_LENGTH,\n\n /// Whether we will have bank conflicts (technically we should find out if the GCD is > 1)\n HAS_CONFLICTS = (CUB_SMEM_BANKS(PTX_ARCH) % SEGMENT_LENGTH == 0),\n\n /// Degree of bank conflicts (e.g., 4-way)\n CONFLICT_DEGREE = (HAS_CONFLICTS) ?\n (MAX_RAKING_THREADS * SEGMENT_LENGTH) / CUB_SMEM_BANKS(PTX_ARCH) :\n 1,\n\n /// Pad each segment length with one element if segment length is not relatively prime to warp size and can't be optimized as a vector load\n USE_SEGMENT_PADDING = ((SEGMENT_LENGTH & 1) == 0) && (SEGMENT_LENGTH > 2),\n\n /// Total number of elements in the raking grid\n GRID_ELEMENTS = RAKING_THREADS * (SEGMENT_LENGTH + USE_SEGMENT_PADDING),\n\n /// Whether or not we need bounds checking during raking (the number of reduction elements is not a multiple of the number of raking threads)\n UNGUARDED = (SHARED_ELEMENTS % RAKING_THREADS == 0),\n };\n\n\n /**\n * \\brief Shared memory storage type\n */\n struct __align__(16) _TempStorage\n {\n T buff[BlockRakingLayout::GRID_ELEMENTS];\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /**\n * \\brief Returns the location for the calling thread to place data into the grid\n */\n static __device__ __forceinline__ T* PlacementPtr(\n TempStorage &temp_storage,\n unsigned int linear_tid)\n {\n // Offset for partial\n unsigned int offset = linear_tid;\n\n // Add in one padding element for every segment\n if (USE_SEGMENT_PADDING > 0)\n {\n offset += offset / SEGMENT_LENGTH;\n }\n\n // Incorporating a block of padding partials every shared memory segment\n return temp_storage.Alias().buff + offset;\n }\n\n\n /**\n * \\brief Returns the location for the calling thread to begin sequential raking\n */\n static __device__ __forceinline__ T* RakingPtr(\n TempStorage &temp_storage,\n unsigned int linear_tid)\n {\n return temp_storage.Alias().buff + (linear_tid * (SEGMENT_LENGTH + USE_SEGMENT_PADDING));\n }\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_reduce.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockReduce class provides [collective](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"specializations/block_reduce_raking.cuh\"\n#include \"specializations/block_reduce_raking_commutative_only.cuh\"\n#include \"specializations/block_reduce_warp_reductions.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_type.cuh\"\n#include \"../thread/thread_operators.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n\n/******************************************************************************\n * Algorithmic variants\n ******************************************************************************/\n\n/**\n * BlockReduceAlgorithm enumerates alternative algorithms for parallel\n * reduction across a CUDA thread block.\n */\nenum BlockReduceAlgorithm\n{\n\n /**\n * \\par Overview\n * An efficient \"raking\" reduction algorithm that only supports commutative\n * reduction operators (true for most operations, e.g., addition).\n *\n * \\par\n * Execution is comprised of three phases:\n * -# Upsweep sequential reduction in registers (if threads contribute more\n * than one input each). Threads in warps other than the first warp place\n * their partial reductions into shared memory.\n * -# Upsweep sequential reduction in shared memory. Threads within the first\n * warp continue to accumulate by raking across segments of shared partial reductions\n * -# A warp-synchronous Kogge-Stone style reduction within the raking warp.\n *\n * \\par\n * \\image html block_reduce.png\n *
\\p BLOCK_REDUCE_RAKING data flow for a hypothetical 16-thread thread block and 4-thread raking warp.
\n *\n * \\par Performance Considerations\n * - This variant performs less communication than BLOCK_REDUCE_RAKING_NON_COMMUTATIVE\n * and is preferable when the reduction operator is commutative. This variant\n * applies fewer reduction operators than BLOCK_REDUCE_WARP_REDUCTIONS, and can provide higher overall\n * throughput across the GPU when suitably occupied. However, turn-around latency may be\n * higher than to BLOCK_REDUCE_WARP_REDUCTIONS and thus less-desirable\n * when the GPU is under-occupied.\n */\n BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY,\n\n\n /**\n * \\par Overview\n * An efficient \"raking\" reduction algorithm that supports commutative\n * (e.g., addition) and non-commutative (e.g., string concatenation) reduction\n * operators. \\blocked.\n *\n * \\par\n * Execution is comprised of three phases:\n * -# Upsweep sequential reduction in registers (if threads contribute more\n * than one input each). Each thread then places the partial reduction\n * of its item(s) into shared memory.\n * -# Upsweep sequential reduction in shared memory. Threads within a\n * single warp rake across segments of shared partial reductions.\n * -# A warp-synchronous Kogge-Stone style reduction within the raking warp.\n *\n * \\par\n * \\image html block_reduce.png\n *
\\p BLOCK_REDUCE_RAKING data flow for a hypothetical 16-thread thread block and 4-thread raking warp.
\n *\n * \\par Performance Considerations\n * - This variant performs more communication than BLOCK_REDUCE_RAKING\n * and is only preferable when the reduction operator is non-commutative. This variant\n * applies fewer reduction operators than BLOCK_REDUCE_WARP_REDUCTIONS, and can provide higher overall\n * throughput across the GPU when suitably occupied. However, turn-around latency may be\n * higher than to BLOCK_REDUCE_WARP_REDUCTIONS and thus less-desirable\n * when the GPU is under-occupied.\n */\n BLOCK_REDUCE_RAKING,\n\n\n /**\n * \\par Overview\n * A quick \"tiled warp-reductions\" reduction algorithm that supports commutative\n * (e.g., addition) and non-commutative (e.g., string concatenation) reduction\n * operators.\n *\n * \\par\n * Execution is comprised of four phases:\n * -# Upsweep sequential reduction in registers (if threads contribute more\n * than one input each). Each thread then places the partial reduction\n * of its item(s) into shared memory.\n * -# Compute a shallow, but inefficient warp-synchronous Kogge-Stone style\n * reduction within each warp.\n * -# A propagation phase where the warp reduction outputs in each warp are\n * updated with the aggregate from each preceding warp.\n *\n * \\par\n * \\image html block_scan_warpscans.png\n *
\\p BLOCK_REDUCE_WARP_REDUCTIONS data flow for a hypothetical 16-thread thread block and 4-thread raking warp.
\n *\n * \\par Performance Considerations\n * - This variant applies more reduction operators than BLOCK_REDUCE_RAKING\n * or BLOCK_REDUCE_RAKING_NON_COMMUTATIVE, which may result in lower overall\n * throughput across the GPU. However turn-around latency may be lower and\n * thus useful when the GPU is under-occupied.\n */\n BLOCK_REDUCE_WARP_REDUCTIONS,\n};\n\n\n/******************************************************************************\n * Block reduce\n ******************************************************************************/\n\n/**\n * \\brief The BlockReduce class provides [collective](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread block. ![](reduce_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam T Data type being reduced\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ALGORITHM [optional] cub::BlockReduceAlgorithm enumerator specifying the underlying algorithm to use (default: cub::BLOCK_REDUCE_WARP_REDUCTIONS)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - A reduction (or fold)\n * uses a binary combining operator to compute a single aggregate from a list of input elements.\n * - \\rowmajor\n * - BlockReduce can be optionally specialized by algorithm to accommodate different latency/throughput workload profiles:\n * -# cub::BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY. An efficient \"raking\" reduction algorithm that only supports commutative reduction operators. [More...](\\ref cub::BlockReduceAlgorithm)\n * -# cub::BLOCK_REDUCE_RAKING. An efficient \"raking\" reduction algorithm that supports commutative and non-commutative reduction operators. [More...](\\ref cub::BlockReduceAlgorithm)\n * -# cub::BLOCK_REDUCE_WARP_REDUCTIONS. A quick \"tiled warp-reductions\" reduction algorithm that supports commutative and non-commutative reduction operators. [More...](\\ref cub::BlockReduceAlgorithm)\n *\n * \\par Performance Considerations\n * - \\granularity\n * - Very efficient (only one synchronization barrier).\n * - Incurs zero bank conflicts for most types\n * - Computation is slightly more efficient (i.e., having lower instruction overhead) for:\n * - Summation (vs. generic reduction)\n * - \\p BLOCK_THREADS is a multiple of the architecture's warp size\n * - Every thread has a valid input (i.e., full vs. partial-tiles)\n * - See cub::BlockReduceAlgorithm for performance details regarding algorithmic alternatives\n *\n * \\par A Simple Example\n * \\blockcollective{BlockReduce}\n * \\par\n * The code snippet below illustrates a sum reduction of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Compute the block-wide sum for thread0\n * int aggregate = BlockReduce(temp_storage).Sum(thread_data);\n *\n * \\endcode\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n BlockReduceAlgorithm ALGORITHM = BLOCK_REDUCE_WARP_REDUCTIONS,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockReduce\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n typedef BlockReduceWarpReductions WarpReductions;\n typedef BlockReduceRakingCommutativeOnly RakingCommutativeOnly;\n typedef BlockReduceRaking Raking;\n\n /// Internal specialization type\n typedef typename If<(ALGORITHM == BLOCK_REDUCE_WARP_REDUCTIONS),\n WarpReductions,\n typename If<(ALGORITHM == BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY),\n RakingCommutativeOnly,\n Raking>::Type>::Type InternalBlockReduce; // BlockReduceRaking\n\n /// Shared memory storage layout type for BlockReduce\n typedef typename InternalBlockReduce::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\npublic:\n\n /// \\smemstorage{BlockReduce}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockReduce()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockReduce(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Generic reductions\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using the specified binary reduction functor. Each thread contributes one input element.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a max reduction of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Each thread obtains an input item\n * int thread_data;\n * ...\n *\n * // Compute the block-wide max for thread0\n * int aggregate = BlockReduce(temp_storage).Reduce(thread_data, cub::Max());\n *\n * \\endcode\n *\n * \\tparam ReductionOp [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n ReductionOp reduction_op) ///< [in] Binary reduction functor \n {\n return InternalBlockReduce(temp_storage).template Reduce(input, BLOCK_THREADS, reduction_op);\n }\n\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using the specified binary reduction functor. Each thread contributes an array of consecutive input elements.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a max reduction of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Compute the block-wide max for thread0\n * int aggregate = BlockReduce(temp_storage).Reduce(thread_data, cub::Max());\n *\n * \\endcode\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ReductionOp [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T (&inputs)[ITEMS_PER_THREAD], ///< [in] Calling thread's input segment\n ReductionOp reduction_op) ///< [in] Binary reduction functor \n {\n // Reduce partials\n T partial = internal::ThreadReduce(inputs, reduction_op);\n return Reduce(partial, reduction_op);\n }\n\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using the specified binary reduction functor. The first \\p num_valid threads each contribute one input element.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a max reduction of a partially-full tile of integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int num_valid, ...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Each thread obtains an input item\n * int thread_data;\n * if (threadIdx.x < num_valid) thread_data = ...\n *\n * // Compute the block-wide max for thread0\n * int aggregate = BlockReduce(temp_storage).Reduce(thread_data, cub::Max(), num_valid);\n *\n * \\endcode\n *\n * \\tparam ReductionOp [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n ReductionOp reduction_op, ///< [in] Binary reduction functor \n int num_valid) ///< [in] Number of threads containing valid elements (may be less than BLOCK_THREADS)\n {\n // Determine if we scan skip bounds checking\n if (num_valid >= BLOCK_THREADS)\n {\n return InternalBlockReduce(temp_storage).template Reduce(input, num_valid, reduction_op);\n }\n else\n {\n return InternalBlockReduce(temp_storage).template Reduce(input, num_valid, reduction_op);\n }\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Summation reductions\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using addition (+) as the reduction operator. Each thread contributes one input element.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sum reduction of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Each thread obtains an input item\n * int thread_data;\n * ...\n *\n * // Compute the block-wide sum for thread0\n * int aggregate = BlockReduce(temp_storage).Sum(thread_data);\n *\n * \\endcode\n *\n */\n __device__ __forceinline__ T Sum(\n T input) ///< [in] Calling thread's input\n {\n return InternalBlockReduce(temp_storage).template Sum(input, BLOCK_THREADS);\n }\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using addition (+) as the reduction operator. Each thread contributes an array of consecutive input elements.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sum reduction of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Compute the block-wide sum for thread0\n * int aggregate = BlockReduce(temp_storage).Sum(thread_data);\n *\n * \\endcode\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n */\n template \n __device__ __forceinline__ T Sum(\n T (&inputs)[ITEMS_PER_THREAD]) ///< [in] Calling thread's input segment\n {\n // Reduce partials\n T partial = internal::ThreadReduce(inputs, cub::Sum());\n return Sum(partial);\n }\n\n\n /**\n * \\brief Computes a block-wide reduction for thread0 using addition (+) as the reduction operator. The first \\p num_valid threads each contribute one input element.\n *\n * \\par\n * - The return value is undefined in threads other than thread0.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sum reduction of a partially-full tile of integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int num_valid, ...)\n * {\n * // Specialize BlockReduce for a 1D block of 128 threads on type int\n * typedef cub::BlockReduce BlockReduce;\n *\n * // Allocate shared memory for BlockReduce\n * __shared__ typename BlockReduce::TempStorage temp_storage;\n *\n * // Each thread obtains an input item (up to num_items)\n * int thread_data;\n * if (threadIdx.x < num_valid)\n * thread_data = ...\n *\n * // Compute the block-wide sum for thread0\n * int aggregate = BlockReduce(temp_storage).Sum(thread_data, num_valid);\n *\n * \\endcode\n *\n */\n __device__ __forceinline__ T Sum(\n T input, ///< [in] Calling thread's input\n int num_valid) ///< [in] Number of threads containing valid elements (may be less than BLOCK_THREADS)\n {\n // Determine if we scan skip bounds checking\n if (num_valid >= BLOCK_THREADS)\n {\n return InternalBlockReduce(temp_storage).template Sum(input, num_valid);\n }\n else\n {\n return InternalBlockReduce(temp_storage).template Sum(input, num_valid);\n }\n }\n\n\n //@} end member group\n};\n\n/**\n * \\example example_block_reduce.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_scan.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockScan class provides [collective](index.html#sec0) methods for computing a parallel prefix sum/scan of items partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"specializations/block_scan_raking.cuh\"\n#include \"specializations/block_scan_warp_scans.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Algorithmic variants\n ******************************************************************************/\n\n/**\n * \\brief BlockScanAlgorithm enumerates alternative algorithms for cub::BlockScan to compute a parallel prefix scan across a CUDA thread block.\n */\nenum BlockScanAlgorithm\n{\n\n /**\n * \\par Overview\n * An efficient \"raking reduce-then-scan\" prefix scan algorithm. Execution is comprised of five phases:\n * -# Upsweep sequential reduction in registers (if threads contribute more than one input each). Each thread then places the partial reduction of its item(s) into shared memory.\n * -# Upsweep sequential reduction in shared memory. Threads within a single warp rake across segments of shared partial reductions.\n * -# A warp-synchronous Kogge-Stone style exclusive scan within the raking warp.\n * -# Downsweep sequential exclusive scan in shared memory. Threads within a single warp rake across segments of shared partial reductions, seeded with the warp-scan output.\n * -# Downsweep sequential scan in registers (if threads contribute more than one input), seeded with the raking scan output.\n *\n * \\par\n * \\image html block_scan_raking.png\n *
\\p BLOCK_SCAN_RAKING data flow for a hypothetical 16-thread thread block and 4-thread raking warp.
\n *\n * \\par Performance Considerations\n * - Although this variant may suffer longer turnaround latencies when the\n * GPU is under-occupied, it can often provide higher overall throughput\n * across the GPU when suitably occupied.\n */\n BLOCK_SCAN_RAKING,\n\n\n /**\n * \\par Overview\n * Similar to cub::BLOCK_SCAN_RAKING, but with fewer shared memory reads at\n * the expense of higher register pressure. Raking threads preserve their\n * \"upsweep\" segment of values in registers while performing warp-synchronous\n * scan, allowing the \"downsweep\" not to re-read them from shared memory.\n */\n BLOCK_SCAN_RAKING_MEMOIZE,\n\n\n /**\n * \\par Overview\n * A quick \"tiled warpscans\" prefix scan algorithm. Execution is comprised of four phases:\n * -# Upsweep sequential reduction in registers (if threads contribute more than one input each). Each thread then places the partial reduction of its item(s) into shared memory.\n * -# Compute a shallow, but inefficient warp-synchronous Kogge-Stone style scan within each warp.\n * -# A propagation phase where the warp scan outputs in each warp are updated with the aggregate from each preceding warp.\n * -# Downsweep sequential scan in registers (if threads contribute more than one input), seeded with the raking scan output.\n *\n * \\par\n * \\image html block_scan_warpscans.png\n *
\\p BLOCK_SCAN_WARP_SCANS data flow for a hypothetical 16-thread thread block and 4-thread raking warp.
\n *\n * \\par Performance Considerations\n * - Although this variant may suffer lower overall throughput across the\n * GPU because due to a heavy reliance on inefficient warpscans, it can\n * often provide lower turnaround latencies when the GPU is under-occupied.\n */\n BLOCK_SCAN_WARP_SCANS,\n};\n\n\n/******************************************************************************\n * Block scan\n ******************************************************************************/\n\n/**\n * \\brief The BlockScan class provides [collective](index.html#sec0) methods for computing a parallel prefix sum/scan of items partitioned across a CUDA thread block. ![](block_scan_logo.png)\n * \\ingroup BlockModule\n *\n * \\tparam T Data type being scanned\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ALGORITHM [optional] cub::BlockScanAlgorithm enumerator specifying the underlying algorithm to use (default: cub::BLOCK_SCAN_RAKING)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - Given a list of input elements and a binary reduction operator, a [prefix scan](http://en.wikipedia.org/wiki/Prefix_sum)\n * produces an output list where each element is computed to be the reduction\n * of the elements occurring earlier in the input list. Prefix sum\n * connotes a prefix scan with the addition operator. The term \\em inclusive indicates\n * that the ith output reduction incorporates the ith input.\n * The term \\em exclusive indicates the ith input is not incorporated into\n * the ith output reduction.\n * - \\rowmajor\n * - BlockScan can be optionally specialized by algorithm to accommodate different workload profiles:\n * -# cub::BLOCK_SCAN_RAKING. An efficient (high throughput) \"raking reduce-then-scan\" prefix scan algorithm. [More...](\\ref cub::BlockScanAlgorithm)\n * -# cub::BLOCK_SCAN_RAKING_MEMOIZE. Similar to cub::BLOCK_SCAN_RAKING, but having higher throughput at the expense of additional register pressure for intermediate storage. [More...](\\ref cub::BlockScanAlgorithm)\n * -# cub::BLOCK_SCAN_WARP_SCANS. A quick (low latency) \"tiled warpscans\" prefix scan algorithm. [More...](\\ref cub::BlockScanAlgorithm)\n *\n * \\par Performance Considerations\n * - \\granularity\n * - Uses special instructions when applicable (e.g., warp \\p SHFL)\n * - Uses synchronization-free communication between warp lanes when applicable\n * - Invokes a minimal number of minimal block-wide synchronization barriers (only\n * one or two depending on algorithm selection)\n * - Incurs zero bank conflicts for most types\n * - Computation is slightly more efficient (i.e., having lower instruction overhead) for:\n * - Prefix sum variants (vs. generic scan)\n * - \\blocksize\n * - See cub::BlockScanAlgorithm for performance details regarding algorithmic alternatives\n *\n * \\par A Simple Example\n * \\blockcollective{BlockScan}\n * \\par\n * The code snippet below illustrates an exclusive prefix sum of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * BlockScan(temp_storage).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * {[1,1,1,1], [1,1,1,1], ..., [1,1,1,1]}.\n * The corresponding output \\p thread_data in those threads will be\n * {[0,1,2,3], [4,5,6,7], ..., [508,509,510,511]}.\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n BlockScanAlgorithm ALGORITHM = BLOCK_SCAN_RAKING,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockScan\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n /**\n * Ensure the template parameterization meets the requirements of the\n * specified algorithm. Currently, the BLOCK_SCAN_WARP_SCANS policy\n * cannot be used with thread block sizes not a multiple of the\n * architectural warp size.\n */\n static const BlockScanAlgorithm SAFE_ALGORITHM =\n ((ALGORITHM == BLOCK_SCAN_WARP_SCANS) && (BLOCK_THREADS % CUB_WARP_THREADS(PTX_ARCH) != 0)) ?\n BLOCK_SCAN_RAKING :\n ALGORITHM;\n\n typedef BlockScanWarpScans WarpScans;\n typedef BlockScanRaking Raking;\n\n /// Define the delegate type for the desired algorithm\n typedef typename If<(SAFE_ALGORITHM == BLOCK_SCAN_WARP_SCANS),\n WarpScans,\n Raking>::Type InternalBlockScan;\n\n /// Shared memory storage layout type for BlockScan\n typedef typename InternalBlockScan::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /******************************************************************************\n * Public types\n ******************************************************************************/\npublic:\n\n /// \\smemstorage{BlockScan}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockScan()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockScan(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Exclusive prefix sum operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element. The value of 0 is applied as the initial value, and is assigned to \\p output in thread0.\n *\n * \\par\n * - \\identityzero\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix sum of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * BlockScan(temp_storage).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 1, 1, ..., 1. The\n * corresponding output \\p thread_data in those threads will be 0, 1, ..., 127.\n *\n */\n __device__ __forceinline__ void ExclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output) ///< [out] Calling thread's output item (may be aliased to \\p input)\n {\n T initial_value = 0;\n ExclusiveScan(input, output, initial_value, cub::Sum());\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element. The value of 0 is applied as the initial value, and is assigned to \\p output in thread0. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\identityzero\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix sum of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * int block_aggregate;\n * BlockScan(temp_storage).ExclusiveSum(thread_data, thread_data, block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 1, 1, ..., 1. The\n * corresponding output \\p thread_data in those threads will be 0, 1, ..., 127.\n * Furthermore the value \\p 128 will be stored in \\p block_aggregate for all threads.\n *\n */\n __device__ __forceinline__ void ExclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n T initial_value = 0;\n ExclusiveScan(input, output, initial_value, cub::Sum(), block_aggregate);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element. Instead of using 0 as the block-wide prefix, the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\identityzero\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an exclusive prefix sum over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total += block_aggregate;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data = d_data[block_offset];\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * BlockScan(temp_storage).ExclusiveSum(\n * thread_data, thread_data, prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * d_data[block_offset] = thread_data;\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 1, 1, 1, 1, 1, 1, 1, 1, ....\n * The corresponding output for the first segment will be 0, 1, ..., 127.\n * The output for the second segment will be 128, 129, ..., 255.\n *\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template \n __device__ __forceinline__ void ExclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n ExclusiveScan(input, output, cub::Sum(), block_prefix_callback_op);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Exclusive prefix sum operations (multiple data per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements. The value of 0 is applied as the initial value, and is assigned to \\p output[0] in thread0.\n *\n * \\par\n * - \\identityzero\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix sum of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * BlockScan(temp_storage).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is { [1,1,1,1], [1,1,1,1], ..., [1,1,1,1] }. The\n * corresponding output \\p thread_data in those threads will be { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n */\n template \n __device__ __forceinline__ void ExclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD]) ///< [out] Calling thread's output items (may be aliased to \\p input)\n {\n T initial_value = 0;\n ExclusiveScan(input, output, initial_value, cub::Sum());\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements. The value of 0 is applied as the initial value, and is assigned to \\p output[0] in thread0. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\identityzero\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix sum of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * int block_aggregate;\n * BlockScan(temp_storage).ExclusiveSum(thread_data, thread_data, block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is { [1,1,1,1], [1,1,1,1], ..., [1,1,1,1] }. The\n * corresponding output \\p thread_data in those threads will be { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n * Furthermore the value \\p 512 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n */\n template \n __device__ __forceinline__ void ExclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n // Reduce consecutive thread items in registers\n T initial_value = 0;\n ExclusiveScan(input, output, initial_value, cub::Sum(), block_aggregate);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements. Instead of using 0 as the block-wide prefix, the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\identityzero\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an exclusive prefix sum over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 512 integer items that are partitioned in a [blocked arrangement](index.html#sec5sec3)\n * across 128 threads where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total += block_aggregate;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockLoad, BlockStore, and BlockScan for a 1D block of 128 threads, 4 ints per thread\n * typedef cub::BlockLoad BlockLoad;\n * typedef cub::BlockStore BlockStore;\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate aliased shared memory for BlockLoad, BlockStore, and BlockScan\n * __shared__ union {\n * typename BlockLoad::TempStorage load;\n * typename BlockScan::TempStorage scan;\n * typename BlockStore::TempStorage store;\n * } temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128 * 4)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage.load).Load(d_data + block_offset, thread_data);\n * CTA_SYNC();\n *\n * // Collectively compute the block-wide exclusive prefix sum\n * int block_aggregate;\n * BlockScan(temp_storage.scan).ExclusiveSum(\n * thread_data, thread_data, prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * BlockStore(temp_storage.store).Store(d_data + block_offset, thread_data);\n * CTA_SYNC();\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 1, 1, 1, 1, 1, 1, 1, 1, ....\n * The corresponding output for the first segment will be 0, 1, 2, 3, ..., 510, 511.\n * The output for the second segment will be 512, 513, 514, 515, ..., 1022, 1023.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n int ITEMS_PER_THREAD,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n ExclusiveScan(input, output, cub::Sum(), block_prefix_callback_op);\n }\n\n\n\n //@} end member group // Exclusive prefix sums\n /******************************************************************//**\n * \\name Exclusive prefix scan operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix max scan of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * BlockScan(temp_storage).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 0, -1, 2, -3, ..., 126, -127. The\n * corresponding output \\p thread_data in those threads will be INT_MIN, 0, 0, 2, ..., 124, 126.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n T initial_value, ///< [in] Initial value to seed the exclusive scan (and is assigned to \\p output[0] in thread0)\n ScanOp scan_op) ///< [in] Binary scan functor \n {\n InternalBlockScan(temp_storage).ExclusiveScan(input, output, initial_value, scan_op);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix max scan of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * int block_aggregate;\n * BlockScan(temp_storage).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max(), block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 0, -1, 2, -3, ..., 126, -127. The\n * corresponding output \\p thread_data in those threads will be INT_MIN, 0, 0, 2, ..., 124, 126.\n * Furthermore the value \\p 126 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n T initial_value, ///< [in] Initial value to seed the exclusive scan (and is assigned to \\p output[0] in thread0)\n ScanOp scan_op, ///< [in] Binary scan functor \n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n InternalBlockScan(temp_storage).ExclusiveScan(input, output, initial_value, scan_op, block_aggregate);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an exclusive prefix max scan over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total = (block_aggregate > old_prefix) ? block_aggregate : old_prefix;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(INT_MIN);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data = d_data[block_offset];\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * BlockScan(temp_storage).ExclusiveScan(\n * thread_data, thread_data, INT_MIN, cub::Max(), prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * d_data[block_offset] = thread_data;\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, -1, 2, -3, 4, -5, ....\n * The corresponding output for the first segment will be INT_MIN, 0, 0, 2, ..., 124, 126.\n * The output for the second segment will be 126, 128, 128, 130, ..., 252, 254.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor \n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n InternalBlockScan(temp_storage).ExclusiveScan(input, output, scan_op, block_prefix_callback_op);\n }\n\n\n //@} end member group // Inclusive prefix sums\n /******************************************************************//**\n * \\name Exclusive prefix scan operations (multiple data per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix max scan of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * BlockScan(temp_storage).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,-1,2,-3], [4,-5,6,-7], ..., [508,-509,510,-511] }.\n * The corresponding output \\p thread_data in those threads will be\n * { [INT_MIN,0,0,2], [2,4,4,6], ..., [506,508,508,510] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void ExclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n T initial_value, ///< [in] Initial value to seed the exclusive scan (and is assigned to \\p output[0] in thread0)\n ScanOp scan_op) ///< [in] Binary scan functor\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_prefix, thread_prefix, initial_value, scan_op);\n\n // Exclusive scan in registers with prefix as seed\n internal::ThreadScanExclusive(input, output, scan_op, thread_prefix);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an exclusive prefix max scan of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * int block_aggregate;\n * BlockScan(temp_storage).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max(), block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is { [0,-1,2,-3], [4,-5,6,-7], ..., [508,-509,510,-511] }. The\n * corresponding output \\p thread_data in those threads will be { [INT_MIN,0,0,2], [2,4,4,6], ..., [506,508,508,510] }.\n * Furthermore the value \\p 510 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void ExclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n T initial_value, ///< [in] Initial value to seed the exclusive scan (and is assigned to \\p output[0] in thread0)\n ScanOp scan_op, ///< [in] Binary scan functor\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_prefix, thread_prefix, initial_value, scan_op, block_aggregate);\n\n // Exclusive scan in registers with prefix as seed\n internal::ThreadScanExclusive(input, output, scan_op, thread_prefix);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an exclusive prefix max scan over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total = (block_aggregate > old_prefix) ? block_aggregate : old_prefix;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockLoad, BlockStore, and BlockScan for a 1D block of 128 threads, 4 ints per thread\n * typedef cub::BlockLoad BlockLoad;\n * typedef cub::BlockStore BlockStore;\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate aliased shared memory for BlockLoad, BlockStore, and BlockScan\n * __shared__ union {\n * typename BlockLoad::TempStorage load;\n * typename BlockScan::TempStorage scan;\n * typename BlockStore::TempStorage store;\n * } temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128 * 4)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage.load).Load(d_data + block_offset, thread_data);\n * CTA_SYNC();\n *\n * // Collectively compute the block-wide exclusive prefix max scan\n * BlockScan(temp_storage.scan).ExclusiveScan(\n * thread_data, thread_data, INT_MIN, cub::Max(), prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * BlockStore(temp_storage.store).Store(d_data + block_offset, thread_data);\n * CTA_SYNC();\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, -1, 2, -3, 4, -5, ....\n * The corresponding output for the first segment will be INT_MIN, 0, 0, 2, 2, 4, ..., 508, 510.\n * The output for the second segment will be 510, 512, 512, 514, 514, 516, ..., 1020, 1022.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_prefix, thread_prefix, scan_op, block_prefix_callback_op);\n\n // Exclusive scan in registers with prefix as seed\n internal::ThreadScanExclusive(input, output, scan_op, thread_prefix);\n }\n\n\n //@} end member group\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document no-initial-value scans\n\n /******************************************************************//**\n * \\name Exclusive prefix scan operations (no initial value, single datum per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for thread0 is undefined.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan functor\n {\n InternalBlockScan(temp_storage).ExclusiveScan(input, output, scan_op);\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n InternalBlockScan(temp_storage).ExclusiveScan(input, output, scan_op, block_aggregate);\n }\n\n //@} end member group\n /******************************************************************//**\n * \\name Exclusive prefix scan operations (no initial value, multiple data per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. With no initial value, the output computed for thread0 is undefined.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void ExclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan functor\n {\n // Reduce consecutive thread items in registers\n T thread_partial = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_partial, thread_partial, scan_op);\n\n // Exclusive scan in registers with prefix\n internal::ThreadScanExclusive(input, output, scan_op, thread_partial, (linear_tid != 0));\n }\n\n\n /**\n * \\brief Computes an exclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void ExclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n // Reduce consecutive thread items in registers\n T thread_partial = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_partial, thread_partial, scan_op, block_aggregate);\n\n // Exclusive scan in registers with prefix\n internal::ThreadScanExclusive(input, output, scan_op, thread_partial, (linear_tid != 0));\n }\n\n\n //@} end member group\n#endif // DOXYGEN_SHOULD_SKIP_THIS // Do not document no-initial-value scans\n\n /******************************************************************//**\n * \\name Inclusive prefix sum operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element.\n *\n * \\par\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix sum of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * BlockScan(temp_storage).InclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 1, 1, ..., 1. The\n * corresponding output \\p thread_data in those threads will be 1, 2, ..., 128.\n *\n */\n __device__ __forceinline__ void InclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output) ///< [out] Calling thread's output item (may be aliased to \\p input)\n {\n InclusiveScan(input, output, cub::Sum());\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix sum of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * int block_aggregate;\n * BlockScan(temp_storage).InclusiveSum(thread_data, thread_data, block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 1, 1, ..., 1. The\n * corresponding output \\p thread_data in those threads will be 1, 2, ..., 128.\n * Furthermore the value \\p 128 will be stored in \\p block_aggregate for all threads.\n *\n */\n __device__ __forceinline__ void InclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n InclusiveScan(input, output, cub::Sum(), block_aggregate);\n }\n\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes one input element. Instead of using 0 as the block-wide prefix, the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an inclusive prefix sum over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total += block_aggregate;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data = d_data[block_offset];\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * BlockScan(temp_storage).InclusiveSum(\n * thread_data, thread_data, prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * d_data[block_offset] = thread_data;\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 1, 1, 1, 1, 1, 1, 1, 1, ....\n * The corresponding output for the first segment will be 1, 2, ..., 128.\n * The output for the second segment will be 129, 130, ..., 256.\n *\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template \n __device__ __forceinline__ void InclusiveSum(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n InclusiveScan(input, output, cub::Sum(), block_prefix_callback_op);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive prefix sum operations (multiple data per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements.\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix sum of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * BlockScan(temp_storage).InclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is { [1,1,1,1], [1,1,1,1], ..., [1,1,1,1] }. The\n * corresponding output \\p thread_data in those threads will be { [1,2,3,4], [5,6,7,8], ..., [509,510,511,512] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n */\n template \n __device__ __forceinline__ void InclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD]) ///< [out] Calling thread's output items (may be aliased to \\p input)\n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveSum(input[0], output[0]);\n }\n else\n {\n // Reduce consecutive thread items in registers\n Sum scan_op;\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveSum(thread_prefix, thread_prefix);\n\n // Inclusive scan in registers with prefix as seed\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix, (linear_tid != 0));\n }\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix sum of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * int block_aggregate;\n * BlockScan(temp_storage).InclusiveSum(thread_data, thread_data, block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [1,1,1,1], [1,1,1,1], ..., [1,1,1,1] }. The\n * corresponding output \\p thread_data in those threads will be\n * { [1,2,3,4], [5,6,7,8], ..., [509,510,511,512] }.\n * Furthermore the value \\p 512 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void InclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveSum(input[0], output[0], block_aggregate);\n }\n else\n {\n // Reduce consecutive thread items in registers\n Sum scan_op;\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveSum(thread_prefix, thread_prefix, block_aggregate);\n\n // Inclusive scan in registers with prefix as seed\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix, (linear_tid != 0));\n }\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using addition (+) as the scan operator. Each thread contributes an array of consecutive input elements. Instead of using 0 as the block-wide prefix, the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an inclusive prefix sum over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 512 integer items that are partitioned in a [blocked arrangement](index.html#sec5sec3)\n * across 128 threads where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total += block_aggregate;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockLoad, BlockStore, and BlockScan for a 1D block of 128 threads, 4 ints per thread\n * typedef cub::BlockLoad BlockLoad;\n * typedef cub::BlockStore BlockStore;\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate aliased shared memory for BlockLoad, BlockStore, and BlockScan\n * __shared__ union {\n * typename BlockLoad::TempStorage load;\n * typename BlockScan::TempStorage scan;\n * typename BlockStore::TempStorage store;\n * } temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128 * 4)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage.load).Load(d_data + block_offset, thread_data);\n * CTA_SYNC();\n *\n * // Collectively compute the block-wide inclusive prefix sum\n * BlockScan(temp_storage.scan).IncluisveSum(\n * thread_data, thread_data, prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * BlockStore(temp_storage.store).Store(d_data + block_offset, thread_data);\n * CTA_SYNC();\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 1, 1, 1, 1, 1, 1, 1, 1, ....\n * The corresponding output for the first segment will be 1, 2, 3, 4, ..., 511, 512.\n * The output for the second segment will be 513, 514, 515, 516, ..., 1023, 1024.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n int ITEMS_PER_THREAD,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveSum(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveSum(input[0], output[0], block_prefix_callback_op);\n }\n else\n {\n // Reduce consecutive thread items in registers\n Sum scan_op;\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveSum(thread_prefix, thread_prefix, block_prefix_callback_op);\n\n // Inclusive scan in registers with prefix as seed\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix);\n }\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive prefix scan operations\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix max scan of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * BlockScan(temp_storage).InclusiveScan(thread_data, thread_data, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 0, -1, 2, -3, ..., 126, -127. The\n * corresponding output \\p thread_data in those threads will be 0, 0, 2, 2, ..., 126, 126.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan functor \n {\n InternalBlockScan(temp_storage).InclusiveScan(input, output, scan_op);\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix max scan of 128 integer items that\n * are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain input item for each thread\n * int thread_data;\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * int block_aggregate;\n * BlockScan(temp_storage).InclusiveScan(thread_data, thread_data, cub::Max(), block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is 0, -1, 2, -3, ..., 126, -127. The\n * corresponding output \\p thread_data in those threads will be 0, 0, 2, 2, ..., 126, 126.\n * Furthermore the value \\p 126 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor \n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n InternalBlockScan(temp_storage).InclusiveScan(input, output, scan_op, block_aggregate);\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - Supports non-commutative scan operators.\n * - \\rowmajor\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an inclusive prefix max scan over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total = (block_aggregate > old_prefix) ? block_aggregate : old_prefix;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(INT_MIN);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data = d_data[block_offset];\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * BlockScan(temp_storage).InclusiveScan(\n * thread_data, thread_data, cub::Max(), prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * d_data[block_offset] = thread_data;\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, -1, 2, -3, 4, -5, ....\n * The corresponding output for the first segment will be 0, 0, 2, 2, ..., 126, 126.\n * The output for the second segment will be 128, 128, 130, 130, ..., 254, 254.\n *\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor \n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n InternalBlockScan(temp_storage).InclusiveScan(input, output, scan_op, block_prefix_callback_op);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive prefix scan operations (multiple data per thread)\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix max scan of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * BlockScan(temp_storage).InclusiveScan(thread_data, thread_data, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is { [0,-1,2,-3], [4,-5,6,-7], ..., [508,-509,510,-511] }. The\n * corresponding output \\p thread_data in those threads will be { [0,0,2,2], [4,4,6,6], ..., [508,508,510,510] }.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void InclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan functor \n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveScan(input[0], output[0], scan_op);\n }\n else\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_prefix, thread_prefix, scan_op);\n\n // Inclusive scan in registers with prefix as seed (first thread does not seed)\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix, (linear_tid != 0));\n }\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates an inclusive prefix max scan of 512 integer items that\n * are partitioned in a [blocked arrangement](index.html#sec5sec3) across 128 threads\n * where each thread owns 4 consecutive items.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize BlockScan for a 1D block of 128 threads on type int\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate shared memory for BlockScan\n * __shared__ typename BlockScan::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * int block_aggregate;\n * BlockScan(temp_storage).InclusiveScan(thread_data, thread_data, cub::Max(), block_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is\n * { [0,-1,2,-3], [4,-5,6,-7], ..., [508,-509,510,-511] }.\n * The corresponding output \\p thread_data in those threads will be\n * { [0,0,2,2], [4,4,6,6], ..., [508,508,510,510] }.\n * Furthermore the value \\p 510 will be stored in \\p block_aggregate for all threads.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp>\n __device__ __forceinline__ void InclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor \n T &block_aggregate) ///< [out] block-wide aggregate reduction of input items\n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveScan(input[0], output[0], scan_op, block_aggregate);\n }\n else\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan (with no initial value)\n ExclusiveScan(thread_prefix, thread_prefix, scan_op, block_aggregate);\n\n // Inclusive scan in registers with prefix as seed (first thread does not seed)\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix, (linear_tid != 0));\n }\n }\n\n\n /**\n * \\brief Computes an inclusive block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes an array of consecutive input elements. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n *\n * \\par\n * - The \\p block_prefix_callback_op functor must implement a member function T operator()(T block_aggregate).\n * The functor's input parameter \\p block_aggregate is the same value also returned by the scan operation.\n * The functor will be invoked by the first warp of threads in the block, however only the return value from\n * lane0 is applied as the block-wide prefix. Can be stateful.\n * - Supports non-commutative scan operators.\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a single thread block that progressively\n * computes an inclusive prefix max scan over multiple \"tiles\" of input using a\n * prefix functor to maintain a running total between block-wide scans. Each tile consists\n * of 128 integer items that are partitioned across 128 threads.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // A stateful callback functor that maintains a running prefix to be applied\n * // during consecutive scan operations.\n * struct BlockPrefixCallbackOp\n * {\n * // Running prefix\n * int running_total;\n *\n * // Constructor\n * __device__ BlockPrefixCallbackOp(int running_total) : running_total(running_total) {}\n *\n * // Callback operator to be entered by the first warp of threads in the block.\n * // Thread-0 is responsible for returning a value for seeding the block-wide scan.\n * __device__ int operator()(int block_aggregate)\n * {\n * int old_prefix = running_total;\n * running_total = (block_aggregate > old_prefix) ? block_aggregate : old_prefix;\n * return old_prefix;\n * }\n * };\n *\n * __global__ void ExampleKernel(int *d_data, int num_items, ...)\n * {\n * // Specialize BlockLoad, BlockStore, and BlockScan for a 1D block of 128 threads, 4 ints per thread\n * typedef cub::BlockLoad BlockLoad;\n * typedef cub::BlockStore BlockStore;\n * typedef cub::BlockScan BlockScan;\n *\n * // Allocate aliased shared memory for BlockLoad, BlockStore, and BlockScan\n * __shared__ union {\n * typename BlockLoad::TempStorage load;\n * typename BlockScan::TempStorage scan;\n * typename BlockStore::TempStorage store;\n * } temp_storage;\n *\n * // Initialize running total\n * BlockPrefixCallbackOp prefix_op(0);\n *\n * // Have the block iterate over segments of items\n * for (int block_offset = 0; block_offset < num_items; block_offset += 128 * 4)\n * {\n * // Load a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * BlockLoad(temp_storage.load).Load(d_data + block_offset, thread_data);\n * CTA_SYNC();\n *\n * // Collectively compute the block-wide inclusive prefix max scan\n * BlockScan(temp_storage.scan).InclusiveScan(\n * thread_data, thread_data, cub::Max(), prefix_op);\n * CTA_SYNC();\n *\n * // Store scanned items to output segment\n * BlockStore(temp_storage.store).Store(d_data + block_offset, thread_data);\n * CTA_SYNC();\n * }\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is 0, -1, 2, -3, 4, -5, ....\n * The corresponding output for the first segment will be 0, 0, 2, 2, 4, 4, ..., 510, 510.\n * The output for the second segment will be 512, 512, 514, 514, 516, 516, ..., 1022, 1022.\n *\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n * \\tparam BlockPrefixCallbackOp [inferred] Call-back functor type having member T operator()(T block_aggregate)\n */\n template <\n int ITEMS_PER_THREAD,\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T (&input)[ITEMS_PER_THREAD], ///< [in] Calling thread's input items\n T (&output)[ITEMS_PER_THREAD], ///< [out] Calling thread's output items (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan functor \n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a block-wide prefix to be applied to the logical input sequence.\n {\n if (ITEMS_PER_THREAD == 1)\n {\n InclusiveScan(input[0], output[0], scan_op, block_prefix_callback_op);\n }\n else\n {\n // Reduce consecutive thread items in registers\n T thread_prefix = internal::ThreadReduce(input, scan_op);\n\n // Exclusive thread block-scan\n ExclusiveScan(thread_prefix, thread_prefix, scan_op, block_prefix_callback_op);\n\n // Inclusive scan in registers with prefix as seed\n internal::ThreadScanInclusive(input, output, scan_op, thread_prefix);\n }\n }\n\n //@} end member group\n\n\n};\n\n/**\n * \\example example_block_scan.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_shuffle.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockShuffle class provides [collective](index.html#sec0) methods for shuffling data partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../util_arch.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_macro.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief The BlockShuffle class provides [collective](index.html#sec0) methods for shuffling data partitioned across a CUDA thread block.\n * \\ingroup BlockModule\n *\n * \\tparam T The data type to be exchanged.\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * It is commonplace for blocks of threads to rearrange data items between\n * threads. The BlockShuffle abstraction allows threads to efficiently shift items\n * either (a) up to their successor or (b) down to their predecessor.\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockShuffle\n{\nprivate:\n\n /******************************************************************************\n * Constants\n ******************************************************************************/\n\n enum\n {\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n LOG_WARP_THREADS = CUB_LOG_WARP_THREADS(PTX_ARCH),\n WARP_THREADS = 1 << LOG_WARP_THREADS,\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n };\n\n /******************************************************************************\n * Type definitions\n ******************************************************************************/\n\n /// Shared memory storage layout type (last element from each thread's input)\n struct _TempStorage\n {\n T prev[BLOCK_THREADS];\n T next[BLOCK_THREADS];\n };\n\n\npublic:\n\n /// \\smemstorage{BlockShuffle}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\nprivate:\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n unsigned int linear_tid;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\npublic:\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockShuffle()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockShuffle(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Shuffle movement\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Each threadi obtains the \\p input provided by threadi+distance. The offset \\p distance may be negative.\n *\n * \\par\n * - \\smemreuse\n */\n __device__ __forceinline__ void Offset(\n T input, ///< [in] The input item from the calling thread (threadi)\n T& output, ///< [out] The \\p input item from the successor (or predecessor) thread threadi+distance (may be aliased to \\p input). This value is only updated for for threadi when 0 <= (i + \\p distance) < BLOCK_THREADS-1\n int distance = 1) ///< [in] Offset distance (may be negative)\n {\n temp_storage[linear_tid].prev = input;\n\n CTA_SYNC();\n\n if ((linear_tid + distance >= 0) && (linear_tid + distance < BLOCK_THREADS))\n output = temp_storage[linear_tid + distance].prev;\n }\n\n\n /**\n * \\brief Each threadi obtains the \\p input provided by threadi+distance.\n *\n * \\par\n * - \\smemreuse\n */\n __device__ __forceinline__ void Rotate(\n T input, ///< [in] The calling thread's input item\n T& output, ///< [out] The \\p input item from thread thread(i+distance>)% (may be aliased to \\p input). This value is not updated for threadBLOCK_THREADS-1\n unsigned int distance = 1) ///< [in] Offset distance (0 < \\p distance < BLOCK_THREADS)\n {\n temp_storage[linear_tid].prev = input;\n\n CTA_SYNC();\n\n unsigned int offset = threadIdx.x + distance;\n if (offset >= BLOCK_THREADS)\n offset -= BLOCK_THREADS;\n\n output = temp_storage[offset].prev;\n }\n\n\n /**\n * \\brief The thread block rotates its [blocked arrangement](index.html#sec5sec3) of \\p input items, shifting it up by one item\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n */\n template \n __device__ __forceinline__ void Up(\n T (&input)[ITEMS_PER_THREAD], ///< [in] The calling thread's input items\n T (&prev)[ITEMS_PER_THREAD]) ///< [out] The corresponding predecessor items (may be aliased to \\p input). The item \\p prev[0] is not updated for thread0.\n {\n temp_storage[linear_tid].prev = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = ITEMS_PER_THREAD - 1; ITEM > 0; --ITEM)\n prev[ITEM] = input[ITEM - 1];\n\n\n if (linear_tid > 0)\n prev[0] = temp_storage[linear_tid - 1].prev;\n }\n\n\n /**\n * \\brief The thread block rotates its [blocked arrangement](index.html#sec5sec3) of \\p input items, shifting it up by one item. All threads receive the \\p input provided by threadBLOCK_THREADS-1.\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n */\n template \n __device__ __forceinline__ void Up(\n T (&input)[ITEMS_PER_THREAD], ///< [in] The calling thread's input items\n T (&prev)[ITEMS_PER_THREAD], ///< [out] The corresponding predecessor items (may be aliased to \\p input). The item \\p prev[0] is not updated for thread0.\n T &block_suffix) ///< [out] The item \\p input[ITEMS_PER_THREAD-1] from threadBLOCK_THREADS-1, provided to all threads\n {\n Up(input, prev);\n block_suffix = temp_storage[BLOCK_THREADS - 1].prev;\n }\n\n\n /**\n * \\brief The thread block rotates its [blocked arrangement](index.html#sec5sec3) of \\p input items, shifting it down by one item\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n */\n template \n __device__ __forceinline__ void Down(\n T (&input)[ITEMS_PER_THREAD], ///< [in] The calling thread's input items\n T (&prev)[ITEMS_PER_THREAD]) ///< [out] The corresponding predecessor items (may be aliased to \\p input). The value \\p prev[0] is not updated for threadBLOCK_THREADS-1.\n {\n temp_storage[linear_tid].prev = input[ITEMS_PER_THREAD - 1];\n\n CTA_SYNC();\n\n #pragma unroll\n for (int ITEM = ITEMS_PER_THREAD - 1; ITEM > 0; --ITEM)\n prev[ITEM] = input[ITEM - 1];\n\n if (linear_tid > 0)\n prev[0] = temp_storage[linear_tid - 1].prev;\n }\n\n\n /**\n * \\brief The thread block rotates its [blocked arrangement](index.html#sec5sec3) of input items, shifting it down by one item. All threads receive \\p input[0] provided by thread0.\n *\n * \\par\n * - \\blocked\n * - \\granularity\n * - \\smemreuse\n */\n template \n __device__ __forceinline__ void Down(\n T (&input)[ITEMS_PER_THREAD], ///< [in] The calling thread's input items\n T (&prev)[ITEMS_PER_THREAD], ///< [out] The corresponding predecessor items (may be aliased to \\p input). The value \\p prev[0] is not updated for threadBLOCK_THREADS-1.\n T &block_prefix) ///< [out] The item \\p input[0] from thread0, provided to all threads\n {\n Up(input, prev);\n block_prefix = temp_storage[BLOCK_THREADS - 1].prev;\n }\n\n //@} end member group\n\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/block_store.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Operations for writing linear segments of data from the CUDA thread block\n */\n\n#pragma once\n\n#include \n\n#include \"block_exchange.cuh\"\n#include \"../util_ptx.cuh\"\n#include \"../util_macro.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIo\n * @{\n */\n\n\n/******************************************************************//**\n * \\name Blocked arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n/**\n * \\brief Store a blocked arrangement of items across a thread block into a linear segment of items.\n *\n * \\blocked\n *\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectBlocked(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n{\n OutputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);\n\n // Store directly in thread-blocked order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n thread_itr[ITEM] = items[ITEM];\n }\n}\n\n\n/**\n * \\brief Store a blocked arrangement of items across a thread block into a linear segment of items, guarded by range\n *\n * \\blocked\n *\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectBlocked(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n{\n OutputIteratorT thread_itr = block_itr + (linear_tid * ITEMS_PER_THREAD);\n\n // Store directly in thread-blocked order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (ITEM + (linear_tid * ITEMS_PER_THREAD) < valid_items)\n {\n thread_itr[ITEM] = items[ITEM];\n }\n }\n}\n\n\n/**\n * \\brief Store a blocked arrangement of items across a thread block into a linear segment of items.\n *\n * \\blocked\n *\n * The output offset (\\p block_ptr + \\p block_offset) must be quad-item aligned,\n * which is the default starting offset returned by \\p cudaMalloc()\n *\n * \\par\n * The following conditions will prevent vectorization and storing will fall back to cub::BLOCK_STORE_DIRECT:\n * - \\p ITEMS_PER_THREAD is odd\n * - The data type \\p T is not a built-in primitive or CUDA vector type (e.g., \\p short, \\p int2, \\p double, \\p float2, etc.)\n *\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n *\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD>\n__device__ __forceinline__ void StoreDirectBlockedVectorized(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n T *block_ptr, ///< [in] Input pointer for storing from\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n{\n enum\n {\n // Maximum CUDA vector size is 4 elements\n MAX_VEC_SIZE = CUB_MIN(4, ITEMS_PER_THREAD),\n\n // Vector size must be a power of two and an even divisor of the items per thread\n VEC_SIZE = ((((MAX_VEC_SIZE - 1) & MAX_VEC_SIZE) == 0) && ((ITEMS_PER_THREAD % MAX_VEC_SIZE) == 0)) ?\n MAX_VEC_SIZE :\n 1,\n\n VECTORS_PER_THREAD = ITEMS_PER_THREAD / VEC_SIZE,\n };\n\n // Vector type\n typedef typename CubVector::Type Vector;\n\n // Alias global pointer\n Vector *block_ptr_vectors = reinterpret_cast(const_cast(block_ptr));\n\n // Alias pointers (use \"raw\" array here which should get optimized away to prevent conservative PTXAS lmem spilling)\n Vector raw_vector[VECTORS_PER_THREAD];\n T *raw_items = reinterpret_cast(raw_vector);\n\n // Copy\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n raw_items[ITEM] = items[ITEM];\n }\n\n // Direct-store using vector types\n StoreDirectBlocked(linear_tid, block_ptr_vectors, raw_vector);\n}\n\n\n\n//@} end member group\n/******************************************************************//**\n * \\name Striped arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n\n/**\n * \\brief Store a striped arrangement of data across the thread block into a linear segment of items.\n *\n * \\striped\n *\n * \\tparam BLOCK_THREADS The thread block size in threads\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n int BLOCK_THREADS,\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n{\n OutputIteratorT thread_itr = block_itr + linear_tid;\n\n // Store directly in striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n thread_itr[(ITEM * BLOCK_THREADS)] = items[ITEM];\n }\n}\n\n\n/**\n * \\brief Store a striped arrangement of data across the thread block into a linear segment of items, guarded by range\n *\n * \\striped\n *\n * \\tparam BLOCK_THREADS The thread block size in threads\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n int BLOCK_THREADS,\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n{\n OutputIteratorT thread_itr = block_itr + linear_tid;\n\n // Store directly in striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if ((ITEM * BLOCK_THREADS) + linear_tid < valid_items)\n {\n thread_itr[(ITEM * BLOCK_THREADS)] = items[ITEM];\n }\n }\n}\n\n\n\n//@} end member group\n/******************************************************************//**\n * \\name Warp-striped arrangement I/O (direct)\n *********************************************************************/\n//@{\n\n\n/**\n * \\brief Store a warp-striped arrangement of data across the thread block into a linear segment of items.\n *\n * \\warpstriped\n *\n * \\par Usage Considerations\n * The number of threads in the thread block must be a multiple of the architecture's warp size.\n *\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectWarpStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [out] Data to load\n{\n int tid = linear_tid & (CUB_PTX_WARP_THREADS - 1);\n int wid = linear_tid >> CUB_PTX_LOG_WARP_THREADS;\n int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;\n\n OutputIteratorT thread_itr = block_itr + warp_offset + tid;\n\n // Store directly in warp-striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n thread_itr[(ITEM * CUB_PTX_WARP_THREADS)] = items[ITEM];\n }\n}\n\n\n/**\n * \\brief Store a warp-striped arrangement of data across the thread block into a linear segment of items, guarded by range\n *\n * \\warpstriped\n *\n * \\par Usage Considerations\n * The number of threads in the thread block must be a multiple of the architecture's warp size.\n *\n * \\tparam T [inferred] The data type to store.\n * \\tparam ITEMS_PER_THREAD [inferred] The number of consecutive items partitioned onto each thread.\n * \\tparam OutputIteratorT [inferred] The random-access iterator type for output \\iterator.\n */\ntemplate <\n typename T,\n int ITEMS_PER_THREAD,\n typename OutputIteratorT>\n__device__ __forceinline__ void StoreDirectWarpStriped(\n int linear_tid, ///< [in] A suitable 1D thread-identifier for the calling thread (e.g., (threadIdx.y * blockDim.x) + linear_tid for 2D thread blocks)\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n{\n int tid = linear_tid & (CUB_PTX_WARP_THREADS - 1);\n int wid = linear_tid >> CUB_PTX_LOG_WARP_THREADS;\n int warp_offset = wid * CUB_PTX_WARP_THREADS * ITEMS_PER_THREAD;\n\n OutputIteratorT thread_itr = block_itr + warp_offset + tid;\n\n // Store directly in warp-striped order\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (warp_offset + tid + (ITEM * CUB_PTX_WARP_THREADS) < valid_items)\n {\n thread_itr[(ITEM * CUB_PTX_WARP_THREADS)] = items[ITEM];\n }\n }\n}\n\n\n//@} end member group\n\n\n/** @} */ // end group UtilIo\n\n\n//-----------------------------------------------------------------------------\n// Generic BlockStore abstraction\n//-----------------------------------------------------------------------------\n\n/**\n * \\brief cub::BlockStoreAlgorithm enumerates alternative algorithms for cub::BlockStore to write a blocked arrangement of items across a CUDA thread block to a linear segment of memory.\n */\nenum BlockStoreAlgorithm\n{\n /**\n * \\par Overview\n *\n * A [blocked arrangement](index.html#sec5sec3) of data is written\n * directly to memory.\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) decreases as the\n * access stride between threads increases (i.e., the number items per thread).\n */\n BLOCK_STORE_DIRECT,\n\n /**\n * \\par Overview\n *\n * A [blocked arrangement](index.html#sec5sec3) of data is written directly\n * to memory using CUDA's built-in vectorized stores as a coalescing optimization.\n * For example, st.global.v4.s32 instructions will be generated\n * when \\p T = \\p int and \\p ITEMS_PER_THREAD % 4 == 0.\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high until the the\n * access stride between threads (i.e., the number items per thread) exceeds the\n * maximum vector store width (typically 4 items or 64B, whichever is lower).\n * - The following conditions will prevent vectorization and writing will fall back to cub::BLOCK_STORE_DIRECT:\n * - \\p ITEMS_PER_THREAD is odd\n * - The \\p OutputIteratorT is not a simple pointer type\n * - The block output offset is not quadword-aligned\n * - The data type \\p T is not a built-in primitive or CUDA vector type (e.g., \\p short, \\p int2, \\p double, \\p float2, etc.)\n */\n BLOCK_STORE_VECTORIZE,\n\n /**\n * \\par Overview\n * A [blocked arrangement](index.html#sec5sec3) is locally\n * transposed and then efficiently written to memory as a [striped arrangement](index.html#sec5sec3).\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items written per thread.\n * - The local reordering incurs slightly longer latencies and throughput than the\n * direct cub::BLOCK_STORE_DIRECT and cub::BLOCK_STORE_VECTORIZE alternatives.\n */\n BLOCK_STORE_TRANSPOSE,\n\n /**\n * \\par Overview\n * A [blocked arrangement](index.html#sec5sec3) is locally\n * transposed and then efficiently written to memory as a\n * [warp-striped arrangement](index.html#sec5sec3)\n *\n * \\par Usage Considerations\n * - BLOCK_THREADS must be a multiple of WARP_THREADS\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items written per thread.\n * - The local reordering incurs slightly longer latencies and throughput than the\n * direct cub::BLOCK_STORE_DIRECT and cub::BLOCK_STORE_VECTORIZE alternatives.\n */\n BLOCK_STORE_WARP_TRANSPOSE,\n\n /**\n * \\par Overview\n * A [blocked arrangement](index.html#sec5sec3) is locally\n * transposed and then efficiently written to memory as a\n * [warp-striped arrangement](index.html#sec5sec3)\n * To reduce the shared memory requirement, only one warp's worth of shared\n * memory is provisioned and is subsequently time-sliced among warps.\n *\n * \\par Usage Considerations\n * - BLOCK_THREADS must be a multiple of WARP_THREADS\n *\n * \\par Performance Considerations\n * - The utilization of memory transactions (coalescing) remains high regardless\n * of items written per thread.\n * - Provisions less shared memory temporary storage, but incurs larger\n * latencies than the BLOCK_STORE_WARP_TRANSPOSE alternative.\n */\n BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED,\n\n};\n\n\n/**\n * \\brief The BlockStore class provides [collective](index.html#sec0) data movement methods for writing a [blocked arrangement](index.html#sec5sec3) of items partitioned across a CUDA thread block to a linear segment of memory. ![](block_store_logo.png)\n * \\ingroup BlockModule\n * \\ingroup UtilIo\n *\n * \\tparam T The type of data to be written.\n * \\tparam BLOCK_DIM_X The thread block length in threads along the X dimension\n * \\tparam ITEMS_PER_THREAD The number of consecutive items partitioned onto each thread.\n * \\tparam ALGORITHM [optional] cub::BlockStoreAlgorithm tuning policy enumeration. default: cub::BLOCK_STORE_DIRECT.\n * \\tparam WARP_TIME_SLICING [optional] Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage). (default: false)\n * \\tparam BLOCK_DIM_Y [optional] The thread block length in threads along the Y dimension (default: 1)\n * \\tparam BLOCK_DIM_Z [optional] The thread block length in threads along the Z dimension (default: 1)\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - The BlockStore class provides a single data movement abstraction that can be specialized\n * to implement different cub::BlockStoreAlgorithm strategies. This facilitates different\n * performance policies for different architectures, data types, granularity sizes, etc.\n * - BlockStore can be optionally specialized by different data movement strategies:\n * -# cub::BLOCK_STORE_DIRECT. A [blocked arrangement](index.html#sec5sec3) of data is written\n * directly to memory. [More...](\\ref cub::BlockStoreAlgorithm)\n * -# cub::BLOCK_STORE_VECTORIZE. A [blocked arrangement](index.html#sec5sec3)\n * of data is written directly to memory using CUDA's built-in vectorized stores as a\n * coalescing optimization. [More...](\\ref cub::BlockStoreAlgorithm)\n * -# cub::BLOCK_STORE_TRANSPOSE. A [blocked arrangement](index.html#sec5sec3)\n * is locally transposed into a [striped arrangement](index.html#sec5sec3) which is\n * then written to memory. [More...](\\ref cub::BlockStoreAlgorithm)\n * -# cub::BLOCK_STORE_WARP_TRANSPOSE. A [blocked arrangement](index.html#sec5sec3)\n * is locally transposed into a [warp-striped arrangement](index.html#sec5sec3) which is\n * then written to memory. [More...](\\ref cub::BlockStoreAlgorithm)\n * - \\rowmajor\n *\n * \\par A Simple Example\n * \\blockcollective{BlockStore}\n * \\par\n * The code snippet below illustrates the storing of a \"blocked\" arrangement\n * of 512 integers across 128 threads (where each thread owns 4 consecutive items)\n * into a linear segment of memory. The store is specialized for \\p BLOCK_STORE_WARP_TRANSPOSE,\n * meaning items are locally reordered among threads so that memory references will be\n * efficiently coalesced using a warp-striped access pattern.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockStore for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockStore BlockStore;\n *\n * // Allocate shared memory for BlockStore\n * __shared__ typename BlockStore::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Store items to linear memory\n * int thread_data[4];\n * BlockStore(temp_storage).Store(d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of \\p thread_data across the block of threads is\n * { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n * The output \\p d_data will be 0, 1, 2, 3, 4, 5, ....\n *\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X,\n int ITEMS_PER_THREAD,\n BlockStoreAlgorithm ALGORITHM = BLOCK_STORE_DIRECT,\n int BLOCK_DIM_Y = 1,\n int BLOCK_DIM_Z = 1,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass BlockStore\n{\nprivate:\n /******************************************************************************\n * Constants and typed definitions\n ******************************************************************************/\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n\n /******************************************************************************\n * Algorithmic variants\n ******************************************************************************/\n\n /// Store helper\n template \n struct StoreInternal;\n\n\n /**\n * BLOCK_STORE_DIRECT specialization of store helper\n */\n template \n struct StoreInternal\n {\n /// Shared memory storage layout type\n typedef NullType TempStorage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ StoreInternal(\n TempStorage &/*temp_storage*/,\n int linear_tid)\n :\n linear_tid(linear_tid)\n {}\n\n /// Store items into a linear segment of memory\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n StoreDirectBlocked(linear_tid, block_itr, items);\n }\n\n /// Store items into a linear segment of memory, guarded by range\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n StoreDirectBlocked(linear_tid, block_itr, items, valid_items);\n }\n };\n\n\n /**\n * BLOCK_STORE_VECTORIZE specialization of store helper\n */\n template \n struct StoreInternal\n {\n /// Shared memory storage layout type\n typedef NullType TempStorage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ StoreInternal(\n TempStorage &/*temp_storage*/,\n int linear_tid)\n :\n linear_tid(linear_tid)\n {}\n\n /// Store items into a linear segment of memory, specialized for native pointer types (attempts vectorization)\n __device__ __forceinline__ void Store(\n T *block_ptr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n StoreDirectBlockedVectorized(linear_tid, block_ptr, items);\n }\n\n /// Store items into a linear segment of memory, specialized for opaque input iterators (skips vectorization)\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n StoreDirectBlocked(linear_tid, block_itr, items);\n }\n\n /// Store items into a linear segment of memory, guarded by range\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n StoreDirectBlocked(linear_tid, block_itr, items, valid_items);\n }\n };\n\n\n /**\n * BLOCK_STORE_TRANSPOSE specialization of store helper\n */\n template \n struct StoreInternal\n {\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ StoreInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Store items into a linear segment of memory\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n BlockExchange(temp_storage).BlockedToStriped(items);\n StoreDirectStriped(linear_tid, block_itr, items);\n }\n\n /// Store items into a linear segment of memory, guarded by range\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n BlockExchange(temp_storage).BlockedToStriped(items);\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n StoreDirectStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n }\n };\n\n\n /**\n * BLOCK_STORE_WARP_TRANSPOSE specialization of store helper\n */\n template \n struct StoreInternal\n {\n enum\n {\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)\n };\n\n // Assert BLOCK_THREADS must be a multiple of WARP_THREADS\n CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), \"BLOCK_THREADS must be a multiple of WARP_THREADS\");\n\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ StoreInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Store items into a linear segment of memory\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n BlockExchange(temp_storage).BlockedToWarpStriped(items);\n StoreDirectWarpStriped(linear_tid, block_itr, items);\n }\n\n /// Store items into a linear segment of memory, guarded by range\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n BlockExchange(temp_storage).BlockedToWarpStriped(items);\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n StoreDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n }\n };\n\n\n /**\n * BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED specialization of store helper\n */\n template \n struct StoreInternal\n {\n enum\n {\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH)\n };\n\n // Assert BLOCK_THREADS must be a multiple of WARP_THREADS\n CUB_STATIC_ASSERT((BLOCK_THREADS % WARP_THREADS == 0), \"BLOCK_THREADS must be a multiple of WARP_THREADS\");\n\n // BlockExchange utility type for keys\n typedef BlockExchange BlockExchange;\n\n /// Shared memory storage layout type\n struct _TempStorage : BlockExchange::TempStorage\n {\n /// Temporary storage for partially-full block guard\n volatile int valid_items;\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\n /// Constructor\n __device__ __forceinline__ StoreInternal(\n TempStorage &temp_storage,\n int linear_tid)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(linear_tid)\n {}\n\n /// Store items into a linear segment of memory\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n BlockExchange(temp_storage).BlockedToWarpStriped(items);\n StoreDirectWarpStriped(linear_tid, block_itr, items);\n }\n\n /// Store items into a linear segment of memory, guarded by range\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n BlockExchange(temp_storage).BlockedToWarpStriped(items);\n if (linear_tid == 0)\n temp_storage.valid_items = valid_items; // Move through volatile smem as a workaround to prevent RF spilling on subsequent loads\n CTA_SYNC();\n StoreDirectWarpStriped(linear_tid, block_itr, items, temp_storage.valid_items);\n }\n };\n\n /******************************************************************************\n * Type definitions\n ******************************************************************************/\n\n /// Internal load implementation to use\n typedef StoreInternal InternalStore;\n\n\n /// Shared memory storage layout type\n typedef typename InternalStore::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Internal storage allocator\n __device__ __forceinline__ _TempStorage& PrivateStorage()\n {\n __shared__ _TempStorage private_storage;\n return private_storage;\n }\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Thread reference to shared storage\n _TempStorage &temp_storage;\n\n /// Linear thread-id\n int linear_tid;\n\npublic:\n\n\n /// \\smemstorage{BlockStore}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using a private static allocation of shared memory as temporary storage.\n */\n __device__ __forceinline__ BlockStore()\n :\n temp_storage(PrivateStorage()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage.\n */\n __device__ __forceinline__ BlockStore(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Data movement\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Store items into a linear segment of memory.\n *\n * \\par\n * - \\blocked\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the storing of a \"blocked\" arrangement\n * of 512 integers across 128 threads (where each thread owns 4 consecutive items)\n * into a linear segment of memory. The store is specialized for \\p BLOCK_STORE_WARP_TRANSPOSE,\n * meaning items are locally reordered among threads so that memory references will be\n * efficiently coalesced using a warp-striped access pattern.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, ...)\n * {\n * // Specialize BlockStore for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockStore BlockStore;\n *\n * // Allocate shared memory for BlockStore\n * __shared__ typename BlockStore::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Store items to linear memory\n * int thread_data[4];\n * BlockStore(temp_storage).Store(d_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of \\p thread_data across the block of threads is\n * { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] }.\n * The output \\p d_data will be 0, 1, 2, 3, 4, 5, ....\n *\n */\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD]) ///< [in] Data to store\n {\n InternalStore(temp_storage, linear_tid).Store(block_itr, items);\n }\n\n /**\n * \\brief Store items into a linear segment of memory, guarded by range.\n *\n * \\par\n * - \\blocked\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the guarded storing of a \"blocked\" arrangement\n * of 512 integers across 128 threads (where each thread owns 4 consecutive items)\n * into a linear segment of memory. The store is specialized for \\p BLOCK_STORE_WARP_TRANSPOSE,\n * meaning items are locally reordered among threads so that memory references will be\n * efficiently coalesced using a warp-striped access pattern.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(int *d_data, int valid_items, ...)\n * {\n * // Specialize BlockStore for a 1D block of 128 threads owning 4 integer items each\n * typedef cub::BlockStore BlockStore;\n *\n * // Allocate shared memory for BlockStore\n * __shared__ typename BlockStore::TempStorage temp_storage;\n *\n * // Obtain a segment of consecutive items that are blocked across threads\n * int thread_data[4];\n * ...\n *\n * // Store items to linear memory\n * int thread_data[4];\n * BlockStore(temp_storage).Store(d_data, thread_data, valid_items);\n *\n * \\endcode\n * \\par\n * Suppose the set of \\p thread_data across the block of threads is\n * { [0,1,2,3], [4,5,6,7], ..., [508,509,510,511] } and \\p valid_items is \\p 5.\n * The output \\p d_data will be 0, 1, 2, 3, 4, ?, ?, ?, ..., with\n * only the first two threads being unmasked to store portions of valid data.\n *\n */\n template \n __device__ __forceinline__ void Store(\n OutputIteratorT block_itr, ///< [in] The thread block's base output iterator for storing to\n T (&items)[ITEMS_PER_THREAD], ///< [in] Data to store\n int valid_items) ///< [in] Number of valid items to write\n {\n InternalStore(temp_storage, linear_tid).Store(block_itr, items, valid_items);\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_histogram_atomic.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockHistogramAtomic class provides atomic-based methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief The BlockHistogramAtomic class provides atomic-based methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block.\n */\ntemplate \nstruct BlockHistogramAtomic\n{\n /// Shared memory storage layout type\n struct TempStorage {};\n\n\n /// Constructor\n __device__ __forceinline__ BlockHistogramAtomic(\n TempStorage &temp_storage)\n {}\n\n\n /// Composite data onto an existing histogram\n template <\n typename T,\n typename CounterT, \n int ITEMS_PER_THREAD>\n __device__ __forceinline__ void Composite(\n T (&items)[ITEMS_PER_THREAD], ///< [in] Calling thread's input values to histogram\n CounterT histogram[BINS]) ///< [out] Reference to shared/device-accessible memory histogram\n {\n // Update histogram\n #pragma unroll\n for (int i = 0; i < ITEMS_PER_THREAD; ++i)\n {\n atomicAdd(histogram + items[i], 1);\n }\n }\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_histogram_sort.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::BlockHistogramSort class provides sorting-based methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../block/block_radix_sort.cuh\"\n#include \"../../block/block_discontinuity.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n\n/**\n * \\brief The BlockHistogramSort class provides sorting-based methods for constructing block-wide histograms from data samples partitioned across a CUDA thread block.\n */\ntemplate <\n typename T, ///< Sample type\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int ITEMS_PER_THREAD, ///< The number of samples per thread\n int BINS, ///< The number of bins into which histogram samples may fall\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockHistogramSort\n{\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n // Parameterize BlockRadixSort type for our thread block\n typedef BlockRadixSort<\n T,\n BLOCK_DIM_X,\n ITEMS_PER_THREAD,\n NullType,\n 4,\n (PTX_ARCH >= 350) ? true : false,\n BLOCK_SCAN_WARP_SCANS,\n cudaSharedMemBankSizeFourByte,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n BlockRadixSortT;\n\n // Parameterize BlockDiscontinuity type for our thread block\n typedef BlockDiscontinuity<\n T,\n BLOCK_DIM_X,\n BLOCK_DIM_Y,\n BLOCK_DIM_Z,\n PTX_ARCH>\n BlockDiscontinuityT;\n\n /// Shared memory\n union _TempStorage\n {\n // Storage for sorting bin values\n typename BlockRadixSortT::TempStorage sort;\n\n struct\n {\n // Storage for detecting discontinuities in the tile of sorted bin values\n typename BlockDiscontinuityT::TempStorage flag;\n\n // Storage for noting begin/end offsets of bin runs in the tile of sorted bin values\n unsigned int run_begin[BINS];\n unsigned int run_end[BINS];\n };\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n\n\n /// Constructor\n __device__ __forceinline__ BlockHistogramSort(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n // Discontinuity functor\n struct DiscontinuityOp\n {\n // Reference to temp_storage\n _TempStorage &temp_storage;\n\n // Constructor\n __device__ __forceinline__ DiscontinuityOp(_TempStorage &temp_storage) :\n temp_storage(temp_storage)\n {}\n\n // Discontinuity predicate\n __device__ __forceinline__ bool operator()(const T &a, const T &b, int b_index)\n {\n if (a != b)\n {\n // Note the begin/end offsets in shared storage\n temp_storage.run_begin[b] = b_index;\n temp_storage.run_end[a] = b_index;\n\n return true;\n }\n else\n {\n return false;\n }\n }\n };\n\n\n // Composite data onto an existing histogram\n template <\n typename CounterT >\n __device__ __forceinline__ void Composite(\n T (&items)[ITEMS_PER_THREAD], ///< [in] Calling thread's input values to histogram\n CounterT histogram[BINS]) ///< [out] Reference to shared/device-accessible memory histogram\n {\n enum { TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD };\n\n // Sort bytes in blocked arrangement\n BlockRadixSortT(temp_storage.sort).Sort(items);\n\n CTA_SYNC();\n\n // Initialize the shared memory's run_begin and run_end for each bin\n int histo_offset = 0;\n\n #pragma unroll\n for(; histo_offset + BLOCK_THREADS <= BINS; histo_offset += BLOCK_THREADS)\n {\n temp_storage.run_begin[histo_offset + linear_tid] = TILE_SIZE;\n temp_storage.run_end[histo_offset + linear_tid] = TILE_SIZE;\n }\n // Finish up with guarded initialization if necessary\n if ((BINS % BLOCK_THREADS != 0) && (histo_offset + linear_tid < BINS))\n {\n temp_storage.run_begin[histo_offset + linear_tid] = TILE_SIZE;\n temp_storage.run_end[histo_offset + linear_tid] = TILE_SIZE;\n }\n\n CTA_SYNC();\n\n int flags[ITEMS_PER_THREAD]; // unused\n\n // Compute head flags to demarcate contiguous runs of the same bin in the sorted tile\n DiscontinuityOp flag_op(temp_storage);\n BlockDiscontinuityT(temp_storage.flag).FlagHeads(flags, items, flag_op);\n\n // Update begin for first item\n if (linear_tid == 0) temp_storage.run_begin[items[0]] = 0;\n\n CTA_SYNC();\n\n // Composite into histogram\n histo_offset = 0;\n\n #pragma unroll\n for(; histo_offset + BLOCK_THREADS <= BINS; histo_offset += BLOCK_THREADS)\n {\n int thread_offset = histo_offset + linear_tid;\n CounterT count = temp_storage.run_end[thread_offset] - temp_storage.run_begin[thread_offset];\n histogram[thread_offset] += count;\n }\n\n // Finish up with guarded composition if necessary\n if ((BINS % BLOCK_THREADS != 0) && (histo_offset + linear_tid < BINS))\n {\n int thread_offset = histo_offset + linear_tid;\n CounterT count = temp_storage.run_end[thread_offset] - temp_storage.run_begin[thread_offset];\n histogram[thread_offset] += count;\n }\n }\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_reduce_raking.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockReduceRaking provides raking-based methods of parallel reduction across a CUDA thread block. Supports non-commutative reduction operators.\n */\n\n#pragma once\n\n#include \"../../block/block_raking_layout.cuh\"\n#include \"../../warp/warp_reduce.cuh\"\n#include \"../../thread/thread_reduce.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief BlockReduceRaking provides raking-based methods of parallel reduction across a CUDA thread block. Supports non-commutative reduction operators.\n *\n * Supports non-commutative binary reduction operators. Unlike commutative\n * reduction operators (e.g., addition), the application of a non-commutative\n * reduction operator (e.g, string concatenation) across a sequence of inputs must\n * honor the relative ordering of items and partial reductions when applying the\n * reduction operator.\n *\n * Compared to the implementation of BlockReduceRaking (which does not support\n * non-commutative operators), this implementation requires a few extra\n * rounds of inter-thread communication.\n */\ntemplate <\n typename T, ///< Data type being reduced\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockReduceRaking\n{\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n /// Layout type for padded thread block raking grid\n typedef BlockRakingLayout BlockRakingLayout;\n\n /// WarpReduce utility type\n typedef typename WarpReduce::InternalWarpReduce WarpReduce;\n\n /// Constants\n enum\n {\n /// Number of raking threads\n RAKING_THREADS = BlockRakingLayout::RAKING_THREADS,\n\n /// Number of raking elements per warp synchronous raking thread\n SEGMENT_LENGTH = BlockRakingLayout::SEGMENT_LENGTH,\n\n /// Cooperative work can be entirely warp synchronous\n WARP_SYNCHRONOUS = (RAKING_THREADS == BLOCK_THREADS),\n\n /// Whether or not warp-synchronous reduction should be unguarded (i.e., the warp-reduction elements is a power of two\n WARP_SYNCHRONOUS_UNGUARDED = PowerOfTwo::VALUE,\n\n /// Whether or not accesses into smem are unguarded\n RAKING_UNGUARDED = BlockRakingLayout::UNGUARDED,\n\n };\n\n\n /// Shared memory storage layout type\n union _TempStorage\n {\n typename WarpReduce::TempStorage warp_storage; ///< Storage for warp-synchronous reduction\n typename BlockRakingLayout::TempStorage raking_grid; ///< Padded thread block raking grid\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n\n\n /// Constructor\n __device__ __forceinline__ BlockReduceRaking(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n template \n __device__ __forceinline__ T RakingReduction(\n ReductionOp reduction_op, ///< [in] Binary scan operator\n T *raking_segment,\n T partial, ///< [in] [lane0 only] Warp-wide aggregate reduction of input items\n int num_valid, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n Int2Type /*iteration*/)\n {\n // Update partial if addend is in range\n if ((IS_FULL_TILE && RAKING_UNGUARDED) || ((linear_tid * SEGMENT_LENGTH) + ITERATION < num_valid))\n {\n T addend = raking_segment[ITERATION];\n partial = reduction_op(partial, addend);\n }\n return RakingReduction(reduction_op, raking_segment, partial, num_valid, Int2Type());\n }\n\n template \n __device__ __forceinline__ T RakingReduction(\n ReductionOp /*reduction_op*/, ///< [in] Binary scan operator\n T * /*raking_segment*/,\n T partial, ///< [in] [lane0 only] Warp-wide aggregate reduction of input items\n int /*num_valid*/, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n Int2Type /*iteration*/)\n {\n return partial;\n }\n\n\n\n /// Computes a thread block-wide reduction using the specified reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template <\n bool IS_FULL_TILE,\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T partial, ///< [in] Calling thread's input partial reductions\n int num_valid, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp synchronous reduction (unguarded if active threads is a power-of-two)\n partial = WarpReduce(temp_storage.warp_storage).template Reduce(\n partial,\n num_valid,\n reduction_op);\n }\n else\n {\n // Place partial into shared memory grid.\n *BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid) = partial;\n\n CTA_SYNC();\n\n // Reduce parallelism to one warp\n if (linear_tid < RAKING_THREADS)\n {\n // Raking reduction in grid\n T *raking_segment = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n partial = raking_segment[0];\n\n partial = RakingReduction(reduction_op, raking_segment, partial, num_valid, Int2Type<1>());\n\n partial = WarpReduce(temp_storage.warp_storage).template Reduce(\n partial,\n num_valid,\n reduction_op);\n\n }\n }\n\n return partial;\n }\n\n\n /// Computes a thread block-wide reduction using addition (+) as the reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template \n __device__ __forceinline__ T Sum(\n T partial, ///< [in] Calling thread's input partial reductions\n int num_valid) ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n {\n cub::Sum reduction_op;\n\n return Reduce(partial, num_valid, reduction_op);\n }\n\n\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_reduce_raking_commutative_only.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockReduceRakingCommutativeOnly provides raking-based methods of parallel reduction across a CUDA thread block. Does not support non-commutative reduction operators.\n */\n\n#pragma once\n\n#include \"block_reduce_raking.cuh\"\n#include \"../../warp/warp_reduce.cuh\"\n#include \"../../thread/thread_reduce.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief BlockReduceRakingCommutativeOnly provides raking-based methods of parallel reduction across a CUDA thread block. Does not support non-commutative reduction operators. Does not support block sizes that are not a multiple of the warp size.\n */\ntemplate <\n typename T, ///< Data type being reduced\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockReduceRakingCommutativeOnly\n{\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n // The fall-back implementation to use when BLOCK_THREADS is not a multiple of the warp size or not all threads have valid values\n typedef BlockReduceRaking FallBack;\n\n /// Constants\n enum\n {\n /// Number of warp threads\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n\n /// Whether or not to use fall-back\n USE_FALLBACK = ((BLOCK_THREADS % WARP_THREADS != 0) || (BLOCK_THREADS <= WARP_THREADS)),\n\n /// Number of raking threads\n RAKING_THREADS = WARP_THREADS,\n\n /// Number of threads actually sharing items with the raking threads\n SHARING_THREADS = CUB_MAX(1, BLOCK_THREADS - RAKING_THREADS),\n\n /// Number of raking elements per warp synchronous raking thread\n SEGMENT_LENGTH = SHARING_THREADS / WARP_THREADS,\n };\n\n /// WarpReduce utility type\n typedef WarpReduce WarpReduce;\n\n /// Layout type for padded thread block raking grid\n typedef BlockRakingLayout BlockRakingLayout;\n\n /// Shared memory storage layout type\n union _TempStorage\n {\n struct\n {\n typename WarpReduce::TempStorage warp_storage; ///< Storage for warp-synchronous reduction\n typename BlockRakingLayout::TempStorage raking_grid; ///< Padded thread block raking grid\n };\n typename FallBack::TempStorage fallback_storage; ///< Fall-back storage for non-commutative block scan\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n\n\n /// Constructor\n __device__ __forceinline__ BlockReduceRakingCommutativeOnly(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n /// Computes a thread block-wide reduction using addition (+) as the reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template \n __device__ __forceinline__ T Sum(\n T partial, ///< [in] Calling thread's input partial reductions\n int num_valid) ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n {\n if (USE_FALLBACK || !FULL_TILE)\n {\n return FallBack(temp_storage.fallback_storage).template Sum(partial, num_valid);\n }\n else\n {\n // Place partial into shared memory grid\n if (linear_tid >= RAKING_THREADS)\n *BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid - RAKING_THREADS) = partial;\n\n CTA_SYNC();\n\n // Reduce parallelism to one warp\n if (linear_tid < RAKING_THREADS)\n {\n // Raking reduction in grid\n T *raking_segment = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n partial = internal::ThreadReduce(raking_segment, cub::Sum(), partial);\n\n // Warpscan\n partial = WarpReduce(temp_storage.warp_storage).Sum(partial);\n }\n }\n\n return partial;\n }\n\n\n /// Computes a thread block-wide reduction using the specified reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template <\n bool FULL_TILE,\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T partial, ///< [in] Calling thread's input partial reductions\n int num_valid, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n if (USE_FALLBACK || !FULL_TILE)\n {\n return FallBack(temp_storage.fallback_storage).template Reduce(partial, num_valid, reduction_op);\n }\n else\n {\n // Place partial into shared memory grid\n if (linear_tid >= RAKING_THREADS)\n *BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid - RAKING_THREADS) = partial;\n\n CTA_SYNC();\n\n // Reduce parallelism to one warp\n if (linear_tid < RAKING_THREADS)\n {\n // Raking reduction in grid\n T *raking_segment = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n partial = internal::ThreadReduce(raking_segment, reduction_op, partial);\n\n // Warpscan\n partial = WarpReduce(temp_storage.warp_storage).Reduce(partial, reduction_op);\n }\n }\n\n return partial;\n }\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_reduce_warp_reductions.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockReduceWarpReductions provides variants of warp-reduction-based parallel reduction across a CUDA thread block. Supports non-commutative reduction operators.\n */\n\n#pragma once\n\n#include \"../../warp/warp_reduce.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_arch.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief BlockReduceWarpReductions provides variants of warp-reduction-based parallel reduction across a CUDA thread block. Supports non-commutative reduction operators.\n */\ntemplate <\n typename T, ///< Data type being reduced\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockReduceWarpReductions\n{\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n /// Number of warp threads\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n\n /// Number of active warps\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n\n /// The logical warp size for warp reductions\n LOGICAL_WARP_SIZE = CUB_MIN(BLOCK_THREADS, WARP_THREADS),\n\n /// Whether or not the logical warp size evenly divides the thread block size\n EVEN_WARP_MULTIPLE = (BLOCK_THREADS % LOGICAL_WARP_SIZE == 0)\n };\n\n\n /// WarpReduce utility type\n typedef typename WarpReduce::InternalWarpReduce WarpReduce;\n\n\n /// Shared memory storage layout type\n struct _TempStorage\n {\n typename WarpReduce::TempStorage warp_reduce[WARPS]; ///< Buffer for warp-synchronous scan\n T warp_aggregates[WARPS]; ///< Shared totals from each warp-synchronous scan\n T block_prefix; ///< Shared prefix for the entire thread block\n };\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n unsigned int warp_id;\n unsigned int lane_id;\n\n\n /// Constructor\n __device__ __forceinline__ BlockReduceWarpReductions(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n warp_id((WARPS == 1) ? 0 : linear_tid / WARP_THREADS),\n lane_id(LaneId())\n {}\n\n\n template \n __device__ __forceinline__ T ApplyWarpAggregates(\n ReductionOp reduction_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [lane0 only] Warp-wide aggregate reduction of input items\n int num_valid, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n Int2Type /*successor_warp*/)\n {\n if (FULL_TILE || (SUCCESSOR_WARP * LOGICAL_WARP_SIZE < num_valid))\n {\n T addend = temp_storage.warp_aggregates[SUCCESSOR_WARP];\n warp_aggregate = reduction_op(warp_aggregate, addend);\n }\n return ApplyWarpAggregates(reduction_op, warp_aggregate, num_valid, Int2Type());\n }\n\n template \n __device__ __forceinline__ T ApplyWarpAggregates(\n ReductionOp /*reduction_op*/, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [lane0 only] Warp-wide aggregate reduction of input items\n int /*num_valid*/, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n Int2Type /*successor_warp*/)\n {\n return warp_aggregate;\n }\n\n\n /// Returns block-wide aggregate in thread0.\n template <\n bool FULL_TILE,\n typename ReductionOp>\n __device__ __forceinline__ T ApplyWarpAggregates(\n ReductionOp reduction_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [lane0 only] Warp-wide aggregate reduction of input items\n int num_valid) ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n {\n // Share lane aggregates\n if (lane_id == 0)\n {\n temp_storage.warp_aggregates[warp_id] = warp_aggregate;\n }\n\n CTA_SYNC();\n\n // Update total aggregate in warp 0, lane 0\n if (linear_tid == 0)\n {\n warp_aggregate = ApplyWarpAggregates(reduction_op, warp_aggregate, num_valid, Int2Type<1>());\n }\n\n return warp_aggregate;\n }\n\n\n /// Computes a thread block-wide reduction using addition (+) as the reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template \n __device__ __forceinline__ T Sum(\n T input, ///< [in] Calling thread's input partial reductions\n int num_valid) ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n {\n cub::Sum reduction_op;\n unsigned int warp_offset = warp_id * LOGICAL_WARP_SIZE;\n unsigned int warp_num_valid = (FULL_TILE && EVEN_WARP_MULTIPLE) ?\n LOGICAL_WARP_SIZE :\n (warp_offset < num_valid) ?\n num_valid - warp_offset :\n 0;\n\n // Warp reduction in every warp\n T warp_aggregate = WarpReduce(temp_storage.warp_reduce[warp_id]).template Reduce<(FULL_TILE && EVEN_WARP_MULTIPLE), 1>(\n input,\n warp_num_valid,\n cub::Sum());\n\n // Update outputs and block_aggregate with warp-wide aggregates from lane-0s\n return ApplyWarpAggregates(reduction_op, warp_aggregate, num_valid);\n }\n\n\n /// Computes a thread block-wide reduction using the specified reduction operator. The first num_valid threads each contribute one reduction partial. The return value is only valid for thread0.\n template <\n bool FULL_TILE,\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input partial reductions\n int num_valid, ///< [in] Number of valid elements (may be less than BLOCK_THREADS)\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n unsigned int warp_offset = warp_id * LOGICAL_WARP_SIZE;\n unsigned int warp_num_valid = (FULL_TILE && EVEN_WARP_MULTIPLE) ?\n LOGICAL_WARP_SIZE :\n (warp_offset < static_cast(num_valid)) ?\n num_valid - warp_offset :\n 0;\n\n // Warp reduction in every warp\n T warp_aggregate = WarpReduce(temp_storage.warp_reduce[warp_id]).template Reduce<(FULL_TILE && EVEN_WARP_MULTIPLE), 1>(\n input,\n warp_num_valid,\n reduction_op);\n\n // Update outputs and block_aggregate with warp-wide aggregates from lane-0s\n return ApplyWarpAggregates(reduction_op, warp_aggregate, num_valid);\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_scan_raking.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n\n/**\n * \\file\n * cub::BlockScanRaking provides variants of raking-based parallel prefix scan across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../util_ptx.cuh\"\n#include \"../../util_arch.cuh\"\n#include \"../../block/block_raking_layout.cuh\"\n#include \"../../thread/thread_reduce.cuh\"\n#include \"../../thread/thread_scan.cuh\"\n#include \"../../warp/warp_scan.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief BlockScanRaking provides variants of raking-based parallel prefix scan across a CUDA thread block.\n */\ntemplate <\n typename T, ///< Data type being scanned\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n bool MEMOIZE, ///< Whether or not to buffer outer raking scan partials to incur fewer shared memory reads at the expense of higher register pressure\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockScanRaking\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n };\n\n /// Layout type for padded thread block raking grid\n typedef BlockRakingLayout BlockRakingLayout;\n\n /// Constants\n enum\n {\n /// Number of raking threads\n RAKING_THREADS = BlockRakingLayout::RAKING_THREADS,\n\n /// Number of raking elements per warp synchronous raking thread\n SEGMENT_LENGTH = BlockRakingLayout::SEGMENT_LENGTH,\n\n /// Cooperative work can be entirely warp synchronous\n WARP_SYNCHRONOUS = (BLOCK_THREADS == RAKING_THREADS),\n };\n\n /// WarpScan utility type\n typedef WarpScan WarpScan;\n\n /// Shared memory storage layout type\n struct _TempStorage\n {\n typename WarpScan::TempStorage warp_scan; ///< Buffer for warp-synchronous scan\n typename BlockRakingLayout::TempStorage raking_grid; ///< Padded thread block raking grid\n T block_aggregate; ///< Block aggregate\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n T cached_segment[SEGMENT_LENGTH];\n\n\n //---------------------------------------------------------------------\n // Utility methods\n //---------------------------------------------------------------------\n\n /// Templated reduction\n template \n __device__ __forceinline__ T GuardedReduce(\n T* raking_ptr, ///< [in] Input array\n ScanOp scan_op, ///< [in] Binary reduction operator\n T raking_partial, ///< [in] Prefix to seed reduction with\n Int2Type /*iteration*/)\n {\n if ((BlockRakingLayout::UNGUARDED) || (((linear_tid * SEGMENT_LENGTH) + ITERATION) < BLOCK_THREADS))\n {\n T addend = raking_ptr[ITERATION];\n raking_partial = scan_op(raking_partial, addend);\n }\n\n return GuardedReduce(raking_ptr, scan_op, raking_partial, Int2Type());\n }\n\n\n /// Templated reduction (base case)\n template \n __device__ __forceinline__ T GuardedReduce(\n T* /*raking_ptr*/, ///< [in] Input array\n ScanOp /*scan_op*/, ///< [in] Binary reduction operator\n T raking_partial, ///< [in] Prefix to seed reduction with\n Int2Type /*iteration*/)\n {\n return raking_partial;\n }\n\n\n /// Templated copy\n template \n __device__ __forceinline__ void CopySegment(\n T* out, ///< [out] Out array\n T* in, ///< [in] Input array\n Int2Type /*iteration*/)\n {\n out[ITERATION] = in[ITERATION];\n CopySegment(out, in, Int2Type());\n }\n\n \n /// Templated copy (base case)\n __device__ __forceinline__ void CopySegment(\n T* /*out*/, ///< [out] Out array\n T* /*in*/, ///< [in] Input array\n Int2Type /*iteration*/)\n {}\n\n\n /// Performs upsweep raking reduction, returning the aggregate\n template \n __device__ __forceinline__ T Upsweep(\n ScanOp scan_op)\n {\n T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n\n // Read data into registers\n CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());\n\n T raking_partial = cached_segment[0];\n\n return GuardedReduce(cached_segment, scan_op, raking_partial, Int2Type<1>());\n }\n\n\n /// Performs exclusive downsweep raking scan\n template \n __device__ __forceinline__ void ExclusiveDownsweep(\n ScanOp scan_op,\n T raking_partial,\n bool apply_prefix = true)\n {\n T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n\n // Read data back into registers\n if (!MEMOIZE)\n {\n CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());\n }\n\n internal::ThreadScanExclusive(cached_segment, cached_segment, scan_op, raking_partial, apply_prefix);\n\n // Write data back to smem\n CopySegment(smem_raking_ptr, cached_segment, Int2Type<0>());\n }\n\n\n /// Performs inclusive downsweep raking scan\n template \n __device__ __forceinline__ void InclusiveDownsweep(\n ScanOp scan_op,\n T raking_partial,\n bool apply_prefix = true)\n {\n T *smem_raking_ptr = BlockRakingLayout::RakingPtr(temp_storage.raking_grid, linear_tid);\n\n // Read data back into registers\n if (!MEMOIZE)\n {\n CopySegment(cached_segment, smem_raking_ptr, Int2Type<0>());\n }\n\n internal::ThreadScanInclusive(cached_segment, cached_segment, scan_op, raking_partial, apply_prefix);\n\n // Write data back to smem\n CopySegment(smem_raking_ptr, cached_segment, Int2Type<0>());\n }\n\n\n //---------------------------------------------------------------------\n // Constructors\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ BlockScanRaking(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z))\n {}\n\n\n //---------------------------------------------------------------------\n // Exclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).ExclusiveScan(input, exclusive_output, scan_op);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Warp-synchronous scan\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, scan_op);\n\n // Exclusive raking downsweep scan\n ExclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n exclusive_output = *placement_ptr;\n }\n }\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, initial_value, scan_op);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Exclusive Warp-synchronous scan\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, initial_value, scan_op);\n\n // Exclusive raking downsweep scan\n ExclusiveDownsweep(scan_op, exclusive_partial);\n }\n\n CTA_SYNC();\n\n // Grab exclusive partial from shared memory\n output = *placement_ptr;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, scan_op, block_aggregate);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial= Upsweep(scan_op);\n\n // Warp-synchronous scan\n T inclusive_partial;\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).Scan(upsweep_partial, inclusive_partial, exclusive_partial, scan_op);\n\n // Exclusive raking downsweep scan\n ExclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));\n\n // Broadcast aggregate to all threads\n if (linear_tid == RAKING_THREADS - 1)\n temp_storage.block_aggregate = inclusive_partial;\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n output = *placement_ptr;\n\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).ExclusiveScan(input, output, initial_value, scan_op, block_aggregate);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Warp-synchronous scan\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, initial_value, scan_op, block_aggregate);\n\n // Exclusive raking downsweep scan\n ExclusiveDownsweep(scan_op, exclusive_partial);\n\n // Broadcast aggregate to other threads\n if (linear_tid == 0)\n temp_storage.block_aggregate = block_aggregate;\n }\n\n CTA_SYNC();\n\n // Grab exclusive partial from shared memory\n output = *placement_ptr;\n\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n T block_aggregate;\n WarpScan warp_scan(temp_storage.warp_scan);\n warp_scan.ExclusiveScan(input, output, scan_op, block_aggregate);\n\n // Obtain warp-wide prefix in lane0, then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = warp_scan.Broadcast(block_prefix, 0);\n\n output = scan_op(block_prefix, output);\n if (linear_tid == 0)\n output = block_prefix;\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n WarpScan warp_scan(temp_storage.warp_scan);\n\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Warp-synchronous scan\n T exclusive_partial, block_aggregate;\n warp_scan.ExclusiveScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);\n\n // Obtain block-wide prefix in lane0, then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = warp_scan.Broadcast(block_prefix, 0);\n\n // Update prefix with warpscan exclusive partial\n T downsweep_prefix = scan_op(block_prefix, exclusive_partial);\n if (linear_tid == 0)\n downsweep_prefix = block_prefix;\n\n // Exclusive raking downsweep scan\n ExclusiveDownsweep(scan_op, downsweep_prefix);\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n output = *placement_ptr;\n }\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).InclusiveScan(input, output, scan_op);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Exclusive Warp-synchronous scan\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).ExclusiveScan(upsweep_partial, exclusive_partial, scan_op);\n\n // Inclusive raking downsweep scan\n InclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n output = *placement_ptr;\n }\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n WarpScan(temp_storage.warp_scan).InclusiveScan(input, output, scan_op, block_aggregate);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Warp-synchronous scan\n T inclusive_partial;\n T exclusive_partial;\n WarpScan(temp_storage.warp_scan).Scan(upsweep_partial, inclusive_partial, exclusive_partial, scan_op);\n\n // Inclusive raking downsweep scan\n InclusiveDownsweep(scan_op, exclusive_partial, (linear_tid != 0));\n\n // Broadcast aggregate to all threads\n if (linear_tid == RAKING_THREADS - 1)\n temp_storage.block_aggregate = inclusive_partial;\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n output = *placement_ptr;\n\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n }\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n if (WARP_SYNCHRONOUS)\n {\n // Short-circuit directly to warp-synchronous scan\n T block_aggregate;\n WarpScan warp_scan(temp_storage.warp_scan);\n warp_scan.InclusiveScan(input, output, scan_op, block_aggregate);\n\n // Obtain warp-wide prefix in lane0, then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = warp_scan.Broadcast(block_prefix, 0);\n\n // Update prefix with exclusive warpscan partial\n output = scan_op(block_prefix, output);\n }\n else\n {\n // Place thread partial into shared memory raking grid\n T *placement_ptr = BlockRakingLayout::PlacementPtr(temp_storage.raking_grid, linear_tid);\n *placement_ptr = input;\n\n CTA_SYNC();\n\n // Reduce parallelism down to just raking threads\n if (linear_tid < RAKING_THREADS)\n {\n WarpScan warp_scan(temp_storage.warp_scan);\n\n // Raking upsweep reduction across shared partials\n T upsweep_partial = Upsweep(scan_op);\n\n // Warp-synchronous scan\n T exclusive_partial, block_aggregate;\n warp_scan.ExclusiveScan(upsweep_partial, exclusive_partial, scan_op, block_aggregate);\n\n // Obtain block-wide prefix in lane0, then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = warp_scan.Broadcast(block_prefix, 0);\n\n // Update prefix with warpscan exclusive partial\n T downsweep_prefix = scan_op(block_prefix, exclusive_partial);\n if (linear_tid == 0)\n downsweep_prefix = block_prefix;\n\n // Inclusive raking downsweep scan\n InclusiveDownsweep(scan_op, downsweep_prefix);\n }\n\n CTA_SYNC();\n\n // Grab thread prefix from shared memory\n output = *placement_ptr;\n }\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_scan_warp_scans.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockScanWarpscans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../util_arch.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../warp/warp_scan.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief BlockScanWarpScans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockScanWarpScans\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// Constants\n enum\n {\n /// Number of warp threads\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n /// Number of active warps\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n };\n\n /// WarpScan utility type\n typedef WarpScan WarpScanT;\n\n /// WarpScan utility type\n typedef WarpScan WarpAggregateScan;\n\n /// Shared memory storage layout type\n\n struct __align__(32) _TempStorage\n {\n T warp_aggregates[WARPS];\n typename WarpScanT::TempStorage warp_scan[WARPS]; ///< Buffer for warp-synchronous scans\n T block_prefix; ///< Shared prefix for the entire thread block\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n unsigned int warp_id;\n unsigned int lane_id;\n\n\n //---------------------------------------------------------------------\n // Constructors\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ BlockScanWarpScans(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n warp_id((WARPS == 1) ? 0 : linear_tid / WARP_THREADS),\n lane_id(LaneId())\n {}\n\n\n //---------------------------------------------------------------------\n // Utility methods\n //---------------------------------------------------------------------\n\n template \n __device__ __forceinline__ void ApplyWarpAggregates(\n T &warp_prefix, ///< [out] The calling thread's partial reduction\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate, ///< [out] Threadblock-wide aggregate reduction of input items\n Int2Type /*addend_warp*/)\n {\n if (warp_id == WARP)\n warp_prefix = block_aggregate;\n\n T addend = temp_storage.warp_aggregates[WARP];\n block_aggregate = scan_op(block_aggregate, addend);\n\n ApplyWarpAggregates(warp_prefix, scan_op, block_aggregate, Int2Type());\n }\n\n template \n __device__ __forceinline__ void ApplyWarpAggregates(\n T &/*warp_prefix*/, ///< [out] The calling thread's partial reduction\n ScanOp /*scan_op*/, ///< [in] Binary scan operator\n T &/*block_aggregate*/, ///< [out] Threadblock-wide aggregate reduction of input items\n Int2Type /*addend_warp*/)\n {}\n\n\n /// Use the warp-wide aggregates to compute the calling warp's prefix. Also returns block-wide aggregate in all threads.\n template \n __device__ __forceinline__ T ComputeWarpPrefix(\n ScanOp scan_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [laneWARP_THREADS - 1 only] Warp-wide aggregate reduction of input items\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Last lane in each warp shares its warp-aggregate\n if (lane_id == WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = warp_aggregate;\n\n CTA_SYNC();\n\n // Accumulate block aggregates and save the one that is our warp's prefix\n T warp_prefix;\n block_aggregate = temp_storage.warp_aggregates[0];\n\n // Use template unrolling (since the PTX backend can't handle unrolling it for SM1x)\n ApplyWarpAggregates(warp_prefix, scan_op, block_aggregate, Int2Type<1>());\n/*\n #pragma unroll\n for (int WARP = 1; WARP < WARPS; ++WARP)\n {\n if (warp_id == WARP)\n warp_prefix = block_aggregate;\n\n T addend = temp_storage.warp_aggregates[WARP];\n block_aggregate = scan_op(block_aggregate, addend);\n }\n*/\n\n return warp_prefix;\n }\n\n\n /// Use the warp-wide aggregates and initial-value to compute the calling warp's prefix. Also returns block-wide aggregate in all threads.\n template \n __device__ __forceinline__ T ComputeWarpPrefix(\n ScanOp scan_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [laneWARP_THREADS - 1 only] Warp-wide aggregate reduction of input items\n T &block_aggregate, ///< [out] Threadblock-wide aggregate reduction of input items\n const T &initial_value) ///< [in] Initial value to seed the exclusive scan\n {\n T warp_prefix = ComputeWarpPrefix(scan_op, warp_aggregate, block_aggregate);\n\n warp_prefix = scan_op(initial_value, warp_prefix);\n\n if (warp_id == 0)\n warp_prefix = initial_value;\n\n return warp_prefix;\n }\n\n //---------------------------------------------------------------------\n // Exclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n // Compute block-wide exclusive scan. The exclusive output from tid0 is invalid.\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, initial_value, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n WarpScanT(temp_storage.warp_scan[warp_id]).Scan(input, inclusive_output, exclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp. Warp prefix for warp0 is invalid.\n T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate);\n\n // Apply warp prefix to our lane's partial\n if (warp_id != 0)\n {\n exclusive_output = scan_op(warp_prefix, exclusive_output);\n if (lane_id == 0)\n exclusive_output = warp_prefix;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n WarpScanT(temp_storage.warp_scan[warp_id]).Scan(input, inclusive_output, exclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp\n T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate, initial_value);\n\n // Apply warp prefix to our lane's partial\n exclusive_output = scan_op(warp_prefix, exclusive_output);\n if (lane_id == 0)\n exclusive_output = warp_prefix;\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n // Compute block-wide exclusive scan. The exclusive output from tid0 is invalid.\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n\n // Use the first warp to determine the thread block prefix, returning the result in lane0\n if (warp_id == 0)\n {\n T block_prefix = block_prefix_callback_op(block_aggregate);\n if (lane_id == 0)\n {\n // Share the prefix with all threads\n temp_storage.block_prefix = block_prefix;\n exclusive_output = block_prefix; // The block prefix is the exclusive output for tid0\n }\n }\n\n CTA_SYNC();\n\n // Incorporate thread block prefix into outputs\n T block_prefix = temp_storage.block_prefix;\n if (linear_tid > 0)\n {\n exclusive_output = scan_op(block_prefix, exclusive_output);\n }\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n InclusiveScan(input, inclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n WarpScanT(temp_storage.warp_scan[warp_id]).InclusiveScan(input, inclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp. Warp prefix for warp0 is invalid.\n T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate);\n\n // Apply warp prefix to our lane's partial\n if (warp_id != 0)\n {\n inclusive_output = scan_op(warp_prefix, inclusive_output);\n }\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n T block_aggregate;\n InclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n\n // Use the first warp to determine the thread block prefix, returning the result in lane0\n if (warp_id == 0)\n {\n T block_prefix = block_prefix_callback_op(block_aggregate);\n if (lane_id == 0)\n {\n // Share the prefix with all threads\n temp_storage.block_prefix = block_prefix;\n }\n }\n\n CTA_SYNC();\n\n // Incorporate thread block prefix into outputs\n T block_prefix = temp_storage.block_prefix;\n exclusive_output = scan_op(block_prefix, exclusive_output);\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_scan_warp_scans2.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockScanWarpscans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../util_arch.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../warp/warp_scan.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief BlockScanWarpScans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockScanWarpScans\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// Constants\n enum\n {\n /// Number of warp threads\n WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n /// Number of active warps\n WARPS = (BLOCK_THREADS + WARP_THREADS - 1) / WARP_THREADS,\n };\n\n /// WarpScan utility type\n typedef WarpScan WarpScanT;\n\n /// WarpScan utility type\n typedef WarpScan WarpAggregateScanT;\n\n /// Shared memory storage layout type\n struct _TempStorage\n {\n typename WarpAggregateScanT::TempStorage inner_scan[WARPS]; ///< Buffer for warp-synchronous scans\n typename WarpScanT::TempStorage warp_scan[WARPS]; ///< Buffer for warp-synchronous scans\n T warp_aggregates[WARPS];\n T block_prefix; ///< Shared prefix for the entire thread block\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n unsigned int warp_id;\n unsigned int lane_id;\n\n\n //---------------------------------------------------------------------\n // Constructors\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ BlockScanWarpScans(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n warp_id((WARPS == 1) ? 0 : linear_tid / WARP_THREADS),\n lane_id(LaneId())\n {}\n\n\n //---------------------------------------------------------------------\n // Utility methods\n //---------------------------------------------------------------------\n\n template \n __device__ __forceinline__ void ApplyWarpAggregates(\n T &warp_prefix, ///< [out] The calling thread's partial reduction\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate, ///< [out] Threadblock-wide aggregate reduction of input items\n Int2Type addend_warp)\n {\n if (warp_id == WARP)\n warp_prefix = block_aggregate;\n\n T addend = temp_storage.warp_aggregates[WARP];\n block_aggregate = scan_op(block_aggregate, addend);\n\n ApplyWarpAggregates(warp_prefix, scan_op, block_aggregate, Int2Type());\n }\n\n template \n __device__ __forceinline__ void ApplyWarpAggregates(\n T &warp_prefix, ///< [out] The calling thread's partial reduction\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate, ///< [out] Threadblock-wide aggregate reduction of input items\n Int2Type addend_warp)\n {}\n\n\n /// Use the warp-wide aggregates to compute the calling warp's prefix. Also returns block-wide aggregate in all threads.\n template \n __device__ __forceinline__ T ComputeWarpPrefix(\n ScanOp scan_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [laneWARP_THREADS - 1 only] Warp-wide aggregate reduction of input items\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Last lane in each warp shares its warp-aggregate\n if (lane_id == WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = warp_aggregate;\n\n CTA_SYNC();\n\n // Accumulate block aggregates and save the one that is our warp's prefix\n T warp_prefix;\n block_aggregate = temp_storage.warp_aggregates[0];\n\n // Use template unrolling (since the PTX backend can't handle unrolling it for SM1x)\n ApplyWarpAggregates(warp_prefix, scan_op, block_aggregate, Int2Type<1>());\n/*\n #pragma unroll\n for (int WARP = 1; WARP < WARPS; ++WARP)\n {\n if (warp_id == WARP)\n warp_prefix = block_aggregate;\n\n T addend = temp_storage.warp_aggregates[WARP];\n block_aggregate = scan_op(block_aggregate, addend);\n }\n*/\n\n return warp_prefix;\n }\n\n\n /// Use the warp-wide aggregates and initial-value to compute the calling warp's prefix. Also returns block-wide aggregate in all threads.\n template \n __device__ __forceinline__ T ComputeWarpPrefix(\n ScanOp scan_op, ///< [in] Binary scan operator\n T warp_aggregate, ///< [in] [laneWARP_THREADS - 1 only] Warp-wide aggregate reduction of input items\n T &block_aggregate, ///< [out] Threadblock-wide aggregate reduction of input items\n const T &initial_value) ///< [in] Initial value to seed the exclusive scan\n {\n T warp_prefix = ComputeWarpPrefix(scan_op, warp_aggregate, block_aggregate);\n\n warp_prefix = scan_op(initial_value, warp_prefix);\n\n if (warp_id == 0)\n warp_prefix = initial_value;\n\n return warp_prefix;\n }\n\n //---------------------------------------------------------------------\n // Exclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n // Compute block-wide exclusive scan. The exclusive output from tid0 is invalid.\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, initial_value, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n WarpScanT my_warp_scan(temp_storage.warp_scan[warp_id]);\n\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n my_warp_scan.Scan(input, inclusive_output, exclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp. Warp prefix for warp0 is invalid.\n// T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate);\n\n//--------------------------------------------------\n // Last lane in each warp shares its warp-aggregate\n if (lane_id == WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n // Get the warp scan partial\n T warp_inclusive, warp_prefix;\n if (lane_id < WARPS)\n {\n // Scan the warpscan partials\n T warp_val = temp_storage.warp_aggregates[lane_id];\n WarpAggregateScanT(temp_storage.inner_scan[warp_id]).Scan(warp_val, warp_inclusive, warp_prefix, scan_op);\n }\n\n warp_prefix = my_warp_scan.Broadcast(warp_prefix, warp_id);\n block_aggregate = my_warp_scan.Broadcast(warp_inclusive, WARPS - 1);\n//--------------------------------------------------\n\n // Apply warp prefix to our lane's partial\n if (warp_id != 0)\n {\n exclusive_output = scan_op(warp_prefix, exclusive_output);\n if (lane_id == 0)\n exclusive_output = warp_prefix;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n WarpScanT my_warp_scan(temp_storage.warp_scan[warp_id]);\n\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n my_warp_scan.Scan(input, inclusive_output, exclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp\n// T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate, initial_value);\n\n//--------------------------------------------------\n // Last lane in each warp shares its warp-aggregate\n if (lane_id == WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n // Get the warp scan partial\n T warp_inclusive, warp_prefix;\n if (lane_id < WARPS)\n {\n // Scan the warpscan partials\n T warp_val = temp_storage.warp_aggregates[lane_id];\n WarpAggregateScanT(temp_storage.inner_scan[warp_id]).Scan(warp_val, warp_inclusive, warp_prefix, initial_value, scan_op);\n }\n\n warp_prefix = my_warp_scan.Broadcast(warp_prefix, warp_id);\n block_aggregate = my_warp_scan.Broadcast(warp_inclusive, WARPS - 1);\n//--------------------------------------------------\n\n // Apply warp prefix to our lane's partial\n exclusive_output = scan_op(warp_prefix, exclusive_output);\n if (lane_id == 0)\n exclusive_output = warp_prefix;\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n // Compute block-wide exclusive scan. The exclusive output from tid0 is invalid.\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n\n // Use the first warp to determine the thread block prefix, returning the result in lane0\n if (warp_id == 0)\n {\n T block_prefix = block_prefix_callback_op(block_aggregate);\n if (lane_id == 0)\n {\n // Share the prefix with all threads\n temp_storage.block_prefix = block_prefix;\n exclusive_output = block_prefix; // The block prefix is the exclusive output for tid0\n }\n }\n\n CTA_SYNC();\n\n // Incorporate thread block prefix into outputs\n T block_prefix = temp_storage.block_prefix;\n if (linear_tid > 0)\n {\n exclusive_output = scan_op(block_prefix, exclusive_output);\n }\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n InclusiveScan(input, inclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n WarpScanT(temp_storage.warp_scan[warp_id]).InclusiveScan(input, inclusive_output, scan_op);\n\n // Compute the warp-wide prefix and block-wide aggregate for each warp. Warp prefix for warp0 is invalid.\n T warp_prefix = ComputeWarpPrefix(scan_op, inclusive_output, block_aggregate);\n\n // Apply warp prefix to our lane's partial\n if (warp_id != 0)\n {\n inclusive_output = scan_op(warp_prefix, inclusive_output);\n }\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n T block_aggregate;\n InclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n\n // Use the first warp to determine the thread block prefix, returning the result in lane0\n if (warp_id == 0)\n {\n T block_prefix = block_prefix_callback_op(block_aggregate);\n if (lane_id == 0)\n {\n // Share the prefix with all threads\n temp_storage.block_prefix = block_prefix;\n }\n }\n\n CTA_SYNC();\n\n // Incorporate thread block prefix into outputs\n T block_prefix = temp_storage.block_prefix;\n exclusive_output = scan_op(block_prefix, exclusive_output);\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/block/specializations/block_scan_warp_scans3.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::BlockScanWarpscans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\n\n#pragma once\n\n#include \"../../util_arch.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../warp/warp_scan.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief BlockScanWarpScans provides warpscan-based variants of parallel prefix scan across a CUDA thread block.\n */\ntemplate <\n typename T,\n int BLOCK_DIM_X, ///< The thread block length in threads along the X dimension\n int BLOCK_DIM_Y, ///< The thread block length in threads along the Y dimension\n int BLOCK_DIM_Z, ///< The thread block length in threads along the Z dimension\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct BlockScanWarpScans\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// Constants\n enum\n {\n /// The thread block size in threads\n BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z,\n\n /// Number of warp threads\n INNER_WARP_THREADS = CUB_WARP_THREADS(PTX_ARCH),\n OUTER_WARP_THREADS = BLOCK_THREADS / INNER_WARP_THREADS,\n\n /// Number of outer scan warps\n OUTER_WARPS = INNER_WARP_THREADS\n };\n\n /// Outer WarpScan utility type\n typedef WarpScan OuterWarpScanT;\n\n /// Inner WarpScan utility type\n typedef WarpScan InnerWarpScanT;\n\n typedef typename OuterWarpScanT::TempStorage OuterScanArray[OUTER_WARPS];\n\n\n /// Shared memory storage layout type\n struct _TempStorage\n {\n union Aliasable\n {\n Uninitialized outer_warp_scan; ///< Buffer for warp-synchronous outer scans\n typename InnerWarpScanT::TempStorage inner_warp_scan; ///< Buffer for warp-synchronous inner scan\n\n } aliasable;\n\n T warp_aggregates[OUTER_WARPS];\n\n T block_aggregate; ///< Shared prefix for the entire thread block\n };\n\n\n /// Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Per-thread fields\n //---------------------------------------------------------------------\n\n // Thread fields\n _TempStorage &temp_storage;\n unsigned int linear_tid;\n unsigned int warp_id;\n unsigned int lane_id;\n\n\n //---------------------------------------------------------------------\n // Constructors\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ BlockScanWarpScans(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n linear_tid(RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z)),\n warp_id((OUTER_WARPS == 1) ? 0 : linear_tid / OUTER_WARP_THREADS),\n lane_id((OUTER_WARPS == 1) ? linear_tid : linear_tid % OUTER_WARP_THREADS)\n {}\n\n\n //---------------------------------------------------------------------\n // Exclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n // Compute block-wide exclusive scan. The exclusive output from tid0 is invalid.\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n ExclusiveScan(input, exclusive_output, initial_value, scan_op, block_aggregate);\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs. With no initial value, the output computed for thread0 is undefined.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n OuterWarpScanT(temp_storage.aliasable.outer_warp_scan.Alias()[warp_id]).Scan(\n input, inclusive_output, exclusive_output, scan_op);\n\n // Share outer warp total\n if (lane_id == OUTER_WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n if (linear_tid < INNER_WARP_THREADS)\n {\n T outer_warp_input = temp_storage.warp_aggregates[linear_tid];\n T outer_warp_exclusive;\n\n InnerWarpScanT(temp_storage.aliasable.inner_warp_scan).ExclusiveScan(\n outer_warp_input, outer_warp_exclusive, scan_op, block_aggregate);\n\n temp_storage.block_aggregate = block_aggregate;\n temp_storage.warp_aggregates[linear_tid] = outer_warp_exclusive;\n }\n\n CTA_SYNC();\n\n if (warp_id != 0)\n {\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n\n // Apply warp prefix to our lane's partial\n T outer_warp_exclusive = temp_storage.warp_aggregates[warp_id];\n exclusive_output = scan_op(outer_warp_exclusive, exclusive_output);\n if (lane_id == 0)\n exclusive_output = outer_warp_exclusive;\n }\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input items\n T &exclusive_output, ///< [out] Calling thread's output items (may be aliased to \\p input)\n const T &initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n OuterWarpScanT(temp_storage.aliasable.outer_warp_scan.Alias()[warp_id]).Scan(\n input, inclusive_output, exclusive_output, scan_op);\n\n // Share outer warp total\n if (lane_id == OUTER_WARP_THREADS - 1)\n {\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n }\n\n CTA_SYNC();\n\n if (linear_tid < INNER_WARP_THREADS)\n {\n T outer_warp_input = temp_storage.warp_aggregates[linear_tid];\n T outer_warp_exclusive;\n\n InnerWarpScanT(temp_storage.aliasable.inner_warp_scan).ExclusiveScan(\n outer_warp_input, outer_warp_exclusive, initial_value, scan_op, block_aggregate);\n\n temp_storage.block_aggregate = block_aggregate;\n temp_storage.warp_aggregates[linear_tid] = outer_warp_exclusive;\n }\n\n CTA_SYNC();\n\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n\n // Apply warp prefix to our lane's partial\n T outer_warp_exclusive = temp_storage.warp_aggregates[warp_id];\n exclusive_output = scan_op(outer_warp_exclusive, exclusive_output);\n if (lane_id == 0)\n exclusive_output = outer_warp_exclusive;\n }\n\n\n /// Computes an exclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. The call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &exclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n T inclusive_output;\n OuterWarpScanT(temp_storage.aliasable.outer_warp_scan.Alias()[warp_id]).Scan(\n input, inclusive_output, exclusive_output, scan_op);\n\n // Share outer warp total\n if (lane_id == OUTER_WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n if (linear_tid < INNER_WARP_THREADS)\n {\n InnerWarpScanT inner_scan(temp_storage.aliasable.inner_warp_scan);\n\n T upsweep = temp_storage.warp_aggregates[linear_tid];\n T downsweep_prefix, block_aggregate;\n\n inner_scan.ExclusiveScan(upsweep, downsweep_prefix, scan_op, block_aggregate);\n\n // Use callback functor to get block prefix in lane0 and then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = inner_scan.Broadcast(block_prefix, 0);\n\n downsweep_prefix = scan_op(block_prefix, downsweep_prefix);\n if (linear_tid == 0)\n downsweep_prefix = block_prefix;\n\n temp_storage.warp_aggregates[linear_tid] = downsweep_prefix;\n }\n\n CTA_SYNC();\n\n // Apply warp prefix to our lane's partial (or assign it if partial is invalid)\n T outer_warp_exclusive = temp_storage.warp_aggregates[warp_id];\n exclusive_output = scan_op(outer_warp_exclusive, exclusive_output);\n if (lane_id == 0)\n exclusive_output = outer_warp_exclusive;\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive scans\n //---------------------------------------------------------------------\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n T block_aggregate;\n InclusiveScan(input, inclusive_output, scan_op, block_aggregate);\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. Also provides every thread with the block-wide \\p block_aggregate of all inputs.\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T &block_aggregate) ///< [out] Threadblock-wide aggregate reduction of input items\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n OuterWarpScanT(temp_storage.aliasable.outer_warp_scan.Alias()[warp_id]).InclusiveScan(\n input, inclusive_output, scan_op);\n\n // Share outer warp total\n if (lane_id == OUTER_WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n if (linear_tid < INNER_WARP_THREADS)\n {\n T outer_warp_input = temp_storage.warp_aggregates[linear_tid];\n T outer_warp_exclusive;\n\n InnerWarpScanT(temp_storage.aliasable.inner_warp_scan).ExclusiveScan(\n outer_warp_input, outer_warp_exclusive, scan_op, block_aggregate);\n\n temp_storage.block_aggregate = block_aggregate;\n temp_storage.warp_aggregates[linear_tid] = outer_warp_exclusive;\n }\n\n CTA_SYNC();\n\n if (warp_id != 0)\n {\n // Retrieve block aggregate\n block_aggregate = temp_storage.block_aggregate;\n\n // Apply warp prefix to our lane's partial\n T outer_warp_exclusive = temp_storage.warp_aggregates[warp_id];\n inclusive_output = scan_op(outer_warp_exclusive, inclusive_output);\n }\n }\n\n\n /// Computes an inclusive thread block-wide prefix scan using the specified binary \\p scan_op functor. Each thread contributes one input element. the call-back functor \\p block_prefix_callback_op is invoked by the first warp in the block, and the value returned by lane0 in that warp is used as the \"seed\" value that logically prefixes the thread block's scan inputs.\n template <\n typename ScanOp,\n typename BlockPrefixCallbackOp>\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item\n T &inclusive_output, ///< [out] Calling thread's output item (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n BlockPrefixCallbackOp &block_prefix_callback_op) ///< [in-out] [warp0 only] Call-back functor for specifying a thread block-wide prefix to be applied to all inputs.\n {\n // Compute warp scan in each warp. The exclusive output from each lane0 is invalid.\n OuterWarpScanT(temp_storage.aliasable.outer_warp_scan.Alias()[warp_id]).InclusiveScan(\n input, inclusive_output, scan_op);\n\n // Share outer warp total\n if (lane_id == OUTER_WARP_THREADS - 1)\n temp_storage.warp_aggregates[warp_id] = inclusive_output;\n\n CTA_SYNC();\n\n if (linear_tid < INNER_WARP_THREADS)\n {\n InnerWarpScanT inner_scan(temp_storage.aliasable.inner_warp_scan);\n\n T upsweep = temp_storage.warp_aggregates[linear_tid];\n T downsweep_prefix, block_aggregate;\n inner_scan.ExclusiveScan(upsweep, downsweep_prefix, scan_op, block_aggregate);\n\n // Use callback functor to get block prefix in lane0 and then broadcast to other lanes\n T block_prefix = block_prefix_callback_op(block_aggregate);\n block_prefix = inner_scan.Broadcast(block_prefix, 0);\n\n downsweep_prefix = scan_op(block_prefix, downsweep_prefix);\n if (linear_tid == 0)\n downsweep_prefix = block_prefix;\n\n temp_storage.warp_aggregates[linear_tid] = downsweep_prefix;\n }\n\n CTA_SYNC();\n\n // Apply warp prefix to our lane's partial\n T outer_warp_exclusive = temp_storage.warp_aggregates[warp_id];\n inclusive_output = scan_op(outer_warp_exclusive, inclusive_output);\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/cub.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * CUB umbrella include file\n */\n\n#pragma once\n\n\n// Block\n#include \"block/block_histogram.cuh\"\n#include \"block/block_discontinuity.cuh\"\n#include \"block/block_exchange.cuh\"\n#include \"block/block_load.cuh\"\n#include \"block/block_radix_rank.cuh\"\n#include \"block/block_radix_sort.cuh\"\n#include \"block/block_reduce.cuh\"\n#include \"block/block_scan.cuh\"\n#include \"block/block_store.cuh\"\n//#include \"block/block_shift.cuh\"\n\n// Device\n#include \"device/device_histogram.cuh\"\n#include \"device/device_partition.cuh\"\n#include \"device/device_radix_sort.cuh\"\n#include \"device/device_reduce.cuh\"\n#include \"device/device_run_length_encode.cuh\"\n#include \"device/device_scan.cuh\"\n#include \"device/device_segmented_radix_sort.cuh\"\n#include \"device/device_segmented_reduce.cuh\"\n#include \"device/device_select.cuh\"\n#include \"device/device_spmv.cuh\"\n\n// Grid\n//#include \"grid/grid_barrier.cuh\"\n#include \"grid/grid_even_share.cuh\"\n#include \"grid/grid_mapping.cuh\"\n#include \"grid/grid_queue.cuh\"\n\n// Thread\n#include \"thread/thread_load.cuh\"\n#include \"thread/thread_operators.cuh\"\n#include \"thread/thread_reduce.cuh\"\n#include \"thread/thread_scan.cuh\"\n#include \"thread/thread_store.cuh\"\n\n// Warp\n#include \"warp/warp_reduce.cuh\"\n#include \"warp/warp_scan.cuh\"\n\n// Iterator\n#include \"iterator/arg_index_input_iterator.cuh\"\n#include \"iterator/cache_modified_input_iterator.cuh\"\n#include \"iterator/cache_modified_output_iterator.cuh\"\n#include \"iterator/constant_input_iterator.cuh\"\n#include \"iterator/counting_input_iterator.cuh\"\n#include \"iterator/tex_obj_input_iterator.cuh\"\n#include \"iterator/tex_ref_input_iterator.cuh\"\n#include \"iterator/transform_input_iterator.cuh\"\n\n// Util\n#include \"util_arch.cuh\"\n#include \"util_debug.cuh\"\n#include \"util_device.cuh\"\n#include \"util_macro.cuh\"\n#include \"util_ptx.cuh\"\n#include \"util_type.cuh\"\n\n" }, { "path": "external/cub/cub/device/device_histogram.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceHistogram provides device-wide parallel operations for constructing histogram(s) from a sequence of samples data residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n#include \n\n#include \"dispatch/dispatch_histogram.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceHistogram provides device-wide parallel operations for constructing histogram(s) from a sequence of samples data residing within device-accessible memory. ![](histogram_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * A histogram\n * counts the number of observations that fall into each of the disjoint categories (known as bins).\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceHistogram}\n *\n */\nstruct DeviceHistogram\n{\n /******************************************************************//**\n * \\name Evenly-segmented bin ranges\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes an intensity histogram from a sequence of data samples using equal-width bins.\n *\n * \\par\n * - The number of histogram bins is (\\p num_levels - 1)\n * - All bins comprise the same width of sample values: (\\p upper_level - \\p lower_level) / (\\p num_levels - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of a six-bin histogram\n * from a sequence of float samples\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples and\n * // output histogram\n * int num_samples; // e.g., 10\n * float* d_samples; // e.g., [2.2, 6.0, 7.1, 2.9, 3.5, 0.3, 2.9, 2.0, 6.1, 999.5]\n * int* d_histogram; // e.g., [ -, -, -, -, -, -, -, -]\n * int num_levels; // e.g., 7 (seven level boundaries for six bins)\n * float lower_level; // e.g., 0.0 (lower sample value boundary of lowest bin)\n * float upper_level; // e.g., 12.0 (upper sample value boundary of upper bin)\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::HistogramEven(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level, num_samples);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::HistogramEven(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level, num_samples);\n *\n * // d_histogram <-- [1, 0, 5, 0, 3, 0, 0, 0];\n *\n * \\endcode\n *\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t HistogramEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of data samples.\n CounterT* d_histogram, ///< [out] The pointer to the histogram counter output array of length num_levels - 1.\n int num_levels, ///< [in] The number of boundaries (levels) for delineating histogram samples. Implies that the number of bins is num_levels - 1.\n LevelT lower_level, ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin.\n LevelT upper_level, ///< [in] The upper sample value bound (exclusive) for the highest histogram bin.\n OffsetT num_samples, ///< [in] The number of input samples (i.e., the length of \\p d_samples)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n CounterT* d_histogram1[1] = {d_histogram};\n int num_levels1[1] = {num_levels};\n LevelT lower_level1[1] = {lower_level};\n LevelT upper_level1[1] = {upper_level};\n\n return MultiHistogramEven<1, 1>(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram1,\n num_levels1,\n lower_level1,\n upper_level1,\n num_samples,\n 1,\n sizeof(SampleT) * num_samples,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes an intensity histogram from a sequence of data samples using equal-width bins.\n *\n * \\par\n * - A two-dimensional region of interest within \\p d_samples can be specified\n * using the \\p num_row_samples, num_rows, and \\p row_stride_bytes parameters.\n * - The row stride must be a whole multiple of the sample data type\n * size, i.e., (row_stride_bytes % sizeof(SampleT)) == 0.\n * - The number of histogram bins is (\\p num_levels - 1)\n * - All bins comprise the same width of sample values: (\\p upper_level - \\p lower_level) / (\\p num_levels - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of a six-bin histogram\n * from a 2x5 region of interest within a flattened 2x7 array of float samples.\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples and\n * // output histogram\n * int num_row_samples; // e.g., 5\n * int num_rows; // e.g., 2;\n * size_t row_stride_bytes; // e.g., 7 * sizeof(float)\n * float* d_samples; // e.g., [2.2, 6.0, 7.1, 2.9, 3.5, -, -,\n * // 0.3, 2.9, 2.0, 6.1, 999.5, -, -]\n * int* d_histogram; // e.g., [ -, -, -, -, -, -, -, -]\n * int num_levels; // e.g., 7 (seven level boundaries for six bins)\n * float lower_level; // e.g., 0.0 (lower sample value boundary of lowest bin)\n * float upper_level; // e.g., 12.0 (upper sample value boundary of upper bin)\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::HistogramEven(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level,\n * num_row_samples, num_rows, row_stride_bytes);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::HistogramEven(d_temp_storage, temp_storage_bytes, d_samples, d_histogram,\n * d_samples, d_histogram, num_levels, lower_level, upper_level,\n * num_row_samples, num_rows, row_stride_bytes);\n *\n * // d_histogram <-- [1, 0, 5, 0, 3, 0, 0, 0];\n *\n * \\endcode\n *\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t HistogramEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of data samples.\n CounterT* d_histogram, ///< [out] The pointer to the histogram counter output array of length num_levels - 1.\n int num_levels, ///< [in] The number of boundaries (levels) for delineating histogram samples. Implies that the number of bins is num_levels - 1.\n LevelT lower_level, ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin.\n LevelT upper_level, ///< [in] The upper sample value bound (exclusive) for the highest histogram bin.\n OffsetT num_row_samples, ///< [in] The number of data samples per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n size_t row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n CounterT* d_histogram1[1] = {d_histogram};\n int num_levels1[1] = {num_levels};\n LevelT lower_level1[1] = {lower_level};\n LevelT upper_level1[1] = {upper_level};\n\n return MultiHistogramEven<1, 1>(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram1,\n num_levels1,\n lower_level1,\n upper_level1,\n num_row_samples,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n\n /**\n * \\brief Computes per-channel intensity histograms from a sequence of multi-channel \"pixel\" data samples using equal-width bins.\n *\n * \\par\n * - The input is a sequence of pixel structures, where each pixel comprises\n * a record of \\p NUM_CHANNELS consecutive data samples (e.g., an RGBA pixel).\n * - Of the \\p NUM_CHANNELS specified, the function will only compute histograms\n * for the first \\p NUM_ACTIVE_CHANNELS (e.g., only RGB histograms from RGBA\n * pixel samples).\n * - The number of histogram bins for channeli is num_levels[i] - 1.\n * - For channeli, the range of values for all histogram bins\n * have the same width: (upper_level[i] - lower_level[i]) / ( num_levels[i] - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of three 256-bin RGB histograms\n * from a quad-channel sequence of RGBA pixels (8 bits per channel per pixel)\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples\n * // and output histograms\n * int num_pixels; // e.g., 5\n * unsigned char* d_samples; // e.g., [(2, 6, 7, 5), (3, 0, 2, 1), (7, 0, 6, 2),\n * // (0, 6, 7, 5), (3, 0, 2, 6)]\n * int* d_histogram[3]; // e.g., three device pointers to three device buffers,\n * // each allocated with 256 integer counters\n * int num_levels[3]; // e.g., {257, 257, 257};\n * unsigned int lower_level[3]; // e.g., {0, 0, 0};\n * unsigned int upper_level[3]; // e.g., {256, 256, 256};\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::MultiHistogramEven<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level, num_pixels);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::MultiHistogramEven<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level, num_pixels);\n *\n * // d_histogram <-- [ [1, 0, 1, 2, 0, 0, 0, 1, 0, 0, 0, ..., 0],\n * // [0, 3, 0, 0, 0, 0, 2, 0, 0, 0, 0, ..., 0],\n * // [0, 0, 2, 0, 0, 0, 1, 2, 0, 0, 0, ..., 0] ]\n *\n * \\endcode\n *\n * \\tparam NUM_CHANNELS Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n * \\tparam NUM_ACTIVE_CHANNELS [inferred] Number of channels actively being histogrammed\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t MultiHistogramEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histogram[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_pixels, ///< [in] The number of multi-channel pixels (i.e., the length of \\p d_samples / NUM_CHANNELS)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n return MultiHistogramEven(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram,\n num_levels,\n lower_level,\n upper_level,\n num_pixels,\n 1,\n sizeof(SampleT) * NUM_CHANNELS * num_pixels,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes per-channel intensity histograms from a sequence of multi-channel \"pixel\" data samples using equal-width bins.\n *\n * \\par\n * - The input is a sequence of pixel structures, where each pixel comprises\n * a record of \\p NUM_CHANNELS consecutive data samples (e.g., an RGBA pixel).\n * - Of the \\p NUM_CHANNELS specified, the function will only compute histograms\n * for the first \\p NUM_ACTIVE_CHANNELS (e.g., only RGB histograms from RGBA\n * pixel samples).\n * - A two-dimensional region of interest within \\p d_samples can be specified\n * using the \\p num_row_samples, num_rows, and \\p row_stride_bytes parameters.\n * - The row stride must be a whole multiple of the sample data type\n * size, i.e., (row_stride_bytes % sizeof(SampleT)) == 0.\n * - The number of histogram bins for channeli is num_levels[i] - 1.\n * - For channeli, the range of values for all histogram bins\n * have the same width: (upper_level[i] - lower_level[i]) / ( num_levels[i] - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of three 256-bin RGB histograms from a 2x3 region of\n * interest of within a flattened 2x4 array of quad-channel RGBA pixels (8 bits per channel per pixel).\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples\n * // and output histograms\n * int num_row_pixels; // e.g., 3\n * int num_rows; // e.g., 2\n * size_t row_stride_bytes; // e.g., 4 * sizeof(unsigned char) * NUM_CHANNELS\n * unsigned char* d_samples; // e.g., [(2, 6, 7, 5), (3, 0, 2, 1), (7, 0, 6, 2), (-, -, -, -),\n * // (0, 6, 7, 5), (3, 0, 2, 6), (1, 1, 1, 1), (-, -, -, -)]\n * int* d_histogram[3]; // e.g., three device pointers to three device buffers,\n * // each allocated with 256 integer counters\n * int num_levels[3]; // e.g., {257, 257, 257};\n * unsigned int lower_level[3]; // e.g., {0, 0, 0};\n * unsigned int upper_level[3]; // e.g., {256, 256, 256};\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::MultiHistogramEven<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level,\n * num_row_pixels, num_rows, row_stride_bytes);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::MultiHistogramEven<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, lower_level, upper_level,\n * num_row_pixels, num_rows, row_stride_bytes);\n *\n * // d_histogram <-- [ [1, 1, 1, 2, 0, 0, 0, 1, 0, 0, 0, ..., 0],\n * // [0, 4, 0, 0, 0, 0, 2, 0, 0, 0, 0, ..., 0],\n * // [0, 1, 2, 0, 0, 0, 1, 2, 0, 0, 0, ..., 0] ]\n *\n * \\endcode\n *\n * \\tparam NUM_CHANNELS Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n * \\tparam NUM_ACTIVE_CHANNELS [inferred] Number of channels actively being histogrammed\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t MultiHistogramEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histogram[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n size_t row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n Int2Type is_byte_sample;\n\n if ((sizeof(OffsetT) > sizeof(int)) &&\n ((unsigned long long) (num_rows * row_stride_bytes) < (unsigned long long) std::numeric_limits::max()))\n {\n // Down-convert OffsetT data type\n\n\n return DipatchHistogram::DispatchEven(\n d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, lower_level, upper_level,\n (int) num_row_pixels, (int) num_rows, (int) (row_stride_bytes / sizeof(SampleT)),\n stream, debug_synchronous, is_byte_sample);\n }\n\n return DipatchHistogram::DispatchEven(\n d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, (OffsetT) (row_stride_bytes / sizeof(SampleT)),\n stream, debug_synchronous, is_byte_sample);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Custom bin ranges\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes an intensity histogram from a sequence of data samples using the specified bin boundary levels.\n *\n * \\par\n * - The number of histogram bins is (\\p num_levels - 1)\n * - The value range for bini is [level[i], level[i+1])\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of an six-bin histogram\n * from a sequence of float samples\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples and\n * // output histogram\n * int num_samples; // e.g., 10\n * float* d_samples; // e.g., [2.2, 6.0, 7.1, 2.9, 3.5, 0.3, 2.9, 2.0, 6.1, 999.5]\n * int* d_histogram; // e.g., [ -, -, -, -, -, -, -, -]\n * int num_levels // e.g., 7 (seven level boundaries for six bins)\n * float* d_levels; // e.g., [0.0, 2.0, 4.0, 6.0, 8.0, 12.0, 16.0]\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::HistogramRange(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_samples);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::HistogramRange(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_samples);\n *\n * // d_histogram <-- [1, 0, 5, 0, 3, 0, 0, 0];\n *\n * \\endcode\n *\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t HistogramRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of data samples.\n CounterT* d_histogram, ///< [out] The pointer to the histogram counter output array of length num_levels - 1.\n int num_levels, ///< [in] The number of boundaries (levels) for delineating histogram samples. Implies that the number of bins is num_levels - 1.\n LevelT* d_levels, ///< [in] The pointer to the array of boundaries (levels). Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_samples, ///< [in] The number of data samples per row in the region of interest\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n CounterT* d_histogram1[1] = {d_histogram};\n int num_levels1[1] = {num_levels};\n LevelT* d_levels1[1] = {d_levels};\n\n return MultiHistogramRange<1, 1>(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram1,\n num_levels1,\n d_levels1,\n num_samples,\n 1,\n sizeof(SampleT) * num_samples,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes an intensity histogram from a sequence of data samples using the specified bin boundary levels.\n *\n * \\par\n * - A two-dimensional region of interest within \\p d_samples can be specified\n * using the \\p num_row_samples, num_rows, and \\p row_stride_bytes parameters.\n * - The row stride must be a whole multiple of the sample data type\n * size, i.e., (row_stride_bytes % sizeof(SampleT)) == 0.\n * - The number of histogram bins is (\\p num_levels - 1)\n * - The value range for bini is [level[i], level[i+1])\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of a six-bin histogram\n * from a 2x5 region of interest within a flattened 2x7 array of float samples.\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples and\n * // output histogram\n * int num_row_samples; // e.g., 5\n * int num_rows; // e.g., 2;\n * int row_stride_bytes; // e.g., 7 * sizeof(float)\n * float* d_samples; // e.g., [2.2, 6.0, 7.1, 2.9, 3.5, -, -,\n * // 0.3, 2.9, 2.0, 6.1, 999.5, -, -]\n * int* d_histogram; // e.g., [ , , , , , , , ]\n * int num_levels // e.g., 7 (seven level boundaries for six bins)\n * float *d_levels; // e.g., [0.0, 2.0, 4.0, 6.0, 8.0, 12.0, 16.0]\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::HistogramRange(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels,\n * num_row_samples, num_rows, row_stride_bytes);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::HistogramRange(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels,\n * num_row_samples, num_rows, row_stride_bytes);\n *\n * // d_histogram <-- [1, 0, 5, 0, 3, 0, 0, 0];\n *\n * \\endcode\n *\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t HistogramRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of data samples.\n CounterT* d_histogram, ///< [out] The pointer to the histogram counter output array of length num_levels - 1.\n int num_levels, ///< [in] The number of boundaries (levels) for delineating histogram samples. Implies that the number of bins is num_levels - 1.\n LevelT* d_levels, ///< [in] The pointer to the array of boundaries (levels). Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_samples, ///< [in] The number of data samples per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n size_t row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n CounterT* d_histogram1[1] = {d_histogram};\n int num_levels1[1] = {num_levels};\n LevelT* d_levels1[1] = {d_levels};\n\n return MultiHistogramRange<1, 1>(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram1,\n num_levels1,\n d_levels1,\n num_row_samples,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n\n /**\n * \\brief Computes per-channel intensity histograms from a sequence of multi-channel \"pixel\" data samples using the specified bin boundary levels.\n *\n * \\par\n * - The input is a sequence of pixel structures, where each pixel comprises\n * a record of \\p NUM_CHANNELS consecutive data samples (e.g., an RGBA pixel).\n * - Of the \\p NUM_CHANNELS specified, the function will only compute histograms\n * for the first \\p NUM_ACTIVE_CHANNELS (e.g., RGB histograms from RGBA\n * pixel samples).\n * - The number of histogram bins for channeli is num_levels[i] - 1.\n * - For channeli, the range of values for all histogram bins\n * have the same width: (upper_level[i] - lower_level[i]) / ( num_levels[i] - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of three 4-bin RGB histograms\n * from a quad-channel sequence of RGBA pixels (8 bits per channel per pixel)\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples\n * // and output histograms\n * int num_pixels; // e.g., 5\n * unsigned char *d_samples; // e.g., [(2, 6, 7, 5),(3, 0, 2, 1),(7, 0, 6, 2),\n * // (0, 6, 7, 5),(3, 0, 2, 6)]\n * unsigned int *d_histogram[3]; // e.g., [[ -, -, -, -],[ -, -, -, -],[ -, -, -, -]];\n * int num_levels[3]; // e.g., {5, 5, 5};\n * unsigned int *d_levels[3]; // e.g., [ [0, 2, 4, 6, 8],\n * // [0, 2, 4, 6, 8],\n * // [0, 2, 4, 6, 8] ];\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::MultiHistogramRange<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_pixels);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::MultiHistogramRange<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_pixels);\n *\n * // d_histogram <-- [ [1, 3, 0, 1],\n * // [3, 0, 0, 2],\n * // [0, 2, 0, 3] ]\n *\n * \\endcode\n *\n * \\tparam NUM_CHANNELS Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n * \\tparam NUM_ACTIVE_CHANNELS [inferred] Number of channels actively being histogrammed\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t MultiHistogramRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histogram[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT* d_levels[NUM_ACTIVE_CHANNELS], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_pixels, ///< [in] The number of multi-channel pixels (i.e., the length of \\p d_samples / NUM_CHANNELS)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n return MultiHistogramRange(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram,\n num_levels,\n d_levels,\n num_pixels,\n 1,\n sizeof(SampleT) * NUM_CHANNELS * num_pixels,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes per-channel intensity histograms from a sequence of multi-channel \"pixel\" data samples using the specified bin boundary levels.\n *\n * \\par\n * - The input is a sequence of pixel structures, where each pixel comprises\n * a record of \\p NUM_CHANNELS consecutive data samples (e.g., an RGBA pixel).\n * - Of the \\p NUM_CHANNELS specified, the function will only compute histograms\n * for the first \\p NUM_ACTIVE_CHANNELS (e.g., RGB histograms from RGBA\n * pixel samples).\n * - A two-dimensional region of interest within \\p d_samples can be specified\n * using the \\p num_row_samples, num_rows, and \\p row_stride_bytes parameters.\n * - The row stride must be a whole multiple of the sample data type\n * size, i.e., (row_stride_bytes % sizeof(SampleT)) == 0.\n * - The number of histogram bins for channeli is num_levels[i] - 1.\n * - For channeli, the range of values for all histogram bins\n * have the same width: (upper_level[i] - lower_level[i]) / ( num_levels[i] - 1)\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the computation of three 4-bin RGB histograms from a 2x3 region of\n * interest of within a flattened 2x4 array of quad-channel RGBA pixels (8 bits per channel per pixel).\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input samples\n * // and output histograms\n * int num_row_pixels; // e.g., 3\n * int num_rows; // e.g., 2\n * size_t row_stride_bytes; // e.g., 4 * sizeof(unsigned char) * NUM_CHANNELS\n * unsigned char* d_samples; // e.g., [(2, 6, 7, 5),(3, 0, 2, 1),(1, 1, 1, 1),(-, -, -, -),\n * // (7, 0, 6, 2),(0, 6, 7, 5),(3, 0, 2, 6),(-, -, -, -)]\n * int* d_histogram[3]; // e.g., [[ -, -, -, -],[ -, -, -, -],[ -, -, -, -]];\n * int num_levels[3]; // e.g., {5, 5, 5};\n * unsigned int* d_levels[3]; // e.g., [ [0, 2, 4, 6, 8],\n * // [0, 2, 4, 6, 8],\n * // [0, 2, 4, 6, 8] ];\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceHistogram::MultiHistogramRange<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_row_pixels, num_rows, row_stride_bytes);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Compute histograms\n * cub::DeviceHistogram::MultiHistogramRange<4, 3>(d_temp_storage, temp_storage_bytes,\n * d_samples, d_histogram, num_levels, d_levels, num_row_pixels, num_rows, row_stride_bytes);\n *\n * // d_histogram <-- [ [2, 3, 0, 1],\n * // [3, 0, 0, 2],\n * // [1, 2, 0, 3] ]\n *\n * \\endcode\n *\n * \\tparam NUM_CHANNELS Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n * \\tparam NUM_ACTIVE_CHANNELS [inferred] Number of channels actively being histogrammed\n * \\tparam SampleIteratorT [inferred] Random-access input iterator type for reading input samples. \\iterator\n * \\tparam CounterT [inferred] Integer type for histogram bin counters\n * \\tparam LevelT [inferred] Type for specifying boundaries (levels)\n * \\tparam OffsetT [inferred] Signed integer type for sequence offsets, list lengths, pointer differences, etc. \\offset_size1\n */\n template <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleIteratorT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t MultiHistogramRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histogram[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT* d_levels[NUM_ACTIVE_CHANNELS], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n size_t row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n Int2Type is_byte_sample;\n\n if ((sizeof(OffsetT) > sizeof(int)) &&\n ((unsigned long long) (num_rows * row_stride_bytes) < (unsigned long long) std::numeric_limits::max()))\n {\n // Down-convert OffsetT data type\n return DipatchHistogram::DispatchRange(\n d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, d_levels,\n (int) num_row_pixels, (int) num_rows, (int) (row_stride_bytes / sizeof(SampleT)),\n stream, debug_synchronous, is_byte_sample);\n }\n\n return DipatchHistogram::DispatchRange(\n d_temp_storage, temp_storage_bytes, d_samples, d_histogram, num_levels, d_levels,\n num_row_pixels, num_rows, (OffsetT) (row_stride_bytes / sizeof(SampleT)),\n stream, debug_synchronous, is_byte_sample);\n }\n\n\n\n //@} end member group\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_partition.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DevicePartition provides device-wide, parallel operations for partitioning sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_select_if.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DevicePartition provides device-wide, parallel operations for partitioning sequences of data items residing within device-accessible memory. ![](partition_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * These operations apply a selection criterion to construct a partitioned output sequence from items selected/unselected from\n * a specified input sequence.\n *\n * \\par Usage Considerations\n * \\cdp_class{DevicePartition}\n *\n * \\par Performance\n * \\linear_performance{partition}\n *\n * \\par\n * The following chart illustrates DevicePartition::If\n * performance across different CUDA architectures for \\p int32 items,\n * where 50% of the items are randomly selected for the first partition.\n * \\plots_below\n *\n * \\image html partition_if_int32_50_percent.png\n *\n */\nstruct DevicePartition\n{\n /**\n * \\brief Uses the \\p d_flags sequence to split the corresponding items from \\p d_in into a partitioned sequence \\p d_out. The total number of items copied into the first partition is written to \\p d_num_selected_out. ![](partition_flags_logo.png)\n *\n * \\par\n * - The value type of \\p d_flags must be castable to \\p bool (e.g., \\p bool, \\p char, \\p int, etc.).\n * - Copies of the selected items are compacted into \\p d_out and maintain their original\n * relative ordering, however copies of the unselected items are compacted into the\n * rear of \\p d_out in reverse order.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the compaction of items selected from an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input, flags, and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [1, 2, 3, 4, 5, 6, 7, 8]\n * char *d_flags; // e.g., [1, 0, 0, 1, 0, 1, 1, 0]\n * int *d_out; // e.g., [ , , , , , , , ]\n * int *d_num_selected_out; // e.g., [ ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run selection\n * cub::DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items);\n *\n * // d_out <-- [1, 4, 6, 7, 8, 5, 3, 2]\n * // d_num_selected_out <-- [4]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam FlagIterator [inferred] Random-access input iterator type for reading selection flags \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing output items \\iterator\n * \\tparam NumSelectedIteratorT [inferred] Output iterator type for recording the number of items selected \\iterator\n */\n template <\n typename InputIteratorT,\n typename FlagIterator,\n typename OutputIteratorT,\n typename NumSelectedIteratorT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Flagged(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n FlagIterator d_flags, ///< [in] Pointer to the input sequence of selection flags\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of partitioned data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the output total number of items selected (i.e., the offset of the unselected partition)\n int num_items, ///< [in] Total number of items to select from\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType SelectOp; // Selection op (not used)\n typedef NullType EqualityOp; // Equality operator (not used)\n\n return DispatchSelectIf::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_flags,\n d_out,\n d_num_selected_out,\n SelectOp(),\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Uses the \\p select_op functor to split the corresponding items from \\p d_in into a partitioned sequence \\p d_out. The total number of items copied into the first partition is written to \\p d_num_selected_out. ![](partition_logo.png)\n *\n * \\par\n * - Copies of the selected items are compacted into \\p d_out and maintain their original\n * relative ordering, however copies of the unselected items are compacted into the\n * rear of \\p d_out in reverse order.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated partition-if performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively. Items are\n * selected for the first partition with 50% probability.\n *\n * \\image html partition_if_int32_50_percent.png\n * \\image html partition_if_int64_50_percent.png\n *\n * \\par\n * The following charts are similar, but 5% selection probability for the first partition:\n *\n * \\image html partition_if_int32_5_percent.png\n * \\image html partition_if_int64_5_percent.png\n *\n * \\par Snippet\n * The code snippet below illustrates the compaction of items selected from an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Functor type for selecting values less than some criteria\n * struct LessThan\n * {\n * int compare;\n *\n * CUB_RUNTIME_FUNCTION __forceinline__\n * LessThan(int compare) : compare(compare) {}\n *\n * CUB_RUNTIME_FUNCTION __forceinline__\n * bool operator()(const int &a) const {\n * return (a < compare);\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [0, 2, 3, 9, 5, 2, 81, 8]\n * int *d_out; // e.g., [ , , , , , , , ]\n * int *d_num_selected_out; // e.g., [ ]\n * LessThan select_op(7);\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run selection\n * cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op);\n *\n * // d_out <-- [0, 2, 3, 5, 2, 8, 81, 9]\n * // d_num_selected_out <-- [5]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing output items \\iterator\n * \\tparam NumSelectedIteratorT [inferred] Output iterator type for recording the number of items selected \\iterator\n * \\tparam SelectOp [inferred] Selection functor type having member bool operator()(const T &a)\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename NumSelectedIteratorT,\n typename SelectOp>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t If(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of partitioned data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the output total number of items selected (i.e., the offset of the unselected partition)\n int num_items, ///< [in] Total number of items to select from\n SelectOp select_op, ///< [in] Unary selection operator\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType* FlagIterator; // FlagT iterator type (not used)\n typedef NullType EqualityOp; // Equality operator (not used)\n\n return DispatchSelectIf::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n NULL,\n d_out,\n d_num_selected_out,\n select_op,\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n};\n\n/**\n * \\example example_device_partition_flagged.cu\n * \\example example_device_partition_if.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_radix_sort.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceRadixSort provides device-wide, parallel operations for computing a radix sort across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_radix_sort.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceRadixSort provides device-wide, parallel operations for computing a radix sort across a sequence of data items residing within device-accessible memory. ![](sorting_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * The [radix sorting method](http://en.wikipedia.org/wiki/Radix_sort) arranges\n * items into ascending (or descending) order. The algorithm relies upon a positional representation for\n * keys, i.e., each key is comprised of an ordered sequence of symbols (e.g., digits,\n * characters, etc.) specified from least-significant to most-significant. For a\n * given input sequence of keys and a set of rules specifying a total ordering\n * of the symbolic alphabet, the radix sorting method produces a lexicographic\n * ordering of those keys.\n *\n * \\par\n * DeviceRadixSort can sort all of the built-in C++ numeric primitive types, e.g.:\n * unsigned char, \\p int, \\p double, etc. Although the direct radix sorting\n * method can only be applied to unsigned integral types, DeviceRadixSort\n * is able to sort signed and floating-point types via simple bit-wise transformations\n * that ensure lexicographic key ordering.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceRadixSort}\n *\n * \\par Performance\n * \\linear_performance{radix sort} The following chart illustrates DeviceRadixSort::SortKeys\n * performance across different CUDA architectures for uniform-random \\p uint32 keys.\n * \\plots_below\n *\n * \\image html lsb_radix_sort_int32_keys.png\n *\n */\nstruct DeviceRadixSort\n{\n\n /******************************************************************//**\n * \\name KeyT-value pairs\n *********************************************************************/\n //@{\n\n /**\n * \\brief Sorts key-value pairs into ascending order. (~2N auxiliary storage required)\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated sorting performance across different\n * CUDA architectures for uniform-random uint32,uint32 and\n * uint64,uint64 pairs, respectively.\n *\n * \\image html lsb_radix_sort_int32_pairs.png\n * \\image html lsb_radix_sort_int64_pairs.png\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [ ... ]\n * int *d_values_in; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_values_out; // e.g., [ ... ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out, num_items);\n *\n * // d_keys_out <-- [0, 3, 5, 6, 7, 8, 9]\n * // d_values_out <-- [5, 4, 3, 1, 2, 0, 6]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n * \\tparam ValueT [inferred] ValueT type\n */\n template <\n typename KeyT,\n typename ValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairs(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] Pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] Pointer to the sorted output sequence of key data\n const ValueT *d_values_in, ///< [in] Pointer to the corresponding input sequence of associated value items\n ValueT *d_values_out, ///< [out] Pointer to the correspondingly-reordered output sequence of associated value items\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values(const_cast(d_values_in), d_values_out);\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts key-value pairs into ascending order. (~N auxiliary storage required)\n *\n * \\par\n * - The sorting operation is given a pair of key buffers and a corresponding\n * pair of associated value buffers. Each pair is managed by a DoubleBuffer\n * structure that indicates which of the two buffers is \"current\" (and thus\n * contains the input data to be sorted).\n * - The contents of both buffers within each pair may be altered by the sorting\n * operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within each DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated sorting performance across different\n * CUDA architectures for uniform-random uint32,uint32 and\n * uint64,uint64 pairs, respectively.\n *\n * \\image html lsb_radix_sort_int32_pairs.png\n * \\image html lsb_radix_sort_int64_pairs.png\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [ ... ]\n * int *d_value_buf; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_value_alt_buf; // e.g., [ ... ]\n * ...\n *\n * // Create a set of DoubleBuffers to wrap pairs of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n * cub::DoubleBuffer d_values(d_value_buf, d_value_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items);\n *\n * // d_keys.Current() <-- [0, 3, 5, 6, 7, 8, 9]\n * // d_values.Current() <-- [5, 4, 3, 1, 2, 0, 6]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n * \\tparam ValueT [inferred] ValueT type\n */\n template <\n typename KeyT,\n typename ValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairs(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer of values whose \"current\" device-accessible buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts key-value pairs into descending order. (~2N auxiliary storage required).\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Performance\n * Performance is similar to DeviceRadixSort::SortPairs.\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [ ... ]\n * int *d_values_in; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_values_out; // e.g., [ ... ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out, num_items);\n *\n * // d_keys_out <-- [9, 8, 7, 6, 5, 3, 0]\n * // d_values_out <-- [6, 0, 2, 1, 3, 4, 5]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n * \\tparam ValueT [inferred] ValueT type\n */\n template <\n typename KeyT,\n typename ValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairsDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] Pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] Pointer to the sorted output sequence of key data\n const ValueT *d_values_in, ///< [in] Pointer to the corresponding input sequence of associated value items\n ValueT *d_values_out, ///< [out] Pointer to the correspondingly-reordered output sequence of associated value items\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values(const_cast(d_values_in), d_values_out);\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts key-value pairs into descending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers and a corresponding\n * pair of associated value buffers. Each pair is managed by a DoubleBuffer\n * structure that indicates which of the two buffers is \"current\" (and thus\n * contains the input data to be sorted).\n * - The contents of both buffers within each pair may be altered by the sorting\n * operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within each DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Performance\n * Performance is similar to DeviceRadixSort::SortPairs.\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [ ... ]\n * int *d_value_buf; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_value_alt_buf; // e.g., [ ... ]\n * ...\n *\n * // Create a set of DoubleBuffers to wrap pairs of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n * cub::DoubleBuffer d_values(d_value_buf, d_value_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items);\n *\n * // d_keys.Current() <-- [9, 8, 7, 6, 5, 3, 0]\n * // d_values.Current() <-- [6, 0, 2, 1, 3, 4, 5]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n * \\tparam ValueT [inferred] ValueT type\n */\n template <\n typename KeyT,\n typename ValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairsDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer of values whose \"current\" device-accessible buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Keys-only\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Sorts keys into ascending order. (~2N auxiliary storage required)\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated sorting performance across different\n * CUDA architectures for uniform-random \\p uint32 and \\p uint64 keys, respectively.\n *\n * \\image html lsb_radix_sort_int32_keys.png\n * \\image html lsb_radix_sort_int64_keys.png\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [ ... ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, num_items);\n *\n * // d_keys_out <-- [0, 3, 5, 6, 7, 8, 9]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n */\n template \n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeys(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] Pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] Pointer to the sorted output sequence of key data\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts keys into ascending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers managed by a\n * DoubleBuffer structure that indicates which of the two buffers is\n * \"current\" (and thus contains the input data to be sorted).\n * - The contents of both buffers may be altered by the sorting operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within the DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated sorting performance across different\n * CUDA architectures for uniform-random \\p uint32 and \\p uint64 keys, respectively.\n *\n * \\image html lsb_radix_sort_int32_keys.png\n * \\image html lsb_radix_sort_int64_keys.png\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [ ... ]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys, num_items);\n *\n * // d_keys.Current() <-- [0, 3, 5, 6, 7, 8, 9]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n */\n template \n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeys(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_values;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n /**\n * \\brief Sorts keys into descending order. (~2N auxiliary storage required).\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Performance\n * Performance is similar to DeviceRadixSort::SortKeys.\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [ ... ]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, num_items);\n *\n * // d_keys_out <-- [9, 8, 7, 6, 5, 3, 0]s\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n */\n template \n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeysDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] Pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] Pointer to the sorted output sequence of key data\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts keys into descending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers managed by a\n * DoubleBuffer structure that indicates which of the two buffers is\n * \"current\" (and thus contains the input data to be sorted).\n * - The contents of both buffers may be altered by the sorting operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within the DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Performance\n * Performance is similar to DeviceRadixSort::SortKeys.\n *\n * \\par Snippet\n * The code snippet below illustrates the sorting of a device vector of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [ ... ]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys, num_items);\n *\n * // d_keys.Current() <-- [9, 8, 7, 6, 5, 3, 0]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] KeyT type\n */\n template \n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeysDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n int num_items, ///< [in] Number of items to sort\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_values;\n\n return DispatchRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n //@} end member group\n\n\n};\n\n/**\n * \\example example_device_radix_sort.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_reduce.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceReduce provides device-wide, parallel operations for computing a reduction across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n#include \n\n#include \"../iterator/arg_index_input_iterator.cuh\"\n#include \"dispatch/dispatch_reduce.cuh\"\n#include \"dispatch/dispatch_reduce_by_key.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceReduce provides device-wide, parallel operations for computing a reduction across a sequence of data items residing within device-accessible memory. ![](reduce_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * A reduction (or fold)\n * uses a binary combining operator to compute a single aggregate from a sequence of input elements.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceReduce}\n *\n * \\par Performance\n * \\linear_performance{reduction, reduce-by-key, and run-length encode}\n *\n * \\par\n * The following chart illustrates DeviceReduce::Sum\n * performance across different CUDA architectures for \\p int32 keys.\n *\n * \\image html reduce_int32.png\n *\n * \\par\n * The following chart illustrates DeviceReduce::ReduceByKey (summation)\n * performance across different CUDA architectures for \\p fp32\n * values. Segments are identified by \\p int32 keys, and have lengths uniformly sampled from [1,1000].\n *\n * \\image html reduce_by_key_fp32_len_500.png\n *\n * \\par\n * \\plots_below\n *\n */\nstruct DeviceReduce\n{\n /**\n * \\brief Computes a device-wide reduction using the specified binary \\p reduction_op functor and initial value \\p init.\n *\n * \\par\n * - Does not support binary reduction operators that are non-commutative.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates a user-defined min-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // CustomMin functor\n * struct CustomMin\n * {\n * template \n * __device__ __forceinline__\n * T operator()(const T &a, const T &b) const {\n * return (b < a) ? b : a;\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-]\n * CustomMin min_op;\n * int init; // e.g., INT_MAX\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, min_op, init);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run reduction\n * cub::DeviceReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, min_op, init);\n *\n * // d_out <-- [0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n * \\tparam ReductionOpT [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b) \n * \\tparam T [inferred] Data element type that is convertible to the \\p value type of \\p InputIteratorT\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename ReductionOpT,\n typename T>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Reduce(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n ReductionOpT reduction_op, ///< [in] Binary reduction functor\n T init, ///< [in] Initial value of the reduction\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_items,\n reduction_op,\n init,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide sum using the addition (\\p +) operator.\n *\n * \\par\n * - Uses \\p 0 as the initial value of the reduction.\n * - Does not support \\p + operators that are non-commutative..\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated sum-reduction performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively.\n *\n * \\image html reduce_int32.png\n * \\image html reduce_int64.png\n *\n * \\par Snippet\n * The code snippet below illustrates the sum-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sum-reduction\n * cub::DeviceReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // d_out <-- [38]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Sum(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_items,\n cub::Sum(),\n OutputT(), // zero-initialize\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide minimum using the less-than ('<') operator.\n *\n * \\par\n * - Uses std::numeric_limits::max() as the initial value of the reduction.\n * - Does not support \\p < operators that are non-commutative.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the min-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::Min(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run min-reduction\n * cub::DeviceReduce::Min(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // d_out <-- [0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Min(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_items,\n cub::Min(),\n Traits::Max(), // replace with std::numeric_limits::max() when C++11 support is more prevalent\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Finds the first device-wide minimum using the less-than ('<') operator, also returning the index of that item.\n *\n * \\par\n * - The output value type of \\p d_out is cub::KeyValuePair (assuming the value type of \\p d_in is \\p T)\n * - The minimum is written to d_out.value and its offset in the input array is written to d_out.key.\n * - The {1, std::numeric_limits::max()} tuple is produced for zero-length inputs\n * - Does not support \\p < operators that are non-commutative.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the argmin-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * KeyValuePair *d_out; // e.g., [{-,-}]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::ArgMin(d_temp_storage, temp_storage_bytes, d_in, d_argmin, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run argmin-reduction\n * cub::DeviceReduce::ArgMin(d_temp_storage, temp_storage_bytes, d_in, d_argmin, num_items);\n *\n * // d_out <-- [{5, 0}]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items (of some type \\p T) \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate (having value type cub::KeyValuePair) \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ArgMin(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input type\n typedef typename std::iterator_traits::value_type InputValueT;\n\n // The output tuple type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n KeyValuePair, // ... then the key value pair OffsetT + InputValueT\n typename std::iterator_traits::value_type>::Type OutputTupleT; // ... else the output iterator's value type\n\n // The output value type\n typedef typename OutputTupleT::Value OutputValueT;\n\n // Wrapped input iterator to produce index-value tuples\n typedef ArgIndexInputIterator ArgIndexInputIteratorT;\n ArgIndexInputIteratorT d_indexed_in(d_in);\n\n // Initial value\n OutputTupleT initial_value(1, Traits::Max()); // replace with std::numeric_limits::max() when C++11 support is more prevalent\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_indexed_in,\n d_out,\n num_items,\n cub::ArgMin(),\n initial_value,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide maximum using the greater-than ('>') operator.\n *\n * \\par\n * - Uses std::numeric_limits::lowest() as the initial value of the reduction.\n * - Does not support \\p > operators that are non-commutative.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the max-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::Max(d_temp_storage, temp_storage_bytes, d_in, d_max, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run max-reduction\n * cub::DeviceReduce::Max(d_temp_storage, temp_storage_bytes, d_in, d_max, num_items);\n *\n * // d_out <-- [9]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Max(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_items,\n cub::Max(),\n Traits::Lowest(), // replace with std::numeric_limits::lowest() when C++11 support is more prevalent\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Finds the first device-wide maximum using the greater-than ('>') operator, also returning the index of that item\n *\n * \\par\n * - The output value type of \\p d_out is cub::KeyValuePair (assuming the value type of \\p d_in is \\p T)\n * - The maximum is written to d_out.value and its offset in the input array is written to d_out.key.\n * - The {1, std::numeric_limits::lowest()} tuple is produced for zero-length inputs\n * - Does not support \\p > operators that are non-commutative.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the argmax-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * KeyValuePair *d_out; // e.g., [{-,-}]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::ArgMax(d_temp_storage, temp_storage_bytes, d_in, d_argmax, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run argmax-reduction\n * cub::DeviceReduce::ArgMax(d_temp_storage, temp_storage_bytes, d_in, d_argmax, num_items);\n *\n * // d_out <-- [{6, 9}]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items (of some type \\p T) \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate (having value type cub::KeyValuePair) \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ArgMax(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input type\n typedef typename std::iterator_traits::value_type InputValueT;\n\n // The output tuple type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n KeyValuePair, // ... then the key value pair OffsetT + InputValueT\n typename std::iterator_traits::value_type>::Type OutputTupleT; // ... else the output iterator's value type\n\n // The output value type\n typedef typename OutputTupleT::Value OutputValueT;\n\n // Wrapped input iterator to produce index-value tuples\n typedef ArgIndexInputIterator ArgIndexInputIteratorT;\n ArgIndexInputIteratorT d_indexed_in(d_in);\n\n // Initial value\n OutputTupleT initial_value(1, Traits::Lowest()); // replace with std::numeric_limits::lowest() when C++11 support is more prevalent\n\n return DispatchReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_indexed_in,\n d_out,\n num_items,\n cub::ArgMax(),\n initial_value,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Reduces segments of values, where segments are demarcated by corresponding runs of identical keys.\n *\n * \\par\n * This operation computes segmented reductions within \\p d_values_in using\n * the specified binary \\p reduction_op functor. The segments are identified by\n * \"runs\" of corresponding keys in \\p d_keys_in, where runs are maximal ranges of\n * consecutive, identical keys. For the ith run encountered,\n * the first key of the run and the corresponding value aggregate of that run are\n * written to d_unique_out[i] and d_aggregates_out[i],\n * respectively. The total number of runs encountered is written to \\p d_num_runs_out.\n *\n * \\par\n * - The == equality operator is used to determine whether keys are equivalent\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Performance\n * The following chart illustrates reduction-by-key (sum) performance across\n * different CUDA architectures for \\p fp32 and \\p fp64 values, respectively. Segments\n * are identified by \\p int32 keys, and have lengths uniformly sampled from [1,1000].\n *\n * \\image html reduce_by_key_fp32_len_500.png\n * \\image html reduce_by_key_fp64_len_500.png\n *\n * \\par\n * The following charts are similar, but with segment lengths uniformly sampled from [1,10]:\n *\n * \\image html reduce_by_key_fp32_len_5.png\n * \\image html reduce_by_key_fp64_len_5.png\n *\n * \\par Snippet\n * The code snippet below illustrates the segmented reduction of \\p int values grouped\n * by runs of associated \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // CustomMin functor\n * struct CustomMin\n * {\n * template \n * CUB_RUNTIME_FUNCTION __forceinline__\n * T operator()(const T &a, const T &b) const {\n * return (b < a) ? b : a;\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_keys_in; // e.g., [0, 2, 2, 9, 5, 5, 5, 8]\n * int *d_values_in; // e.g., [0, 7, 1, 6, 2, 5, 3, 4]\n * int *d_unique_out; // e.g., [-, -, -, -, -, -, -, -]\n * int *d_aggregates_out; // e.g., [-, -, -, -, -, -, -, -]\n * int *d_num_runs_out; // e.g., [-]\n * CustomMin reduction_op;\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceReduce::ReduceByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_unique_out, d_values_in, d_aggregates_out, d_num_runs_out, reduction_op, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run reduce-by-key\n * cub::DeviceReduce::ReduceByKey(d_temp_storage, temp_storage_bytes, d_keys_in, d_unique_out, d_values_in, d_aggregates_out, d_num_runs_out, reduction_op, num_items);\n *\n * // d_unique_out <-- [0, 2, 9, 5, 8]\n * // d_aggregates_out <-- [0, 1, 6, 2, 4]\n * // d_num_runs_out <-- [5]\n *\n * \\endcode\n *\n * \\tparam KeysInputIteratorT [inferred] Random-access input iterator type for reading input keys \\iterator\n * \\tparam UniqueOutputIteratorT [inferred] Random-access output iterator type for writing unique output keys \\iterator\n * \\tparam ValuesInputIteratorT [inferred] Random-access input iterator type for reading input values \\iterator\n * \\tparam AggregatesOutputIterator [inferred] Random-access output iterator type for writing output value aggregates \\iterator\n * \\tparam NumRunsOutputIteratorT [inferred] Output iterator type for recording the number of runs encountered \\iterator\n * \\tparam ReductionOpT [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b) \n */\n template <\n typename KeysInputIteratorT,\n typename UniqueOutputIteratorT,\n typename ValuesInputIteratorT,\n typename AggregatesOutputIteratorT,\n typename NumRunsOutputIteratorT,\n typename ReductionOpT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t ReduceByKey(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n KeysInputIteratorT d_keys_in, ///< [in] Pointer to the input sequence of keys\n UniqueOutputIteratorT d_unique_out, ///< [out] Pointer to the output sequence of unique keys (one key per run)\n ValuesInputIteratorT d_values_in, ///< [in] Pointer to the input sequence of corresponding values\n AggregatesOutputIteratorT d_aggregates_out, ///< [out] Pointer to the output sequence of value aggregates (one aggregate per run)\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs encountered (i.e., the length of d_unique_out)\n ReductionOpT reduction_op, ///< [in] Binary reduction functor\n int num_items, ///< [in] Total number of associated key+value pairs (i.e., the length of \\p d_in_keys and \\p d_in_values)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // FlagT iterator type (not used)\n\n // Selection op (not used)\n\n // Default == operator\n typedef Equality EqualityOp;\n\n return DispatchReduceByKey::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys_in,\n d_unique_out,\n d_values_in,\n d_aggregates_out,\n d_num_runs_out,\n EqualityOp(),\n reduction_op,\n num_items,\n stream,\n debug_synchronous);\n }\n\n};\n\n/**\n * \\example example_device_reduce.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_run_length_encode.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceRunLengthEncode provides device-wide, parallel operations for computing a run-length encoding across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_rle.cuh\"\n#include \"dispatch/dispatch_reduce_by_key.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceRunLengthEncode provides device-wide, parallel operations for demarcating \"runs\" of same-valued items within a sequence residing within device-accessible memory. ![](run_length_encode_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * A run-length encoding\n * computes a simple compressed representation of a sequence of input elements such that each\n * maximal \"run\" of consecutive same-valued data items is encoded as a single data value along with a\n * count of the elements in that run.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceRunLengthEncode}\n *\n * \\par Performance\n * \\linear_performance{run-length encode}\n *\n * \\par\n * The following chart illustrates DeviceRunLengthEncode::RunLengthEncode performance across\n * different CUDA architectures for \\p int32 items.\n * Segments have lengths uniformly sampled from [1,1000].\n *\n * \\image html rle_int32_len_500.png\n *\n * \\par\n * \\plots_below\n *\n */\nstruct DeviceRunLengthEncode\n{\n\n /**\n * \\brief Computes a run-length encoding of the sequence \\p d_in.\n *\n * \\par\n * - For the ith run encountered, the first key of the run and its length are written to\n * d_unique_out[i] and d_counts_out[i],\n * respectively.\n * - The total number of runs encountered is written to \\p d_num_runs_out.\n * - The == equality operator is used to determine whether values are equivalent\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated encode performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively. Segments have\n * lengths uniformly sampled from [1,1000].\n *\n * \\image html rle_int32_len_500.png\n * \\image html rle_int64_len_500.png\n *\n * \\par\n * The following charts are similar, but with segment lengths uniformly sampled from [1,10]:\n *\n * \\image html rle_int32_len_5.png\n * \\image html rle_int64_len_5.png\n *\n * \\par Snippet\n * The code snippet below illustrates the run-length encoding of a sequence of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [0, 2, 2, 9, 5, 5, 5, 8]\n * int *d_unique_out; // e.g., [ , , , , , , , ]\n * int *d_counts_out; // e.g., [ , , , , , , , ]\n * int *d_num_runs_out; // e.g., [ ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRunLengthEncode::Encode(d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_counts_out, d_num_runs_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run encoding\n * cub::DeviceRunLengthEncode::Encode(d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_counts_out, d_num_runs_out, num_items);\n *\n * // d_unique_out <-- [0, 2, 9, 5, 8]\n * // d_counts_out <-- [1, 2, 1, 3, 1]\n * // d_num_runs_out <-- [5]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam UniqueOutputIteratorT [inferred] Random-access output iterator type for writing unique output items \\iterator\n * \\tparam LengthsOutputIteratorT [inferred] Random-access output iterator type for writing output counts \\iterator\n * \\tparam NumRunsOutputIteratorT [inferred] Output iterator type for recording the number of runs encountered \\iterator\n */\n template <\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsOutputIteratorT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Encode(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of keys\n UniqueOutputIteratorT d_unique_out, ///< [out] Pointer to the output sequence of unique keys (one key per run)\n LengthsOutputIteratorT d_counts_out, ///< [out] Pointer to the output sequence of run-lengths (one count per run)\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs\n int num_items, ///< [in] Total number of associated key+value pairs (i.e., the length of \\p d_in_keys and \\p d_in_values)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType* FlagIterator; // FlagT iterator type (not used)\n typedef NullType SelectOp; // Selection op (not used)\n typedef Equality EqualityOp; // Default == operator\n typedef cub::Sum ReductionOp; // Value reduction operator\n\n // The lengths output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // LengthT = (if output iterator's value type is void) ?\n OffsetT, // ... then the OffsetT type,\n typename std::iterator_traits::value_type>::Type LengthT; // ... else the output iterator's value type\n\n // Generator type for providing 1s values for run-length reduction\n typedef ConstantInputIterator LengthsInputIteratorT;\n\n return DispatchReduceByKey::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_unique_out,\n LengthsInputIteratorT((LengthT) 1),\n d_counts_out,\n d_num_runs_out,\n EqualityOp(),\n ReductionOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Enumerates the starting offsets and lengths of all non-trivial runs (of length > 1) of same-valued keys in the sequence \\p d_in.\n *\n * \\par\n * - For the ith non-trivial run, the run's starting offset\n * and its length are written to d_offsets_out[i] and\n * d_lengths_out[i], respectively.\n * - The total number of runs encountered is written to \\p d_num_runs_out.\n * - The == equality operator is used to determine whether values are equivalent\n * - \\devicestorage\n *\n * \\par Performance\n *\n * \\par Snippet\n * The code snippet below illustrates the identification of non-trivial runs within a sequence of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [0, 2, 2, 9, 5, 5, 5, 8]\n * int *d_offsets_out; // e.g., [ , , , , , , , ]\n * int *d_lengths_out; // e.g., [ , , , , , , , ]\n * int *d_num_runs_out; // e.g., [ ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceRunLengthEncode::NonTrivialRuns(d_temp_storage, temp_storage_bytes, d_in, d_offsets_out, d_lengths_out, d_num_runs_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run encoding\n * cub::DeviceRunLengthEncode::NonTrivialRuns(d_temp_storage, temp_storage_bytes, d_in, d_offsets_out, d_lengths_out, d_num_runs_out, num_items);\n *\n * // d_offsets_out <-- [1, 4]\n * // d_lengths_out <-- [2, 3]\n * // d_num_runs_out <-- [2]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OffsetsOutputIteratorT [inferred] Random-access output iterator type for writing run-offset values \\iterator\n * \\tparam LengthsOutputIteratorT [inferred] Random-access output iterator type for writing run-length values \\iterator\n * \\tparam NumRunsOutputIteratorT [inferred] Output iterator type for recording the number of runs encountered \\iterator\n */\n template <\n typename InputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsOutputIteratorT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t NonTrivialRuns(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to input sequence of data items\n OffsetsOutputIteratorT d_offsets_out, ///< [out] Pointer to output sequence of run-offsets (one offset per non-trivial run)\n LengthsOutputIteratorT d_lengths_out, ///< [out] Pointer to output sequence of run-lengths (one count per non-trivial run)\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs (i.e., length of \\p d_offsets_out)\n int num_items, ///< [in] Total number of associated key+value pairs (i.e., the length of \\p d_in_keys and \\p d_in_values)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef Equality EqualityOp; // Default == operator\n\n return DeviceRleDispatch::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_offsets_out,\n d_lengths_out,\n d_num_runs_out,\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_scan.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceScan provides device-wide, parallel operations for computing a prefix scan across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_scan.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceScan provides device-wide, parallel operations for computing a prefix scan across a sequence of data items residing within device-accessible memory. ![](device_scan.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * Given a sequence of input elements and a binary reduction operator, a [prefix scan](http://en.wikipedia.org/wiki/Prefix_sum)\n * produces an output sequence where each element is computed to be the reduction\n * of the elements occurring earlier in the input sequence. Prefix sum\n * connotes a prefix scan with the addition operator. The term \\em inclusive indicates\n * that the ith output reduction incorporates the ith input.\n * The term \\em exclusive indicates the ith input is not incorporated into\n * the ith output reduction.\n *\n * \\par\n * As of CUB 1.0.1 (2013), CUB's device-wide scan APIs have implemented our \"decoupled look-back\" algorithm\n * for performing global prefix scan with only a single pass through the\n * input data, as described in our 2016 technical report [1]. The central\n * idea is to leverage a small, constant factor of redundant work in order to overlap the latencies\n * of global prefix propagation with local computation. As such, our algorithm requires only\n * ~2n data movement (n inputs are read, n outputs are written), and typically\n * proceeds at \"memcpy\" speeds.\n *\n * \\par\n * [1] [Duane Merrill and Michael Garland. \"Single-pass Parallel Prefix Scan with Decoupled Look-back\", NVIDIA Technical Report NVR-2016-002, 2016.](https://research.nvidia.com/publication/single-pass-parallel-prefix-scan-decoupled-look-back)\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceScan}\n *\n * \\par Performance\n * \\linear_performance{prefix scan}\n *\n * \\par\n * The following chart illustrates DeviceScan::ExclusiveSum\n * performance across different CUDA architectures for \\p int32 keys.\n * \\plots_below\n *\n * \\image html scan_int32.png\n *\n */\nstruct DeviceScan\n{\n /******************************************************************//**\n * \\name Exclusive scans\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes a device-wide exclusive prefix sum. The value of 0 is applied as the initial value, and is assigned to *d_out.\n *\n * \\par\n * - Supports non-commutative sum operators.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated exclusive sum performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively.\n *\n * \\image html scan_int32.png\n * \\image html scan_int64.png\n *\n * \\par Snippet\n * The code snippet below illustrates the exclusive prefix sum of an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [ , , , , , , ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run exclusive prefix sum\n * cub::DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // d_out s<-- [0, 8, 14, 21, 26, 29, 29]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading scan inputs \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing scan outputs \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ExclusiveSum(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n int num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Initial value\n OutputT init_value = 0;\n\n return DispatchScan::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n Sum(),\n init_value,\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide exclusive prefix scan using the specified binary \\p scan_op functor. The \\p init_value value is applied as the initial value, and is assigned to *d_out.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the exclusive prefix min-scan of an \\p int device vector\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // CustomMin functor\n * struct CustomMin\n * {\n * template \n * CUB_RUNTIME_FUNCTION __forceinline__\n * T operator()(const T &a, const T &b) const {\n * return (b < a) ? b : a;\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [ , , , , , , ]\n * CustomMin min_op\n * ...\n *\n * // Determine temporary device storage requirements for exclusive prefix scan\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceScan::ExclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, (int) MAX_INT, num_items);\n *\n * // Allocate temporary storage for exclusive prefix scan\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run exclusive prefix min-scan\n * cub::DeviceScan::ExclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, (int) MAX_INT, num_items);\n *\n * // d_out <-- [2147483647, 8, 6, 6, 5, 3, 0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading scan inputs \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing scan outputs \\iterator\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n * \\tparam Identity [inferred] Type of the \\p identity value used Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename ScanOpT,\n typename InitValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ExclusiveScan(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n ScanOpT scan_op, ///< [in] Binary scan functor\n InitValueT init_value, ///< [in] Initial value to seed the exclusive scan (and is assigned to *d_out)\n int num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchScan::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n init_value,\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive scans\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes a device-wide inclusive prefix sum.\n *\n * \\par\n * - Supports non-commutative sum operators.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the inclusive prefix sum of an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [ , , , , , , ]\n * ...\n *\n * // Determine temporary device storage requirements for inclusive prefix sum\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // Allocate temporary storage for inclusive prefix sum\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run inclusive prefix sum\n * cub::DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);\n *\n * // d_out <-- [8, 14, 21, 26, 29, 29, 38]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading scan inputs \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing scan outputs \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t InclusiveSum(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n int num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchScan::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n Sum(),\n NullType(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide inclusive prefix scan using the specified binary \\p scan_op functor.\n *\n * \\par\n * - Supports non-commutative scan operators.\n * - Provides \"run-to-run\" determinism for pseudo-associative reduction\n * (e.g., addition of floating point types) on the same GPU device.\n * However, results for pseudo-associative reduction may be inconsistent\n * from one device to a another device of a different compute-capability\n * because CUB can employ different tile-sizing for different architectures.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the inclusive prefix min-scan of an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // CustomMin functor\n * struct CustomMin\n * {\n * template \n * CUB_RUNTIME_FUNCTION __forceinline__\n * T operator()(const T &a, const T &b) const {\n * return (b < a) ? b : a;\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 7\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [ , , , , , , ]\n * CustomMin min_op;\n * ...\n *\n * // Determine temporary device storage requirements for inclusive prefix scan\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceScan::InclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, num_items);\n *\n * // Allocate temporary storage for inclusive prefix scan\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run inclusive prefix min-scan\n * cub::DeviceScan::InclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, min_op, num_items);\n *\n * // d_out <-- [8, 6, 6, 5, 3, 0, 0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading scan inputs \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing scan outputs \\iterator\n * \\tparam ScanOp [inferred] Binary scan functor type having member T operator()(const T &a, const T &b)\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename ScanOpT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t InclusiveScan(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n ScanOpT scan_op, ///< [in] Binary scan functor\n int num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchScan::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n NullType(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n //@} end member group\n\n};\n\n/**\n * \\example example_device_scan.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_segmented_radix_sort.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSegmentedRadixSort provides device-wide, parallel operations for computing a batched radix sort across multiple, non-overlapping sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_radix_sort.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceSegmentedRadixSort provides device-wide, parallel operations for computing a batched radix sort across multiple, non-overlapping sequences of data items residing within device-accessible memory. ![](segmented_sorting_logo.png)\n * \\ingroup SegmentedModule\n *\n * \\par Overview\n * The [radix sorting method](http://en.wikipedia.org/wiki/Radix_sort) arranges\n * items into ascending (or descending) order. The algorithm relies upon a positional representation for\n * keys, i.e., each key is comprised of an ordered sequence of symbols (e.g., digits,\n * characters, etc.) specified from least-significant to most-significant. For a\n * given input sequence of keys and a set of rules specifying a total ordering\n * of the symbolic alphabet, the radix sorting method produces a lexicographic\n * ordering of those keys.\n *\n * \\par\n * DeviceSegmentedRadixSort can sort all of the built-in C++ numeric primitive types, e.g.:\n * unsigned char, \\p int, \\p double, etc. Although the direct radix sorting\n * method can only be applied to unsigned integral types, DeviceSegmentedRadixSort\n * is able to sort signed and floating-point types via simple bit-wise transformations\n * that ensure lexicographic key ordering.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceSegmentedRadixSort}\n *\n */\nstruct DeviceSegmentedRadixSort\n{\n\n /******************************************************************//**\n * \\name Key-value pairs\n *********************************************************************/\n //@{\n\n /**\n * \\brief Sorts segments of key-value pairs into ascending order. (~2N auxiliary storage required)\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [-, -, -, -, -, -, -]\n * int *d_values_in; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_values_out; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys_out <-- [6, 7, 8, 0, 3, 5, 9]\n * // d_values_out <-- [1, 2, 0, 5, 4, 3, 6]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam ValueT [inferred] Value type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename ValueT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairs(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] %Device-accessible pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] %Device-accessible pointer to the sorted output sequence of key data\n const ValueT *d_values_in, ///< [in] %Device-accessible pointer to the corresponding input sequence of associated value items\n ValueT *d_values_out, ///< [out] %Device-accessible pointer to the correspondingly-reordered output sequence of associated value items\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values(const_cast(d_values_in), d_values_out);\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts segments of key-value pairs into ascending order. (~N auxiliary storage required)\n *\n * \\par\n * - The sorting operation is given a pair of key buffers and a corresponding\n * pair of associated value buffers. Each pair is managed by a DoubleBuffer\n * structure that indicates which of the two buffers is \"current\" (and thus\n * contains the input data to be sorted).\n * - The contents of both buffers within each pair may be altered by the sorting\n * operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within each DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * int *d_value_buf; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_value_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Create a set of DoubleBuffers to wrap pairs of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n * cub::DoubleBuffer d_values(d_value_buf, d_value_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys.Current() <-- [6, 7, 8, 0, 3, 5, 9]\n * // d_values.Current() <-- [5, 4, 3, 1, 2, 0, 6]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam ValueT [inferred] Value type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename ValueT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairs(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer of values whose \"current\" device-accessible buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts segments of key-value pairs into descending order. (~2N auxiliary storage required).\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [-, -, -, -, -, -, -]\n * int *d_values_in; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_values_out; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes,\n * d_keys_in, d_keys_out, d_values_in, d_values_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys_out <-- [8, 7, 6, 9, 5, 3, 0]\n * // d_values_out <-- [0, 2, 1, 6, 3, 4, 5]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam ValueT [inferred] Value type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename ValueT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairsDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] %Device-accessible pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] %Device-accessible pointer to the sorted output sequence of key data\n const ValueT *d_values_in, ///< [in] %Device-accessible pointer to the corresponding input sequence of associated value items\n ValueT *d_values_out, ///< [out] %Device-accessible pointer to the correspondingly-reordered output sequence of associated value items\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values(const_cast(d_values_in), d_values_out);\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts segments of key-value pairs into descending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers and a corresponding\n * pair of associated value buffers. Each pair is managed by a DoubleBuffer\n * structure that indicates which of the two buffers is \"current\" (and thus\n * contains the input data to be sorted).\n * - The contents of both buffers within each pair may be altered by the sorting\n * operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within each DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys\n * with associated vector of \\p int values.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * int *d_value_buf; // e.g., [0, 1, 2, 3, 4, 5, 6]\n * int *d_value_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Create a set of DoubleBuffers to wrap pairs of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n * cub::DoubleBuffer d_values(d_value_buf, d_value_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, d_keys, d_values,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, d_keys, d_values,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys.Current() <-- [8, 7, 6, 9, 5, 3, 0]\n * // d_values.Current() <-- [0, 2, 1, 6, 3, 4, 5]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam ValueT [inferred] Value type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename ValueT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortPairsDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer of values whose \"current\" device-accessible buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Keys-only\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Sorts segments of keys into ascending order. (~2N auxiliary storage required)\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys_out <-- [6, 7, 8, 0, 3, 5, 9]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeys(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] %Device-accessible pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] %Device-accessible pointer to the sorted output sequence of key data\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts segments of keys into ascending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers managed by a\n * DoubleBuffer structure that indicates which of the two buffers is\n * \"current\" (and thus contains the input data to be sorted).\n * - The contents of both buffers may be altered by the sorting operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within the DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys.Current() <-- [6, 7, 8, 0, 3, 5, 9]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeys(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_values;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n /**\n * \\brief Sorts segments of keys into descending order. (~2N auxiliary storage required).\n *\n * \\par\n * - The contents of the input data are not altered by the sorting operation\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageNP For sorting using only O(P) temporary storage, see the sorting interface using DoubleBuffer wrappers below.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_keys_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_keys_out; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys_out <-- [8, 7, 6, 9, 5, 3, 0]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeysDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n const KeyT *d_keys_in, ///< [in] %Device-accessible pointer to the input data of key data to sort\n KeyT *d_keys_out, ///< [out] %Device-accessible pointer to the sorted output sequence of key data\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n DoubleBuffer d_keys(const_cast(d_keys_in), d_keys_out);\n DoubleBuffer d_values;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n false,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Sorts segments of keys into descending order. (~N auxiliary storage required).\n *\n * \\par\n * - The sorting operation is given a pair of key buffers managed by a\n * DoubleBuffer structure that indicates which of the two buffers is\n * \"current\" (and thus contains the input data to be sorted).\n * - The contents of both buffers may be altered by the sorting operation.\n * - Upon completion, the sorting operation will update the \"current\" indicator\n * within the DoubleBuffer wrapper to reference which of the two buffers\n * now contains the sorted output sequence (a function of the number of key bits\n * specified and the targeted device architecture).\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - An optional bit subrange [begin_bit, end_bit) of differentiating key bits can be specified. This can reduce overall sorting overhead and yield a corresponding performance improvement.\n * - \\devicestorageP\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the batched sorting of three segments (with one zero-length segment) of \\p int keys.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for sorting data\n * int num_items; // e.g., 7\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_key_buf; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_key_alt_buf; // e.g., [-, -, -, -, -, -, -]\n * ...\n *\n * // Create a DoubleBuffer to wrap the pair of device pointers\n * cub::DoubleBuffer d_keys(d_key_buf, d_key_alt_buf);\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sorting operation\n * cub::DeviceSegmentedRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys,\n * num_items, num_segments, d_offsets, d_offsets + 1);\n *\n * // d_keys.Current() <-- [8, 7, 6, 9, 5, 3, 0]\n *\n * \\endcode\n *\n * \\tparam KeyT [inferred] Key type\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename KeyT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t SortKeysDescending(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Reference to the double-buffer of keys whose \"current\" device-accessible buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n int num_items, ///< [in] The total number of items to sort (across all segments)\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit = 0, ///< [in] [optional] The least-significant bit index (inclusive) needed for key comparison\n int end_bit = sizeof(KeyT) * 8, ///< [in] [optional] The most-significant bit index (exclusive) needed for key comparison (e.g., sizeof(unsigned int) * 8)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Null value type\n DoubleBuffer d_values;\n\n return DispatchSegmentedRadixSort::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys,\n d_values,\n num_items,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n begin_bit,\n end_bit,\n true,\n stream,\n debug_synchronous);\n }\n\n\n //@} end member group\n\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_segmented_reduce.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSegmentedReduce provides device-wide, parallel operations for computing a batched reduction across multiple sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../iterator/arg_index_input_iterator.cuh\"\n#include \"dispatch/dispatch_reduce.cuh\"\n#include \"dispatch/dispatch_reduce_by_key.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceSegmentedReduce provides device-wide, parallel operations for computing a reduction across multiple sequences of data items residing within device-accessible memory. ![](reduce_logo.png)\n * \\ingroup SegmentedModule\n *\n * \\par Overview\n * A reduction (or fold)\n * uses a binary combining operator to compute a single aggregate from a sequence of input elements.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceSegmentedReduce}\n *\n */\nstruct DeviceSegmentedReduce\n{\n /**\n * \\brief Computes a device-wide segmented reduction using the specified binary \\p reduction_op functor.\n *\n * \\par\n * - Does not support binary reduction operators that are non-commutative.\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates a custom min-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // CustomMin functor\n * struct CustomMin\n * {\n * template \n * CUB_RUNTIME_FUNCTION __forceinline__\n * T operator()(const T &a, const T &b) const {\n * return (b < a) ? b : a;\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-, -, -]\n * CustomMin min_op;\n * int initial_value; // e.g., INT_MAX\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1, min_op, initial_value);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run reduction\n * cub::DeviceSegmentedReduce::Reduce(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1, min_op, initial_value);\n *\n * // d_out <-- [6, INT_MAX, 0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n * \\tparam ReductionOp [inferred] Binary reduction functor type having member T operator()(const T &a, const T &b)\n * \\tparam T [inferred] Data element type that is convertible to the \\p value type of \\p InputIteratorT\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT,\n typename ReductionOp,\n typename T>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Reduce(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n ReductionOp reduction_op, ///< [in] Binary reduction functor \n T initial_value, ///< [in] Initial value of the reduction for each segment\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n reduction_op,\n initial_value,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide segmented sum using the addition ('+') operator.\n *\n * \\par\n * - Uses \\p 0 as the initial value of the reduction for each segment.\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - Does not support \\p + operators that are non-commutative..\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the sum reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run sum-reduction\n * cub::DeviceSegmentedReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // d_out <-- [21, 0, 17]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Sum(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n cub::Sum(),\n OutputT(), // zero-initialize\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide segmented minimum using the less-than ('<') operator.\n *\n * \\par\n * - Uses std::numeric_limits::max() as the initial value of the reduction for each segment.\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - Does not support \\p < operators that are non-commutative.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the min-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::Min(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run min-reduction\n * cub::DeviceSegmentedReduce::Min(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // d_out <-- [6, INT_MAX, 0]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Min(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n cub::Min(),\n Traits::Max(), // replace with std::numeric_limits::max() when C++11 support is more prevalent\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Finds the first device-wide minimum in each segment using the less-than ('<') operator, also returning the in-segment index of that item.\n *\n * \\par\n * - The output value type of \\p d_out is cub::KeyValuePair (assuming the value type of \\p d_in is \\p T)\n * - The minimum of the ith segment is written to d_out[i].value and its offset in that segment is written to d_out[i].key.\n * - The {1, std::numeric_limits::max()} tuple is produced for zero-length inputs\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - Does not support \\p < operators that are non-commutative.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the argmin-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * KeyValuePair *d_out; // e.g., [{-,-}, {-,-}, {-,-}]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::ArgMin(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run argmin-reduction\n * cub::DeviceSegmentedReduce::ArgMin(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // d_out <-- [{1,6}, {1,INT_MAX}, {2,0}]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items (of some type \\p T) \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate (having value type KeyValuePair) \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ArgMin(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input type\n typedef typename std::iterator_traits::value_type InputValueT;\n\n // The output tuple type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n KeyValuePair, // ... then the key value pair OffsetT + InputValueT\n typename std::iterator_traits::value_type>::Type OutputTupleT; // ... else the output iterator's value type\n\n // The output value type\n typedef typename OutputTupleT::Value OutputValueT;\n\n // Wrapped input iterator to produce index-value tuples\n typedef ArgIndexInputIterator ArgIndexInputIteratorT;\n ArgIndexInputIteratorT d_indexed_in(d_in);\n\n // Initial value\n OutputTupleT initial_value(1, Traits::Max()); // replace with std::numeric_limits::max() when C++11 support is more prevalent\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_indexed_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n cub::ArgMin(),\n initial_value,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide segmented maximum using the greater-than ('>') operator.\n *\n * \\par\n * - Uses std::numeric_limits::lowest() as the initial value of the reduction.\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - Does not support \\p > operators that are non-commutative.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the max-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * int *d_out; // e.g., [-, -, -]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::Max(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run max-reduction\n * cub::DeviceSegmentedReduce::Max(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // d_out <-- [8, INT_MIN, 9]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t Max(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n cub::Max(),\n Traits::Lowest(), // replace with std::numeric_limits::lowest() when C++11 support is more prevalent\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Finds the first device-wide maximum in each segment using the greater-than ('>') operator, also returning the in-segment index of that item\n *\n * \\par\n * - The output value type of \\p d_out is cub::KeyValuePair (assuming the value type of \\p d_in is \\p T)\n * - The maximum of the ith segment is written to d_out[i].value and its offset in that segment is written to d_out[i].key.\n * - The {1, std::numeric_limits::lowest()} tuple is produced for zero-length inputs\n * - When input a contiguous sequence of segments, a single sequence\n * \\p segment_offsets (of length num_segments+1) can be aliased\n * for both the \\p d_begin_offsets and \\p d_end_offsets parameters (where\n * the latter is specified as segment_offsets+1).\n * - Does not support \\p > operators that are non-commutative.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the argmax-reduction of a device vector of \\p int data elements.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_segments; // e.g., 3\n * int *d_offsets; // e.g., [0, 3, 3, 7]\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * KeyValuePair *d_out; // e.g., [{-,-}, {-,-}, {-,-}]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSegmentedReduce::ArgMax(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run argmax-reduction\n * cub::DeviceSegmentedReduce::ArgMax(d_temp_storage, temp_storage_bytes, d_in, d_out,\n * num_segments, d_offsets, d_offsets + 1);\n *\n * // d_out <-- [{0,8}, {1,INT_MIN}, {3,9}]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items (of some type \\p T) \\iterator\n * \\tparam OutputIteratorT [inferred] Output iterator type for recording the reduced aggregate (having value type KeyValuePair) \\iterator\n * \\tparam OffsetIteratorT [inferred] Random-access input iterator type for reading segment offsets \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t ArgMax(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // The input type\n typedef typename std::iterator_traits::value_type InputValueT;\n\n // The output tuple type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n KeyValuePair, // ... then the key value pair OffsetT + InputValueT\n typename std::iterator_traits::value_type>::Type OutputTupleT; // ... else the output iterator's value type\n\n // The output value type\n typedef typename OutputTupleT::Value OutputValueT;\n\n // Wrapped input iterator to produce index-value tuples\n typedef ArgIndexInputIterator ArgIndexInputIteratorT;\n ArgIndexInputIteratorT d_indexed_in(d_in);\n\n // Initial value\n OutputTupleT initial_value(1, Traits::Lowest()); // replace with std::numeric_limits::lowest() when C++11 support is more prevalent\n\n return DispatchSegmentedReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_indexed_in,\n d_out,\n num_segments,\n d_begin_offsets,\n d_end_offsets,\n cub::ArgMax(),\n initial_value,\n stream,\n debug_synchronous);\n }\n\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_select.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSelect provides device-wide, parallel operations for compacting selected items from sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch/dispatch_select_if.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceSelect provides device-wide, parallel operations for compacting selected items from sequences of data items residing within device-accessible memory. ![](select_logo.png)\n * \\ingroup SingleModule\n *\n * \\par Overview\n * These operations apply a selection criterion to selectively copy\n * items from a specified input sequence to a compact output sequence.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceSelect}\n *\n * \\par Performance\n * \\linear_performance{select-flagged, select-if, and select-unique}\n *\n * \\par\n * The following chart illustrates DeviceSelect::If\n * performance across different CUDA architectures for \\p int32 items,\n * where 50% of the items are randomly selected.\n *\n * \\image html select_if_int32_50_percent.png\n *\n * \\par\n * The following chart illustrates DeviceSelect::Unique\n * performance across different CUDA architectures for \\p int32 items\n * where segments have lengths uniformly sampled from [1,1000].\n *\n * \\image html select_unique_int32_len_500.png\n *\n * \\par\n * \\plots_below\n *\n */\nstruct DeviceSelect\n{\n /**\n * \\brief Uses the \\p d_flags sequence to selectively copy the corresponding items from \\p d_in into \\p d_out. The total number of items selected is written to \\p d_num_selected_out. ![](select_flags_logo.png)\n *\n * \\par\n * - The value type of \\p d_flags must be castable to \\p bool (e.g., \\p bool, \\p char, \\p int, etc.).\n * - Copies of the selected items are compacted into \\p d_out and maintain their original relative ordering.\n * - \\devicestorage\n *\n * \\par Snippet\n * The code snippet below illustrates the compaction of items selected from an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input, flags, and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [1, 2, 3, 4, 5, 6, 7, 8]\n * char *d_flags; // e.g., [1, 0, 0, 1, 0, 1, 1, 0]\n * int *d_out; // e.g., [ , , , , , , , ]\n * int *d_num_selected_out; // e.g., [ ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSelect::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run selection\n * cub::DeviceSelect::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items);\n *\n * // d_out <-- [1, 4, 6, 7]\n * // d_num_selected_out <-- [4]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam FlagIterator [inferred] Random-access input iterator type for reading selection flags \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing selected items \\iterator\n * \\tparam NumSelectedIteratorT [inferred] Output iterator type for recording the number of items selected \\iterator\n */\n template <\n typename InputIteratorT,\n typename FlagIterator,\n typename OutputIteratorT,\n typename NumSelectedIteratorT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Flagged(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n FlagIterator d_flags, ///< [in] Pointer to the input sequence of selection flags\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the output total number of items selected (i.e., length of \\p d_out)\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType SelectOp; // Selection op (not used)\n typedef NullType EqualityOp; // Equality operator (not used)\n\n return DispatchSelectIf::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_flags,\n d_out,\n d_num_selected_out,\n SelectOp(),\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Uses the \\p select_op functor to selectively copy items from \\p d_in into \\p d_out. The total number of items selected is written to \\p d_num_selected_out. ![](select_logo.png)\n *\n * \\par\n * - Copies of the selected items are compacted into \\p d_out and maintain their original relative ordering.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated select-if performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively. Items are\n * selected with 50% probability.\n *\n * \\image html select_if_int32_50_percent.png\n * \\image html select_if_int64_50_percent.png\n *\n * \\par\n * The following charts are similar, but 5% selection probability:\n *\n * \\image html select_if_int32_5_percent.png\n * \\image html select_if_int64_5_percent.png\n *\n * \\par Snippet\n * The code snippet below illustrates the compaction of items selected from an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Functor type for selecting values less than some criteria\n * struct LessThan\n * {\n * int compare;\n *\n * CUB_RUNTIME_FUNCTION __forceinline__\n * LessThan(int compare) : compare(compare) {}\n *\n * CUB_RUNTIME_FUNCTION __forceinline__\n * bool operator()(const int &a) const {\n * return (a < compare);\n * }\n * };\n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [0, 2, 3, 9, 5, 2, 81, 8]\n * int *d_out; // e.g., [ , , , , , , , ]\n * int *d_num_selected_out; // e.g., [ ]\n * LessThan select_op(7);\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run selection\n * cub::DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op);\n *\n * // d_out <-- [0, 2, 3, 5, 2]\n * // d_num_selected_out <-- [5]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing selected items \\iterator\n * \\tparam NumSelectedIteratorT [inferred] Output iterator type for recording the number of items selected \\iterator\n * \\tparam SelectOp [inferred] Selection operator type having member bool operator()(const T &a)\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename NumSelectedIteratorT,\n typename SelectOp>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t If(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the output total number of items selected (i.e., length of \\p d_out)\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n SelectOp select_op, ///< [in] Unary selection operator\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType* FlagIterator; // FlagT iterator type (not used)\n typedef NullType EqualityOp; // Equality operator (not used)\n\n return DispatchSelectIf::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n NULL,\n d_out,\n d_num_selected_out,\n select_op,\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Given an input sequence \\p d_in having runs of consecutive equal-valued keys, only the first key from each run is selectively copied to \\p d_out. The total number of items selected is written to \\p d_num_selected_out. ![](unique_logo.png)\n *\n * \\par\n * - The == equality operator is used to determine whether keys are equivalent\n * - Copies of the selected items are compacted into \\p d_out and maintain their original relative ordering.\n * - \\devicestorage\n *\n * \\par Performance\n * The following charts illustrate saturated select-unique performance across different\n * CUDA architectures for \\p int32 and \\p int64 items, respectively. Segments have\n * lengths uniformly sampled from [1,1000].\n *\n * \\image html select_unique_int32_len_500.png\n * \\image html select_unique_int64_len_500.png\n *\n * \\par\n * The following charts are similar, but with segment lengths uniformly sampled from [1,10]:\n *\n * \\image html select_unique_int32_len_5.png\n * \\image html select_unique_int64_len_5.png\n *\n * \\par Snippet\n * The code snippet below illustrates the compaction of items selected from an \\p int device vector.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input and output\n * int num_items; // e.g., 8\n * int *d_in; // e.g., [0, 2, 2, 9, 5, 5, 5, 8]\n * int *d_out; // e.g., [ , , , , , , , ]\n * int *d_num_selected_out; // e.g., [ ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void *d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSelect::Unique(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run selection\n * cub::DeviceSelect::Unique(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items);\n *\n * // d_out <-- [0, 2, 9, 5, 8]\n * // d_num_selected_out <-- [5]\n *\n * \\endcode\n *\n * \\tparam InputIteratorT [inferred] Random-access input iterator type for reading input items \\iterator\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing selected items \\iterator\n * \\tparam NumSelectedIteratorT [inferred] Output iterator type for recording the number of items selected \\iterator\n */\n template <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename NumSelectedIteratorT>\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Unique(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the output total number of items selected (i.e., length of \\p d_out)\n int num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n typedef int OffsetT; // Signed integer type for global offsets\n typedef NullType* FlagIterator; // FlagT iterator type (not used)\n typedef NullType SelectOp; // Selection op (not used)\n typedef Equality EqualityOp; // Default == operator\n\n return DispatchSelectIf::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n NULL,\n d_out,\n d_num_selected_out,\n SelectOp(),\n EqualityOp(),\n num_items,\n stream,\n debug_synchronous);\n }\n\n};\n\n/**\n * \\example example_device_select_flagged.cu\n * \\example example_device_select_if.cu\n * \\example example_device_select_unique.cu\n */\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/device_spmv.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSpmv provides device-wide parallel operations for performing sparse-matrix * vector multiplication (SpMV).\n */\n\n#pragma once\n\n#include \n#include \n#include \n\n#include \"dispatch/dispatch_spmv_orig.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief DeviceSpmv provides device-wide parallel operations for performing sparse-matrix * dense-vector multiplication (SpMV).\n * \\ingroup SingleModule\n *\n * \\par Overview\n * The [SpMV computation](http://en.wikipedia.org/wiki/Sparse_matrix-vector_multiplication)\n * performs the matrix-vector operation\n * y = alpha*A*x + beta*y,\n * where:\n * - A is an mxn sparse matrix whose non-zero structure is specified in\n * [compressed-storage-row (CSR) format](http://en.wikipedia.org/wiki/Sparse_matrix#Compressed_row_Storage_.28CRS_or_CSR.29)\n * (i.e., three arrays: values, row_offsets, and column_indices)\n * - x and y are dense vectors\n * - alpha and beta are scalar multiplicands\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceSpmv}\n *\n */\nstruct DeviceSpmv\n{\n /******************************************************************//**\n * \\name CSR matrix operations\n *********************************************************************/\n //@{\n\n /**\n * \\brief This function performs the matrix-vector operation y = A*x.\n *\n * \\par Snippet\n * The code snippet below illustrates SpMV upon a 9x9 CSR matrix A\n * representing a 3x3 lattice (24 non-zeros).\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize device-accessible pointers for input matrix A, input vector x,\n * // and output vector y\n * int num_rows = 9;\n * int num_cols = 9;\n * int num_nonzeros = 24;\n *\n * float* d_values; // e.g., [1, 1, 1, 1, 1, 1, 1, 1,\n * // 1, 1, 1, 1, 1, 1, 1, 1,\n * // 1, 1, 1, 1, 1, 1, 1, 1]\n *\n * int* d_column_indices; // e.g., [1, 3, 0, 2, 4, 1, 5, 0,\n * // 4, 6, 1, 3, 5, 7, 2, 4,\n * // 8, 3, 7, 4, 6, 8, 5, 7]\n *\n * int* d_row_offsets; // e.g., [0, 2, 5, 7, 10, 14, 17, 19, 22, 24]\n *\n * float* d_vector_x; // e.g., [1, 1, 1, 1, 1, 1, 1, 1, 1]\n * float* d_vector_y; // e.g., [ , , , , , , , , ]\n * ...\n *\n * // Determine temporary device storage requirements\n * void* d_temp_storage = NULL;\n * size_t temp_storage_bytes = 0;\n * cub::DeviceSpmv::CsrMV(d_temp_storage, temp_storage_bytes, d_values,\n * d_row_offsets, d_column_indices, d_vector_x, d_vector_y,\n * num_rows, num_cols, num_nonzeros, alpha, beta);\n *\n * // Allocate temporary storage\n * cudaMalloc(&d_temp_storage, temp_storage_bytes);\n *\n * // Run SpMV\n * cub::DeviceSpmv::CsrMV(d_temp_storage, temp_storage_bytes, d_values,\n * d_row_offsets, d_column_indices, d_vector_x, d_vector_y,\n * num_rows, num_cols, num_nonzeros, alpha, beta);\n *\n * // d_vector_y <-- [2, 3, 2, 3, 4, 3, 2, 3, 2]\n *\n * \\endcode\n *\n * \\tparam ValueT [inferred] Matrix and vector value type (e.g., /p float, /p double, etc.)\n */\n template <\n typename ValueT>\n CUB_RUNTIME_FUNCTION\n static cudaError_t CsrMV(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n ValueT* d_values, ///< [in] Pointer to the array of \\p num_nonzeros values of the corresponding nonzero elements of matrix A.\n int* d_row_offsets, ///< [in] Pointer to the array of \\p m + 1 offsets demarcating the start of every row in \\p d_column_indices and \\p d_values (with the final entry being equal to \\p num_nonzeros)\n int* d_column_indices, ///< [in] Pointer to the array of \\p num_nonzeros column-indices of the corresponding nonzero elements of matrix A. (Indices are zero-valued.)\n ValueT* d_vector_x, ///< [in] Pointer to the array of \\p num_cols values corresponding to the dense input vector x\n ValueT* d_vector_y, ///< [out] Pointer to the array of \\p num_rows values corresponding to the dense output vector y\n int num_rows, ///< [in] number of rows of matrix A.\n int num_cols, ///< [in] number of columns of matrix A.\n int num_nonzeros, ///< [in] number of nonzero elements of matrix A.\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n SpmvParams spmv_params;\n spmv_params.d_values = d_values;\n spmv_params.d_row_end_offsets = d_row_offsets + 1;\n spmv_params.d_column_indices = d_column_indices;\n spmv_params.d_vector_x = d_vector_x;\n spmv_params.d_vector_y = d_vector_y;\n spmv_params.num_rows = num_rows;\n spmv_params.num_cols = num_cols;\n spmv_params.num_nonzeros = num_nonzeros;\n spmv_params.alpha = 1.0;\n spmv_params.beta = 0.0;\n\n return DispatchSpmv::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n spmv_params,\n stream,\n debug_synchronous);\n }\n\n //@} end member group\n};\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_histogram.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceHistogram provides device-wide parallel operations for constructing histogram(s) from a sequence of samples data residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n#include \n\n#include \"../../agent/agent_histogram.cuh\"\n#include \"../../util_debug.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../thread/thread_search.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n\n/******************************************************************************\n * Histogram kernel entry points\n *****************************************************************************/\n\n/**\n * Histogram initialization kernel entry point\n */\ntemplate <\n int NUM_ACTIVE_CHANNELS, ///< Number of channels actively being histogrammed\n typename CounterT, ///< Integer type for counting sample occurrences per histogram bin\n typename OffsetT> ///< Signed integer type for global offsets\n__global__ void DeviceHistogramInitKernel(\n ArrayWrapper num_output_bins_wrapper, ///< Number of output histogram bins per channel\n ArrayWrapper d_output_histograms_wrapper, ///< Histogram counter data having logical dimensions CounterT[NUM_ACTIVE_CHANNELS][num_bins.array[CHANNEL]]\n GridQueue tile_queue) ///< Drain queue descriptor for dynamically mapping tile data onto thread blocks\n{\n if ((threadIdx.x == 0) && (blockIdx.x == 0))\n tile_queue.ResetDrain();\n\n int output_bin = (blockIdx.x * blockDim.x) + threadIdx.x;\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n if (output_bin < num_output_bins_wrapper.array[CHANNEL])\n d_output_histograms_wrapper.array[CHANNEL][output_bin] = 0;\n }\n}\n\n\n/**\n * Histogram privatized sweep kernel entry point (multi-block). Computes privatized histograms, one per thread block.\n */\ntemplate <\n typename AgentHistogramPolicyT, ///< Parameterized AgentHistogramPolicy tuning policy type\n int PRIVATIZED_SMEM_BINS, ///< Maximum number of histogram bins per channel (e.g., up to 256)\n int NUM_CHANNELS, ///< Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n int NUM_ACTIVE_CHANNELS, ///< Number of channels actively being histogrammed\n typename SampleIteratorT, ///< The input iterator type. \\iterator.\n typename CounterT, ///< Integer type for counting sample occurrences per histogram bin\n typename PrivatizedDecodeOpT, ///< The transform operator type for determining privatized counter indices from samples, one for each channel\n typename OutputDecodeOpT, ///< The transform operator type for determining output bin-ids from privatized counter indices, one for each channel\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(AgentHistogramPolicyT::BLOCK_THREADS))\n__global__ void DeviceHistogramSweepKernel(\n SampleIteratorT d_samples, ///< Input data to reduce\n ArrayWrapper num_output_bins_wrapper, ///< The number bins per final output histogram\n ArrayWrapper num_privatized_bins_wrapper, ///< The number bins per privatized histogram\n ArrayWrapper d_output_histograms_wrapper, ///< Reference to final output histograms\n ArrayWrapper d_privatized_histograms_wrapper, ///< Reference to privatized histograms\n ArrayWrapper output_decode_op_wrapper, ///< The transform operator for determining output bin-ids from privatized counter indices, one for each channel\n ArrayWrapper privatized_decode_op_wrapper, ///< The transform operator for determining privatized counter indices from samples, one for each channel\n OffsetT num_row_pixels, ///< The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< The number of rows in the region of interest\n OffsetT row_stride_samples, ///< The number of samples between starts of consecutive rows in the region of interest\n int tiles_per_row, ///< Number of image tiles per row\n GridQueue tile_queue) ///< Drain queue descriptor for dynamically mapping tile data onto thread blocks\n{\n // Thread block type for compositing input tiles\n typedef AgentHistogram<\n AgentHistogramPolicyT,\n PRIVATIZED_SMEM_BINS,\n NUM_CHANNELS,\n NUM_ACTIVE_CHANNELS,\n SampleIteratorT,\n CounterT,\n PrivatizedDecodeOpT,\n OutputDecodeOpT,\n OffsetT>\n AgentHistogramT;\n\n // Shared memory for AgentHistogram\n __shared__ typename AgentHistogramT::TempStorage temp_storage;\n\n AgentHistogramT agent(\n temp_storage,\n d_samples,\n num_output_bins_wrapper.array,\n num_privatized_bins_wrapper.array,\n d_output_histograms_wrapper.array,\n d_privatized_histograms_wrapper.array,\n output_decode_op_wrapper.array,\n privatized_decode_op_wrapper.array);\n\n // Initialize counters\n agent.InitBinCounters();\n\n // Consume input tiles\n agent.ConsumeTiles(\n num_row_pixels,\n num_rows,\n row_stride_samples,\n tiles_per_row,\n tile_queue);\n\n // Store output to global (if necessary)\n agent.StoreOutput();\n\n}\n\n\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceHistogram\n */\ntemplate <\n int NUM_CHANNELS, ///< Number of channels interleaved in the input data (may be greater than the number of channels being actively histogrammed)\n int NUM_ACTIVE_CHANNELS, ///< Number of channels actively being histogrammed\n typename SampleIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename CounterT, ///< Integer type for counting sample occurrences per histogram bin\n typename LevelT, ///< Type for specifying bin level boundaries\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DipatchHistogram\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n /// The sample value type of the input iterator\n typedef typename std::iterator_traits::value_type SampleT;\n\n enum\n {\n // Maximum number of bins per channel for which we will use a privatized smem strategy\n MAX_PRIVATIZED_SMEM_BINS = 256\n };\n\n\n //---------------------------------------------------------------------\n // Transform functors for converting samples to bin-ids\n //---------------------------------------------------------------------\n\n // Searches for bin given a list of bin-boundary levels\n template \n struct SearchTransform\n {\n LevelIteratorT d_levels; // Pointer to levels array\n int num_output_levels; // Number of levels in array\n\n // Initializer\n __host__ __device__ __forceinline__ void Init(\n LevelIteratorT d_levels, // Pointer to levels array\n int num_output_levels) // Number of levels in array\n {\n this->d_levels = d_levels;\n this->num_output_levels = num_output_levels;\n }\n\n // Method for converting samples to bin-ids\n template \n __host__ __device__ __forceinline__ void BinSelect(_SampleT sample, int &bin, bool valid)\n {\n /// Level iterator wrapper type\n typedef typename If::VALUE,\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n LevelIteratorT>::Type // Directly use the supplied input iterator type\n WrappedLevelIteratorT;\n\n WrappedLevelIteratorT wrapped_levels(d_levels);\n\n int num_bins = num_output_levels - 1;\n if (valid)\n {\n bin = UpperBound(wrapped_levels, num_output_levels, (LevelT) sample) - 1;\n if (bin >= num_bins)\n bin = -1;\n }\n }\n };\n\n\n // Scales samples to evenly-spaced bins\n struct ScaleTransform\n {\n int num_bins; // Number of levels in array\n LevelT max; // Max sample level (exclusive)\n LevelT min; // Min sample level (inclusive)\n LevelT scale; // Bin scaling factor\n\n // Initializer\n template \n __host__ __device__ __forceinline__ void Init(\n int num_output_levels, // Number of levels in array\n _LevelT max, // Max sample level (exclusive)\n _LevelT min, // Min sample level (inclusive)\n _LevelT scale) // Bin scaling factor\n {\n this->num_bins = num_output_levels - 1;\n this->max = max;\n this->min = min;\n this->scale = scale;\n }\n\n // Initializer (float specialization)\n __host__ __device__ __forceinline__ void Init(\n int num_output_levels, // Number of levels in array\n float max, // Max sample level (exclusive)\n float min, // Min sample level (inclusive)\n float scale) // Bin scaling factor\n {\n this->num_bins = num_output_levels - 1;\n this->max = max;\n this->min = min;\n this->scale = float(1.0) / scale;\n }\n\n // Initializer (double specialization)\n __host__ __device__ __forceinline__ void Init(\n int num_output_levels, // Number of levels in array\n double max, // Max sample level (exclusive)\n double min, // Min sample level (inclusive)\n double scale) // Bin scaling factor\n {\n this->num_bins = num_output_levels - 1;\n this->max = max;\n this->min = min;\n this->scale = double(1.0) / scale;\n }\n\n // Method for converting samples to bin-ids\n template \n __host__ __device__ __forceinline__ void BinSelect(_SampleT sample, int &bin, bool valid)\n {\n LevelT level_sample = (LevelT) sample;\n\n if (valid && (level_sample >= min) && (level_sample < max))\n bin = (int) ((level_sample - min) / scale);\n }\n\n // Method for converting samples to bin-ids (float specialization)\n template \n __host__ __device__ __forceinline__ void BinSelect(float sample, int &bin, bool valid)\n {\n LevelT level_sample = (LevelT) sample;\n\n if (valid && (level_sample >= min) && (level_sample < max))\n bin = (int) ((level_sample - min) * scale);\n }\n\n // Method for converting samples to bin-ids (double specialization)\n template \n __host__ __device__ __forceinline__ void BinSelect(double sample, int &bin, bool valid)\n {\n LevelT level_sample = (LevelT) sample;\n\n if (valid && (level_sample >= min) && (level_sample < max))\n bin = (int) ((level_sample - min) * scale);\n }\n };\n\n\n // Pass-through bin transform operator\n struct PassThruTransform\n {\n // Method for converting samples to bin-ids\n template \n __host__ __device__ __forceinline__ void BinSelect(_SampleT sample, int &bin, bool valid)\n {\n if (valid)\n bin = (int) sample;\n }\n };\n\n\n\n //---------------------------------------------------------------------\n // Tuning policies\n //---------------------------------------------------------------------\n\n template \n struct TScale\n {\n enum\n {\n V_SCALE = (sizeof(SampleT) + sizeof(int) - 1) / sizeof(int),\n VALUE = CUB_MAX((NOMINAL_ITEMS_PER_THREAD / NUM_ACTIVE_CHANNELS / V_SCALE), 1)\n };\n };\n\n\n /// SM11\n struct Policy110\n {\n // HistogramSweepPolicy\n typedef AgentHistogramPolicy<\n 512,\n (NUM_CHANNELS == 1) ? 8 : 2,\n BLOCK_LOAD_DIRECT,\n LOAD_DEFAULT,\n true,\n GMEM,\n false>\n HistogramSweepPolicy;\n };\n\n /// SM20\n struct Policy200\n {\n // HistogramSweepPolicy\n typedef AgentHistogramPolicy<\n (NUM_CHANNELS == 1) ? 256 : 128,\n (NUM_CHANNELS == 1) ? 8 : 3,\n (NUM_CHANNELS == 1) ? BLOCK_LOAD_DIRECT : BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n true,\n SMEM,\n false>\n HistogramSweepPolicy;\n };\n\n /// SM30\n struct Policy300\n {\n // HistogramSweepPolicy\n typedef AgentHistogramPolicy<\n 512,\n (NUM_CHANNELS == 1) ? 8 : 2,\n BLOCK_LOAD_DIRECT,\n LOAD_DEFAULT,\n true,\n GMEM,\n false>\n HistogramSweepPolicy;\n };\n\n /// SM35\n struct Policy350\n {\n // HistogramSweepPolicy\n typedef AgentHistogramPolicy<\n 128,\n TScale<8>::VALUE,\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n true,\n BLEND,\n true>\n HistogramSweepPolicy;\n };\n\n /// SM50\n struct Policy500\n {\n // HistogramSweepPolicy\n typedef AgentHistogramPolicy<\n 384,\n TScale<16>::VALUE,\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n true,\n SMEM,\n false>\n HistogramSweepPolicy;\n };\n\n\n\n //---------------------------------------------------------------------\n // Tuning policies of current PTX compiler pass\n //---------------------------------------------------------------------\n\n#if (CUB_PTX_ARCH >= 500)\n typedef Policy500 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#else\n typedef Policy110 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxHistogramSweepPolicy : PtxPolicy::HistogramSweepPolicy {};\n\n\n //---------------------------------------------------------------------\n // Utilities\n //---------------------------------------------------------------------\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t InitConfigs(\n int ptx_version,\n KernelConfig &histogram_sweep_config)\n {\n #if (CUB_PTX_ARCH > 0)\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n return histogram_sweep_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 500)\n {\n return histogram_sweep_config.template Init();\n }\n else if (ptx_version >= 350)\n {\n return histogram_sweep_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n return histogram_sweep_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n return histogram_sweep_config.template Init();\n }\n else if (ptx_version >= 110)\n {\n return histogram_sweep_config.template Init();\n }\n else\n {\n // No global atomic support\n return cudaErrorNotSupported;\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration\n */\n struct KernelConfig\n {\n int block_threads;\n int pixels_per_thread;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Init()\n {\n block_threads = BlockPolicy::BLOCK_THREADS;\n pixels_per_thread = BlockPolicy::PIXELS_PER_THREAD;\n\n return cudaSuccess;\n }\n };\n\n\n //---------------------------------------------------------------------\n // Dispatch entrypoints\n //---------------------------------------------------------------------\n\n /**\n * Privatization-based dispatch routine\n */\n template <\n typename PrivatizedDecodeOpT, ///< The transform operator type for determining privatized counter indices from samples, one for each channel\n typename OutputDecodeOpT, ///< The transform operator type for determining output bin-ids from privatized counter indices, one for each channel\n typename DeviceHistogramInitKernelT, ///< Function type of cub::DeviceHistogramInitKernel\n typename DeviceHistogramSweepKernelT> ///< Function type of cub::DeviceHistogramSweepKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t PrivatizedDispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of sample items. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_output_histograms[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_output_levels[i] - 1.\n int num_privatized_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of bin level boundaries for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n PrivatizedDecodeOpT privatized_decode_op[NUM_ACTIVE_CHANNELS], ///< [in] Transform operators for determining bin-ids from samples, one for each channel\n int num_output_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of bin level boundaries for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n OutputDecodeOpT output_decode_op[NUM_ACTIVE_CHANNELS], ///< [in] Transform operators for determining bin-ids from samples, one for each channel\n int max_num_output_bins, ///< [in] Maximum number of output bins in any channel\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_samples, ///< [in] The number of samples between starts of consecutive rows in the region of interest\n DeviceHistogramInitKernelT histogram_init_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceHistogramInitKernel\n DeviceHistogramSweepKernelT histogram_sweep_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceHistogramSweepKernel\n KernelConfig histogram_sweep_config, ///< [in] Dispatch parameters that match the policy that \\p histogram_sweep_kernel was compiled for\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n #ifndef CUB_RUNTIME_ENABLED\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported);\n\n #else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Get SM occupancy for histogram_sweep_kernel\n int histogram_sweep_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n histogram_sweep_sm_occupancy,\n histogram_sweep_kernel,\n histogram_sweep_config.block_threads))) break;\n\n // Get device occupancy for histogram_sweep_kernel\n int histogram_sweep_occupancy = histogram_sweep_sm_occupancy * sm_count;\n\n if (num_row_pixels * NUM_CHANNELS == row_stride_samples)\n {\n // Treat as a single linear array of samples\n num_row_pixels *= num_rows;\n num_rows = 1;\n row_stride_samples = num_row_pixels * NUM_CHANNELS;\n }\n\n // Get grid dimensions, trying to keep total blocks ~histogram_sweep_occupancy\n int pixels_per_tile = histogram_sweep_config.block_threads * histogram_sweep_config.pixels_per_thread;\n int tiles_per_row = int(num_row_pixels + pixels_per_tile - 1) / pixels_per_tile;\n int blocks_per_row = CUB_MIN(histogram_sweep_occupancy, tiles_per_row);\n int blocks_per_col = (blocks_per_row > 0) ?\n int(CUB_MIN(histogram_sweep_occupancy / blocks_per_row, num_rows)) :\n 0;\n int num_thread_blocks = blocks_per_row * blocks_per_col;\n\n dim3 sweep_grid_dims;\n sweep_grid_dims.x = (unsigned int) blocks_per_row;\n sweep_grid_dims.y = (unsigned int) blocks_per_col;\n sweep_grid_dims.z = 1;\n\n // Temporary storage allocation requirements\n const int NUM_ALLOCATIONS = NUM_ACTIVE_CHANNELS + 1;\n void* allocations[NUM_ALLOCATIONS];\n size_t allocation_sizes[NUM_ALLOCATIONS];\n\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n allocation_sizes[CHANNEL] = size_t(num_thread_blocks) * (num_privatized_levels[CHANNEL] - 1) * sizeof(CounterT);\n\n allocation_sizes[NUM_ALLOCATIONS - 1] = GridQueue::AllocationSize();\n\n // Alias the temporary allocations from the single storage blob (or compute the necessary size of the blob)\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Construct the grid queue descriptor\n GridQueue tile_queue(allocations[NUM_ALLOCATIONS - 1]);\n\n // Setup array wrapper for histogram channel output (because we can't pass static arrays as kernel parameters)\n ArrayWrapper d_output_histograms_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n d_output_histograms_wrapper.array[CHANNEL] = d_output_histograms[CHANNEL];\n\n // Setup array wrapper for privatized per-block histogram channel output (because we can't pass static arrays as kernel parameters)\n ArrayWrapper d_privatized_histograms_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n d_privatized_histograms_wrapper.array[CHANNEL] = (CounterT*) allocations[CHANNEL];\n\n // Setup array wrapper for sweep bin transforms (because we can't pass static arrays as kernel parameters)\n ArrayWrapper privatized_decode_op_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n privatized_decode_op_wrapper.array[CHANNEL] = privatized_decode_op[CHANNEL];\n\n // Setup array wrapper for aggregation bin transforms (because we can't pass static arrays as kernel parameters)\n ArrayWrapper output_decode_op_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n output_decode_op_wrapper.array[CHANNEL] = output_decode_op[CHANNEL];\n\n // Setup array wrapper for num privatized bins (because we can't pass static arrays as kernel parameters)\n ArrayWrapper num_privatized_bins_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n num_privatized_bins_wrapper.array[CHANNEL] = num_privatized_levels[CHANNEL] - 1;\n\n // Setup array wrapper for num output bins (because we can't pass static arrays as kernel parameters)\n ArrayWrapper num_output_bins_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n num_output_bins_wrapper.array[CHANNEL] = num_output_levels[CHANNEL] - 1;\n\n int histogram_init_block_threads = 256;\n int histogram_init_grid_dims = (max_num_output_bins + histogram_init_block_threads - 1) / histogram_init_block_threads;\n\n // Log DeviceHistogramInitKernel configuration\n if (debug_synchronous) _CubLog(\"Invoking DeviceHistogramInitKernel<<<%d, %d, 0, %lld>>>()\\n\",\n histogram_init_grid_dims, histogram_init_block_threads, (long long) stream);\n\n // Invoke histogram_init_kernel\n histogram_init_kernel<<>>(\n num_output_bins_wrapper,\n d_output_histograms_wrapper,\n tile_queue);\n\n // Return if empty problem\n if ((blocks_per_row == 0) || (blocks_per_col == 0))\n break;\n\n // Log histogram_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking histogram_sweep_kernel<<<{%d, %d, %d}, %d, 0, %lld>>>(), %d pixels per thread, %d SM occupancy\\n\",\n sweep_grid_dims.x, sweep_grid_dims.y, sweep_grid_dims.z,\n histogram_sweep_config.block_threads, (long long) stream, histogram_sweep_config.pixels_per_thread, histogram_sweep_sm_occupancy);\n\n // Invoke histogram_sweep_kernel\n histogram_sweep_kernel<<>>(\n d_samples,\n num_output_bins_wrapper,\n num_privatized_bins_wrapper,\n d_output_histograms_wrapper,\n d_privatized_histograms_wrapper,\n output_decode_op_wrapper,\n privatized_decode_op_wrapper,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n tiles_per_row,\n tile_queue);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n }\n while (0);\n\n return error;\n\n #endif // CUB_RUNTIME_ENABLED\n }\n\n\n\n /**\n * Dispatch routine for HistogramRange, specialized for sample types larger than 8bit\n */\n CUB_RUNTIME_FUNCTION\n static cudaError_t DispatchRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_output_histograms[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_output_levels[i] - 1.\n int num_output_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n LevelT *d_levels[NUM_ACTIVE_CHANNELS], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_samples, ///< [in] The number of samples between starts of consecutive rows in the region of interest\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n Int2Type is_byte_sample) ///< [in] Marker type indicating whether or not SampleT is a 8b type\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel dispatch configurations\n KernelConfig histogram_sweep_config;\n if (CubDebug(error = InitConfigs(ptx_version, histogram_sweep_config)))\n break;\n\n // Use the search transform op for converting samples to privatized bins\n typedef SearchTransform PrivatizedDecodeOpT;\n\n // Use the pass-thru transform op for converting privatized bins to output bins\n typedef PassThruTransform OutputDecodeOpT;\n\n PrivatizedDecodeOpT privatized_decode_op[NUM_ACTIVE_CHANNELS];\n OutputDecodeOpT output_decode_op[NUM_ACTIVE_CHANNELS];\n int max_levels = num_output_levels[0];\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n privatized_decode_op[channel].Init(d_levels[channel], num_output_levels[channel]);\n if (num_output_levels[channel] > max_levels)\n max_levels = num_output_levels[channel];\n }\n int max_num_output_bins = max_levels - 1;\n\n // Dispatch\n if (max_num_output_bins > MAX_PRIVATIZED_SMEM_BINS)\n {\n // Too many bins to keep in shared memory.\n const int PRIVATIZED_SMEM_BINS = 0;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_output_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n }\n else\n {\n // Dispatch shared-privatized approach\n const int PRIVATIZED_SMEM_BINS = MAX_PRIVATIZED_SMEM_BINS;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_output_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n }\n\n } while (0);\n\n return error;\n }\n\n\n /**\n * Dispatch routine for HistogramRange, specialized for 8-bit sample types (computes 256-bin privatized histograms and then reduces to user-specified levels)\n */\n CUB_RUNTIME_FUNCTION\n static cudaError_t DispatchRange(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_output_histograms[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_output_levels[i] - 1.\n int num_output_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n LevelT *d_levels[NUM_ACTIVE_CHANNELS], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_samples, ///< [in] The number of samples between starts of consecutive rows in the region of interest\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n Int2Type is_byte_sample) ///< [in] Marker type indicating whether or not SampleT is a 8b type\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel dispatch configurations\n KernelConfig histogram_sweep_config;\n if (CubDebug(error = InitConfigs(ptx_version, histogram_sweep_config)))\n break;\n\n // Use the pass-thru transform op for converting samples to privatized bins\n typedef PassThruTransform PrivatizedDecodeOpT;\n\n // Use the search transform op for converting privatized bins to output bins\n typedef SearchTransform OutputDecodeOpT;\n\n int num_privatized_levels[NUM_ACTIVE_CHANNELS];\n PrivatizedDecodeOpT privatized_decode_op[NUM_ACTIVE_CHANNELS];\n OutputDecodeOpT output_decode_op[NUM_ACTIVE_CHANNELS];\n int max_levels = num_output_levels[0]; // Maximum number of levels in any channel\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n num_privatized_levels[channel] = 257;\n output_decode_op[channel].Init(d_levels[channel], num_output_levels[channel]);\n\n if (num_output_levels[channel] > max_levels)\n max_levels = num_output_levels[channel];\n }\n int max_num_output_bins = max_levels - 1;\n\n const int PRIVATIZED_SMEM_BINS = 256;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_privatized_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n\n } while (0);\n\n return error;\n }\n\n\n /**\n * Dispatch routine for HistogramEven, specialized for sample types larger than 8-bit\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t DispatchEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of sample items. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_output_histograms[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_output_levels[i] - 1.\n int num_output_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of bin level boundaries for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_samples, ///< [in] The number of samples between starts of consecutive rows in the region of interest\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n Int2Type is_byte_sample) ///< [in] Marker type indicating whether or not SampleT is a 8b type\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel dispatch configurations\n KernelConfig histogram_sweep_config;\n if (CubDebug(error = InitConfigs(ptx_version, histogram_sweep_config)))\n break;\n\n // Use the scale transform op for converting samples to privatized bins\n typedef ScaleTransform PrivatizedDecodeOpT;\n\n // Use the pass-thru transform op for converting privatized bins to output bins\n typedef PassThruTransform OutputDecodeOpT;\n\n PrivatizedDecodeOpT privatized_decode_op[NUM_ACTIVE_CHANNELS];\n OutputDecodeOpT output_decode_op[NUM_ACTIVE_CHANNELS];\n int max_levels = num_output_levels[0];\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n int bins = num_output_levels[channel] - 1;\n LevelT scale = (upper_level[channel] - lower_level[channel]) / bins;\n\n privatized_decode_op[channel].Init(num_output_levels[channel], upper_level[channel], lower_level[channel], scale);\n\n if (num_output_levels[channel] > max_levels)\n max_levels = num_output_levels[channel];\n }\n int max_num_output_bins = max_levels - 1;\n\n if (max_num_output_bins > MAX_PRIVATIZED_SMEM_BINS)\n {\n // Dispatch shared-privatized approach\n const int PRIVATIZED_SMEM_BINS = 0;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_output_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n }\n else\n {\n // Dispatch shared-privatized approach\n const int PRIVATIZED_SMEM_BINS = MAX_PRIVATIZED_SMEM_BINS;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_output_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n }\n }\n while (0);\n\n return error;\n }\n\n\n /**\n * Dispatch routine for HistogramEven, specialized for 8-bit sample types (computes 256-bin privatized histograms and then reduces to user-specified levels)\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t DispatchEven(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SampleIteratorT d_samples, ///< [in] The pointer to the input sequence of sample items. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT* d_output_histograms[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_output_levels[i] - 1.\n int num_output_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of bin level boundaries for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_output_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_samples, ///< [in] The number of samples between starts of consecutive rows in the region of interest\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n Int2Type is_byte_sample) ///< [in] Marker type indicating whether or not SampleT is a 8b type\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel dispatch configurations\n KernelConfig histogram_sweep_config;\n if (CubDebug(error = InitConfigs(ptx_version, histogram_sweep_config)))\n break;\n\n // Use the pass-thru transform op for converting samples to privatized bins\n typedef PassThruTransform PrivatizedDecodeOpT;\n\n // Use the scale transform op for converting privatized bins to output bins\n typedef ScaleTransform OutputDecodeOpT;\n\n int num_privatized_levels[NUM_ACTIVE_CHANNELS];\n PrivatizedDecodeOpT privatized_decode_op[NUM_ACTIVE_CHANNELS];\n OutputDecodeOpT output_decode_op[NUM_ACTIVE_CHANNELS];\n int max_levels = num_output_levels[0];\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n num_privatized_levels[channel] = 257;\n\n int bins = num_output_levels[channel] - 1;\n LevelT scale = (upper_level[channel] - lower_level[channel]) / bins;\n output_decode_op[channel].Init(num_output_levels[channel], upper_level[channel], lower_level[channel], scale);\n\n if (num_output_levels[channel] > max_levels)\n max_levels = num_output_levels[channel];\n }\n int max_num_output_bins = max_levels - 1;\n\n const int PRIVATIZED_SMEM_BINS = 256;\n\n if (CubDebug(error = PrivatizedDispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_output_histograms,\n num_privatized_levels,\n privatized_decode_op,\n num_output_levels,\n output_decode_op,\n max_num_output_bins,\n num_row_pixels,\n num_rows,\n row_stride_samples,\n DeviceHistogramInitKernel,\n DeviceHistogramSweepKernel,\n histogram_sweep_config,\n stream,\n debug_synchronous))) break;\n\n }\n while (0);\n\n return error;\n }\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_radix_sort.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceRadixSort provides device-wide, parallel operations for computing a radix sort across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../../agent/agent_radix_sort_upsweep.cuh\"\n#include \"../../agent/agent_radix_sort_downsweep.cuh\"\n#include \"../../agent/agent_scan.cuh\"\n#include \"../../block/block_radix_sort.cuh\"\n#include \"../../grid/grid_even_share.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_debug.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Upsweep digit-counting kernel entry point (multi-block). Computes privatized digit histograms, one per block.\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n bool ALT_DIGIT_BITS, ///< Whether or not to use the alternate (lower-bits) policy\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int((ALT_DIGIT_BITS) ?\n ChainedPolicyT::ActivePolicy::AltUpsweepPolicy::BLOCK_THREADS :\n ChainedPolicyT::ActivePolicy::UpsweepPolicy::BLOCK_THREADS))\n__global__ void DeviceRadixSortUpsweepKernel(\n const KeyT *d_keys, ///< [in] Input keys buffer\n OffsetT *d_spine, ///< [out] Privatized (per block) digit histograms (striped, i.e., 0s counts from each block, then 1s counts from each block, etc.)\n OffsetT /*num_items*/, ///< [in] Total number of input data items\n int current_bit, ///< [in] Bit position of current radix digit\n int num_bits, ///< [in] Number of bits of current radix digit\n GridEvenShare even_share) ///< [in] Even-share descriptor for mapan equal number of tiles onto each thread block\n{\n enum {\n TILE_ITEMS = ChainedPolicyT::ActivePolicy::AltUpsweepPolicy::BLOCK_THREADS *\n ChainedPolicyT::ActivePolicy::AltUpsweepPolicy::ITEMS_PER_THREAD\n };\n\n // Parameterize AgentRadixSortUpsweep type for the current configuration\n typedef AgentRadixSortUpsweep<\n typename If<(ALT_DIGIT_BITS),\n typename ChainedPolicyT::ActivePolicy::AltUpsweepPolicy,\n typename ChainedPolicyT::ActivePolicy::UpsweepPolicy>::Type,\n KeyT,\n OffsetT>\n AgentRadixSortUpsweepT;\n\n // Shared memory storage\n __shared__ typename AgentRadixSortUpsweepT::TempStorage temp_storage;\n\n // Initialize GRID_MAPPING_RAKE even-share descriptor for this thread block\n even_share.template BlockInit();\n\n AgentRadixSortUpsweepT upsweep(temp_storage, d_keys, current_bit, num_bits);\n\n upsweep.ProcessRegion(even_share.block_offset, even_share.block_end);\n\n CTA_SYNC();\n\n // Write out digit counts (striped)\n upsweep.ExtractCounts(d_spine, gridDim.x, blockIdx.x);\n}\n\n\n/**\n * Spine scan kernel entry point (single-block). Computes an exclusive prefix sum over the privatized digit histograms\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::ScanPolicy::BLOCK_THREADS), 1)\n__global__ void RadixSortScanBinsKernel(\n OffsetT *d_spine, ///< [in,out] Privatized (per block) digit histograms (striped, i.e., 0s counts from each block, then 1s counts from each block, etc.)\n int num_counts) ///< [in] Total number of bin-counts\n{\n // Parameterize the AgentScan type for the current configuration\n typedef AgentScan<\n typename ChainedPolicyT::ActivePolicy::ScanPolicy,\n OffsetT*,\n OffsetT*,\n cub::Sum,\n OffsetT,\n OffsetT>\n AgentScanT;\n\n // Shared memory storage\n __shared__ typename AgentScanT::TempStorage temp_storage;\n\n // Block scan instance\n AgentScanT block_scan(temp_storage, d_spine, d_spine, cub::Sum(), OffsetT(0)) ;\n\n // Process full input tiles\n int block_offset = 0;\n BlockScanRunningPrefixOp prefix_op(0, Sum());\n while (block_offset + AgentScanT::TILE_ITEMS <= num_counts)\n {\n block_scan.template ConsumeTile(block_offset, prefix_op);\n block_offset += AgentScanT::TILE_ITEMS;\n }\n}\n\n\n/**\n * Downsweep pass kernel entry point (multi-block). Scatters keys (and values) into corresponding bins for the current digit place.\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n bool ALT_DIGIT_BITS, ///< Whether or not to use the alternate (lower-bits) policy\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int((ALT_DIGIT_BITS) ?\n ChainedPolicyT::ActivePolicy::AltDownsweepPolicy::BLOCK_THREADS :\n ChainedPolicyT::ActivePolicy::DownsweepPolicy::BLOCK_THREADS))\n__global__ void DeviceRadixSortDownsweepKernel(\n const KeyT *d_keys_in, ///< [in] Input keys buffer\n KeyT *d_keys_out, ///< [in] Output keys buffer\n const ValueT *d_values_in, ///< [in] Input values buffer\n ValueT *d_values_out, ///< [in] Output values buffer\n OffsetT *d_spine, ///< [in] Scan of privatized (per block) digit histograms (striped, i.e., 0s counts from each block, then 1s counts from each block, etc.)\n OffsetT num_items, ///< [in] Total number of input data items\n int current_bit, ///< [in] Bit position of current radix digit\n int num_bits, ///< [in] Number of bits of current radix digit\n GridEvenShare even_share) ///< [in] Even-share descriptor for mapan equal number of tiles onto each thread block\n{\n enum {\n TILE_ITEMS = ChainedPolicyT::ActivePolicy::AltUpsweepPolicy::BLOCK_THREADS *\n ChainedPolicyT::ActivePolicy::AltUpsweepPolicy::ITEMS_PER_THREAD\n };\n\n // Parameterize AgentRadixSortDownsweep type for the current configuration\n typedef AgentRadixSortDownsweep<\n typename If<(ALT_DIGIT_BITS),\n typename ChainedPolicyT::ActivePolicy::AltDownsweepPolicy,\n typename ChainedPolicyT::ActivePolicy::DownsweepPolicy>::Type,\n IS_DESCENDING,\n KeyT,\n ValueT,\n OffsetT>\n AgentRadixSortDownsweepT;\n\n // Shared memory storage\n __shared__ typename AgentRadixSortDownsweepT::TempStorage temp_storage;\n\n // Initialize even-share descriptor for this thread block\n even_share.template BlockInit();\n\n // Process input tiles\n AgentRadixSortDownsweepT(temp_storage, num_items, d_spine, d_keys_in, d_keys_out, d_values_in, d_values_out, current_bit, num_bits).ProcessRegion(\n even_share.block_offset,\n even_share.block_end);\n}\n\n\n/**\n * Single pass kernel entry point (single-block). Fully sorts a tile of input.\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::SingleTilePolicy::BLOCK_THREADS), 1)\n__global__ void DeviceRadixSortSingleTileKernel(\n const KeyT *d_keys_in, ///< [in] Input keys buffer\n KeyT *d_keys_out, ///< [in] Output keys buffer\n const ValueT *d_values_in, ///< [in] Input values buffer\n ValueT *d_values_out, ///< [in] Output values buffer\n OffsetT num_items, ///< [in] Total number of input data items\n int current_bit, ///< [in] Bit position of current radix digit\n int end_bit) ///< [in] The past-the-end (most-significant) bit index needed for key comparison\n{\n // Constants\n enum\n {\n BLOCK_THREADS = ChainedPolicyT::ActivePolicy::SingleTilePolicy::BLOCK_THREADS,\n ITEMS_PER_THREAD = ChainedPolicyT::ActivePolicy::SingleTilePolicy::ITEMS_PER_THREAD,\n KEYS_ONLY = Equals::VALUE,\n };\n\n // BlockRadixSort type\n typedef BlockRadixSort<\n KeyT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n ValueT,\n ChainedPolicyT::ActivePolicy::SingleTilePolicy::RADIX_BITS,\n (ChainedPolicyT::ActivePolicy::SingleTilePolicy::RANK_ALGORITHM == RADIX_RANK_MEMOIZE),\n ChainedPolicyT::ActivePolicy::SingleTilePolicy::SCAN_ALGORITHM>\n BlockRadixSortT;\n\n // BlockLoad type (keys)\n typedef BlockLoad<\n KeyT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n ChainedPolicyT::ActivePolicy::SingleTilePolicy::LOAD_ALGORITHM> BlockLoadKeys;\n\n // BlockLoad type (values)\n typedef BlockLoad<\n ValueT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n ChainedPolicyT::ActivePolicy::SingleTilePolicy::LOAD_ALGORITHM> BlockLoadValues;\n\n // Unsigned word for key bits\n typedef typename Traits::UnsignedBits UnsignedBitsT;\n\n // Shared memory storage\n __shared__ union TempStorage\n {\n typename BlockRadixSortT::TempStorage sort;\n typename BlockLoadKeys::TempStorage load_keys;\n typename BlockLoadValues::TempStorage load_values;\n\n } temp_storage;\n\n // Keys and values for the block\n KeyT keys[ITEMS_PER_THREAD];\n ValueT values[ITEMS_PER_THREAD];\n\n // Get default (min/max) value for out-of-bounds keys\n UnsignedBitsT default_key_bits = (IS_DESCENDING) ? Traits::LOWEST_KEY : Traits::MAX_KEY;\n KeyT default_key = reinterpret_cast(default_key_bits);\n\n // Load keys\n BlockLoadKeys(temp_storage.load_keys).Load(d_keys_in, keys, num_items, default_key);\n\n CTA_SYNC();\n\n // Load values\n if (!KEYS_ONLY)\n {\n BlockLoadValues(temp_storage.load_values).Load(d_values_in, values, num_items);\n\n CTA_SYNC();\n }\n\n // Sort tile\n BlockRadixSortT(temp_storage.sort).SortBlockedToStriped(\n keys,\n values,\n current_bit,\n end_bit,\n Int2Type(),\n Int2Type());\n\n // Store keys and values\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n int item_offset = ITEM * BLOCK_THREADS + threadIdx.x;\n if (item_offset < num_items)\n {\n d_keys_out[item_offset] = keys[ITEM];\n if (!KEYS_ONLY)\n d_values_out[item_offset] = values[ITEM];\n }\n }\n}\n\n\n/**\n * Segmented radix sorting pass (one block per segment)\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n bool ALT_DIGIT_BITS, ///< Whether or not to use the alternate (lower-bits) policy\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetIteratorT, ///< Random-access input iterator type for reading segment offsets \\iterator\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int((ALT_DIGIT_BITS) ?\n ChainedPolicyT::ActivePolicy::AltSegmentedPolicy::BLOCK_THREADS :\n ChainedPolicyT::ActivePolicy::SegmentedPolicy::BLOCK_THREADS))\n__global__ void DeviceSegmentedRadixSortKernel(\n const KeyT *d_keys_in, ///< [in] Input keys buffer\n KeyT *d_keys_out, ///< [in] Output keys buffer\n const ValueT *d_values_in, ///< [in] Input values buffer\n ValueT *d_values_out, ///< [in] Output values buffer\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int /*num_segments*/, ///< [in] The number of segments that comprise the sorting data\n int current_bit, ///< [in] Bit position of current radix digit\n int pass_bits) ///< [in] Number of bits of current radix digit\n{\n //\n // Constants\n //\n\n typedef typename If<(ALT_DIGIT_BITS),\n typename ChainedPolicyT::ActivePolicy::AltSegmentedPolicy,\n typename ChainedPolicyT::ActivePolicy::SegmentedPolicy>::Type SegmentedPolicyT;\n\n enum\n {\n BLOCK_THREADS = SegmentedPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = SegmentedPolicyT::ITEMS_PER_THREAD,\n RADIX_BITS = SegmentedPolicyT::RADIX_BITS,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n RADIX_DIGITS = 1 << RADIX_BITS,\n KEYS_ONLY = Equals::VALUE,\n };\n\n // Upsweep type\n typedef AgentRadixSortUpsweep<\n AgentRadixSortUpsweepPolicy,\n KeyT,\n OffsetT>\n BlockUpsweepT;\n\n // Digit-scan type\n typedef BlockScan DigitScanT;\n\n // Downsweep type\n typedef AgentRadixSortDownsweep BlockDownsweepT;\n\n enum\n {\n /// Number of bin-starting offsets tracked per thread\n BINS_TRACKED_PER_THREAD = BlockDownsweepT::BINS_TRACKED_PER_THREAD\n };\n\n //\n // Process input tiles\n //\n\n // Shared memory storage\n __shared__ union\n {\n typename BlockUpsweepT::TempStorage upsweep;\n typename BlockDownsweepT::TempStorage downsweep;\n struct\n {\n volatile OffsetT reverse_counts_in[RADIX_DIGITS];\n volatile OffsetT reverse_counts_out[RADIX_DIGITS];\n typename DigitScanT::TempStorage scan;\n };\n\n } temp_storage;\n\n OffsetT segment_begin = d_begin_offsets[blockIdx.x];\n OffsetT segment_end = d_end_offsets[blockIdx.x];\n OffsetT num_items = segment_end - segment_begin;\n\n // Check if empty segment\n if (num_items <= 0)\n return;\n\n // Upsweep\n BlockUpsweepT upsweep(temp_storage.upsweep, d_keys_in, current_bit, pass_bits);\n upsweep.ProcessRegion(segment_begin, segment_end);\n\n CTA_SYNC();\n\n // The count of each digit value in this pass (valid in the first RADIX_DIGITS threads)\n OffsetT bin_count[BINS_TRACKED_PER_THREAD];\n upsweep.ExtractCounts(bin_count);\n\n CTA_SYNC();\n\n if (IS_DESCENDING)\n {\n // Reverse bin counts\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n temp_storage.reverse_counts_in[bin_idx] = bin_count[track];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n bin_count[track] = temp_storage.reverse_counts_in[RADIX_DIGITS - bin_idx - 1];\n }\n }\n\n // Scan\n OffsetT bin_offset[BINS_TRACKED_PER_THREAD]; // The global scatter base offset for each digit value in this pass (valid in the first RADIX_DIGITS threads)\n DigitScanT(temp_storage.scan).ExclusiveSum(bin_count, bin_offset);\n\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n bin_offset[track] += segment_begin;\n }\n\n if (IS_DESCENDING)\n {\n // Reverse bin offsets\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n temp_storage.reverse_counts_out[threadIdx.x] = bin_offset[track];\n }\n\n CTA_SYNC();\n\n #pragma unroll\n for (int track = 0; track < BINS_TRACKED_PER_THREAD; ++track)\n {\n int bin_idx = (threadIdx.x * BINS_TRACKED_PER_THREAD) + track;\n\n if ((BLOCK_THREADS == RADIX_DIGITS) || (bin_idx < RADIX_DIGITS))\n bin_offset[track] = temp_storage.reverse_counts_out[RADIX_DIGITS - bin_idx - 1];\n }\n }\n\n CTA_SYNC();\n\n // Downsweep\n BlockDownsweepT downsweep(temp_storage.downsweep, bin_offset, num_items, d_keys_in, d_keys_out, d_values_in, d_values_out, current_bit, pass_bits);\n downsweep.ProcessRegion(segment_begin, segment_end);\n}\n\n\n\n/******************************************************************************\n * Policy\n ******************************************************************************/\n\n/**\n * Tuning policy for kernel specialization\n */\ntemplate <\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DeviceRadixSortPolicy\n{\n //------------------------------------------------------------------------------\n // Constants\n //------------------------------------------------------------------------------\n\n enum\n {\n // Whether this is a keys-only (or key-value) sort\n KEYS_ONLY = (Equals::VALUE),\n\n // Relative size of KeyT type to a 4-byte word\n SCALE_FACTOR_4B = (CUB_MAX(sizeof(KeyT), sizeof(ValueT)) + 3) / 4,\n };\n\n //------------------------------------------------------------------------------\n // Architecture-specific tuning policies\n //------------------------------------------------------------------------------\n\n /// SM13\n struct Policy130 : ChainedPolicy<130, Policy130, Policy130>\n {\n enum {\n PRIMARY_RADIX_BITS = 5,\n ALT_RADIX_BITS = PRIMARY_RADIX_BITS - 1,\n };\n\n // Keys-only upsweep policies\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 19 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyKeys;\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 15 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyKeys;\n\n // Key-value pairs upsweep policies\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 19 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyPairs;\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 15 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyPairs;\n\n // Upsweep policies\n typedef typename If::Type UpsweepPolicy;\n typedef typename If::Type AltUpsweepPolicy;\n\n // Scan policy\n typedef AgentScanPolicy <256, 4, BLOCK_LOAD_VECTORIZE, LOAD_DEFAULT, BLOCK_STORE_VECTORIZE, BLOCK_SCAN_WARP_SCANS> ScanPolicy;\n\n // Keys-only downsweep policies\n typedef AgentRadixSortDownsweepPolicy <64, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyKeys;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 15 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyKeys;\n\n // Key-value pairs downsweep policies\n typedef AgentRadixSortDownsweepPolicy <64, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyPairs;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 15 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyPairs;\n\n // Downsweep policies\n typedef typename If::Type DownsweepPolicy;\n typedef typename If::Type AltDownsweepPolicy;\n\n // Single-tile policy\n typedef DownsweepPolicy SingleTilePolicy;\n\n // Segmented policies\n typedef DownsweepPolicy SegmentedPolicy;\n typedef AltDownsweepPolicy AltSegmentedPolicy;\n };\n\n /// SM20\n struct Policy200 : ChainedPolicy<200, Policy200, Policy130>\n {\n enum {\n PRIMARY_RADIX_BITS = 5,\n ALT_RADIX_BITS = PRIMARY_RADIX_BITS - 1,\n };\n\n // Keys-only upsweep policies\n typedef AgentRadixSortUpsweepPolicy <64, CUB_MAX(1, 18 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyKeys;\n typedef AgentRadixSortUpsweepPolicy <64, CUB_MAX(1, 18 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyKeys;\n\n // Key-value pairs upsweep policies\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 13 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyPairs;\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 13 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyPairs;\n\n // Upsweep policies\n typedef typename If::Type UpsweepPolicy;\n typedef typename If::Type AltUpsweepPolicy;\n\n // Scan policy\n typedef AgentScanPolicy <512, 4, BLOCK_LOAD_VECTORIZE, LOAD_DEFAULT, BLOCK_STORE_VECTORIZE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Keys-only downsweep policies\n typedef AgentRadixSortDownsweepPolicy <64, CUB_MAX(1, 18 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyKeys;\n typedef AgentRadixSortDownsweepPolicy <64, CUB_MAX(1, 18 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyKeys;\n\n // Key-value pairs downsweep policies\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 13 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyPairs;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 13 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyPairs;\n\n // Downsweep policies\n typedef typename If::Type DownsweepPolicy;\n typedef typename If::Type AltDownsweepPolicy;\n\n // Single-tile policy\n typedef DownsweepPolicy SingleTilePolicy;\n\n // Segmented policies\n typedef DownsweepPolicy SegmentedPolicy;\n typedef AltDownsweepPolicy AltSegmentedPolicy;\n };\n\n /// SM30\n struct Policy300 : ChainedPolicy<300, Policy300, Policy200>\n {\n enum {\n PRIMARY_RADIX_BITS = 5,\n ALT_RADIX_BITS = PRIMARY_RADIX_BITS - 1,\n };\n\n // Keys-only upsweep policies\n typedef AgentRadixSortUpsweepPolicy <256, CUB_MAX(1, 7 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyKeys;\n typedef AgentRadixSortUpsweepPolicy <256, CUB_MAX(1, 7 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyKeys;\n\n // Key-value pairs upsweep policies\n typedef AgentRadixSortUpsweepPolicy <256, CUB_MAX(1, 5 / SCALE_FACTOR_4B), LOAD_DEFAULT, PRIMARY_RADIX_BITS> UpsweepPolicyPairs;\n typedef AgentRadixSortUpsweepPolicy <256, CUB_MAX(1, 5 / SCALE_FACTOR_4B), LOAD_DEFAULT, ALT_RADIX_BITS> AltUpsweepPolicyPairs;\n\n // Upsweep policies\n typedef typename If::Type UpsweepPolicy;\n typedef typename If::Type AltUpsweepPolicy;\n\n // Scan policy\n typedef AgentScanPolicy <1024, 4, BLOCK_LOAD_VECTORIZE, LOAD_DEFAULT, BLOCK_STORE_VECTORIZE, BLOCK_SCAN_WARP_SCANS> ScanPolicy;\n\n // Keys-only downsweep policies\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 14 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyKeys;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 14 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyKeys;\n\n // Key-value pairs downsweep policies\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 10 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyPairs;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 10 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, ALT_RADIX_BITS> AltDownsweepPolicyPairs;\n\n // Downsweep policies\n typedef typename If::Type DownsweepPolicy;\n typedef typename If::Type AltDownsweepPolicy;\n\n // Single-tile policy\n typedef DownsweepPolicy SingleTilePolicy;\n\n // Segmented policies\n typedef DownsweepPolicy SegmentedPolicy;\n typedef AltDownsweepPolicy AltSegmentedPolicy;\n };\n\n\n /// SM35\n struct Policy350 : ChainedPolicy<350, Policy350, Policy300>\n {\n enum {\n PRIMARY_RADIX_BITS = 6, // 1.72B 32b keys/s, 1.17B 32b pairs/s, 1.55B 32b segmented keys/s (K40m)\n };\n\n // Scan policy\n typedef AgentScanPolicy <1024, 4, BLOCK_LOAD_VECTORIZE, LOAD_DEFAULT, BLOCK_STORE_VECTORIZE, BLOCK_SCAN_WARP_SCANS> ScanPolicy;\n\n // Keys-only downsweep policies\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 9 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_LDG, RADIX_RANK_MATCH, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicyKeys;\n typedef AgentRadixSortDownsweepPolicy <64, CUB_MAX(1, 18 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicyKeys;\n\n // Key-value pairs downsweep policies\n typedef DownsweepPolicyKeys DownsweepPolicyPairs;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 15 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicyPairs;\n\n // Downsweep policies\n typedef typename If::Type DownsweepPolicy;\n typedef typename If::Type AltDownsweepPolicy;\n\n // Upsweep policies\n typedef DownsweepPolicy UpsweepPolicy;\n typedef AltDownsweepPolicy AltUpsweepPolicy;\n\n // Single-tile policy\n typedef DownsweepPolicy SingleTilePolicy;\n\n // Segmented policies\n typedef DownsweepPolicy SegmentedPolicy;\n typedef AltDownsweepPolicy AltSegmentedPolicy;\n\n\n };\n\n\n /// SM50\n struct Policy500 : ChainedPolicy<500, Policy500, Policy350>\n {\n enum {\n PRIMARY_RADIX_BITS = 7, // 3.5B 32b keys/s, 1.92B 32b pairs/s (TitanX)\n SINGLE_TILE_RADIX_BITS = 6,\n SEGMENTED_RADIX_BITS = 6, // 3.1B 32b segmented keys/s (TitanX)\n };\n\n // ScanPolicy\n typedef AgentScanPolicy <512, 23, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Downsweep policies\n typedef AgentRadixSortDownsweepPolicy <160, CUB_MAX(1, 39 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_BASIC, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 16 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_RAKING_MEMOIZE, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicy;\n\n // Upsweep policies\n typedef DownsweepPolicy UpsweepPolicy;\n typedef AltDownsweepPolicy AltUpsweepPolicy;\n\n // Single-tile policy\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SINGLE_TILE_RADIX_BITS> SingleTilePolicy;\n\n // Segmented policies\n typedef AgentRadixSortDownsweepPolicy <192, CUB_MAX(1, 31 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS> SegmentedPolicy;\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 11 / SCALE_FACTOR_4B), BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS - 1> AltSegmentedPolicy;\n };\n\n\n /// SM60 (GP100)\n struct Policy600 : ChainedPolicy<600, Policy600, Policy500>\n {\n enum {\n PRIMARY_RADIX_BITS = 7, // 6.9B 32b keys/s (Quadro P100)\n SINGLE_TILE_RADIX_BITS = 6,\n SEGMENTED_RADIX_BITS = 6, // 5.9B 32b segmented keys/s (Quadro P100)\n };\n\n // ScanPolicy\n typedef AgentScanPolicy <512, 23, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Downsweep policies\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 25 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MATCH, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <192, CUB_MAX(1, 39 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicy;\n\n // Upsweep policies\n typedef DownsweepPolicy UpsweepPolicy;\n typedef AltDownsweepPolicy AltUpsweepPolicy;\n\n // Single-tile policy\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SINGLE_TILE_RADIX_BITS> SingleTilePolicy;\n\n // Segmented policies\n typedef AgentRadixSortDownsweepPolicy <192, CUB_MAX(1, 39 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS> SegmentedPolicy;\n typedef AgentRadixSortDownsweepPolicy <384, CUB_MAX(1, 11 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS - 1> AltSegmentedPolicy;\n\n };\n\n\n /// SM61 (GP104)\n struct Policy610 : ChainedPolicy<610, Policy610, Policy600>\n {\n enum {\n PRIMARY_RADIX_BITS = 7, // 3.4B 32b keys/s, 1.83B 32b pairs/s (1080)\n SINGLE_TILE_RADIX_BITS = 6,\n SEGMENTED_RADIX_BITS = 6, // 3.3B 32b segmented keys/s (1080)\n };\n\n // ScanPolicy\n typedef AgentScanPolicy <512, 23, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Downsweep policies\n typedef AgentRadixSortDownsweepPolicy <384, CUB_MAX(1, 31 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_DEFAULT, RADIX_RANK_MATCH, BLOCK_SCAN_RAKING_MEMOIZE, PRIMARY_RADIX_BITS> DownsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 35 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_RAKING_MEMOIZE, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicy;\n\n // Upsweep policies\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 16 / SCALE_FACTOR_4B), LOAD_LDG, PRIMARY_RADIX_BITS> UpsweepPolicy;\n typedef AgentRadixSortUpsweepPolicy <128, CUB_MAX(1, 16 / SCALE_FACTOR_4B), LOAD_LDG, PRIMARY_RADIX_BITS - 1> AltUpsweepPolicy;\n\n // Single-tile policy\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SINGLE_TILE_RADIX_BITS> SingleTilePolicy;\n\n // Segmented policies\n typedef AgentRadixSortDownsweepPolicy <192, CUB_MAX(1, 39 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS> SegmentedPolicy;\n typedef AgentRadixSortDownsweepPolicy <384, CUB_MAX(1, 11 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS - 1> AltSegmentedPolicy;\n };\n\n\n /// SM62 (Tegra, less RF)\n struct Policy620 : ChainedPolicy<620, Policy620, Policy610>\n {\n enum {\n PRIMARY_RADIX_BITS = 5,\n ALT_RADIX_BITS = PRIMARY_RADIX_BITS - 1,\n };\n\n // ScanPolicy\n typedef AgentScanPolicy <512, 23, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Downsweep policies\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 16 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_RAKING_MEMOIZE, PRIMARY_RADIX_BITS> DownsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 16 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_RAKING_MEMOIZE, ALT_RADIX_BITS> AltDownsweepPolicy;\n\n // Upsweep policies\n typedef DownsweepPolicy UpsweepPolicy;\n typedef AltDownsweepPolicy AltUpsweepPolicy;\n\n // Single-tile policy\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> SingleTilePolicy;\n\n // Segmented policies\n typedef DownsweepPolicy SegmentedPolicy;\n typedef AltDownsweepPolicy AltSegmentedPolicy;\n };\n\n\n /// SM70 (GV100)\n struct Policy700 : ChainedPolicy<700, Policy700, Policy620>\n {\n enum {\n PRIMARY_RADIX_BITS = 6, // 7.62B 32b keys/s (GV100)\n SINGLE_TILE_RADIX_BITS = 6,\n SEGMENTED_RADIX_BITS = 6, // 8.7B 32b segmented keys/s (GV100)\n };\n\n // ScanPolicy\n typedef AgentScanPolicy <512, 23, BLOCK_LOAD_WARP_TRANSPOSE, LOAD_DEFAULT, BLOCK_STORE_WARP_TRANSPOSE, BLOCK_SCAN_RAKING_MEMOIZE> ScanPolicy;\n\n // Downsweep policies\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 47 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> DownsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <384, CUB_MAX(1, 29 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS - 1> AltDownsweepPolicy;\n\n // Upsweep policies\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 47 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MATCH, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS> UpsweepPolicy;\n typedef AgentRadixSortDownsweepPolicy <128, CUB_MAX(1, 29 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MATCH, BLOCK_SCAN_WARP_SCANS, PRIMARY_RADIX_BITS - 1> AltUpsweepPolicy;\n\n // Single-tile policy\n typedef AgentRadixSortDownsweepPolicy <256, CUB_MAX(1, 19 / SCALE_FACTOR_4B), BLOCK_LOAD_DIRECT, LOAD_LDG, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SINGLE_TILE_RADIX_BITS> SingleTilePolicy;\n\n // Segmented policies\n typedef AgentRadixSortDownsweepPolicy <192, CUB_MAX(1, 39 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS> SegmentedPolicy;\n typedef AgentRadixSortDownsweepPolicy <384, CUB_MAX(1, 11 / SCALE_FACTOR_4B), BLOCK_LOAD_TRANSPOSE, LOAD_DEFAULT, RADIX_RANK_MEMOIZE, BLOCK_SCAN_WARP_SCANS, SEGMENTED_RADIX_BITS - 1> AltSegmentedPolicy;\n };\n\n\n /// MaxPolicy\n typedef Policy700 MaxPolicy;\n\n\n};\n\n\n\n/******************************************************************************\n * Single-problem dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for device-wide radix sort\n */\ntemplate <\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DispatchRadixSort :\n DeviceRadixSortPolicy\n{\n //------------------------------------------------------------------------------\n // Constants\n //------------------------------------------------------------------------------\n\n enum\n {\n // Whether this is a keys-only (or key-value) sort\n KEYS_ONLY = (Equals::VALUE),\n };\n\n\n //------------------------------------------------------------------------------\n // Problem state\n //------------------------------------------------------------------------------\n\n void *d_temp_storage; ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes; ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys; ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values; ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n OffsetT num_items; ///< [in] Number of items to sort\n int begin_bit; ///< [in] The beginning (least-significant) bit index needed for key comparison\n int end_bit; ///< [in] The past-the-end (most-significant) bit index needed for key comparison\n cudaStream_t stream; ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous; ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int ptx_version; ///< [in] PTX version\n bool is_overwrite_okay; ///< [in] Whether is okay to overwrite source buffers\n\n\n //------------------------------------------------------------------------------\n // Constructor\n //------------------------------------------------------------------------------\n\n /// Constructor\n CUB_RUNTIME_FUNCTION __forceinline__\n DispatchRadixSort(\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n OffsetT num_items,\n int begin_bit,\n int end_bit,\n bool is_overwrite_okay,\n cudaStream_t stream,\n bool debug_synchronous,\n int ptx_version)\n :\n d_temp_storage(d_temp_storage),\n temp_storage_bytes(temp_storage_bytes),\n d_keys(d_keys),\n d_values(d_values),\n num_items(num_items),\n begin_bit(begin_bit),\n end_bit(end_bit),\n stream(stream),\n debug_synchronous(debug_synchronous),\n ptx_version(ptx_version),\n is_overwrite_okay(is_overwrite_okay)\n {}\n\n\n //------------------------------------------------------------------------------\n // Small-problem (single tile) invocation\n //------------------------------------------------------------------------------\n\n /// Invoke a single block to sort in-core\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename SingleTileKernelT> ///< Function type of cub::DeviceRadixSortSingleTileKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokeSingleTile(\n SingleTileKernelT single_tile_kernel) ///< [in] Kernel function pointer to parameterization of cub::DeviceRadixSortSingleTileKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void)single_tile_kernel;\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n cudaError error = cudaSuccess;\n do\n {\n // Return if the caller is simply requesting the size of the storage allocation\n if (d_temp_storage == NULL)\n {\n temp_storage_bytes = 1;\n break;\n }\n\n // Return if empty problem\n if (num_items == 0)\n break;\n\n // Log single_tile_kernel configuration\n if (debug_synchronous)\n _CubLog(\"Invoking single_tile_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\\n\",\n 1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, (long long) stream,\n ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD, 1, begin_bit, ActivePolicyT::SingleTilePolicy::RADIX_BITS);\n\n // Invoke upsweep_kernel with same grid size as downsweep_kernel\n single_tile_kernel<<<1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, 0, stream>>>(\n d_keys.Current(),\n d_keys.Alternate(),\n d_values.Current(),\n d_values.Alternate(),\n num_items,\n begin_bit,\n end_bit);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Update selector\n d_keys.selector ^= 1;\n d_values.selector ^= 1;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n //------------------------------------------------------------------------------\n // Normal problem size invocation\n //------------------------------------------------------------------------------\n\n /**\n * Invoke a three-kernel sorting pass at the current bit.\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePass(\n const KeyT *d_keys_in,\n KeyT *d_keys_out,\n const ValueT *d_values_in,\n ValueT *d_values_out,\n OffsetT *d_spine,\n int spine_length,\n int ¤t_bit,\n PassConfigT &pass_config)\n {\n cudaError error = cudaSuccess;\n do\n {\n int pass_bits = CUB_MIN(pass_config.radix_bits, (end_bit - current_bit));\n\n // Log upsweep_kernel configuration\n if (debug_synchronous)\n _CubLog(\"Invoking upsweep_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\\n\",\n pass_config.even_share.grid_size, pass_config.upsweep_config.block_threads, (long long) stream,\n pass_config.upsweep_config.items_per_thread, pass_config.upsweep_config.sm_occupancy, current_bit, pass_bits);\n\n\t\t\t\t\t\t/*printf(\"running upsweep kernel with pass_bits %d\\n\", pass_bits);*/\n\t\t\t\t\t\t/*printf(\"Invoking upsweep_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\\n\",*/\n\t\t\t\t\t\t\t\t/*pass_config.even_share.grid_size, pass_config.upsweep_config.block_threads, (long long) stream,*/\n\t\t\t\t\t\t\t\t/*pass_config.upsweep_config.items_per_thread, pass_config.upsweep_config.sm_occupancy, current_bit, pass_bits);*/\n\n // Invoke upsweep_kernel with same grid size as downsweep_kernel\n pass_config.upsweep_kernel<<>>(\n d_keys_in,\n d_spine,\n num_items,\n current_bit,\n pass_bits,\n pass_config.even_share);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Log scan_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking scan_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread\\n\",\n 1, pass_config.scan_config.block_threads, (long long) stream, pass_config.scan_config.items_per_thread);\n\n // Invoke scan_kernel\n pass_config.scan_kernel<<<1, pass_config.scan_config.block_threads, 0, stream>>>(\n d_spine,\n spine_length);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Log downsweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking downsweep_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n pass_config.even_share.grid_size, pass_config.downsweep_config.block_threads, (long long) stream,\n pass_config.downsweep_config.items_per_thread, pass_config.downsweep_config.sm_occupancy);\n\n // Invoke downsweep_kernel\n pass_config.downsweep_kernel<<>>(\n d_keys_in,\n d_keys_out,\n d_values_in,\n d_values_out,\n d_spine,\n num_items,\n current_bit,\n pass_bits,\n pass_config.even_share);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Update current bit\n current_bit += pass_bits;\n }\n while (0);\n\n return error;\n }\n\n\n\n /// Pass configuration structure\n template <\n typename UpsweepKernelT,\n typename ScanKernelT,\n typename DownsweepKernelT>\n struct PassConfig\n {\n UpsweepKernelT upsweep_kernel;\n KernelConfig upsweep_config;\n ScanKernelT scan_kernel;\n KernelConfig scan_config;\n DownsweepKernelT downsweep_kernel;\n KernelConfig downsweep_config;\n int radix_bits;\n int radix_digits;\n int max_downsweep_grid_size;\n GridEvenShare even_share;\n\n /// Initialize pass configuration\n template <\n typename UpsweepPolicyT,\n typename ScanPolicyT,\n typename DownsweepPolicyT>\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InitPassConfig(\n UpsweepKernelT upsweep_kernel,\n ScanKernelT scan_kernel,\n DownsweepKernelT downsweep_kernel,\n int ptx_version,\n int sm_count,\n int num_items)\n {\n cudaError error = cudaSuccess;\n do\n {\n this->upsweep_kernel = upsweep_kernel;\n this->scan_kernel = scan_kernel;\n this->downsweep_kernel = downsweep_kernel;\n radix_bits = DownsweepPolicyT::RADIX_BITS;\n radix_digits = 1 << radix_bits;\n\n if (CubDebug(error = upsweep_config.Init(upsweep_kernel))) break;\n if (CubDebug(error = scan_config.Init(scan_kernel))) break;\n if (CubDebug(error = downsweep_config.Init(downsweep_kernel))) break;\n\n max_downsweep_grid_size = (downsweep_config.sm_occupancy * sm_count) * CUB_SUBSCRIPTION_FACTOR(ptx_version);\n\n even_share.DispatchInit(\n num_items,\n max_downsweep_grid_size,\n CUB_MAX(downsweep_config.tile_size, upsweep_config.tile_size));\n\n }\n while (0);\n return error;\n }\n\n };\n\n\n /// Invocation (run multiple digit passes)\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename UpsweepKernelT, ///< Function type of cub::DeviceRadixSortUpsweepKernel\n typename ScanKernelT, ///< Function type of cub::SpineScanKernel\n typename DownsweepKernelT> ///< Function type of cub::DeviceRadixSortDownsweepKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePasses(\n UpsweepKernelT upsweep_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceRadixSortUpsweepKernel\n UpsweepKernelT alt_upsweep_kernel, ///< [in] Alternate kernel function pointer to parameterization of cub::DeviceRadixSortUpsweepKernel\n ScanKernelT scan_kernel, ///< [in] Kernel function pointer to parameterization of cub::SpineScanKernel\n DownsweepKernelT downsweep_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceRadixSortDownsweepKernel\n DownsweepKernelT alt_downsweep_kernel) ///< [in] Alternate kernel function pointer to parameterization of cub::DeviceRadixSortDownsweepKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void)upsweep_kernel;\n (void)alt_upsweep_kernel;\n (void)scan_kernel;\n (void)downsweep_kernel;\n (void)alt_downsweep_kernel;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Init regular and alternate-digit kernel configurations\n PassConfig pass_config, alt_pass_config;\n if ((error = pass_config.template InitPassConfig<\n typename ActivePolicyT::UpsweepPolicy,\n typename ActivePolicyT::ScanPolicy,\n typename ActivePolicyT::DownsweepPolicy>(\n upsweep_kernel, scan_kernel, downsweep_kernel, ptx_version, sm_count, num_items))) break;\n\n if ((error = alt_pass_config.template InitPassConfig<\n typename ActivePolicyT::AltUpsweepPolicy,\n typename ActivePolicyT::ScanPolicy,\n typename ActivePolicyT::AltDownsweepPolicy>(\n alt_upsweep_kernel, scan_kernel, alt_downsweep_kernel, ptx_version, sm_count, num_items))) break;\n\n // Get maximum spine length\n int max_grid_size = CUB_MAX(pass_config.max_downsweep_grid_size, alt_pass_config.max_downsweep_grid_size);\n int spine_length = (max_grid_size * pass_config.radix_digits) + pass_config.scan_config.tile_size;\n\n // Temporary storage allocation requirements\n void* allocations[3];\n size_t allocation_sizes[3] =\n {\n spine_length * sizeof(OffsetT), // bytes needed for privatized block digit histograms\n (is_overwrite_okay) ? 0 : num_items * sizeof(KeyT), // bytes needed for 3rd keys buffer\n (is_overwrite_okay || (KEYS_ONLY)) ? 0 : num_items * sizeof(ValueT), // bytes needed for 3rd values buffer\n };\n\n // Alias the temporary allocations from the single storage blob (or compute the necessary size of the blob)\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n\n // Return if the caller is simply requesting the size of the storage allocation\n if (d_temp_storage == NULL)\n return cudaSuccess;\n\n // Pass planning. Run passes of the alternate digit-size configuration until we have an even multiple of our preferred digit size\n int num_bits = end_bit - begin_bit;\n int num_passes = (num_bits + pass_config.radix_bits - 1) / pass_config.radix_bits;\n bool is_num_passes_odd = num_passes & 1;\n int max_alt_passes = (num_passes * pass_config.radix_bits) - num_bits;\n int alt_end_bit = CUB_MIN(end_bit, begin_bit + (max_alt_passes * alt_pass_config.radix_bits));\n\n // Alias the temporary storage allocations\n OffsetT *d_spine = static_cast(allocations[0]);\n\n DoubleBuffer d_keys_remaining_passes(\n (is_overwrite_okay || is_num_passes_odd) ? d_keys.Alternate() : static_cast(allocations[1]),\n (is_overwrite_okay) ? d_keys.Current() : (is_num_passes_odd) ? static_cast(allocations[1]) : d_keys.Alternate());\n\n DoubleBuffer d_values_remaining_passes(\n (is_overwrite_okay || is_num_passes_odd) ? d_values.Alternate() : static_cast(allocations[2]),\n (is_overwrite_okay) ? d_values.Current() : (is_num_passes_odd) ? static_cast(allocations[2]) : d_values.Alternate());\n\n // Run first pass, consuming from the input's current buffers\n int current_bit = begin_bit;\n if (CubDebug(error = InvokePass(\n d_keys.Current(), d_keys_remaining_passes.Current(),\n d_values.Current(), d_values_remaining_passes.Current(),\n d_spine, spine_length, current_bit,\n (current_bit < alt_end_bit) ? alt_pass_config : pass_config))) break;\n\n // Run remaining passes\n while (current_bit < end_bit)\n {\n if (CubDebug(error = InvokePass(\n d_keys_remaining_passes.d_buffers[d_keys_remaining_passes.selector], d_keys_remaining_passes.d_buffers[d_keys_remaining_passes.selector ^ 1],\n d_values_remaining_passes.d_buffers[d_keys_remaining_passes.selector], d_values_remaining_passes.d_buffers[d_keys_remaining_passes.selector ^ 1],\n d_spine, spine_length, current_bit,\n (current_bit < alt_end_bit) ? alt_pass_config : pass_config))) break;;\n\n // Invert selectors\n d_keys_remaining_passes.selector ^= 1;\n d_values_remaining_passes.selector ^= 1;\n }\n\n // Update selector\n if (!is_overwrite_okay) {\n num_passes = 1; // Sorted data always ends up in the other vector\n }\n\n d_keys.selector = (d_keys.selector + num_passes) & 1;\n d_values.selector = (d_values.selector + num_passes) & 1;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n //------------------------------------------------------------------------------\n // Chained policy invocation\n //------------------------------------------------------------------------------\n\n /// Invocation\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Invoke()\n {\n typedef typename DispatchRadixSort::MaxPolicy MaxPolicyT;\n typedef typename ActivePolicyT::SingleTilePolicy SingleTilePolicyT;\n\n // Force kernel code-generation in all compiler passes\n if (num_items <= (SingleTilePolicyT::BLOCK_THREADS * SingleTilePolicyT::ITEMS_PER_THREAD))\n {\n // Small, single tile size\n return InvokeSingleTile(\n DeviceRadixSortSingleTileKernel);\n }\n else\n {\n // Regular size\n return InvokePasses(\n DeviceRadixSortUpsweepKernel< MaxPolicyT, false, IS_DESCENDING, KeyT, OffsetT>,\n DeviceRadixSortUpsweepKernel< MaxPolicyT, true, IS_DESCENDING, KeyT, OffsetT>,\n RadixSortScanBinsKernel< MaxPolicyT, OffsetT>,\n DeviceRadixSortDownsweepKernel< MaxPolicyT, false, IS_DESCENDING, KeyT, ValueT, OffsetT>,\n DeviceRadixSortDownsweepKernel< MaxPolicyT, true, IS_DESCENDING, KeyT, ValueT, OffsetT>);\n }\n }\n\n\n //------------------------------------------------------------------------------\n // Dispatch entrypoints\n //------------------------------------------------------------------------------\n\n /**\n * Internal dispatch routine\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n OffsetT num_items, ///< [in] Number of items to sort\n int begin_bit, ///< [in] The beginning (least-significant) bit index needed for key comparison\n int end_bit, ///< [in] The past-the-end (most-significant) bit index needed for key comparison\n bool is_overwrite_okay, ///< [in] Whether is okay to overwrite source buffers\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n typedef typename DispatchRadixSort::MaxPolicy MaxPolicyT;\n\n cudaError_t error;\n do {\n // Get PTX version\n int ptx_version;\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n\n // Create dispatch functor\n DispatchRadixSort dispatch(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, begin_bit, end_bit, is_overwrite_okay,\n stream, debug_synchronous, ptx_version);\n\n // Dispatch to chained policy\n if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;\n\n } while (0);\n\n return error;\n }\n};\n\n\n\n\n/******************************************************************************\n * Segmented dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for segmented device-wide radix sort\n */\ntemplate <\n bool IS_DESCENDING, ///< Whether or not the sorted-order is high-to-low\n typename KeyT, ///< Key type\n typename ValueT, ///< Value type\n typename OffsetIteratorT, ///< Random-access input iterator type for reading segment offsets \\iterator\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DispatchSegmentedRadixSort :\n DeviceRadixSortPolicy\n{\n //------------------------------------------------------------------------------\n // Constants\n //------------------------------------------------------------------------------\n\n enum\n {\n // Whether this is a keys-only (or key-value) sort\n KEYS_ONLY = (Equals::VALUE),\n };\n\n\n //------------------------------------------------------------------------------\n // Parameter members\n //------------------------------------------------------------------------------\n\n void *d_temp_storage; ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes; ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys; ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values; ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n OffsetT num_items; ///< [in] Number of items to sort\n OffsetT num_segments; ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets; ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets; ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit; ///< [in] The beginning (least-significant) bit index needed for key comparison\n int end_bit; ///< [in] The past-the-end (most-significant) bit index needed for key comparison\n cudaStream_t stream; ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous; ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int ptx_version; ///< [in] PTX version\n bool is_overwrite_okay; ///< [in] Whether is okay to overwrite source buffers\n\n\n //------------------------------------------------------------------------------\n // Constructors\n //------------------------------------------------------------------------------\n\n /// Constructor\n CUB_RUNTIME_FUNCTION __forceinline__\n DispatchSegmentedRadixSort(\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n OffsetT num_items,\n OffsetT num_segments,\n OffsetIteratorT d_begin_offsets,\n OffsetIteratorT d_end_offsets,\n int begin_bit,\n int end_bit,\n bool is_overwrite_okay,\n cudaStream_t stream,\n bool debug_synchronous,\n int ptx_version)\n :\n d_temp_storage(d_temp_storage),\n temp_storage_bytes(temp_storage_bytes),\n d_keys(d_keys),\n d_values(d_values),\n num_items(num_items),\n num_segments(num_segments),\n d_begin_offsets(d_begin_offsets),\n d_end_offsets(d_end_offsets),\n begin_bit(begin_bit),\n end_bit(end_bit),\n is_overwrite_okay(is_overwrite_okay),\n stream(stream),\n debug_synchronous(debug_synchronous),\n ptx_version(ptx_version)\n {}\n\n\n //------------------------------------------------------------------------------\n // Multi-segment invocation\n //------------------------------------------------------------------------------\n\n /// Invoke a three-kernel sorting pass at the current bit.\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePass(\n const KeyT *d_keys_in,\n KeyT *d_keys_out,\n const ValueT *d_values_in,\n ValueT *d_values_out,\n int ¤t_bit,\n PassConfigT &pass_config)\n {\n cudaError error = cudaSuccess;\n do\n {\n int pass_bits = CUB_MIN(pass_config.radix_bits, (end_bit - current_bit));\n\n // Log kernel configuration\n if (debug_synchronous)\n _CubLog(\"Invoking segmented_kernels<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy, current bit %d, bit_grain %d\\n\",\n num_segments, pass_config.segmented_config.block_threads, (long long) stream,\n pass_config.segmented_config.items_per_thread, pass_config.segmented_config.sm_occupancy, current_bit, pass_bits);\n\n pass_config.segmented_kernel<<>>(\n d_keys_in, d_keys_out,\n d_values_in, d_values_out,\n d_begin_offsets, d_end_offsets, num_segments,\n current_bit, pass_bits);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Update current bit\n current_bit += pass_bits;\n }\n while (0);\n\n return error;\n }\n\n\n /// PassConfig data structure\n template \n struct PassConfig\n {\n SegmentedKernelT segmented_kernel;\n KernelConfig segmented_config;\n int radix_bits;\n int radix_digits;\n\n /// Initialize pass configuration\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InitPassConfig(SegmentedKernelT segmented_kernel)\n {\n this->segmented_kernel = segmented_kernel;\n this->radix_bits = SegmentedPolicyT::RADIX_BITS;\n this->radix_digits = 1 << radix_bits;\n\n return CubDebug(segmented_config.Init(segmented_kernel));\n }\n };\n\n\n /// Invocation (run multiple digit passes)\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename SegmentedKernelT> ///< Function type of cub::DeviceSegmentedRadixSortKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePasses(\n SegmentedKernelT segmented_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceSegmentedRadixSortKernel\n SegmentedKernelT alt_segmented_kernel) ///< [in] Alternate kernel function pointer to parameterization of cub::DeviceSegmentedRadixSortKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void)segmented_kernel;\n (void)alt_segmented_kernel;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Init regular and alternate kernel configurations\n PassConfig pass_config, alt_pass_config;\n if ((error = pass_config.template InitPassConfig(segmented_kernel))) break;\n if ((error = alt_pass_config.template InitPassConfig(alt_segmented_kernel))) break;\n\n // Temporary storage allocation requirements\n void* allocations[2];\n size_t allocation_sizes[2] =\n {\n (is_overwrite_okay) ? 0 : num_items * sizeof(KeyT), // bytes needed for 3rd keys buffer\n (is_overwrite_okay || (KEYS_ONLY)) ? 0 : num_items * sizeof(ValueT), // bytes needed for 3rd values buffer\n };\n\n // Alias the temporary allocations from the single storage blob (or compute the necessary size of the blob)\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n\n // Return if the caller is simply requesting the size of the storage allocation\n if (d_temp_storage == NULL)\n {\n if (temp_storage_bytes == 0)\n temp_storage_bytes = 1;\n return cudaSuccess;\n }\n\n // Pass planning. Run passes of the alternate digit-size configuration until we have an even multiple of our preferred digit size\n int radix_bits = ActivePolicyT::SegmentedPolicy::RADIX_BITS;\n int alt_radix_bits = ActivePolicyT::AltSegmentedPolicy::RADIX_BITS;\n int num_bits = end_bit - begin_bit;\n int num_passes = (num_bits + radix_bits - 1) / radix_bits;\n bool is_num_passes_odd = num_passes & 1;\n int max_alt_passes = (num_passes * radix_bits) - num_bits;\n int alt_end_bit = CUB_MIN(end_bit, begin_bit + (max_alt_passes * alt_radix_bits));\n\n DoubleBuffer d_keys_remaining_passes(\n (is_overwrite_okay || is_num_passes_odd) ? d_keys.Alternate() : static_cast(allocations[0]),\n (is_overwrite_okay) ? d_keys.Current() : (is_num_passes_odd) ? static_cast(allocations[0]) : d_keys.Alternate());\n\n DoubleBuffer d_values_remaining_passes(\n (is_overwrite_okay || is_num_passes_odd) ? d_values.Alternate() : static_cast(allocations[1]),\n (is_overwrite_okay) ? d_values.Current() : (is_num_passes_odd) ? static_cast(allocations[1]) : d_values.Alternate());\n\n // Run first pass, consuming from the input's current buffers\n int current_bit = begin_bit;\n\n if (CubDebug(error = InvokePass(\n d_keys.Current(), d_keys_remaining_passes.Current(),\n d_values.Current(), d_values_remaining_passes.Current(),\n current_bit,\n (current_bit < alt_end_bit) ? alt_pass_config : pass_config))) break;\n\n // Run remaining passes\n while (current_bit < end_bit)\n {\n if (CubDebug(error = InvokePass(\n d_keys_remaining_passes.d_buffers[d_keys_remaining_passes.selector], d_keys_remaining_passes.d_buffers[d_keys_remaining_passes.selector ^ 1],\n d_values_remaining_passes.d_buffers[d_keys_remaining_passes.selector], d_values_remaining_passes.d_buffers[d_keys_remaining_passes.selector ^ 1],\n current_bit,\n (current_bit < alt_end_bit) ? alt_pass_config : pass_config))) break;\n\n // Invert selectors and update current bit\n d_keys_remaining_passes.selector ^= 1;\n d_values_remaining_passes.selector ^= 1;\n }\n\n // Update selector\n if (!is_overwrite_okay) {\n num_passes = 1; // Sorted data always ends up in the other vector\n }\n\n d_keys.selector = (d_keys.selector + num_passes) & 1;\n d_values.selector = (d_values.selector + num_passes) & 1;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n //------------------------------------------------------------------------------\n // Chained policy invocation\n //------------------------------------------------------------------------------\n\n /// Invocation\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Invoke()\n {\n typedef typename DispatchSegmentedRadixSort::MaxPolicy MaxPolicyT;\n\n // Force kernel code-generation in all compiler passes\n return InvokePasses(\n DeviceSegmentedRadixSortKernel,\n DeviceSegmentedRadixSortKernel);\n }\n\n\n //------------------------------------------------------------------------------\n // Dispatch entrypoints\n //------------------------------------------------------------------------------\n\n\n /// Internal dispatch routine\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n DoubleBuffer &d_keys, ///< [in,out] Double-buffer whose current buffer contains the unsorted input keys and, upon return, is updated to point to the sorted output keys\n DoubleBuffer &d_values, ///< [in,out] Double-buffer whose current buffer contains the unsorted input values and, upon return, is updated to point to the sorted output values\n int num_items, ///< [in] Number of items to sort\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int begin_bit, ///< [in] The beginning (least-significant) bit index needed for key comparison\n int end_bit, ///< [in] The past-the-end (most-significant) bit index needed for key comparison\n bool is_overwrite_okay, ///< [in] Whether is okay to overwrite source buffers\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n typedef typename DispatchSegmentedRadixSort::MaxPolicy MaxPolicyT;\n\n cudaError_t error;\n do {\n // Get PTX version\n int ptx_version;\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n\n // Create dispatch functor\n DispatchSegmentedRadixSort dispatch(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, num_segments, d_begin_offsets, d_end_offsets,\n begin_bit, end_bit, is_overwrite_okay,\n stream, debug_synchronous, ptx_version);\n\n // Dispatch to chained policy\n if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;\n\n } while (0);\n\n return error;\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_reduce.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceReduce provides device-wide, parallel operations for computing a reduction across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../../agent/agent_reduce.cuh\"\n#include \"../../iterator/arg_index_input_iterator.cuh\"\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../grid/grid_even_share.cuh\"\n#include \"../../iterator/arg_index_input_iterator.cuh\"\n#include \"../../util_debug.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Reduce region kernel entry point (multi-block). Computes privatized reductions, one per thread block.\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OutputIteratorT, ///< Output iterator type for recording the reduced aggregate \\iterator\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT> ///< Binary reduction functor type having member T operator()(const T &a, const T &b)\n__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS))\n__global__ void DeviceReduceKernel(\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n OffsetT num_items, ///< [in] Total number of input data items\n GridEvenShare even_share, ///< [in] Even-share descriptor for mapping an equal number of tiles onto each thread block\n ReductionOpT reduction_op) ///< [in] Binary reduction functor\n{\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Thread block type for reducing input tiles\n typedef AgentReduce<\n typename ChainedPolicyT::ActivePolicy::ReducePolicy,\n InputIteratorT,\n OutputIteratorT,\n OffsetT,\n ReductionOpT>\n AgentReduceT;\n\n // Shared memory storage\n __shared__ typename AgentReduceT::TempStorage temp_storage;\n\n // Consume input tiles\n OutputT block_aggregate = AgentReduceT(temp_storage, d_in, reduction_op).ConsumeTiles(even_share);\n\n // Output result\n if (threadIdx.x == 0)\n d_out[blockIdx.x] = block_aggregate;\n}\n\n\n/**\n * Reduce a single tile kernel entry point (single-block). Can be used to aggregate privatized thread block reductions from a previous multi-block reduction pass.\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OutputIteratorT, ///< Output iterator type for recording the reduced aggregate \\iterator\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT, ///< Binary reduction functor type having member T operator()(const T &a, const T &b)\n typename OuputT> ///< Data element type that is convertible to the \\p value type of \\p OutputIteratorT\n__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::SingleTilePolicy::BLOCK_THREADS), 1)\n__global__ void DeviceReduceSingleTileKernel(\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n OffsetT num_items, ///< [in] Total number of input data items\n ReductionOpT reduction_op, ///< [in] Binary reduction functor\n OuputT init) ///< [in] The initial value of the reduction\n{\n // Thread block type for reducing input tiles\n typedef AgentReduce<\n typename ChainedPolicyT::ActivePolicy::SingleTilePolicy,\n InputIteratorT,\n OutputIteratorT,\n OffsetT,\n ReductionOpT>\n AgentReduceT;\n\n // Shared memory storage\n __shared__ typename AgentReduceT::TempStorage temp_storage;\n\n // Check if empty problem\n if (num_items == 0)\n {\n if (threadIdx.x == 0)\n *d_out = init;\n return;\n }\n\n // Consume input tiles\n OuputT block_aggregate = AgentReduceT(temp_storage, d_in, reduction_op).ConsumeRange(\n OffsetT(0),\n num_items);\n\n // Output result\n if (threadIdx.x == 0)\n *d_out = reduction_op(init, block_aggregate);\n}\n\n\n/// Normalize input iterator to segment offset\ntemplate \n__device__ __forceinline__\nvoid NormalizeReductionOutput(\n T &/*val*/,\n OffsetT /*base_offset*/,\n IteratorT /*itr*/)\n{}\n\n\n/// Normalize input iterator to segment offset (specialized for arg-index)\ntemplate \n__device__ __forceinline__\nvoid NormalizeReductionOutput(\n KeyValuePairT &val,\n OffsetT base_offset,\n ArgIndexInputIterator /*itr*/)\n{\n val.key -= base_offset;\n}\n\n\n/**\n * Segmented reduction (one block per segment)\n */\ntemplate <\n typename ChainedPolicyT, ///< Chained tuning policy\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OutputIteratorT, ///< Output iterator type for recording the reduced aggregate \\iterator\n typename OffsetIteratorT, ///< Random-access input iterator type for reading segment offsets \\iterator\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT, ///< Binary reduction functor type having member T operator()(const T &a, const T &b)\n typename OutputT> ///< Data element type that is convertible to the \\p value type of \\p OutputIteratorT\n__launch_bounds__ (int(ChainedPolicyT::ActivePolicy::ReducePolicy::BLOCK_THREADS))\n__global__ void DeviceSegmentedReduceKernel(\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n int /*num_segments*/, ///< [in] The number of segments that comprise the sorting data\n ReductionOpT reduction_op, ///< [in] Binary reduction functor \n OutputT init) ///< [in] The initial value of the reduction\n{\n // Thread block type for reducing input tiles\n typedef AgentReduce<\n typename ChainedPolicyT::ActivePolicy::ReducePolicy,\n InputIteratorT,\n OutputIteratorT,\n OffsetT,\n ReductionOpT>\n AgentReduceT;\n\n // Shared memory storage\n __shared__ typename AgentReduceT::TempStorage temp_storage;\n\n OffsetT segment_begin = d_begin_offsets[blockIdx.x];\n OffsetT segment_end = d_end_offsets[blockIdx.x];\n\n // Check if empty problem\n if (segment_begin == segment_end)\n {\n if (threadIdx.x == 0)\n d_out[blockIdx.x] = init;\n return;\n }\n\n // Consume input tiles\n OutputT block_aggregate = AgentReduceT(temp_storage, d_in, reduction_op).ConsumeRange(\n segment_begin,\n segment_end);\n\n // Normalize as needed\n NormalizeReductionOutput(block_aggregate, segment_begin, d_in);\n\n if (threadIdx.x == 0)\n d_out[blockIdx.x] = reduction_op(init, block_aggregate);;\n}\n\n\n\n\n/******************************************************************************\n * Policy\n ******************************************************************************/\n\ntemplate <\n typename OuputT, ///< Data type\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT> ///< Binary reduction functor type having member T operator()(const T &a, const T &b) \nstruct DeviceReducePolicy\n{\n //------------------------------------------------------------------------------\n // Architecture-specific tuning policies\n //------------------------------------------------------------------------------\n\n /// SM13\n struct Policy130 : ChainedPolicy<130, Policy130, Policy130>\n {\n // ReducePolicy\n typedef AgentReducePolicy<\n CUB_NOMINAL_CONFIG(128, 8, OuputT), ///< Threads per block, items per thread\n 2, ///< Number of items per vectorized load\n BLOCK_REDUCE_RAKING, ///< Cooperative block-wide reduction algorithm to use\n LOAD_DEFAULT> ///< Cache load modifier\n ReducePolicy;\n\n // SingleTilePolicy\n typedef ReducePolicy SingleTilePolicy;\n\n // SegmentedReducePolicy\n typedef ReducePolicy SegmentedReducePolicy;\n };\n\n\n /// SM20\n struct Policy200 : ChainedPolicy<200, Policy200, Policy130>\n {\n // ReducePolicy (GTX 580: 178.9 GB/s @ 48M 4B items, 158.1 GB/s @ 192M 1B items)\n typedef AgentReducePolicy<\n CUB_NOMINAL_CONFIG(128, 8, OuputT), ///< Threads per block, items per thread\n 4, ///< Number of items per vectorized load\n BLOCK_REDUCE_RAKING, ///< Cooperative block-wide reduction algorithm to use\n LOAD_DEFAULT> ///< Cache load modifier\n ReducePolicy;\n\n // SingleTilePolicy\n typedef ReducePolicy SingleTilePolicy;\n\n // SegmentedReducePolicy\n typedef ReducePolicy SegmentedReducePolicy;\n };\n\n\n /// SM30\n struct Policy300 : ChainedPolicy<300, Policy300, Policy200>\n {\n // ReducePolicy (GTX670: 154.0 @ 48M 4B items)\n typedef AgentReducePolicy<\n CUB_NOMINAL_CONFIG(256, 20, OuputT), ///< Threads per block, items per thread\n 2, ///< Number of items per vectorized load\n BLOCK_REDUCE_WARP_REDUCTIONS, ///< Cooperative block-wide reduction algorithm to use\n LOAD_DEFAULT> ///< Cache load modifier\n ReducePolicy;\n\n // SingleTilePolicy\n typedef ReducePolicy SingleTilePolicy;\n\n // SegmentedReducePolicy\n typedef ReducePolicy SegmentedReducePolicy;\n };\n\n\n /// SM35\n struct Policy350 : ChainedPolicy<350, Policy350, Policy300>\n {\n // ReducePolicy (GTX Titan: 255.1 GB/s @ 48M 4B items; 228.7 GB/s @ 192M 1B items)\n typedef AgentReducePolicy<\n CUB_NOMINAL_CONFIG(256, 20, OuputT), ///< Threads per block, items per thread\n 4, ///< Number of items per vectorized load\n BLOCK_REDUCE_WARP_REDUCTIONS, ///< Cooperative block-wide reduction algorithm to use\n LOAD_LDG> ///< Cache load modifier\n ReducePolicy;\n\n // SingleTilePolicy\n typedef ReducePolicy SingleTilePolicy;\n\n // SegmentedReducePolicy\n typedef ReducePolicy SegmentedReducePolicy;\n };\n\n /// SM60\n struct Policy600 : ChainedPolicy<600, Policy600, Policy350>\n {\n // ReducePolicy (P100: 591 GB/s @ 64M 4B items; 583 GB/s @ 256M 1B items)\n typedef AgentReducePolicy<\n CUB_NOMINAL_CONFIG(256, 16, OuputT), ///< Threads per block, items per thread\n 4, ///< Number of items per vectorized load\n BLOCK_REDUCE_WARP_REDUCTIONS, ///< Cooperative block-wide reduction algorithm to use\n LOAD_LDG> ///< Cache load modifier\n ReducePolicy;\n\n // SingleTilePolicy\n typedef ReducePolicy SingleTilePolicy;\n\n // SegmentedReducePolicy\n typedef ReducePolicy SegmentedReducePolicy;\n };\n\n\n /// MaxPolicy\n typedef Policy600 MaxPolicy;\n\n};\n\n\n\n/******************************************************************************\n * Single-problem dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for device-wide reduction\n */\ntemplate <\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OutputIteratorT, ///< Output iterator type for recording the reduced aggregate \\iterator\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT> ///< Binary reduction functor type having member T operator()(const T &a, const T &b) \nstruct DispatchReduce :\n DeviceReducePolicy<\n typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type, // ... else the output iterator's value type\n OffsetT,\n ReductionOpT>\n{\n //------------------------------------------------------------------------------\n // Constants\n //------------------------------------------------------------------------------\n\n // Data type of output iterator\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n\n //------------------------------------------------------------------------------\n // Problem state\n //------------------------------------------------------------------------------\n\n void *d_temp_storage; ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes; ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in; ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out; ///< [out] Pointer to the output aggregate\n OffsetT num_items; ///< [in] Total number of input items (i.e., length of \\p d_in)\n ReductionOpT reduction_op; ///< [in] Binary reduction functor \n OutputT init; ///< [in] The initial value of the reduction\n cudaStream_t stream; ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous; ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int ptx_version; ///< [in] PTX version\n\n //------------------------------------------------------------------------------\n // Constructor\n //------------------------------------------------------------------------------\n\n /// Constructor\n CUB_RUNTIME_FUNCTION __forceinline__\n DispatchReduce(\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n OffsetT num_items,\n ReductionOpT reduction_op,\n OutputT init,\n cudaStream_t stream,\n bool debug_synchronous,\n int ptx_version)\n :\n d_temp_storage(d_temp_storage),\n temp_storage_bytes(temp_storage_bytes),\n d_in(d_in),\n d_out(d_out),\n num_items(num_items),\n reduction_op(reduction_op),\n init(init),\n stream(stream),\n debug_synchronous(debug_synchronous),\n ptx_version(ptx_version)\n {}\n\n\n //------------------------------------------------------------------------------\n // Small-problem (single tile) invocation\n //------------------------------------------------------------------------------\n\n /// Invoke a single block block to reduce in-core\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename SingleTileKernelT> ///< Function type of cub::DeviceReduceSingleTileKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokeSingleTile(\n SingleTileKernelT single_tile_kernel) ///< [in] Kernel function pointer to parameterization of cub::DeviceReduceSingleTileKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void)single_tile_kernel;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n cudaError error = cudaSuccess;\n do\n {\n // Return if the caller is simply requesting the size of the storage allocation\n if (d_temp_storage == NULL)\n {\n temp_storage_bytes = 1;\n break;\n }\n\n // Log single_reduce_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking DeviceReduceSingleTileKernel<<<1, %d, 0, %lld>>>(), %d items per thread\\n\",\n ActivePolicyT::SingleTilePolicy::BLOCK_THREADS,\n (long long) stream,\n ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD);\n\n // Invoke single_reduce_sweep_kernel\n single_tile_kernel<<<1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, 0, stream>>>(\n d_in,\n d_out,\n num_items,\n reduction_op,\n init);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n //------------------------------------------------------------------------------\n // Normal problem size invocation (two-pass)\n //------------------------------------------------------------------------------\n\n /// Invoke two-passes to reduce\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename ReduceKernelT, ///< Function type of cub::DeviceReduceKernel\n typename SingleTileKernelT> ///< Function type of cub::DeviceReduceSingleTileKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePasses(\n ReduceKernelT reduce_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceReduceKernel\n SingleTileKernelT single_tile_kernel) ///< [in] Kernel function pointer to parameterization of cub::DeviceReduceSingleTileKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void) reduce_kernel;\n (void) single_tile_kernel;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Init regular kernel configuration\n KernelConfig reduce_config;\n if (CubDebug(error = reduce_config.Init(reduce_kernel))) break;\n int reduce_device_occupancy = reduce_config.sm_occupancy * sm_count;\n\n // Even-share work distribution\n int max_blocks = reduce_device_occupancy * CUB_SUBSCRIPTION_FACTOR(ptx_version);\n GridEvenShare even_share;\n even_share.DispatchInit(num_items, max_blocks, reduce_config.tile_size);\n\n // Temporary storage allocation requirements\n void* allocations[1];\n size_t allocation_sizes[1] =\n {\n max_blocks * sizeof(OutputT) // bytes needed for privatized block reductions\n };\n\n // Alias the temporary allocations from the single storage blob (or compute the necessary size of the blob)\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n return cudaSuccess;\n }\n\n // Alias the allocation for the privatized per-block reductions\n OutputT *d_block_reductions = (OutputT*) allocations[0];\n\n // Get grid size for device_reduce_sweep_kernel\n int reduce_grid_size = even_share.grid_size;\n\n // Log device_reduce_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking DeviceReduceKernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n reduce_grid_size,\n ActivePolicyT::ReducePolicy::BLOCK_THREADS,\n (long long) stream,\n ActivePolicyT::ReducePolicy::ITEMS_PER_THREAD,\n reduce_config.sm_occupancy);\n\n // Invoke DeviceReduceKernel\n reduce_kernel<<>>(\n d_in,\n d_block_reductions,\n num_items,\n even_share,\n reduction_op);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Log single_reduce_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking DeviceReduceSingleTileKernel<<<1, %d, 0, %lld>>>(), %d items per thread\\n\",\n ActivePolicyT::SingleTilePolicy::BLOCK_THREADS,\n (long long) stream,\n ActivePolicyT::SingleTilePolicy::ITEMS_PER_THREAD);\n\n // Invoke DeviceReduceSingleTileKernel\n single_tile_kernel<<<1, ActivePolicyT::SingleTilePolicy::BLOCK_THREADS, 0, stream>>>(\n d_block_reductions,\n d_out,\n reduce_grid_size,\n reduction_op,\n init);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n\n }\n\n\n //------------------------------------------------------------------------------\n // Chained policy invocation\n //------------------------------------------------------------------------------\n\n /// Invocation\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Invoke()\n {\n typedef typename ActivePolicyT::SingleTilePolicy SingleTilePolicyT;\n typedef typename DispatchReduce::MaxPolicy MaxPolicyT;\n\n // Force kernel code-generation in all compiler passes\n if (num_items <= (SingleTilePolicyT::BLOCK_THREADS * SingleTilePolicyT::ITEMS_PER_THREAD))\n {\n // Small, single tile size\n return InvokeSingleTile(\n DeviceReduceSingleTileKernel);\n }\n else\n {\n // Regular size\n return InvokePasses(\n DeviceReduceKernel,\n DeviceReduceSingleTileKernel);\n }\n }\n\n\n //------------------------------------------------------------------------------\n // Dispatch entrypoints\n //------------------------------------------------------------------------------\n\n /**\n * Internal dispatch routine for computing a device-wide reduction\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n ReductionOpT reduction_op, ///< [in] Binary reduction functor \n OutputT init, ///< [in] The initial value of the reduction\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n typedef typename DispatchReduce::MaxPolicy MaxPolicyT;\n\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n\n // Create dispatch functor\n DispatchReduce dispatch(\n d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, reduction_op, init,\n stream, debug_synchronous, ptx_version);\n\n // Dispatch to chained policy\n if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n\n\n/******************************************************************************\n * Segmented dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for device-wide reduction\n */\ntemplate <\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OutputIteratorT, ///< Output iterator type for recording the reduced aggregate \\iterator\n typename OffsetIteratorT, ///< Random-access input iterator type for reading segment offsets \\iterator\n typename OffsetT, ///< Signed integer type for global offsets\n typename ReductionOpT> ///< Binary reduction functor type having member T operator()(const T &a, const T &b) \nstruct DispatchSegmentedReduce :\n DeviceReducePolicy<\n typename std::iterator_traits::value_type,\n OffsetT,\n ReductionOpT>\n{\n //------------------------------------------------------------------------------\n // Constants\n //------------------------------------------------------------------------------\n\n /// The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n\n //------------------------------------------------------------------------------\n // Problem state\n //------------------------------------------------------------------------------\n\n void *d_temp_storage; ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes; ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in; ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out; ///< [out] Pointer to the output aggregate\n OffsetT num_segments; ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets; ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets; ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n ReductionOpT reduction_op; ///< [in] Binary reduction functor \n OutputT init; ///< [in] The initial value of the reduction\n cudaStream_t stream; ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous; ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int ptx_version; ///< [in] PTX version\n\n //------------------------------------------------------------------------------\n // Constructor\n //------------------------------------------------------------------------------\n\n /// Constructor\n CUB_RUNTIME_FUNCTION __forceinline__\n DispatchSegmentedReduce(\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n OffsetT num_segments,\n OffsetIteratorT d_begin_offsets,\n OffsetIteratorT d_end_offsets,\n ReductionOpT reduction_op,\n OutputT init,\n cudaStream_t stream,\n bool debug_synchronous,\n int ptx_version)\n :\n d_temp_storage(d_temp_storage),\n temp_storage_bytes(temp_storage_bytes),\n d_in(d_in),\n d_out(d_out),\n num_segments(num_segments),\n d_begin_offsets(d_begin_offsets),\n d_end_offsets(d_end_offsets),\n reduction_op(reduction_op),\n init(init),\n stream(stream),\n debug_synchronous(debug_synchronous),\n ptx_version(ptx_version)\n {}\n\n\n\n //------------------------------------------------------------------------------\n // Chained policy invocation\n //------------------------------------------------------------------------------\n\n /// Invocation\n template <\n typename ActivePolicyT, ///< Umbrella policy active for the target device\n typename DeviceSegmentedReduceKernelT> ///< Function type of cub::DeviceSegmentedReduceKernel\n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t InvokePasses(\n DeviceSegmentedReduceKernelT segmented_reduce_kernel) ///< [in] Kernel function pointer to parameterization of cub::DeviceSegmentedReduceKernel\n {\n#ifndef CUB_RUNTIME_ENABLED\n (void)segmented_reduce_kernel;\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n#else\n cudaError error = cudaSuccess;\n do\n {\n // Return if the caller is simply requesting the size of the storage allocation\n if (d_temp_storage == NULL)\n {\n temp_storage_bytes = 1;\n return cudaSuccess;\n }\n\n // Init kernel configuration\n KernelConfig segmented_reduce_config;\n if (CubDebug(error = segmented_reduce_config.Init(segmented_reduce_kernel))) break;\n\n // Log device_reduce_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking SegmentedDeviceReduceKernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n num_segments,\n ActivePolicyT::SegmentedReducePolicy::BLOCK_THREADS,\n (long long) stream,\n ActivePolicyT::SegmentedReducePolicy::ITEMS_PER_THREAD,\n segmented_reduce_config.sm_occupancy);\n\n // Invoke DeviceReduceKernel\n segmented_reduce_kernel<<>>(\n d_in,\n d_out,\n d_begin_offsets,\n d_end_offsets,\n num_segments,\n reduction_op,\n init);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n\n }\n\n\n /// Invocation\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Invoke()\n {\n typedef typename DispatchSegmentedReduce::MaxPolicy MaxPolicyT;\n\n // Force kernel code-generation in all compiler passes\n return InvokePasses(\n DeviceSegmentedReduceKernel);\n }\n\n\n //------------------------------------------------------------------------------\n // Dispatch entrypoints\n //------------------------------------------------------------------------------\n\n /**\n * Internal dispatch routine for computing a device-wide reduction\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output aggregate\n int num_segments, ///< [in] The number of segments that comprise the sorting data\n OffsetIteratorT d_begin_offsets, ///< [in] Pointer to the sequence of beginning offsets of length \\p num_segments, such that d_begin_offsets[i] is the first element of the ith data segment in d_keys_* and d_values_*\n OffsetIteratorT d_end_offsets, ///< [in] Pointer to the sequence of ending offsets of length \\p num_segments, such that d_end_offsets[i]-1 is the last element of the ith data segment in d_keys_* and d_values_*. If d_end_offsets[i]-1 <= d_begin_offsets[i], the ith is considered empty.\n ReductionOpT reduction_op, ///< [in] Binary reduction functor \n OutputT init, ///< [in] The initial value of the reduction\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n typedef typename DispatchSegmentedReduce::MaxPolicy MaxPolicyT;\n\n if (num_segments <= 0)\n return cudaSuccess;\n\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n\n // Create dispatch functor\n DispatchSegmentedReduce dispatch(\n d_temp_storage, temp_storage_bytes,\n d_in, d_out,\n num_segments, d_begin_offsets, d_end_offsets,\n reduction_op, init,\n stream, debug_synchronous, ptx_version);\n\n // Dispatch to chained policy\n if (CubDebug(error = MaxPolicyT::Invoke(ptx_version, dispatch))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_reduce_by_key.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceReduceByKey provides device-wide, parallel operations for reducing segments of values residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch_scan.cuh\"\n#include \"../../agent/agent_reduce_by_key.cuh\"\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Multi-block reduce-by-key sweep kernel entry point\n */\ntemplate <\n typename AgentReduceByKeyPolicyT, ///< Parameterized AgentReduceByKeyPolicyT tuning policy type\n typename KeysInputIteratorT, ///< Random-access input iterator type for keys\n typename UniqueOutputIteratorT, ///< Random-access output iterator type for keys\n typename ValuesInputIteratorT, ///< Random-access input iterator type for values\n typename AggregatesOutputIteratorT, ///< Random-access output iterator type for values\n typename NumRunsOutputIteratorT, ///< Output iterator type for recording number of segments encountered\n typename ScanTileStateT, ///< Tile status interface type\n typename EqualityOpT, ///< KeyT equality operator type\n typename ReductionOpT, ///< ValueT reduction operator type\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(AgentReduceByKeyPolicyT::BLOCK_THREADS))\n__global__ void DeviceReduceByKeyKernel(\n KeysInputIteratorT d_keys_in, ///< Pointer to the input sequence of keys\n UniqueOutputIteratorT d_unique_out, ///< Pointer to the output sequence of unique keys (one key per run)\n ValuesInputIteratorT d_values_in, ///< Pointer to the input sequence of corresponding values\n AggregatesOutputIteratorT d_aggregates_out, ///< Pointer to the output sequence of value aggregates (one aggregate per run)\n NumRunsOutputIteratorT d_num_runs_out, ///< Pointer to total number of runs encountered (i.e., the length of d_unique_out)\n ScanTileStateT tile_state, ///< Tile status interface\n int start_tile, ///< The starting tile for the current grid\n EqualityOpT equality_op, ///< KeyT equality operator\n ReductionOpT reduction_op, ///< ValueT reduction operator\n OffsetT num_items) ///< Total number of items to select from\n{\n // Thread block type for reducing tiles of value segments\n typedef AgentReduceByKey<\n AgentReduceByKeyPolicyT,\n KeysInputIteratorT,\n UniqueOutputIteratorT,\n ValuesInputIteratorT,\n AggregatesOutputIteratorT,\n NumRunsOutputIteratorT,\n EqualityOpT,\n ReductionOpT,\n OffsetT>\n AgentReduceByKeyT;\n\n // Shared memory for AgentReduceByKey\n __shared__ typename AgentReduceByKeyT::TempStorage temp_storage;\n\n // Process tiles\n AgentReduceByKeyT(temp_storage, d_keys_in, d_unique_out, d_values_in, d_aggregates_out, d_num_runs_out, equality_op, reduction_op).ConsumeRange(\n num_items,\n tile_state,\n start_tile);\n}\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceReduceByKey\n */\ntemplate <\n typename KeysInputIteratorT, ///< Random-access input iterator type for keys\n typename UniqueOutputIteratorT, ///< Random-access output iterator type for keys\n typename ValuesInputIteratorT, ///< Random-access input iterator type for values\n typename AggregatesOutputIteratorT, ///< Random-access output iterator type for values\n typename NumRunsOutputIteratorT, ///< Output iterator type for recording number of segments encountered\n typename EqualityOpT, ///< KeyT equality operator type\n typename ReductionOpT, ///< ValueT reduction operator type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DispatchReduceByKey\n{\n //-------------------------------------------------------------------------\n // Types and constants\n //-------------------------------------------------------------------------\n\n // The input keys type\n typedef typename std::iterator_traits::value_type KeyInputT;\n\n // The output keys type\n typedef typename If<(Equals::value_type, void>::VALUE), // KeyOutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type KeyOutputT; // ... else the output iterator's value type\n\n // The input values type\n typedef typename std::iterator_traits::value_type ValueInputT;\n\n // The output values type\n typedef typename If<(Equals::value_type, void>::VALUE), // ValueOutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type ValueOutputT; // ... else the output iterator's value type\n\n enum\n {\n INIT_KERNEL_THREADS = 128,\n MAX_INPUT_BYTES = CUB_MAX(sizeof(KeyOutputT), sizeof(ValueOutputT)),\n COMBINED_INPUT_BYTES = sizeof(KeyOutputT) + sizeof(ValueOutputT),\n };\n\n // Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n\n //-------------------------------------------------------------------------\n // Tuning policies\n //-------------------------------------------------------------------------\n\n /// SM35\n struct Policy350\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 6,\n ITEMS_PER_THREAD = (MAX_INPUT_BYTES <= 8) ? 6 : CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, ((NOMINAL_4B_ITEMS_PER_THREAD * 8) + COMBINED_INPUT_BYTES - 1) / COMBINED_INPUT_BYTES)),\n };\n\n typedef AgentReduceByKeyPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n BLOCK_SCAN_WARP_SCANS>\n ReduceByKeyPolicyT;\n };\n\n /// SM30\n struct Policy300\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 6,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, ((NOMINAL_4B_ITEMS_PER_THREAD * 8) + COMBINED_INPUT_BYTES - 1) / COMBINED_INPUT_BYTES)),\n };\n\n typedef AgentReduceByKeyPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n ReduceByKeyPolicyT;\n };\n\n /// SM20\n struct Policy200\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 11,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, ((NOMINAL_4B_ITEMS_PER_THREAD * 8) + COMBINED_INPUT_BYTES - 1) / COMBINED_INPUT_BYTES)),\n };\n\n typedef AgentReduceByKeyPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n ReduceByKeyPolicyT;\n };\n\n /// SM13\n struct Policy130\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 7,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, ((NOMINAL_4B_ITEMS_PER_THREAD * 8) + COMBINED_INPUT_BYTES - 1) / COMBINED_INPUT_BYTES)),\n };\n\n typedef AgentReduceByKeyPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n ReduceByKeyPolicyT;\n };\n\n /// SM11\n struct Policy110\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 5,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 8) / COMBINED_INPUT_BYTES)),\n };\n\n typedef AgentReduceByKeyPolicy<\n 64,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_RAKING>\n ReduceByKeyPolicyT;\n };\n\n\n /******************************************************************************\n * Tuning policies of current PTX compiler pass\n ******************************************************************************/\n\n#if (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 130)\n typedef Policy130 PtxPolicy;\n\n#else\n typedef Policy110 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxReduceByKeyPolicy : PtxPolicy::ReduceByKeyPolicyT {};\n\n\n /******************************************************************************\n * Utilities\n ******************************************************************************/\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static void InitConfigs(\n int ptx_version,\n KernelConfig &reduce_by_key_config)\n {\n #if (CUB_PTX_ARCH > 0)\n (void)ptx_version;\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n reduce_by_key_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 350)\n {\n reduce_by_key_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n reduce_by_key_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n reduce_by_key_config.template Init();\n }\n else if (ptx_version >= 130)\n {\n reduce_by_key_config.template Init();\n }\n else\n {\n reduce_by_key_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration.\n */\n struct KernelConfig\n {\n int block_threads;\n int items_per_thread;\n int tile_items;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n void Init()\n {\n block_threads = PolicyT::BLOCK_THREADS;\n items_per_thread = PolicyT::ITEMS_PER_THREAD;\n tile_items = block_threads * items_per_thread;\n }\n };\n\n\n //---------------------------------------------------------------------\n // Dispatch entrypoints\n //---------------------------------------------------------------------\n\n /**\n * Internal dispatch routine for computing a device-wide reduce-by-key using the\n * specified kernel functions.\n */\n template <\n typename ScanInitKernelT, ///< Function type of cub::DeviceScanInitKernel\n typename ReduceByKeyKernelT> ///< Function type of cub::DeviceReduceByKeyKernelT\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n KeysInputIteratorT d_keys_in, ///< [in] Pointer to the input sequence of keys\n UniqueOutputIteratorT d_unique_out, ///< [out] Pointer to the output sequence of unique keys (one key per run)\n ValuesInputIteratorT d_values_in, ///< [in] Pointer to the input sequence of corresponding values\n AggregatesOutputIteratorT d_aggregates_out, ///< [out] Pointer to the output sequence of value aggregates (one aggregate per run)\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs encountered (i.e., the length of d_unique_out)\n EqualityOpT equality_op, ///< [in] KeyT equality operator\n ReductionOpT reduction_op, ///< [in] ValueT reduction operator\n OffsetT num_items, ///< [in] Total number of items to select from\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int /*ptx_version*/, ///< [in] PTX version of dispatch kernels\n ScanInitKernelT init_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceScanInitKernel\n ReduceByKeyKernelT reduce_by_key_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceReduceByKeyKernel\n KernelConfig reduce_by_key_config) ///< [in] Dispatch parameters that match the policy that \\p reduce_by_key_kernel was compiled for\n {\n\n#ifndef CUB_RUNTIME_ENABLED\n (void)d_temp_storage;\n (void)temp_storage_bytes;\n (void)d_keys_in;\n (void)d_unique_out;\n (void)d_values_in;\n (void)d_aggregates_out;\n (void)d_num_runs_out;\n (void)equality_op;\n (void)reduction_op;\n (void)num_items;\n (void)stream;\n (void)debug_synchronous;\n (void)init_kernel;\n (void)reduce_by_key_kernel;\n (void)reduce_by_key_config;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported);\n\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Number of input tiles\n int tile_size = reduce_by_key_config.block_threads * reduce_by_key_config.items_per_thread;\n int num_tiles = (num_items + tile_size - 1) / tile_size;\n\n // Specify temporary storage allocation requirements\n size_t allocation_sizes[1];\n if (CubDebug(error = ScanTileStateT::AllocationSize(num_tiles, allocation_sizes[0]))) break; // bytes needed for tile status descriptors\n\n // Compute allocation pointers into the single storage blob (or compute the necessary size of the blob)\n void* allocations[1];\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Construct the tile status interface\n ScanTileStateT tile_state;\n if (CubDebug(error = tile_state.Init(num_tiles, allocations[0], allocation_sizes[0]))) break;\n\n // Log init_kernel configuration\n int init_grid_size = CUB_MAX(1, (num_tiles + INIT_KERNEL_THREADS - 1) / INIT_KERNEL_THREADS);\n if (debug_synchronous) _CubLog(\"Invoking init_kernel<<<%d, %d, 0, %lld>>>()\\n\", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);\n\n // Invoke init_kernel to initialize tile descriptors\n init_kernel<<>>(\n tile_state,\n num_tiles,\n d_num_runs_out);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Return if empty problem\n if (num_items == 0)\n break;\n\n // Get SM occupancy for reduce_by_key_kernel\n int reduce_by_key_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n reduce_by_key_sm_occupancy, // out\n reduce_by_key_kernel,\n reduce_by_key_config.block_threads))) break;\n\n // Get max x-dimension of grid\n int max_dim_x;\n if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;\n\n // Run grids in epochs (in case number of tiles exceeds max x-dimension\n int scan_grid_size = CUB_MIN(num_tiles, max_dim_x);\n for (int start_tile = 0; start_tile < num_tiles; start_tile += scan_grid_size)\n {\n // Log reduce_by_key_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking %d reduce_by_key_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n start_tile, scan_grid_size, reduce_by_key_config.block_threads, (long long) stream, reduce_by_key_config.items_per_thread, reduce_by_key_sm_occupancy);\n\n // Invoke reduce_by_key_kernel\n reduce_by_key_kernel<<>>(\n d_keys_in,\n d_unique_out,\n d_values_in,\n d_aggregates_out,\n d_num_runs_out,\n tile_state,\n start_tile,\n equality_op,\n reduction_op,\n num_items);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n KeysInputIteratorT d_keys_in, ///< [in] Pointer to the input sequence of keys\n UniqueOutputIteratorT d_unique_out, ///< [out] Pointer to the output sequence of unique keys (one key per run)\n ValuesInputIteratorT d_values_in, ///< [in] Pointer to the input sequence of corresponding values\n AggregatesOutputIteratorT d_aggregates_out, ///< [out] Pointer to the output sequence of value aggregates (one aggregate per run)\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs encountered (i.e., the length of d_unique_out)\n EqualityOpT equality_op, ///< [in] KeyT equality operator\n ReductionOpT reduction_op, ///< [in] ValueT reduction operator\n OffsetT num_items, ///< [in] Total number of items to select from\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel kernel dispatch configurations\n KernelConfig reduce_by_key_config;\n InitConfigs(ptx_version, reduce_by_key_config);\n\n // Dispatch\n if (CubDebug(error = Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_keys_in,\n d_unique_out,\n d_values_in,\n d_aggregates_out,\n d_num_runs_out,\n equality_op,\n reduction_op,\n num_items,\n stream,\n debug_synchronous,\n ptx_version,\n DeviceCompactInitKernel,\n DeviceReduceByKeyKernel,\n reduce_by_key_config))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_rle.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceRle provides device-wide, parallel operations for run-length-encoding sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch_scan.cuh\"\n#include \"../../agent/agent_rle.cuh\"\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Select kernel entry point (multi-block)\n *\n * Performs functor-based selection if SelectOp functor type != NullType\n * Otherwise performs flag-based selection if FlagIterator's value type != NullType\n * Otherwise performs discontinuity selection (keep unique)\n */\ntemplate <\n typename AgentRlePolicyT, ///< Parameterized AgentRlePolicyT tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OffsetsOutputIteratorT, ///< Random-access output iterator type for writing run-offset values \\iterator\n typename LengthsOutputIteratorT, ///< Random-access output iterator type for writing run-length values \\iterator\n typename NumRunsOutputIteratorT, ///< Output iterator type for recording the number of runs encountered \\iterator\n typename ScanTileStateT, ///< Tile status interface type\n typename EqualityOpT, ///< T equality operator type\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(AgentRlePolicyT::BLOCK_THREADS))\n__global__ void DeviceRleSweepKernel(\n InputIteratorT d_in, ///< [in] Pointer to input sequence of data items\n OffsetsOutputIteratorT d_offsets_out, ///< [out] Pointer to output sequence of run-offsets\n LengthsOutputIteratorT d_lengths_out, ///< [out] Pointer to output sequence of run-lengths\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs (i.e., length of \\p d_offsets_out)\n ScanTileStateT tile_status, ///< [in] Tile status interface\n EqualityOpT equality_op, ///< [in] Equality operator for input items\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n int num_tiles) ///< [in] Total number of tiles for the entire problem\n{\n // Thread block type for selecting data from input tiles\n typedef AgentRle<\n AgentRlePolicyT,\n InputIteratorT,\n OffsetsOutputIteratorT,\n LengthsOutputIteratorT,\n EqualityOpT,\n OffsetT> AgentRleT;\n\n // Shared memory for AgentRle\n __shared__ typename AgentRleT::TempStorage temp_storage;\n\n // Process tiles\n AgentRleT(temp_storage, d_in, d_offsets_out, d_lengths_out, equality_op, num_items).ConsumeRange(\n num_tiles,\n tile_status,\n d_num_runs_out);\n}\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceRle\n */\ntemplate <\n typename InputIteratorT, ///< Random-access input iterator type for reading input items \\iterator\n typename OffsetsOutputIteratorT, ///< Random-access output iterator type for writing run-offset values \\iterator\n typename LengthsOutputIteratorT, ///< Random-access output iterator type for writing run-length values \\iterator\n typename NumRunsOutputIteratorT, ///< Output iterator type for recording the number of runs encountered \\iterator\n typename EqualityOpT, ///< T equality operator type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DeviceRleDispatch\n{\n /******************************************************************************\n * Types and constants\n ******************************************************************************/\n\n // The input value type\n typedef typename std::iterator_traits::value_type T;\n\n // The lengths output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // LengthT = (if output iterator's value type is void) ?\n OffsetT, // ... then the OffsetT type,\n typename std::iterator_traits::value_type>::Type LengthT; // ... else the output iterator's value type\n\n enum\n {\n INIT_KERNEL_THREADS = 128,\n };\n\n // Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n\n /******************************************************************************\n * Tuning policies\n ******************************************************************************/\n\n /// SM35\n struct Policy350\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 15,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(T)))),\n };\n\n typedef AgentRlePolicy<\n 96,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n true,\n BLOCK_SCAN_WARP_SCANS>\n RleSweepPolicy;\n };\n\n /// SM30\n struct Policy300\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 5,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(T)))),\n };\n\n typedef AgentRlePolicy<\n 256,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n true,\n BLOCK_SCAN_RAKING_MEMOIZE>\n RleSweepPolicy;\n };\n\n /// SM20\n struct Policy200\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 15,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(T)))),\n };\n\n typedef AgentRlePolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n false,\n BLOCK_SCAN_WARP_SCANS>\n RleSweepPolicy;\n };\n\n /// SM13\n struct Policy130\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 9,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(T)))),\n };\n\n typedef AgentRlePolicy<\n 64,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n true,\n BLOCK_SCAN_RAKING_MEMOIZE>\n RleSweepPolicy;\n };\n\n /// SM10\n struct Policy100\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 9,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(T)))),\n };\n\n typedef AgentRlePolicy<\n 256,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n true,\n BLOCK_SCAN_RAKING_MEMOIZE>\n RleSweepPolicy;\n };\n\n\n /******************************************************************************\n * Tuning policies of current PTX compiler pass\n ******************************************************************************/\n\n#if (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 130)\n typedef Policy130 PtxPolicy;\n\n#else\n typedef Policy100 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxRleSweepPolicy : PtxPolicy::RleSweepPolicy {};\n\n\n /******************************************************************************\n * Utilities\n ******************************************************************************/\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static void InitConfigs(\n int ptx_version,\n KernelConfig& device_rle_config)\n {\n #if (CUB_PTX_ARCH > 0)\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n device_rle_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 350)\n {\n device_rle_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n device_rle_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n device_rle_config.template Init();\n }\n else if (ptx_version >= 130)\n {\n device_rle_config.template Init();\n }\n else\n {\n device_rle_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration. Mirrors the constants within AgentRlePolicyT.\n */\n struct KernelConfig\n {\n int block_threads;\n int items_per_thread;\n BlockLoadAlgorithm load_policy;\n bool store_warp_time_slicing;\n BlockScanAlgorithm scan_algorithm;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n void Init()\n {\n block_threads = AgentRlePolicyT::BLOCK_THREADS;\n items_per_thread = AgentRlePolicyT::ITEMS_PER_THREAD;\n load_policy = AgentRlePolicyT::LOAD_ALGORITHM;\n store_warp_time_slicing = AgentRlePolicyT::STORE_WARP_TIME_SLICING;\n scan_algorithm = AgentRlePolicyT::SCAN_ALGORITHM;\n }\n\n CUB_RUNTIME_FUNCTION __forceinline__\n void Print()\n {\n printf(\"%d, %d, %d, %d, %d\",\n block_threads,\n items_per_thread,\n load_policy,\n store_warp_time_slicing,\n scan_algorithm);\n }\n };\n\n\n /******************************************************************************\n * Dispatch entrypoints\n ******************************************************************************/\n\n /**\n * Internal dispatch routine for computing a device-wide run-length-encode using the\n * specified kernel functions.\n */\n template <\n typename DeviceScanInitKernelPtr, ///< Function type of cub::DeviceScanInitKernel\n typename DeviceRleSweepKernelPtr> ///< Function type of cub::DeviceRleSweepKernelPtr\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OffsetsOutputIteratorT d_offsets_out, ///< [out] Pointer to the output sequence of run-offsets\n LengthsOutputIteratorT d_lengths_out, ///< [out] Pointer to the output sequence of run-lengths\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to the total number of runs encountered (i.e., length of \\p d_offsets_out)\n EqualityOpT equality_op, ///< [in] Equality operator for input items\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int ptx_version, ///< [in] PTX version of dispatch kernels\n DeviceScanInitKernelPtr device_scan_init_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceScanInitKernel\n DeviceRleSweepKernelPtr device_rle_sweep_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceRleSweepKernel\n KernelConfig device_rle_config) ///< [in] Dispatch parameters that match the policy that \\p device_rle_sweep_kernel was compiled for\n {\n\n#ifndef CUB_RUNTIME_ENABLED\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported);\n\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Number of input tiles\n int tile_size = device_rle_config.block_threads * device_rle_config.items_per_thread;\n int num_tiles = (num_items + tile_size - 1) / tile_size;\n\n // Specify temporary storage allocation requirements\n size_t allocation_sizes[1];\n if (CubDebug(error = ScanTileStateT::AllocationSize(num_tiles, allocation_sizes[0]))) break; // bytes needed for tile status descriptors\n\n // Compute allocation pointers into the single storage blob (or compute the necessary size of the blob)\n void* allocations[1];\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Construct the tile status interface\n ScanTileStateT tile_status;\n if (CubDebug(error = tile_status.Init(num_tiles, allocations[0], allocation_sizes[0]))) break;\n\n // Log device_scan_init_kernel configuration\n int init_grid_size = CUB_MAX(1, (num_tiles + INIT_KERNEL_THREADS - 1) / INIT_KERNEL_THREADS);\n if (debug_synchronous) _CubLog(\"Invoking device_scan_init_kernel<<<%d, %d, 0, %lld>>>()\\n\", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);\n\n // Invoke device_scan_init_kernel to initialize tile descriptors and queue descriptors\n device_scan_init_kernel<<>>(\n tile_status,\n num_tiles,\n d_num_runs_out);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Return if empty problem\n if (num_items == 0)\n break;\n\n // Get SM occupancy for device_rle_sweep_kernel\n int device_rle_kernel_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n device_rle_kernel_sm_occupancy, // out\n device_rle_sweep_kernel,\n device_rle_config.block_threads))) break;\n\n // Get max x-dimension of grid\n int max_dim_x;\n if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;\n\n // Get grid size for scanning tiles\n dim3 scan_grid_size;\n scan_grid_size.z = 1;\n scan_grid_size.y = ((unsigned int) num_tiles + max_dim_x - 1) / max_dim_x;\n scan_grid_size.x = CUB_MIN(num_tiles, max_dim_x);\n\n // Log device_rle_sweep_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking device_rle_sweep_kernel<<<{%d,%d,%d}, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n scan_grid_size.x, scan_grid_size.y, scan_grid_size.z, device_rle_config.block_threads, (long long) stream, device_rle_config.items_per_thread, device_rle_kernel_sm_occupancy);\n\n // Invoke device_rle_sweep_kernel\n device_rle_sweep_kernel<<>>(\n d_in,\n d_offsets_out,\n d_lengths_out,\n d_num_runs_out,\n tile_status,\n equality_op,\n num_items,\n num_tiles);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to input sequence of data items\n OffsetsOutputIteratorT d_offsets_out, ///< [out] Pointer to output sequence of run-offsets\n LengthsOutputIteratorT d_lengths_out, ///< [out] Pointer to output sequence of run-lengths\n NumRunsOutputIteratorT d_num_runs_out, ///< [out] Pointer to total number of runs (i.e., length of \\p d_offsets_out)\n EqualityOpT equality_op, ///< [in] Equality operator for input items\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel kernel dispatch configurations\n KernelConfig device_rle_config;\n InitConfigs(ptx_version, device_rle_config);\n\n // Dispatch\n if (CubDebug(error = Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_offsets_out,\n d_lengths_out,\n d_num_runs_out,\n equality_op,\n num_items,\n stream,\n debug_synchronous,\n ptx_version,\n DeviceCompactInitKernel,\n DeviceRleSweepKernel,\n device_rle_config))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_scan.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceScan provides device-wide, parallel operations for computing a prefix scan across a sequence of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../../agent/agent_scan.cuh\"\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_arch.cuh\"\n#include \"../../util_debug.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Initialization kernel for tile status initialization (multi-block)\n */\ntemplate <\n typename ScanTileStateT> ///< Tile status interface type\n__global__ void DeviceScanInitKernel(\n ScanTileStateT tile_state, ///< [in] Tile status interface\n int num_tiles) ///< [in] Number of tiles\n{\n // Initialize tile status\n tile_state.InitializeStatus(num_tiles);\n}\n\n/**\n * Initialization kernel for tile status initialization (multi-block)\n */\ntemplate <\n typename ScanTileStateT, ///< Tile status interface type\n typename NumSelectedIteratorT> ///< Output iterator type for recording the number of items selected\n__global__ void DeviceCompactInitKernel(\n ScanTileStateT tile_state, ///< [in] Tile status interface\n int num_tiles, ///< [in] Number of tiles\n NumSelectedIteratorT d_num_selected_out) ///< [out] Pointer to the total number of items selected (i.e., length of \\p d_selected_out)\n{\n // Initialize tile status\n tile_state.InitializeStatus(num_tiles);\n\n // Initialize d_num_selected_out\n if ((blockIdx.x == 0) && (threadIdx.x == 0))\n *d_num_selected_out = 0;\n}\n\n\n/**\n * Scan kernel entry point (multi-block)\n */\ntemplate <\n typename ScanPolicyT, ///< Parameterized ScanPolicyT tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type for reading scan inputs \\iterator\n typename OutputIteratorT, ///< Random-access output iterator type for writing scan outputs \\iterator\n typename ScanTileStateT, ///< Tile status interface type\n typename ScanOpT, ///< Binary scan functor type having member T operator()(const T &a, const T &b)\n typename InitValueT, ///< Initial value to seed the exclusive scan (cub::NullType for inclusive scans)\n typename OffsetT> ///< Signed integer type for global offsets\n__launch_bounds__ (int(ScanPolicyT::BLOCK_THREADS))\n__global__ void DeviceScanKernel(\n InputIteratorT d_in, ///< Input data\n OutputIteratorT d_out, ///< Output data\n ScanTileStateT tile_state, ///< Tile status interface\n int start_tile, ///< The starting tile for the current grid\n ScanOpT scan_op, ///< Binary scan functor \n InitValueT init_value, ///< Initial value to seed the exclusive scan\n OffsetT num_items) ///< Total number of scan items for the entire problem\n{\n // Thread block type for scanning input tiles\n typedef AgentScan<\n ScanPolicyT,\n InputIteratorT,\n OutputIteratorT,\n ScanOpT,\n InitValueT,\n OffsetT> AgentScanT;\n\n // Shared memory for AgentScan\n __shared__ typename AgentScanT::TempStorage temp_storage;\n\n // Process tiles\n AgentScanT(temp_storage, d_in, d_out, scan_op, init_value).ConsumeRange(\n num_items,\n tile_state,\n start_tile);\n}\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceScan\n */\ntemplate <\n typename InputIteratorT, ///< Random-access input iterator type for reading scan inputs \\iterator\n typename OutputIteratorT, ///< Random-access output iterator type for writing scan outputs \\iterator\n typename ScanOpT, ///< Binary scan functor type having member T operator()(const T &a, const T &b)\n typename InitValueT, ///< The init_value element type for ScanOpT (cub::NullType for inclusive scans)\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DispatchScan\n{\n //---------------------------------------------------------------------\n // Constants and Types\n //---------------------------------------------------------------------\n\n enum\n {\n INIT_KERNEL_THREADS = 128\n };\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Tile status descriptor interface type\n typedef ScanTileState ScanTileStateT;\n\n\n //---------------------------------------------------------------------\n // Tuning policies\n //---------------------------------------------------------------------\n\n /// SM600\n struct Policy600\n {\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(128, 15, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_STORE_TRANSPOSE,\n BLOCK_SCAN_WARP_SCANS>\n ScanPolicyT;\n };\n\n\n /// SM520\n struct Policy520\n {\n // Titan X: 32.47B items/s @ 48M 32-bit T\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(128, 12, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n BLOCK_STORE_WARP_TRANSPOSE,\n BLOCK_SCAN_WARP_SCANS>\n ScanPolicyT;\n };\n\n\n /// SM35\n struct Policy350\n {\n // GTX Titan: 29.5B items/s (232.4 GB/s) @ 48M 32-bit T\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(128, 12, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n BLOCK_STORE_WARP_TRANSPOSE_TIMESLICED,\n BLOCK_SCAN_RAKING>\n ScanPolicyT;\n };\n\n /// SM30\n struct Policy300\n {\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(256, 9, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_STORE_WARP_TRANSPOSE,\n BLOCK_SCAN_WARP_SCANS>\n ScanPolicyT;\n };\n\n /// SM20\n struct Policy200\n {\n // GTX 580: 20.3B items/s (162.3 GB/s) @ 48M 32-bit T\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(128, 12, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_STORE_WARP_TRANSPOSE,\n BLOCK_SCAN_WARP_SCANS>\n ScanPolicyT;\n };\n\n /// SM13\n struct Policy130\n {\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(96, 21, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_STORE_WARP_TRANSPOSE,\n BLOCK_SCAN_RAKING_MEMOIZE>\n ScanPolicyT;\n };\n\n /// SM10\n struct Policy100\n {\n typedef AgentScanPolicy<\n CUB_NOMINAL_CONFIG(64, 9, OutputT), ///< Threads per block, items per thread\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_STORE_WARP_TRANSPOSE,\n BLOCK_SCAN_WARP_SCANS>\n ScanPolicyT;\n };\n\n\n //---------------------------------------------------------------------\n // Tuning policies of current PTX compiler pass\n //---------------------------------------------------------------------\n\n#if (CUB_PTX_ARCH >= 600)\n typedef Policy600 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 520)\n typedef Policy520 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 130)\n typedef Policy130 PtxPolicy;\n\n#else\n typedef Policy100 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxAgentScanPolicy : PtxPolicy::ScanPolicyT {};\n\n\n //---------------------------------------------------------------------\n // Utilities\n //---------------------------------------------------------------------\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static void InitConfigs(\n int ptx_version,\n KernelConfig &scan_kernel_config)\n {\n #if (CUB_PTX_ARCH > 0)\n (void)ptx_version;\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n scan_kernel_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 600)\n {\n scan_kernel_config.template Init();\n }\n else if (ptx_version >= 520)\n {\n scan_kernel_config.template Init();\n }\n else if (ptx_version >= 350)\n {\n scan_kernel_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n scan_kernel_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n scan_kernel_config.template Init();\n }\n else if (ptx_version >= 130)\n {\n scan_kernel_config.template Init();\n }\n else\n {\n scan_kernel_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration.\n */\n struct KernelConfig\n {\n int block_threads;\n int items_per_thread;\n int tile_items;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n void Init()\n {\n block_threads = PolicyT::BLOCK_THREADS;\n items_per_thread = PolicyT::ITEMS_PER_THREAD;\n tile_items = block_threads * items_per_thread;\n }\n };\n\n\n //---------------------------------------------------------------------\n // Dispatch entrypoints\n //---------------------------------------------------------------------\n\n /**\n * Internal dispatch routine for computing a device-wide prefix scan using the\n * specified kernel functions.\n */\n template <\n typename ScanInitKernelPtrT, ///< Function type of cub::DeviceScanInitKernel\n typename ScanSweepKernelPtrT> ///< Function type of cub::DeviceScanKernelPtrT\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n ScanOpT scan_op, ///< [in] Binary scan functor \n InitValueT init_value, ///< [in] Initial value to seed the exclusive scan\n OffsetT num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int /*ptx_version*/, ///< [in] PTX version of dispatch kernels\n ScanInitKernelPtrT init_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceScanInitKernel\n ScanSweepKernelPtrT scan_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceScanKernel\n KernelConfig scan_kernel_config) ///< [in] Dispatch parameters that match the policy that \\p scan_kernel was compiled for\n {\n\n#ifndef CUB_RUNTIME_ENABLED\n (void)d_temp_storage;\n (void)temp_storage_bytes;\n (void)d_in;\n (void)d_out;\n (void)scan_op;\n (void)init_value;\n (void)num_items;\n (void)stream;\n (void)debug_synchronous;\n (void)init_kernel;\n (void)scan_kernel;\n (void)scan_kernel_config;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported);\n\n#else\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Number of input tiles\n int tile_size = scan_kernel_config.block_threads * scan_kernel_config.items_per_thread;\n int num_tiles = (num_items + tile_size - 1) / tile_size;\n\n // Specify temporary storage allocation requirements\n size_t allocation_sizes[1];\n if (CubDebug(error = ScanTileStateT::AllocationSize(num_tiles, allocation_sizes[0]))) break; // bytes needed for tile status descriptors\n\n // Compute allocation pointers into the single storage blob (or compute the necessary size of the blob)\n void* allocations[1];\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Return if empty problem\n if (num_items == 0)\n break;\n\n // Construct the tile status interface\n ScanTileStateT tile_state;\n if (CubDebug(error = tile_state.Init(num_tiles, allocations[0], allocation_sizes[0]))) break;\n\n // Log init_kernel configuration\n int init_grid_size = (num_tiles + INIT_KERNEL_THREADS - 1) / INIT_KERNEL_THREADS;\n if (debug_synchronous) _CubLog(\"Invoking init_kernel<<<%d, %d, 0, %lld>>>()\\n\", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);\n\n // Invoke init_kernel to initialize tile descriptors\n init_kernel<<>>(\n tile_state,\n num_tiles);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Get SM occupancy for scan_kernel\n int scan_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n scan_sm_occupancy, // out\n scan_kernel,\n scan_kernel_config.block_threads))) break;\n\n // Get max x-dimension of grid\n int max_dim_x;\n if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;\n\n // Run grids in epochs (in case number of tiles exceeds max x-dimension\n int scan_grid_size = CUB_MIN(num_tiles, max_dim_x);\n for (int start_tile = 0; start_tile < num_tiles; start_tile += scan_grid_size)\n {\n // Log scan_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking %d scan_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n start_tile, scan_grid_size, scan_kernel_config.block_threads, (long long) stream, scan_kernel_config.items_per_thread, scan_sm_occupancy);\n\n // Invoke scan_kernel\n scan_kernel<<>>(\n d_in,\n d_out,\n tile_state,\n start_tile,\n scan_op,\n init_value,\n num_items);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n OutputIteratorT d_out, ///< [out] Pointer to the output sequence of data items\n ScanOpT scan_op, ///< [in] Binary scan functor \n InitValueT init_value, ///< [in] Initial value to seed the exclusive scan\n OffsetT num_items, ///< [in] Total number of input items (i.e., the length of \\p d_in)\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n\n // Get kernel kernel dispatch configurations\n KernelConfig scan_kernel_config;\n InitConfigs(ptx_version, scan_kernel_config);\n\n // Dispatch\n if (CubDebug(error = Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n init_value,\n num_items,\n stream,\n debug_synchronous,\n ptx_version,\n DeviceScanInitKernel,\n DeviceScanKernel,\n scan_kernel_config))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_select_if.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSelect provides device-wide, parallel operations for selecting items from sequences of data items residing within device-accessible memory.\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"dispatch_scan.cuh\"\n#include \"../../agent/agent_select_if.cuh\"\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/******************************************************************************\n * Kernel entry points\n *****************************************************************************/\n\n/**\n * Select kernel entry point (multi-block)\n *\n * Performs functor-based selection if SelectOpT functor type != NullType\n * Otherwise performs flag-based selection if FlagsInputIterator's value type != NullType\n * Otherwise performs discontinuity selection (keep unique)\n */\ntemplate <\n typename AgentSelectIfPolicyT, ///< Parameterized AgentSelectIfPolicyT tuning policy type\n typename InputIteratorT, ///< Random-access input iterator type for reading input items\n typename FlagsInputIteratorT, ///< Random-access input iterator type for reading selection flags (NullType* if a selection functor or discontinuity flagging is to be used for selection)\n typename SelectedOutputIteratorT, ///< Random-access output iterator type for writing selected items\n typename NumSelectedIteratorT, ///< Output iterator type for recording the number of items selected\n typename ScanTileStateT, ///< Tile status interface type\n typename SelectOpT, ///< Selection operator type (NullType if selection flags or discontinuity flagging is to be used for selection)\n typename EqualityOpT, ///< Equality operator type (NullType if selection functor or selection flags is to be used for selection)\n typename OffsetT, ///< Signed integer type for global offsets\n bool KEEP_REJECTS> ///< Whether or not we push rejected items to the back of the output\n__launch_bounds__ (int(AgentSelectIfPolicyT::BLOCK_THREADS))\n__global__ void DeviceSelectSweepKernel(\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n FlagsInputIteratorT d_flags, ///< [in] Pointer to the input sequence of selection flags (if applicable)\n SelectedOutputIteratorT d_selected_out, ///< [out] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [out] Pointer to the total number of items selected (i.e., length of \\p d_selected_out)\n ScanTileStateT tile_status, ///< [in] Tile status interface\n SelectOpT select_op, ///< [in] Selection operator\n EqualityOpT equality_op, ///< [in] Equality operator\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n int num_tiles) ///< [in] Total number of tiles for the entire problem\n{\n // Thread block type for selecting data from input tiles\n typedef AgentSelectIf<\n AgentSelectIfPolicyT,\n InputIteratorT,\n FlagsInputIteratorT,\n SelectedOutputIteratorT,\n SelectOpT,\n EqualityOpT,\n OffsetT,\n KEEP_REJECTS> AgentSelectIfT;\n\n // Shared memory for AgentSelectIf\n __shared__ typename AgentSelectIfT::TempStorage temp_storage;\n\n // Process tiles\n AgentSelectIfT(temp_storage, d_in, d_flags, d_selected_out, select_op, equality_op, num_items).ConsumeRange(\n num_tiles,\n tile_status,\n d_num_selected_out);\n}\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceSelect\n */\ntemplate <\n typename InputIteratorT, ///< Random-access input iterator type for reading input items\n typename FlagsInputIteratorT, ///< Random-access input iterator type for reading selection flags (NullType* if a selection functor or discontinuity flagging is to be used for selection)\n typename SelectedOutputIteratorT, ///< Random-access output iterator type for writing selected items\n typename NumSelectedIteratorT, ///< Output iterator type for recording the number of items selected\n typename SelectOpT, ///< Selection operator type (NullType if selection flags or discontinuity flagging is to be used for selection)\n typename EqualityOpT, ///< Equality operator type (NullType if selection functor or selection flags is to be used for selection)\n typename OffsetT, ///< Signed integer type for global offsets\n bool KEEP_REJECTS> ///< Whether or not we push rejected items to the back of the output\nstruct DispatchSelectIf\n{\n /******************************************************************************\n * Types and constants\n ******************************************************************************/\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // The flag value type\n typedef typename std::iterator_traits::value_type FlagT;\n\n enum\n {\n INIT_KERNEL_THREADS = 128,\n };\n\n // Tile status descriptor interface type\n typedef ScanTileState ScanTileStateT;\n\n\n /******************************************************************************\n * Tuning policies\n ******************************************************************************/\n\n /// SM35\n struct Policy350\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 10,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(OutputT)))),\n };\n\n typedef AgentSelectIfPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_DIRECT,\n LOAD_LDG,\n BLOCK_SCAN_WARP_SCANS>\n SelectIfPolicyT;\n };\n\n /// SM30\n struct Policy300\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 7,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(3, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(OutputT)))),\n };\n\n typedef AgentSelectIfPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n SelectIfPolicyT;\n };\n\n /// SM20\n struct Policy200\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = (KEEP_REJECTS) ? 7 : 15,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(OutputT)))),\n };\n\n typedef AgentSelectIfPolicy<\n 128,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n SelectIfPolicyT;\n };\n\n /// SM13\n struct Policy130\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 9,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(OutputT)))),\n };\n\n typedef AgentSelectIfPolicy<\n 64,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_RAKING_MEMOIZE>\n SelectIfPolicyT;\n };\n\n /// SM10\n struct Policy100\n {\n enum {\n NOMINAL_4B_ITEMS_PER_THREAD = 9,\n ITEMS_PER_THREAD = CUB_MIN(NOMINAL_4B_ITEMS_PER_THREAD, CUB_MAX(1, (NOMINAL_4B_ITEMS_PER_THREAD * 4 / sizeof(OutputT)))),\n };\n\n typedef AgentSelectIfPolicy<\n 64,\n ITEMS_PER_THREAD,\n BLOCK_LOAD_WARP_TRANSPOSE,\n LOAD_DEFAULT,\n BLOCK_SCAN_RAKING>\n SelectIfPolicyT;\n };\n\n\n /******************************************************************************\n * Tuning policies of current PTX compiler pass\n ******************************************************************************/\n\n#if (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 130)\n typedef Policy130 PtxPolicy;\n\n#else\n typedef Policy100 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxSelectIfPolicyT : PtxPolicy::SelectIfPolicyT {};\n\n\n /******************************************************************************\n * Utilities\n ******************************************************************************/\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static void InitConfigs(\n int ptx_version,\n KernelConfig &select_if_config)\n {\n #if (CUB_PTX_ARCH > 0)\n (void)ptx_version;\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n select_if_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 350)\n {\n select_if_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n select_if_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n select_if_config.template Init();\n }\n else if (ptx_version >= 130)\n {\n select_if_config.template Init();\n }\n else\n {\n select_if_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration.\n */\n struct KernelConfig\n {\n int block_threads;\n int items_per_thread;\n int tile_items;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n void Init()\n {\n block_threads = PolicyT::BLOCK_THREADS;\n items_per_thread = PolicyT::ITEMS_PER_THREAD;\n tile_items = block_threads * items_per_thread;\n }\n };\n\n\n /******************************************************************************\n * Dispatch entrypoints\n ******************************************************************************/\n\n /**\n * Internal dispatch routine for computing a device-wide selection using the\n * specified kernel functions.\n */\n template <\n typename ScanInitKernelPtrT, ///< Function type of cub::DeviceScanInitKernel\n typename SelectIfKernelPtrT> ///< Function type of cub::SelectIfKernelPtrT\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n FlagsInputIteratorT d_flags, ///< [in] Pointer to the input sequence of selection flags (if applicable)\n SelectedOutputIteratorT d_selected_out, ///< [in] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [in] Pointer to the total number of items selected (i.e., length of \\p d_selected_out)\n SelectOpT select_op, ///< [in] Selection operator\n EqualityOpT equality_op, ///< [in] Equality operator\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int /*ptx_version*/, ///< [in] PTX version of dispatch kernels\n ScanInitKernelPtrT scan_init_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceScanInitKernel\n SelectIfKernelPtrT select_if_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceSelectSweepKernel\n KernelConfig select_if_config) ///< [in] Dispatch parameters that match the policy that \\p select_if_kernel was compiled for\n {\n\n#ifndef CUB_RUNTIME_ENABLED\n (void)d_temp_storage;\n (void)temp_storage_bytes;\n (void)d_in;\n (void)d_flags;\n (void)d_selected_out;\n (void)d_num_selected_out;\n (void)select_op;\n (void)equality_op;\n (void)num_items;\n (void)stream;\n (void)debug_synchronous;\n (void)scan_init_kernel;\n (void)select_if_kernel;\n (void)select_if_config;\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported);\n\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Number of input tiles\n int tile_size = select_if_config.block_threads * select_if_config.items_per_thread;\n int num_tiles = (num_items + tile_size - 1) / tile_size;\n\n // Specify temporary storage allocation requirements\n size_t allocation_sizes[1];\n if (CubDebug(error = ScanTileStateT::AllocationSize(num_tiles, allocation_sizes[0]))) break; // bytes needed for tile status descriptors\n\n // Compute allocation pointers into the single storage blob (or compute the necessary size of the blob)\n void* allocations[1];\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Construct the tile status interface\n ScanTileStateT tile_status;\n if (CubDebug(error = tile_status.Init(num_tiles, allocations[0], allocation_sizes[0]))) break;\n\n // Log scan_init_kernel configuration\n int init_grid_size = CUB_MAX(1, (num_tiles + INIT_KERNEL_THREADS - 1) / INIT_KERNEL_THREADS);\n if (debug_synchronous) _CubLog(\"Invoking scan_init_kernel<<<%d, %d, 0, %lld>>>()\\n\", init_grid_size, INIT_KERNEL_THREADS, (long long) stream);\n\n // Invoke scan_init_kernel to initialize tile descriptors\n scan_init_kernel<<>>(\n tile_status,\n num_tiles,\n d_num_selected_out);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Return if empty problem\n if (num_items == 0)\n break;\n\n // Get SM occupancy for select_if_kernel\n int range_select_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n range_select_sm_occupancy, // out\n select_if_kernel,\n select_if_config.block_threads))) break;\n\n // Get max x-dimension of grid\n int max_dim_x;\n if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;\n\n // Get grid size for scanning tiles\n dim3 scan_grid_size;\n scan_grid_size.z = 1;\n scan_grid_size.y = ((unsigned int) num_tiles + max_dim_x - 1) / max_dim_x;\n scan_grid_size.x = CUB_MIN(num_tiles, max_dim_x);\n\n // Log select_if_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking select_if_kernel<<<{%d,%d,%d}, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n scan_grid_size.x, scan_grid_size.y, scan_grid_size.z, select_if_config.block_threads, (long long) stream, select_if_config.items_per_thread, range_select_sm_occupancy);\n\n // Invoke select_if_kernel\n select_if_kernel<<>>(\n d_in,\n d_flags,\n d_selected_out,\n d_num_selected_out,\n tile_status,\n select_op,\n equality_op,\n num_items,\n num_tiles);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n InputIteratorT d_in, ///< [in] Pointer to the input sequence of data items\n FlagsInputIteratorT d_flags, ///< [in] Pointer to the input sequence of selection flags (if applicable)\n SelectedOutputIteratorT d_selected_out, ///< [in] Pointer to the output sequence of selected data items\n NumSelectedIteratorT d_num_selected_out, ///< [in] Pointer to the total number of items selected (i.e., length of \\p d_selected_out)\n SelectOpT select_op, ///< [in] Selection operator\n EqualityOpT equality_op, ///< [in] Equality operator\n OffsetT num_items, ///< [in] Total number of input items (i.e., length of \\p d_in)\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel kernel dispatch configurations\n KernelConfig select_if_config;\n InitConfigs(ptx_version, select_if_config);\n\n // Dispatch\n if (CubDebug(error = Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_flags,\n d_selected_out,\n d_num_selected_out,\n select_op,\n equality_op,\n num_items,\n stream,\n debug_synchronous,\n ptx_version,\n DeviceCompactInitKernel,\n DeviceSelectSweepKernel,\n select_if_config))) break;\n }\n while (0);\n\n return error;\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/device/dispatch/dispatch_spmv_orig.cuh", "content": "\n/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::DeviceSpmv provides device-wide parallel operations for performing sparse-matrix * vector multiplication (SpMV).\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../../agent/single_pass_scan_operators.cuh\"\n#include \"../../agent/agent_segment_fixup.cuh\"\n#include \"../../agent/agent_spmv_orig.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_debug.cuh\"\n#include \"../../util_device.cuh\"\n#include \"../../thread/thread_search.cuh\"\n#include \"../../grid/grid_queue.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * SpMV kernel entry points\n *****************************************************************************/\n\n/**\n * Spmv search kernel. Identifies merge path starting coordinates for each tile.\n */\ntemplate <\n typename AgentSpmvPolicyT, ///< Parameterized SpmvPolicy tuning policy type\n typename ValueT, ///< Matrix and vector value type\n typename OffsetT> ///< Signed integer type for sequence offsets\n__global__ void DeviceSpmv1ColKernel(\n SpmvParams spmv_params) ///< [in] SpMV input parameter bundle\n{\n typedef CacheModifiedInputIterator<\n AgentSpmvPolicyT::VECTOR_VALUES_LOAD_MODIFIER,\n ValueT,\n OffsetT>\n VectorValueIteratorT;\n\n VectorValueIteratorT wrapped_vector_x(spmv_params.d_vector_x);\n\n int row_idx = (blockIdx.x * blockDim.x) + threadIdx.x;\n if (row_idx < spmv_params.num_rows)\n {\n OffsetT end_nonzero_idx = spmv_params.d_row_end_offsets[row_idx];\n OffsetT nonzero_idx = spmv_params.d_row_end_offsets[row_idx - 1];\n\n ValueT value = 0.0;\n if (end_nonzero_idx != nonzero_idx)\n {\n value = spmv_params.d_values[nonzero_idx] * wrapped_vector_x[spmv_params.d_column_indices[nonzero_idx]];\n }\n\n spmv_params.d_vector_y[row_idx] = value;\n }\n}\n\n\n/**\n * Spmv search kernel. Identifies merge path starting coordinates for each tile.\n */\ntemplate <\n typename SpmvPolicyT, ///< Parameterized SpmvPolicy tuning policy type\n typename OffsetT, ///< Signed integer type for sequence offsets\n typename CoordinateT, ///< Merge path coordinate type\n typename SpmvParamsT> ///< SpmvParams type\n__global__ void DeviceSpmvSearchKernel(\n int num_merge_tiles, ///< [in] Number of SpMV merge tiles (spmv grid size)\n CoordinateT* d_tile_coordinates, ///< [out] Pointer to the temporary array of tile starting coordinates\n SpmvParamsT spmv_params) ///< [in] SpMV input parameter bundle\n{\n /// Constants\n enum\n {\n BLOCK_THREADS = SpmvPolicyT::BLOCK_THREADS,\n ITEMS_PER_THREAD = SpmvPolicyT::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n typedef CacheModifiedInputIterator<\n SpmvPolicyT::ROW_OFFSETS_SEARCH_LOAD_MODIFIER,\n OffsetT,\n OffsetT>\n RowOffsetsSearchIteratorT;\n\n // Find the starting coordinate for all tiles (plus the end coordinate of the last one)\n int tile_idx = (blockIdx.x * blockDim.x) + threadIdx.x;\n if (tile_idx < num_merge_tiles + 1)\n {\n OffsetT diagonal = (tile_idx * TILE_ITEMS);\n CoordinateT tile_coordinate;\n CountingInputIterator nonzero_indices(0);\n\n // Search the merge path\n MergePathSearch(\n diagonal,\n RowOffsetsSearchIteratorT(spmv_params.d_row_end_offsets),\n nonzero_indices,\n spmv_params.num_rows,\n spmv_params.num_nonzeros,\n tile_coordinate);\n\n // Output starting offset\n d_tile_coordinates[tile_idx] = tile_coordinate;\n }\n}\n\n\n/**\n * Spmv agent entry point\n */\ntemplate <\n typename SpmvPolicyT, ///< Parameterized SpmvPolicy tuning policy type\n typename ScanTileStateT, ///< Tile status interface type\n typename ValueT, ///< Matrix and vector value type\n typename OffsetT, ///< Signed integer type for sequence offsets\n typename CoordinateT, ///< Merge path coordinate type\n bool HAS_ALPHA, ///< Whether the input parameter Alpha is 1\n bool HAS_BETA> ///< Whether the input parameter Beta is 0\n__launch_bounds__ (int(SpmvPolicyT::BLOCK_THREADS))\n__global__ void DeviceSpmvKernel(\n SpmvParams spmv_params, ///< [in] SpMV input parameter bundle\n CoordinateT* d_tile_coordinates, ///< [in] Pointer to the temporary array of tile starting coordinates\n KeyValuePair* d_tile_carry_pairs, ///< [out] Pointer to the temporary array carry-out dot product row-ids, one per block\n int num_tiles, ///< [in] Number of merge tiles\n ScanTileStateT tile_state, ///< [in] Tile status interface for fixup reduce-by-key kernel\n int num_segment_fixup_tiles) ///< [in] Number of reduce-by-key tiles (fixup grid size)\n{\n // Spmv agent type specialization\n typedef AgentSpmv<\n SpmvPolicyT,\n ValueT,\n OffsetT,\n HAS_ALPHA,\n HAS_BETA>\n AgentSpmvT;\n\n // Shared memory for AgentSpmv\n __shared__ typename AgentSpmvT::TempStorage temp_storage;\n\n AgentSpmvT(temp_storage, spmv_params).ConsumeTile(\n d_tile_coordinates,\n d_tile_carry_pairs,\n num_tiles);\n\n // Initialize fixup tile status\n tile_state.InitializeStatus(num_segment_fixup_tiles);\n\n}\n\n\n/**\n * Multi-block reduce-by-key sweep kernel entry point\n */\ntemplate <\n typename AgentSegmentFixupPolicyT, ///< Parameterized AgentSegmentFixupPolicy tuning policy type\n typename PairsInputIteratorT, ///< Random-access input iterator type for keys\n typename AggregatesOutputIteratorT, ///< Random-access output iterator type for values\n typename OffsetT, ///< Signed integer type for global offsets\n typename ScanTileStateT> ///< Tile status interface type\n__launch_bounds__ (int(AgentSegmentFixupPolicyT::BLOCK_THREADS))\n__global__ void DeviceSegmentFixupKernel(\n PairsInputIteratorT d_pairs_in, ///< [in] Pointer to the array carry-out dot product row-ids, one per spmv block\n AggregatesOutputIteratorT d_aggregates_out, ///< [in,out] Output value aggregates\n OffsetT num_items, ///< [in] Total number of items to select from\n int num_tiles, ///< [in] Total number of tiles for the entire problem\n ScanTileStateT tile_state) ///< [in] Tile status interface\n{\n // Thread block type for reducing tiles of value segments\n typedef AgentSegmentFixup<\n AgentSegmentFixupPolicyT,\n PairsInputIteratorT,\n AggregatesOutputIteratorT,\n cub::Equality,\n cub::Sum,\n OffsetT>\n AgentSegmentFixupT;\n\n // Shared memory for AgentSegmentFixup\n __shared__ typename AgentSegmentFixupT::TempStorage temp_storage;\n\n // Process tiles\n AgentSegmentFixupT(temp_storage, d_pairs_in, d_aggregates_out, cub::Equality(), cub::Sum()).ConsumeRange(\n num_items,\n num_tiles,\n tile_state);\n}\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceSpmv\n */\ntemplate <\n typename ValueT, ///< Matrix and vector value type\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DispatchSpmv\n{\n //---------------------------------------------------------------------\n // Constants and Types\n //---------------------------------------------------------------------\n\n enum\n {\n INIT_KERNEL_THREADS = 128\n };\n\n // SpmvParams bundle type\n typedef SpmvParams SpmvParamsT;\n\n // 2D merge path coordinate type\n typedef typename CubVector::Type CoordinateT;\n\n // Tile status descriptor interface type\n typedef ReduceByKeyScanTileState ScanTileStateT;\n\n // Tuple type for scanning (pairs accumulated segment-value with segment-index)\n typedef KeyValuePair KeyValuePairT;\n\n\n //---------------------------------------------------------------------\n // Tuning policies\n //---------------------------------------------------------------------\n\n /// SM11\n struct Policy110\n {\n typedef AgentSpmvPolicy<\n 128,\n 1,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n false,\n BLOCK_SCAN_WARP_SCANS>\n SpmvPolicyT;\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 4,\n BLOCK_LOAD_VECTORIZE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n SegmentFixupPolicyT;\n };\n\n /// SM20\n struct Policy200 \n {\n typedef AgentSpmvPolicy<\n 96,\n 18,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n false,\n BLOCK_SCAN_RAKING>\n SpmvPolicyT;\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 4,\n BLOCK_LOAD_VECTORIZE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n SegmentFixupPolicyT;\n\n };\n\n\n\n /// SM30\n struct Policy300 \n {\n typedef AgentSpmvPolicy<\n 96,\n 6,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n LOAD_DEFAULT,\n false,\n BLOCK_SCAN_WARP_SCANS>\n SpmvPolicyT;\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 4,\n BLOCK_LOAD_VECTORIZE,\n LOAD_DEFAULT,\n BLOCK_SCAN_WARP_SCANS>\n SegmentFixupPolicyT;\n\n };\n\n\n /// SM35\n struct Policy350\n {\n typedef AgentSpmvPolicy<\n (sizeof(ValueT) > 4) ? 96 : 128,\n (sizeof(ValueT) > 4) ? 4 : 7,\n LOAD_LDG,\n LOAD_CA,\n LOAD_LDG,\n LOAD_LDG,\n LOAD_LDG,\n (sizeof(ValueT) > 4) ? true : false,\n BLOCK_SCAN_WARP_SCANS>\n SpmvPolicyT;\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 3,\n BLOCK_LOAD_VECTORIZE,\n LOAD_LDG,\n BLOCK_SCAN_WARP_SCANS>\n SegmentFixupPolicyT;\n };\n\n\n /// SM37\n struct Policy370\n {\n\n typedef AgentSpmvPolicy<\n (sizeof(ValueT) > 4) ? 128 : 128,\n (sizeof(ValueT) > 4) ? 9 : 14,\n LOAD_LDG,\n LOAD_CA,\n LOAD_LDG,\n LOAD_LDG,\n LOAD_LDG,\n false, \n BLOCK_SCAN_WARP_SCANS>\n SpmvPolicyT;\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 3,\n BLOCK_LOAD_VECTORIZE,\n LOAD_LDG,\n BLOCK_SCAN_WARP_SCANS>\n SegmentFixupPolicyT;\n };\n\n /// SM50\n struct Policy500\n {\n typedef AgentSpmvPolicy<\n (sizeof(ValueT) > 4) ? 64 : 128,\n (sizeof(ValueT) > 4) ? 6 : 7,\n LOAD_LDG,\n LOAD_DEFAULT,\n (sizeof(ValueT) > 4) ? LOAD_LDG : LOAD_DEFAULT,\n (sizeof(ValueT) > 4) ? LOAD_LDG : LOAD_DEFAULT,\n LOAD_LDG,\n (sizeof(ValueT) > 4) ? true : false,\n (sizeof(ValueT) > 4) ? BLOCK_SCAN_WARP_SCANS : BLOCK_SCAN_RAKING_MEMOIZE>\n SpmvPolicyT;\n\n\n typedef AgentSegmentFixupPolicy<\n 128,\n 3,\n BLOCK_LOAD_VECTORIZE,\n LOAD_LDG,\n BLOCK_SCAN_RAKING_MEMOIZE>\n SegmentFixupPolicyT;\n };\n\n\n\n //---------------------------------------------------------------------\n // Tuning policies of current PTX compiler pass\n //---------------------------------------------------------------------\n\n#if (CUB_PTX_ARCH >= 500)\n typedef Policy500 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 370)\n typedef Policy370 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#else\n typedef Policy110 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxSpmvPolicyT : PtxPolicy::SpmvPolicyT {};\n struct PtxSegmentFixupPolicy : PtxPolicy::SegmentFixupPolicyT {};\n\n\n //---------------------------------------------------------------------\n // Utilities\n //---------------------------------------------------------------------\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static void InitConfigs(\n int ptx_version,\n KernelConfig &spmv_config,\n KernelConfig &segment_fixup_config)\n {\n #if (CUB_PTX_ARCH > 0)\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n spmv_config.template Init();\n segment_fixup_config.template Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 500)\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n }\n else if (ptx_version >= 370)\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n }\n else if (ptx_version >= 350)\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n }\n else if (ptx_version >= 300)\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n\n }\n else if (ptx_version >= 200)\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n }\n else\n {\n spmv_config.template Init();\n segment_fixup_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * Kernel kernel dispatch configuration.\n */\n struct KernelConfig\n {\n int block_threads;\n int items_per_thread;\n int tile_items;\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n void Init()\n {\n block_threads = PolicyT::BLOCK_THREADS;\n items_per_thread = PolicyT::ITEMS_PER_THREAD;\n tile_items = block_threads * items_per_thread;\n }\n };\n\n\n //---------------------------------------------------------------------\n // Dispatch entrypoints\n //---------------------------------------------------------------------\n\n /**\n * Internal dispatch routine for computing a device-wide reduction using the\n * specified kernel functions.\n *\n * If the input is larger than a single tile, this method uses two-passes of\n * kernel invocations.\n */\n template <\n typename Spmv1ColKernelT, ///< Function type of cub::DeviceSpmv1ColKernel\n typename SpmvSearchKernelT, ///< Function type of cub::AgentSpmvSearchKernel\n typename SpmvKernelT, ///< Function type of cub::AgentSpmvKernel\n typename SegmentFixupKernelT> ///< Function type of cub::DeviceSegmentFixupKernelT\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SpmvParamsT& spmv_params, ///< SpMV input parameter bundle\n cudaStream_t stream, ///< [in] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n Spmv1ColKernelT spmv_1col_kernel, ///< [in] Kernel function pointer to parameterization of DeviceSpmv1ColKernel\n SpmvSearchKernelT spmv_search_kernel, ///< [in] Kernel function pointer to parameterization of AgentSpmvSearchKernel\n SpmvKernelT spmv_kernel, ///< [in] Kernel function pointer to parameterization of AgentSpmvKernel\n SegmentFixupKernelT segment_fixup_kernel, ///< [in] Kernel function pointer to parameterization of cub::DeviceSegmentFixupKernel\n KernelConfig spmv_config, ///< [in] Dispatch parameters that match the policy that \\p spmv_kernel was compiled for\n KernelConfig segment_fixup_config) ///< [in] Dispatch parameters that match the policy that \\p segment_fixup_kernel was compiled for\n {\n#ifndef CUB_RUNTIME_ENABLED\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n\n#else\n cudaError error = cudaSuccess;\n do\n {\n if (spmv_params.num_cols == 1)\n {\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n temp_storage_bytes = 1;\n break;\n }\n\n // Get search/init grid dims\n int degen_col_kernel_block_size = INIT_KERNEL_THREADS;\n int degen_col_kernel_grid_size = (spmv_params.num_rows + degen_col_kernel_block_size - 1) / degen_col_kernel_block_size;\n\n if (debug_synchronous) _CubLog(\"Invoking spmv_1col_kernel<<<%d, %d, 0, %lld>>>()\\n\",\n degen_col_kernel_grid_size, degen_col_kernel_block_size, (long long) stream);\n\n // Invoke spmv_search_kernel\n spmv_1col_kernel<<>>(\n spmv_params);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n break;\n }\n\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Get max x-dimension of grid\n int max_dim_x;\n if (CubDebug(error = cudaDeviceGetAttribute(&max_dim_x, cudaDevAttrMaxGridDimX, device_ordinal))) break;;\n\n // Total number of spmv work items\n int num_merge_items = spmv_params.num_rows + spmv_params.num_nonzeros;\n\n // Tile sizes of kernels\n int merge_tile_size = spmv_config.block_threads * spmv_config.items_per_thread;\n int segment_fixup_tile_size = segment_fixup_config.block_threads * segment_fixup_config.items_per_thread;\n\n // Number of tiles for kernels\n unsigned int num_merge_tiles = (num_merge_items + merge_tile_size - 1) / merge_tile_size;\n unsigned int num_segment_fixup_tiles = (num_merge_tiles + segment_fixup_tile_size - 1) / segment_fixup_tile_size;\n\n // Get SM occupancy for kernels\n int spmv_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n spmv_sm_occupancy,\n spmv_kernel,\n spmv_config.block_threads))) break;\n\n int segment_fixup_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n segment_fixup_sm_occupancy,\n segment_fixup_kernel,\n segment_fixup_config.block_threads))) break;\n\n // Get grid dimensions\n dim3 spmv_grid_size(\n CUB_MIN(num_merge_tiles, max_dim_x),\n (num_merge_tiles + max_dim_x - 1) / max_dim_x,\n 1);\n\n dim3 segment_fixup_grid_size(\n CUB_MIN(num_segment_fixup_tiles, max_dim_x),\n (num_segment_fixup_tiles + max_dim_x - 1) / max_dim_x,\n 1);\n\n // Get the temporary storage allocation requirements\n size_t allocation_sizes[3];\n if (CubDebug(error = ScanTileStateT::AllocationSize(num_segment_fixup_tiles, allocation_sizes[0]))) break; // bytes needed for reduce-by-key tile status descriptors\n allocation_sizes[1] = num_merge_tiles * sizeof(KeyValuePairT); // bytes needed for block carry-out pairs\n allocation_sizes[2] = (num_merge_tiles + 1) * sizeof(CoordinateT); // bytes needed for tile starting coordinates\n\n // Alias the temporary allocations from the single storage blob (or compute the necessary size of the blob)\n void* allocations[3];\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n break;\n }\n\n // Construct the tile status interface\n ScanTileStateT tile_state;\n if (CubDebug(error = tile_state.Init(num_segment_fixup_tiles, allocations[0], allocation_sizes[0]))) break;\n\n // Alias the other allocations\n KeyValuePairT* d_tile_carry_pairs = (KeyValuePairT*) allocations[1]; // Agent carry-out pairs\n CoordinateT* d_tile_coordinates = (CoordinateT*) allocations[2]; // Agent starting coordinates\n\n // Get search/init grid dims\n int search_block_size = INIT_KERNEL_THREADS;\n int search_grid_size = (num_merge_tiles + 1 + search_block_size - 1) / search_block_size;\n\n#if (CUB_PTX_ARCH == 0)\n // Init textures\n if (CubDebug(error = spmv_params.t_vector_x.BindTexture(spmv_params.d_vector_x))) break;\n#endif\n\n if (search_grid_size < sm_count)\n// if (num_merge_tiles < spmv_sm_occupancy * sm_count)\n {\n // Not enough spmv tiles to saturate the device: have spmv blocks search their own staring coords\n d_tile_coordinates = NULL;\n }\n else\n {\n // Use separate search kernel if we have enough spmv tiles to saturate the device\n\n // Log spmv_search_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking spmv_search_kernel<<<%d, %d, 0, %lld>>>()\\n\",\n search_grid_size, search_block_size, (long long) stream);\n\n // Invoke spmv_search_kernel\n spmv_search_kernel<<>>(\n num_merge_tiles,\n d_tile_coordinates,\n spmv_params);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n\n // Log spmv_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking spmv_kernel<<<{%d,%d,%d}, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n spmv_grid_size.x, spmv_grid_size.y, spmv_grid_size.z, spmv_config.block_threads, (long long) stream, spmv_config.items_per_thread, spmv_sm_occupancy);\n\n // Invoke spmv_kernel\n spmv_kernel<<>>(\n spmv_params,\n d_tile_coordinates,\n d_tile_carry_pairs,\n num_merge_tiles,\n tile_state,\n num_segment_fixup_tiles);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n\n // Run reduce-by-key fixup if necessary\n if (num_merge_tiles > 1)\n {\n // Log segment_fixup_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking segment_fixup_kernel<<<{%d,%d,%d}, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n segment_fixup_grid_size.x, segment_fixup_grid_size.y, segment_fixup_grid_size.z, segment_fixup_config.block_threads, (long long) stream, segment_fixup_config.items_per_thread, segment_fixup_sm_occupancy);\n\n // Invoke segment_fixup_kernel\n segment_fixup_kernel<<>>(\n d_tile_carry_pairs,\n spmv_params.d_vector_y,\n num_merge_tiles,\n num_segment_fixup_tiles,\n tile_state);\n\n // Check for failure to launch\n if (CubDebug(error = cudaPeekAtLastError())) break;\n\n // Sync the stream if specified to flush runtime errors\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n\n#if (CUB_PTX_ARCH == 0)\n // Free textures\n if (CubDebug(error = spmv_params.t_vector_x.UnbindTexture())) break;\n#endif\n }\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine for computing a device-wide reduction\n */\n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t& temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation\n SpmvParamsT& spmv_params, ///< SpMV input parameter bundle\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. May cause significant slowdown. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel kernel dispatch configurations\n KernelConfig spmv_config, segment_fixup_config;\n InitConfigs(ptx_version, spmv_config, segment_fixup_config);\n\n if (CubDebug(error = Dispatch(\n d_temp_storage, temp_storage_bytes, spmv_params, stream, debug_synchronous,\n DeviceSpmv1ColKernel,\n DeviceSpmvSearchKernel,\n DeviceSpmvKernel,\n DeviceSegmentFixupKernel,\n spmv_config, segment_fixup_config))) break;\n\n/*\n // Dispatch\n if (spmv_params.beta == 0.0)\n {\n if (spmv_params.alpha == 1.0)\n {\n // Dispatch y = A*x\n if (CubDebug(error = Dispatch(\n d_temp_storage, temp_storage_bytes, spmv_params, stream, debug_synchronous,\n DeviceSpmv1ColKernel,\n DeviceSpmvSearchKernel,\n DeviceSpmvKernel,\n DeviceSegmentFixupKernel,\n spmv_config, segment_fixup_config))) break;\n }\n else\n {\n // Dispatch y = alpha*A*x\n if (CubDebug(error = Dispatch(\n d_temp_storage, temp_storage_bytes, spmv_params, stream, debug_synchronous,\n DeviceSpmvSearchKernel,\n DeviceSpmvKernel,\n DeviceSegmentFixupKernel,\n spmv_config, segment_fixup_config))) break;\n }\n }\n else\n {\n if (spmv_params.alpha == 1.0)\n {\n // Dispatch y = A*x + beta*y\n if (CubDebug(error = Dispatch(\n d_temp_storage, temp_storage_bytes, spmv_params, stream, debug_synchronous,\n DeviceSpmvSearchKernel,\n DeviceSpmvKernel,\n DeviceSegmentFixupKernel,\n spmv_config, segment_fixup_config))) break;\n }\n else\n {\n // Dispatch y = alpha*A*x + beta*y\n if (CubDebug(error = Dispatch(\n d_temp_storage, temp_storage_bytes, spmv_params, stream, debug_synchronous,\n DeviceSpmvSearchKernel,\n DeviceSpmvKernel,\n DeviceSegmentFixupKernel,\n spmv_config, segment_fixup_config))) break;\n }\n }\n*/\n }\n while (0);\n\n return error;\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/grid/grid_barrier.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::GridBarrier implements a software global barrier among thread blocks within a CUDA grid\n */\n\n#pragma once\n\n#include \"../util_debug.cuh\"\n#include \"../util_namespace.cuh\"\n#include \"../thread/thread_load.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup GridModule\n * @{\n */\n\n\n/**\n * \\brief GridBarrier implements a software global barrier among thread blocks within a CUDA grid\n */\nclass GridBarrier\n{\nprotected :\n\n typedef unsigned int SyncFlag;\n\n // Counters in global device memory\n SyncFlag* d_sync;\n\npublic:\n\n /**\n * Constructor\n */\n GridBarrier() : d_sync(NULL) {}\n\n\n /**\n * Synchronize\n */\n __device__ __forceinline__ void Sync() const\n {\n volatile SyncFlag *d_vol_sync = d_sync;\n\n // Threadfence and syncthreads to make sure global writes are visible before\n // thread-0 reports in with its sync counter\n __threadfence();\n CTA_SYNC();\n\n if (blockIdx.x == 0)\n {\n // Report in ourselves\n if (threadIdx.x == 0)\n {\n d_vol_sync[blockIdx.x] = 1;\n }\n\n CTA_SYNC();\n\n // Wait for everyone else to report in\n for (int peer_block = threadIdx.x; peer_block < gridDim.x; peer_block += blockDim.x)\n {\n while (ThreadLoad(d_sync + peer_block) == 0)\n {\n __threadfence_block();\n }\n }\n\n CTA_SYNC();\n\n // Let everyone know it's safe to proceed\n for (int peer_block = threadIdx.x; peer_block < gridDim.x; peer_block += blockDim.x)\n {\n d_vol_sync[peer_block] = 0;\n }\n }\n else\n {\n if (threadIdx.x == 0)\n {\n // Report in\n d_vol_sync[blockIdx.x] = 1;\n\n // Wait for acknowledgment\n while (ThreadLoad(d_sync + blockIdx.x) == 1)\n {\n __threadfence_block();\n }\n }\n\n CTA_SYNC();\n }\n }\n};\n\n\n/**\n * \\brief GridBarrierLifetime extends GridBarrier to provide lifetime management of the temporary device storage needed for cooperation.\n *\n * Uses RAII for lifetime, i.e., device resources are reclaimed when\n * the destructor is called.\n */\nclass GridBarrierLifetime : public GridBarrier\n{\nprotected:\n\n // Number of bytes backed by d_sync\n size_t sync_bytes;\n\npublic:\n\n /**\n * Constructor\n */\n GridBarrierLifetime() : GridBarrier(), sync_bytes(0) {}\n\n\n /**\n * DeviceFrees and resets the progress counters\n */\n cudaError_t HostReset()\n {\n cudaError_t retval = cudaSuccess;\n if (d_sync)\n {\n CubDebug(retval = cudaFree(d_sync));\n d_sync = NULL;\n }\n sync_bytes = 0;\n return retval;\n }\n\n\n /**\n * Destructor\n */\n virtual ~GridBarrierLifetime()\n {\n HostReset();\n }\n\n\n /**\n * Sets up the progress counters for the next kernel launch (lazily\n * allocating and initializing them if necessary)\n */\n cudaError_t Setup(int sweep_grid_size)\n {\n cudaError_t retval = cudaSuccess;\n do {\n size_t new_sync_bytes = sweep_grid_size * sizeof(SyncFlag);\n if (new_sync_bytes > sync_bytes)\n {\n if (d_sync)\n {\n if (CubDebug(retval = cudaFree(d_sync))) break;\n }\n\n sync_bytes = new_sync_bytes;\n\n // Allocate and initialize to zero\n if (CubDebug(retval = cudaMalloc((void**) &d_sync, sync_bytes))) break;\n if (CubDebug(retval = cudaMemset(d_sync, 0, new_sync_bytes))) break;\n }\n } while (0);\n\n return retval;\n }\n};\n\n\n/** @} */ // end group GridModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/grid/grid_even_share.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::GridEvenShare is a descriptor utility for distributing input among CUDA thread blocks in an \"even-share\" fashion. Each thread block gets roughly the same number of fixed-size work units (grains).\n */\n\n\n#pragma once\n\n#include \"../util_namespace.cuh\"\n#include \"../util_macro.cuh\"\n#include \"grid_mapping.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup GridModule\n * @{\n */\n\n\n/**\n * \\brief GridEvenShare is a descriptor utility for distributing input among\n * CUDA thread blocks in an \"even-share\" fashion. Each thread block gets roughly\n * the same number of input tiles.\n *\n * \\par Overview\n * Each thread block is assigned a consecutive sequence of input tiles. To help\n * preserve alignment and eliminate the overhead of guarded loads for all but the\n * last thread block, to GridEvenShare assigns one of three different amounts of\n * work to a given thread block: \"big\", \"normal\", or \"last\". The \"big\" workloads\n * are one scheduling grain larger than \"normal\". The \"last\" work unit for the\n * last thread block may be partially-full if the input is not an even multiple of\n * the scheduling grain size.\n *\n * \\par\n * Before invoking a child grid, a parent thread will typically construct an\n * instance of GridEvenShare. The instance can be passed to child thread blocks\n * which can initialize their per-thread block offsets using \\p BlockInit().\n */\ntemplate \nstruct GridEvenShare\n{\nprivate:\n\n OffsetT total_tiles;\n int big_shares;\n OffsetT big_share_items;\n OffsetT normal_share_items;\n OffsetT normal_base_offset;\n\npublic:\n\n /// Total number of input items\n OffsetT num_items;\n\n /// Grid size in thread blocks\n int grid_size;\n\n /// OffsetT into input marking the beginning of the owning thread block's segment of input tiles\n OffsetT block_offset;\n\n /// OffsetT into input of marking the end (one-past) of the owning thread block's segment of input tiles\n OffsetT block_end;\n\n /// Stride between input tiles\n OffsetT block_stride;\n\n\n /**\n * \\brief Constructor.\n */\n __host__ __device__ __forceinline__ GridEvenShare() :\n total_tiles(0),\n big_shares(0),\n big_share_items(0),\n normal_share_items(0),\n normal_base_offset(0),\n num_items(0),\n grid_size(0),\n block_offset(0),\n block_end(0),\n block_stride(0)\n {}\n\n\n /**\n * \\brief Dispatch initializer. To be called prior prior to kernel launch.\n */\n __host__ __device__ __forceinline__ void DispatchInit(\n OffsetT num_items, ///< Total number of input items\n int max_grid_size, ///< Maximum grid size allowable (actual grid size may be less if not warranted by the the number of input items)\n int tile_items) ///< Number of data items per input tile\n {\n this->block_offset = num_items; // Initialize past-the-end\n this->block_end = num_items; // Initialize past-the-end\n this->num_items = num_items;\n this->total_tiles = (num_items + tile_items - 1) / tile_items;\n this->grid_size = CUB_MIN(total_tiles, max_grid_size);\n OffsetT avg_tiles_per_block = total_tiles / grid_size;\n this->big_shares = total_tiles - (avg_tiles_per_block * grid_size); // leftover grains go to big blocks\n this->normal_share_items = avg_tiles_per_block * tile_items;\n this->normal_base_offset = big_shares * tile_items;\n this->big_share_items = normal_share_items + tile_items;\n }\n\n\n /**\n * \\brief Initializes ranges for the specified thread block index. Specialized\n * for a \"raking\" access pattern in which each thread block is assigned a\n * consecutive sequence of input tiles.\n */\n template \n __device__ __forceinline__ void BlockInit(\n int block_id,\n Int2Type /*strategy_tag*/)\n {\n block_stride = TILE_ITEMS;\n if (block_id < big_shares)\n {\n // This thread block gets a big share of grains (avg_tiles_per_block + 1)\n block_offset = (block_id * big_share_items);\n block_end = block_offset + big_share_items;\n }\n else if (block_id < total_tiles)\n {\n // This thread block gets a normal share of grains (avg_tiles_per_block)\n block_offset = normal_base_offset + (block_id * normal_share_items);\n block_end = CUB_MIN(num_items, block_offset + normal_share_items);\n }\n // Else default past-the-end\n }\n\n\n /**\n * \\brief Block-initialization, specialized for a \"raking\" access\n * pattern in which each thread block is assigned a consecutive sequence\n * of input tiles.\n */\n template \n __device__ __forceinline__ void BlockInit(\n int block_id,\n Int2Type /*strategy_tag*/)\n {\n block_stride = grid_size * TILE_ITEMS;\n block_offset = (block_id * TILE_ITEMS);\n block_end = num_items;\n }\n\n\n /**\n * \\brief Block-initialization, specialized for \"strip mining\" access\n * pattern in which the input tiles assigned to each thread block are\n * separated by a stride equal to the the extent of the grid.\n */\n template <\n int TILE_ITEMS,\n GridMappingStrategy STRATEGY>\n __device__ __forceinline__ void BlockInit()\n {\n BlockInit(blockIdx.x, Int2Type());\n }\n\n\n /**\n * \\brief Block-initialization, specialized for a \"raking\" access\n * pattern in which each thread block is assigned a consecutive sequence\n * of input tiles.\n */\n template \n __device__ __forceinline__ void BlockInit(\n OffsetT block_offset, ///< [in] Threadblock begin offset (inclusive)\n OffsetT block_end) ///< [in] Threadblock end offset (exclusive)\n {\n this->block_offset = block_offset;\n this->block_end = block_end;\n this->block_stride = TILE_ITEMS;\n }\n\n\n};\n\n\n\n\n\n/** @} */ // end group GridModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/grid/grid_mapping.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::GridMappingStrategy enumerates alternative strategies for mapping constant-sized tiles of device-wide data onto a grid of CUDA thread blocks.\n */\n\n#pragma once\n\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup GridModule\n * @{\n */\n\n\n/******************************************************************************\n * Mapping policies\n *****************************************************************************/\n\n\n/**\n * \\brief cub::GridMappingStrategy enumerates alternative strategies for mapping constant-sized tiles of device-wide data onto a grid of CUDA thread blocks.\n */\nenum GridMappingStrategy\n{\n /**\n * \\brief An a \"raking\" access pattern in which each thread block is\n * assigned a consecutive sequence of input tiles\n *\n * \\par Overview\n * The input is evenly partitioned into \\p p segments, where \\p p is\n * constant and corresponds loosely to the number of thread blocks that may\n * actively reside on the target device. Each segment is comprised of\n * consecutive tiles, where a tile is a small, constant-sized unit of input\n * to be processed to completion before the thread block terminates or\n * obtains more work. The kernel invokes \\p p thread blocks, each\n * of which iteratively consumes a segment of n/p elements\n * in tile-size increments.\n */\n GRID_MAPPING_RAKE,\n\n /**\n * \\brief An a \"strip mining\" access pattern in which the input tiles assigned\n * to each thread block are separated by a stride equal to the the extent of\n * the grid.\n *\n * \\par Overview\n * The input is evenly partitioned into \\p p sets, where \\p p is\n * constant and corresponds loosely to the number of thread blocks that may\n * actively reside on the target device. Each set is comprised of\n * data tiles separated by stride \\p tiles, where a tile is a small,\n * constant-sized unit of input to be processed to completion before the\n * thread block terminates or obtains more work. The kernel invokes \\p p\n * thread blocks, each of which iteratively consumes a segment of\n * n/p elements in tile-size increments.\n */\n GRID_MAPPING_STRIP_MINE,\n\n /**\n * \\brief A dynamic \"queue-based\" strategy for assigning input tiles to thread blocks.\n *\n * \\par Overview\n * The input is treated as a queue to be dynamically consumed by a grid of\n * thread blocks. Work is atomically dequeued in tiles, where a tile is a\n * unit of input to be processed to completion before the thread block\n * terminates or obtains more work. The grid size \\p p is constant,\n * loosely corresponding to the number of thread blocks that may actively\n * reside on the target device.\n */\n GRID_MAPPING_DYNAMIC,\n};\n\n\n/** @} */ // end group GridModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/grid/grid_queue.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::GridQueue is a descriptor utility for dynamic queue management.\n */\n\n#pragma once\n\n#include \"../util_namespace.cuh\"\n#include \"../util_debug.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup GridModule\n * @{\n */\n\n\n/**\n * \\brief GridQueue is a descriptor utility for dynamic queue management.\n *\n * \\par Overview\n * GridQueue descriptors provides abstractions for \"filling\" or\n * \"draining\" globally-shared vectors.\n *\n * \\par\n * A \"filling\" GridQueue works by atomically-adding to a zero-initialized counter,\n * returning a unique offset for the calling thread to write its items.\n * The GridQueue maintains the total \"fill-size\". The fill counter must be reset\n * using GridQueue::ResetFill by the host or kernel instance prior to the kernel instance that\n * will be filling.\n *\n * \\par\n * Similarly, a \"draining\" GridQueue works by works by atomically-incrementing a\n * zero-initialized counter, returning a unique offset for the calling thread to\n * read its items. Threads can safely drain until the array's logical fill-size is\n * exceeded. The drain counter must be reset using GridQueue::ResetDrain or\n * GridQueue::FillAndResetDrain by the host or kernel instance prior to the kernel instance that\n * will be filling. (For dynamic work distribution of existing data, the corresponding fill-size\n * is simply the number of elements in the array.)\n *\n * \\par\n * Iterative work management can be implemented simply with a pair of flip-flopping\n * work buffers, each with an associated set of fill and drain GridQueue descriptors.\n *\n * \\tparam OffsetT Signed integer type for global offsets\n */\ntemplate \nclass GridQueue\n{\nprivate:\n\n /// Counter indices\n enum\n {\n FILL = 0,\n DRAIN = 1,\n };\n\n /// Pair of counters\n OffsetT *d_counters;\n\npublic:\n\n /// Returns the device allocation size in bytes needed to construct a GridQueue instance\n __host__ __device__ __forceinline__\n static size_t AllocationSize()\n {\n return sizeof(OffsetT) * 2;\n }\n\n\n /// Constructs an invalid GridQueue descriptor\n __host__ __device__ __forceinline__ GridQueue()\n :\n d_counters(NULL)\n {}\n\n\n /// Constructs a GridQueue descriptor around the device storage allocation\n __host__ __device__ __forceinline__ GridQueue(\n void *d_storage) ///< Device allocation to back the GridQueue. Must be at least as big as AllocationSize().\n :\n d_counters((OffsetT*) d_storage)\n {}\n\n\n /// This operation sets the fill-size and resets the drain counter, preparing the GridQueue for draining in the next kernel instance. To be called by the host or by a kernel prior to that which will be draining.\n __host__ __device__ __forceinline__ cudaError_t FillAndResetDrain(\n OffsetT fill_size,\n cudaStream_t stream = 0)\n {\n#if (CUB_PTX_ARCH > 0)\n (void)stream;\n d_counters[FILL] = fill_size;\n d_counters[DRAIN] = 0;\n return cudaSuccess;\n#else\n OffsetT counters[2];\n counters[FILL] = fill_size;\n counters[DRAIN] = 0;\n return CubDebug(cudaMemcpyAsync(d_counters, counters, sizeof(OffsetT) * 2, cudaMemcpyHostToDevice, stream));\n#endif\n }\n\n\n /// This operation resets the drain so that it may advance to meet the existing fill-size. To be called by the host or by a kernel prior to that which will be draining.\n __host__ __device__ __forceinline__ cudaError_t ResetDrain(cudaStream_t stream = 0)\n {\n#if (CUB_PTX_ARCH > 0)\n (void)stream;\n d_counters[DRAIN] = 0;\n return cudaSuccess;\n#else\n return CubDebug(cudaMemsetAsync(d_counters + DRAIN, 0, sizeof(OffsetT), stream));\n#endif\n }\n\n\n /// This operation resets the fill counter. To be called by the host or by a kernel prior to that which will be filling.\n __host__ __device__ __forceinline__ cudaError_t ResetFill(cudaStream_t stream = 0)\n {\n#if (CUB_PTX_ARCH > 0)\n (void)stream;\n d_counters[FILL] = 0;\n return cudaSuccess;\n#else\n return CubDebug(cudaMemsetAsync(d_counters + FILL, 0, sizeof(OffsetT), stream));\n#endif\n }\n\n\n /// Returns the fill-size established by the parent or by the previous kernel.\n __host__ __device__ __forceinline__ cudaError_t FillSize(\n OffsetT &fill_size,\n cudaStream_t stream = 0)\n {\n#if (CUB_PTX_ARCH > 0)\n (void)stream;\n fill_size = d_counters[FILL];\n return cudaSuccess;\n#else\n return CubDebug(cudaMemcpyAsync(&fill_size, d_counters + FILL, sizeof(OffsetT), cudaMemcpyDeviceToHost, stream));\n#endif\n }\n\n\n /// Drain \\p num_items from the queue. Returns offset from which to read items. To be called from CUDA kernel.\n __device__ __forceinline__ OffsetT Drain(OffsetT num_items)\n {\n return atomicAdd(d_counters + DRAIN, num_items);\n }\n\n\n /// Fill \\p num_items into the queue. Returns offset from which to write items. To be called from CUDA kernel.\n __device__ __forceinline__ OffsetT Fill(OffsetT num_items)\n {\n return atomicAdd(d_counters + FILL, num_items);\n }\n};\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/**\n * Reset grid queue (call with 1 block of 1 thread)\n */\ntemplate \n__global__ void FillAndResetDrainKernel(\n GridQueue grid_queue,\n OffsetT num_items)\n{\n grid_queue.FillAndResetDrain(num_items);\n}\n\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/** @} */ // end group GridModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n\n" }, { "path": "external/cub/cub/host/mutex.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Simple portable mutex\n */\n\n\n#pragma once\n\n#if (__cplusplus > 199711L) || (defined(_MSC_VER) && _MSC_VER >= 1800)\n #include \n#else\n #if defined(_WIN32) || defined(_WIN64)\n #include \n\n #define WIN32_LEAN_AND_MEAN\n #define NOMINMAX\n #include \n #undef WIN32_LEAN_AND_MEAN\n #undef NOMINMAX\n\n /**\n * Compiler read/write barrier\n */\n #pragma intrinsic(_ReadWriteBarrier)\n\n #endif\n#endif\n\n#include \"../util_namespace.cuh\"\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * Simple portable mutex\n * - Wraps std::mutex when compiled with C++11 or newer (supported on all platforms)\n * - Uses GNU/Windows spinlock mechanisms for pre C++11 (supported on x86/x64 when compiled with cl.exe or g++)\n */\nstruct Mutex\n{\n#if (__cplusplus > 199711L) || (defined(_MSC_VER) && _MSC_VER >= 1800)\n\n std::mutex mtx;\n\n void Lock()\n {\n mtx.lock();\n }\n\n void Unlock()\n {\n mtx.unlock();\n }\n\n void TryLock()\n {\n mtx.try_lock();\n }\n\n#else //__cplusplus > 199711L\n\n #if defined(_MSC_VER)\n\n // Microsoft VC++\n typedef long Spinlock;\n\n #else\n\n // GNU g++\n typedef int Spinlock;\n\n /**\n * Compiler read/write barrier\n */\n __forceinline__ void _ReadWriteBarrier()\n {\n __sync_synchronize();\n }\n\n /**\n * Atomic exchange\n */\n __forceinline__ long _InterlockedExchange(volatile int * const Target, const int Value)\n {\n // NOTE: __sync_lock_test_and_set would be an acquire barrier, so we force a full barrier\n _ReadWriteBarrier();\n return __sync_lock_test_and_set(Target, Value);\n }\n\n /**\n * Pause instruction to prevent excess processor bus usage\n */\n __forceinline__ void YieldProcessor()\n {\n }\n\n #endif // defined(_MSC_VER)\n\n /// Lock member\n volatile Spinlock lock;\n\n /**\n * Constructor\n */\n Mutex() : lock(0) {}\n\n /**\n * Return when the specified spinlock has been acquired\n */\n __forceinline__ void Lock()\n {\n while (1)\n {\n if (!_InterlockedExchange(&lock, 1)) return;\n while (lock) YieldProcessor();\n }\n }\n\n\n /**\n * Release the specified spinlock\n */\n __forceinline__ void Unlock()\n {\n _ReadWriteBarrier();\n lock = 0;\n }\n\n#endif // __cplusplus > 199711L\n\n};\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n" }, { "path": "external/cub/cub/iterator/arg_index_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n#include \n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A random-access input wrapper for pairing dereferenced values with their corresponding indices (forming \\p KeyValuePair tuples).\n *\n * \\par Overview\n * - ArgIndexInputIteratorTwraps a random access input iterator \\p itr of type \\p InputIteratorT.\n * Dereferencing an ArgIndexInputIteratorTat offset \\p i produces a \\p KeyValuePair value whose\n * \\p key field is \\p i and whose \\p value field is itr[i].\n * - Can be used with any data type.\n * - Can be constructed, manipulated, and exchanged within and between host and device\n * functions. Wrapped host memory can only be dereferenced on the host, and wrapped\n * device memory can only be dereferenced on the device.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p ArgIndexInputIteratorTto\n * dereference an array of doubles\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize a device array\n * double *d_in; // e.g., [8.0, 6.0, 7.0, 5.0, 3.0, 0.0, 9.0]\n *\n * // Create an iterator wrapper\n * cub::ArgIndexInputIterator itr(d_in);\n *\n * // Within device code:\n * typedef typename cub::ArgIndexInputIterator::value_type Tuple;\n * Tuple item_offset_pair.key = *itr;\n * printf(\"%f @ %d\\n\",\n * item_offset_pair.value,\n * item_offset_pair.key); // 8.0 @ 0\n *\n * itr = itr + 6;\n * item_offset_pair.key = *itr;\n * printf(\"%f @ %d\\n\",\n * item_offset_pair.value,\n * item_offset_pair.key); // 9.0 @ 6\n *\n * \\endcode\n *\n * \\tparam InputIteratorT The value type of the wrapped input iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n * \\tparam OutputValueT The paired value type of the tuple (Default: value type of input iterator)\n */\ntemplate <\n typename InputIteratorT,\n typename OffsetT = ptrdiff_t,\n typename OutputValueT = typename std::iterator_traits::value_type>\nclass ArgIndexInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef ArgIndexInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef KeyValuePair value_type; ///< The type of the element the iterator can point to\n typedef value_type* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef value_type reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::any_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n InputIteratorT itr;\n difference_type offset;\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ ArgIndexInputIterator(\n InputIteratorT itr, ///< Input iterator to wrap\n difference_type offset = 0) ///< OffsetT (in items) from \\p itr denoting the position of the iterator\n :\n itr(itr),\n offset(offset)\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n offset++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n offset++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n value_type retval;\n retval.value = itr[offset];\n retval.key = offset;\n return retval;\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(itr, offset + n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n offset += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(itr, offset - n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n offset -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return offset - other.offset;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n self_type offset = (*this) + n;\n return *offset;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &(*(*this));\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return ((itr == rhs.itr) && (offset == rhs.offset));\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return ((itr != rhs.itr) || (offset != rhs.offset));\n }\n\n /// Normalize\n __host__ __device__ __forceinline__ void normalize()\n {\n itr += offset;\n offset = 0;\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& /*itr*/)\n {\n return os;\n }\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/cache_modified_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A random-access input wrapper for dereferencing array values using a PTX cache load modifier.\n *\n * \\par Overview\n * - CacheModifiedInputIteratorTis a random-access input iterator that wraps a native\n * device pointer of type ValueType*. \\p ValueType references are\n * made by reading \\p ValueType values through loads modified by \\p MODIFIER.\n * - Can be used to load any data type from memory using PTX cache load modifiers (e.g., \"LOAD_LDG\",\n * \"LOAD_CG\", \"LOAD_CA\", \"LOAD_CS\", \"LOAD_CV\", etc.).\n * - Can be constructed, manipulated, and exchanged within and between host and device\n * functions, but can only be dereferenced within device functions.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p CacheModifiedInputIteratorTto\n * dereference a device array of double using the \"ldg\" PTX load modifier\n * (i.e., load values through texture cache).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize a device array\n * double *d_in; // e.g., [8.0, 6.0, 7.0, 5.0, 3.0, 0.0, 9.0]\n *\n * // Create an iterator wrapper\n * cub::CacheModifiedInputIterator itr(d_in);\n *\n * // Within device code:\n * printf(\"%f\\n\", itr[0]); // 8.0\n * printf(\"%f\\n\", itr[1]); // 6.0\n * printf(\"%f\\n\", itr[6]); // 9.0\n *\n * \\endcode\n *\n * \\tparam CacheLoadModifier The cub::CacheLoadModifier to use when accessing data\n * \\tparam ValueType The value type of this iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n CacheLoadModifier MODIFIER,\n typename ValueType,\n typename OffsetT = ptrdiff_t>\nclass CacheModifiedInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef CacheModifiedInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef ValueType value_type; ///< The type of the element the iterator can point to\n typedef ValueType* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef ValueType reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::device_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\n\npublic:\n\n /// Wrapped native pointer\n ValueType* ptr;\n\n /// Constructor\n template \n __host__ __device__ __forceinline__ CacheModifiedInputIterator(\n QualifiedValueType* ptr) ///< Native pointer to wrap\n :\n ptr(const_cast::Type *>(ptr))\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n ptr++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n ptr++;\n return *this;\n }\n\n /// Indirection\n __device__ __forceinline__ reference operator*() const\n {\n return ThreadLoad(ptr);\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(ptr + n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n ptr += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(ptr - n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n ptr -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return ptr - other.ptr;\n }\n\n /// Array subscript\n template \n __device__ __forceinline__ reference operator[](Distance n) const\n {\n return ThreadLoad(ptr + n);\n }\n\n /// Structure dereference\n __device__ __forceinline__ pointer operator->()\n {\n return &ThreadLoad(ptr);\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (ptr == rhs.ptr);\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (ptr != rhs.ptr);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& /*itr*/)\n {\n return os;\n }\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/cache_modified_output_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A random-access output wrapper for storing array values using a PTX cache-modifier.\n *\n * \\par Overview\n * - CacheModifiedOutputIterator is a random-access output iterator that wraps a native\n * device pointer of type ValueType*. \\p ValueType references are\n * made by writing \\p ValueType values through stores modified by \\p MODIFIER.\n * - Can be used to store any data type to memory using PTX cache store modifiers (e.g., \"STORE_WB\",\n * \"STORE_CG\", \"STORE_CS\", \"STORE_WT\", etc.).\n * - Can be constructed, manipulated, and exchanged within and between host and device\n * functions, but can only be dereferenced within device functions.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p CacheModifiedOutputIterator to\n * dereference a device array of doubles using the \"wt\" PTX load modifier\n * (i.e., write-through to system memory).\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize a device array\n * double *d_out; // e.g., [, , , , , , ]\n *\n * // Create an iterator wrapper\n * cub::CacheModifiedOutputIterator itr(d_out);\n *\n * // Within device code:\n * itr[0] = 8.0;\n * itr[1] = 66.0;\n * itr[55] = 24.0;\n *\n * \\endcode\n *\n * \\par Usage Considerations\n * - Can only be dereferenced within device code\n *\n * \\tparam CacheStoreModifier The cub::CacheStoreModifier to use when accessing data\n * \\tparam ValueType The value type of this iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n CacheStoreModifier MODIFIER,\n typename ValueType,\n typename OffsetT = ptrdiff_t>\nclass CacheModifiedOutputIterator\n{\nprivate:\n\n // Proxy object\n struct Reference\n {\n ValueType* ptr;\n\n /// Constructor\n __host__ __device__ __forceinline__ Reference(ValueType* ptr) : ptr(ptr) {}\n\n /// Assignment\n __device__ __forceinline__ ValueType operator =(ValueType val)\n {\n ThreadStore(ptr, val);\n return val;\n }\n };\n\npublic:\n\n // Required iterator traits\n typedef CacheModifiedOutputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef void value_type; ///< The type of the element the iterator can point to\n typedef void pointer; ///< The type of a pointer to an element the iterator can point to\n typedef Reference reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::device_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n ValueType* ptr;\n\npublic:\n\n /// Constructor\n template \n __host__ __device__ __forceinline__ CacheModifiedOutputIterator(\n QualifiedValueType* ptr) ///< Native pointer to wrap\n :\n ptr(const_cast::Type *>(ptr))\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n ptr++;\n return retval;\n }\n\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n ptr++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n return Reference(ptr);\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(ptr + n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n ptr += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(ptr - n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n ptr -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return ptr - other.ptr;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n return Reference(ptr + n);\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (ptr == rhs.ptr);\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (ptr != rhs.ptr);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n return os;\n }\n};\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/constant_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A random-access input generator for dereferencing a sequence of homogeneous values\n *\n * \\par Overview\n * - Read references to a ConstantInputIteratorTiterator always return the supplied constant\n * of type \\p ValueType.\n * - Can be used with any data type.\n * - Can be constructed, manipulated, dereferenced, and exchanged within and between host and device\n * functions.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p ConstantInputIteratorTto\n * dereference a sequence of homogeneous doubles.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * cub::ConstantInputIterator itr(5.0);\n *\n * printf(\"%f\\n\", itr[0]); // 5.0\n * printf(\"%f\\n\", itr[1]); // 5.0\n * printf(\"%f\\n\", itr[2]); // 5.0\n * printf(\"%f\\n\", itr[50]); // 5.0\n *\n * \\endcode\n *\n * \\tparam ValueType The value type of this iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n typename ValueType,\n typename OffsetT = ptrdiff_t>\nclass ConstantInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef ConstantInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef ValueType value_type; ///< The type of the element the iterator can point to\n typedef ValueType* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef ValueType reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::any_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n ValueType val;\n OffsetT offset;\n#ifdef _WIN32\n OffsetT pad[CUB_MAX(1, (16 / sizeof(OffsetT) - 1))]; // Workaround for win32 parameter-passing bug (ulonglong2 argmin DeviceReduce)\n#endif\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ ConstantInputIterator(\n ValueType val, ///< Starting value for the iterator instance to report\n OffsetT offset = 0) ///< Base offset\n :\n val(val),\n offset(offset)\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n offset++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n offset++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n return val;\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(val, offset + n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n offset += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(val, offset - n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n offset -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return offset - other.offset;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance /*n*/) const\n {\n return val;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &val;\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (offset == rhs.offset) && ((val == rhs.val));\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (offset != rhs.offset) || (val!= rhs.val);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n os << \"[\" << itr.val << \",\" << itr.offset << \"]\";\n return os;\n }\n\n};\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/counting_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n/**\n * \\brief A random-access input generator for dereferencing a sequence of incrementing integer values.\n *\n * \\par Overview\n * - After initializing a CountingInputIteratorTto a certain integer \\p base, read references\n * at \\p offset will return the value \\p base + \\p offset.\n * - Can be constructed, manipulated, dereferenced, and exchanged within and between host and device\n * functions.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p CountingInputIteratorTto\n * dereference a sequence of incrementing integers.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * cub::CountingInputIterator itr(5);\n *\n * printf(\"%d\\n\", itr[0]); // 5\n * printf(\"%d\\n\", itr[1]); // 6\n * printf(\"%d\\n\", itr[2]); // 7\n * printf(\"%d\\n\", itr[50]); // 55\n *\n * \\endcode\n *\n * \\tparam ValueType The value type of this iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n typename ValueType,\n typename OffsetT = ptrdiff_t>\nclass CountingInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef CountingInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef ValueType value_type; ///< The type of the element the iterator can point to\n typedef ValueType* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef ValueType reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::any_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n ValueType val;\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ CountingInputIterator(\n const ValueType &val) ///< Starting value for the iterator instance to report\n :\n val(val)\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n val++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n val++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n return val;\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(val + (ValueType) n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n val += (ValueType) n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(val - (ValueType) n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n val -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return (difference_type) (val - other.val);\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n return val + (ValueType) n;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &val;\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (val == rhs.val);\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (val != rhs.val);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n os << \"[\" << itr.val << \"]\";\n return os;\n }\n\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/discard_output_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../util_namespace.cuh\"\n#include \"../util_macro.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A discard iterator\n */\ntemplate \nclass DiscardOutputIterator\n{\npublic:\n\n // Required iterator traits\n typedef DiscardOutputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef void value_type; ///< The type of the element the iterator can point to\n typedef void pointer; ///< The type of a pointer to an element the iterator can point to\n typedef void reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::any_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n OffsetT offset;\n\n#if defined(_WIN32) || !defined(_WIN64)\n // Workaround for win32 parameter-passing bug (ulonglong2 argmin DeviceReduce)\n OffsetT pad[CUB_MAX(1, (16 / sizeof(OffsetT) - 1))];\n#endif\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ DiscardOutputIterator(\n OffsetT offset = 0) ///< Base offset\n :\n offset(offset)\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n offset++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n offset++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ self_type& operator*()\n {\n // return self reference, which can be assigned to anything\n return *this;\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(offset + n);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n offset += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(offset - n);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n offset -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return offset - other.offset;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ self_type& operator[](Distance n)\n {\n // return self reference, which can be assigned to anything\n return *this;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return;\n }\n\n /// Assignment to self (no-op)\n __host__ __device__ __forceinline__ void operator=(self_type const& other)\n {\n offset = other.offset;\n }\n\n /// Assignment to anything else (no-op)\n template\n __host__ __device__ __forceinline__ void operator=(T const&)\n {}\n\n /// Cast to void* operator\n __host__ __device__ __forceinline__ operator void*() const { return NULL; }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (offset == rhs.offset);\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (offset != rhs.offset);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n os << \"[\" << itr.offset << \"]\";\n return os;\n }\n\n};\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/tex_obj_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_debug.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n\n/**\n * \\brief A random-access input wrapper for dereferencing array values through texture cache. Uses newer Kepler-style texture objects.\n *\n * \\par Overview\n * - TexObjInputIteratorTwraps a native device pointer of type ValueType*. References\n * to elements are to be loaded through texture cache.\n * - Can be used to load any data type from memory through texture cache.\n * - Can be manipulated and exchanged within and between host and device\n * functions, can only be constructed within host functions, and can only be\n * dereferenced within device functions.\n * - With regard to nested/dynamic parallelism, TexObjInputIteratorTiterators may only be\n * created by the host thread, but can be used by any descendant kernel.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p TexRefInputIteratorTto\n * dereference a device array of doubles through texture cache.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize a device array\n * int num_items; // e.g., 7\n * double *d_in; // e.g., [8.0, 6.0, 7.0, 5.0, 3.0, 0.0, 9.0]\n *\n * // Create an iterator wrapper\n * cub::TexObjInputIterator itr;\n * itr.BindTexture(d_in, sizeof(double) * num_items);\n * ...\n *\n * // Within device code:\n * printf(\"%f\\n\", itr[0]); // 8.0\n * printf(\"%f\\n\", itr[1]); // 6.0\n * printf(\"%f\\n\", itr[6]); // 9.0\n *\n * ...\n * itr.UnbindTexture();\n *\n * \\endcode\n *\n * \\tparam T The value type of this iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n typename T,\n typename OffsetT = ptrdiff_t>\nclass TexObjInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef TexObjInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef T value_type; ///< The type of the element the iterator can point to\n typedef T* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef T reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::device_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n // Largest texture word we can use in device\n typedef typename UnitWord::TextureWord TextureWord;\n\n // Number of texture words per T\n enum {\n TEXTURE_MULTIPLE = sizeof(T) / sizeof(TextureWord)\n };\n\nprivate:\n\n T* ptr;\n difference_type tex_offset;\n cudaTextureObject_t tex_obj;\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ TexObjInputIterator()\n :\n ptr(NULL),\n tex_offset(0),\n tex_obj(0)\n {}\n\n /// Use this iterator to bind \\p ptr with a texture reference\n template \n cudaError_t BindTexture(\n QualifiedT *ptr, ///< Native pointer to wrap that is aligned to cudaDeviceProp::textureAlignment\n size_t bytes = size_t(-1), ///< Number of bytes in the range\n size_t tex_offset = 0) ///< OffsetT (in items) from \\p ptr denoting the position of the iterator\n {\n this->ptr = const_cast::Type *>(ptr);\n this->tex_offset = tex_offset;\n\n cudaChannelFormatDesc channel_desc = cudaCreateChannelDesc();\n cudaResourceDesc res_desc;\n cudaTextureDesc tex_desc;\n memset(&res_desc, 0, sizeof(cudaResourceDesc));\n memset(&tex_desc, 0, sizeof(cudaTextureDesc));\n res_desc.resType = cudaResourceTypeLinear;\n res_desc.res.linear.devPtr = this->ptr;\n res_desc.res.linear.desc = channel_desc;\n res_desc.res.linear.sizeInBytes = bytes;\n tex_desc.readMode = cudaReadModeElementType;\n return cudaCreateTextureObject(&tex_obj, &res_desc, &tex_desc, NULL);\n }\n\n /// Unbind this iterator from its texture reference\n cudaError_t UnbindTexture()\n {\n return cudaDestroyTextureObject(tex_obj);\n }\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n tex_offset++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n tex_offset++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n#if (CUB_PTX_ARCH == 0)\n // Simply dereference the pointer on the host\n return ptr[tex_offset];\n#else\n // Move array of uninitialized words, then alias and assign to return value\n TextureWord words[TEXTURE_MULTIPLE];\n\n #pragma unroll\n for (int i = 0; i < TEXTURE_MULTIPLE; ++i)\n {\n words[i] = tex1Dfetch(\n tex_obj,\n (tex_offset * TEXTURE_MULTIPLE) + i);\n }\n\n // Load from words\n return *reinterpret_cast(words);\n#endif\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval;\n retval.ptr = ptr;\n retval.tex_obj = tex_obj;\n retval.tex_offset = tex_offset + n;\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n tex_offset += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval;\n retval.ptr = ptr;\n retval.tex_obj = tex_obj;\n retval.tex_offset = tex_offset - n;\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n tex_offset -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return tex_offset - other.tex_offset;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n self_type offset = (*this) + n;\n return *offset;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &(*(*this));\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return ((ptr == rhs.ptr) && (tex_offset == rhs.tex_offset) && (tex_obj == rhs.tex_obj));\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return ((ptr != rhs.ptr) || (tex_offset != rhs.tex_offset) || (tex_obj != rhs.tex_obj));\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n return os;\n }\n\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/iterator/tex_ref_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_debug.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (CUDA_VERSION >= 5050) || defined(DOXYGEN_ACTIVE) // This iterator is compatible with CUDA 5.5 and newer\n\n#if (THRUST_VERSION >= 100700) // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/******************************************************************************\n * Static file-scope Tesla/Fermi-style texture references\n *****************************************************************************/\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n// Anonymous namespace\nnamespace {\n\n/// Global texture reference specialized by type\ntemplate \nstruct IteratorTexRef\n{\n /// And by unique ID\n template \n struct TexId\n {\n // Largest texture word we can use in device\n typedef typename UnitWord::DeviceWord DeviceWord;\n typedef typename UnitWord::TextureWord TextureWord;\n\n // Number of texture words per T\n enum {\n DEVICE_MULTIPLE = sizeof(T) / sizeof(DeviceWord),\n TEXTURE_MULTIPLE = sizeof(T) / sizeof(TextureWord)\n };\n\n // Texture reference type\n typedef texture TexRef;\n\n // Texture reference\n static TexRef ref;\n\n /// Bind texture\n static cudaError_t BindTexture(void *d_in, size_t &offset)\n {\n if (d_in)\n {\n cudaChannelFormatDesc tex_desc = cudaCreateChannelDesc();\n ref.channelDesc = tex_desc;\n return (CubDebug(cudaBindTexture(&offset, ref, d_in)));\n }\n\n return cudaSuccess;\n }\n\n /// Unbind texture\n static cudaError_t UnbindTexture()\n {\n return CubDebug(cudaUnbindTexture(ref));\n }\n\n /// Fetch element\n template \n static __device__ __forceinline__ T Fetch(Distance tex_offset)\n {\n DeviceWord temp[DEVICE_MULTIPLE];\n TextureWord *words = reinterpret_cast(temp);\n\n #pragma unroll\n for (int i = 0; i < TEXTURE_MULTIPLE; ++i)\n {\n words[i] = tex1Dfetch(ref, (tex_offset * TEXTURE_MULTIPLE) + i);\n }\n\n return reinterpret_cast(temp);\n }\n };\n};\n\n// Texture reference definitions\ntemplate \ntemplate \ntypename IteratorTexRef::template TexId::TexRef IteratorTexRef::template TexId::ref = 0;\n\n\n} // Anonymous namespace\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n\n/**\n * \\brief A random-access input wrapper for dereferencing array values through texture cache. Uses older Tesla/Fermi-style texture references.\n *\n * \\par Overview\n * - TexRefInputIteratorTwraps a native device pointer of type ValueType*. References\n * to elements are to be loaded through texture cache.\n * - Can be used to load any data type from memory through texture cache.\n * - Can be manipulated and exchanged within and between host and device\n * functions, can only be constructed within host functions, and can only be\n * dereferenced within device functions.\n * - The \\p UNIQUE_ID template parameter is used to statically name the underlying texture\n * reference. Only one TexRefInputIteratorTinstance can be bound at any given time for a\n * specific combination of (1) data type \\p T, (2) \\p UNIQUE_ID, (3) host\n * thread, and (4) compilation .o unit.\n * - With regard to nested/dynamic parallelism, TexRefInputIteratorTiterators may only be\n * created by the host thread and used by a top-level kernel (i.e. the one which is launched\n * from the host).\n * - Compatible with Thrust API v1.7 or newer.\n * - Compatible with CUDA toolkit v5.5 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p TexRefInputIteratorTto\n * dereference a device array of doubles through texture cache.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Declare, allocate, and initialize a device array\n * int num_items; // e.g., 7\n * double *d_in; // e.g., [8.0, 6.0, 7.0, 5.0, 3.0, 0.0, 9.0]\n *\n * // Create an iterator wrapper\n * cub::TexRefInputIterator itr;\n * itr.BindTexture(d_in, sizeof(double) * num_items);\n * ...\n *\n * // Within device code:\n * printf(\"%f\\n\", itr[0]); // 8.0\n * printf(\"%f\\n\", itr[1]); // 6.0\n * printf(\"%f\\n\", itr[6]); // 9.0\n *\n * ...\n * itr.UnbindTexture();\n *\n * \\endcode\n *\n * \\tparam T The value type of this iterator\n * \\tparam UNIQUE_ID A globally-unique identifier (within the compilation unit) to name the underlying texture reference\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n */\ntemplate <\n typename T,\n int UNIQUE_ID,\n typename OffsetT = ptrdiff_t>\nclass TexRefInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef TexRefInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef T value_type; ///< The type of the element the iterator can point to\n typedef T* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef T reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::device_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n T* ptr;\n difference_type tex_offset;\n\n // Texture reference wrapper (old Tesla/Fermi-style textures)\n typedef typename IteratorTexRef::template TexId TexId;\n\npublic:\n/*\n /// Constructor\n __host__ __device__ __forceinline__ TexRefInputIterator()\n :\n ptr(NULL),\n tex_offset(0)\n {}\n*/\n /// Use this iterator to bind \\p ptr with a texture reference\n template \n cudaError_t BindTexture(\n QualifiedT *ptr, ///< Native pointer to wrap that is aligned to cudaDeviceProp::textureAlignment\n size_t bytes = size_t(-1), ///< Number of bytes in the range\n size_t tex_offset = 0) ///< OffsetT (in items) from \\p ptr denoting the position of the iterator\n {\n this->ptr = const_cast::Type *>(ptr);\n size_t offset;\n cudaError_t retval = TexId::BindTexture(this->ptr + tex_offset, offset);\n this->tex_offset = (difference_type) (offset / sizeof(QualifiedT));\n return retval;\n }\n\n /// Unbind this iterator from its texture reference\n cudaError_t UnbindTexture()\n {\n return TexId::UnbindTexture();\n }\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n tex_offset++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n tex_offset++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n#if (CUB_PTX_ARCH == 0)\n // Simply dereference the pointer on the host\n return ptr[tex_offset];\n#else\n // Use the texture reference\n return TexId::Fetch(tex_offset);\n#endif\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval;\n retval.ptr = ptr;\n retval.tex_offset = tex_offset + n;\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n tex_offset += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval;\n retval.ptr = ptr;\n retval.tex_offset = tex_offset - n;\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n tex_offset -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return tex_offset - other.tex_offset;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n self_type offset = (*this) + n;\n return *offset;\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &(*(*this));\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return ((ptr == rhs.ptr) && (tex_offset == rhs.tex_offset));\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return ((ptr != rhs.ptr) || (tex_offset != rhs.tex_offset));\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n return os;\n }\n\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n\n#endif // CUDA_VERSION\n" }, { "path": "external/cub/cub/iterator/transform_input_iterator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Random-access iterator types\n */\n\n#pragma once\n\n#include \n#include \n\n#include \"../thread/thread_load.cuh\"\n#include \"../thread/thread_store.cuh\"\n#include \"../util_device.cuh\"\n#include \"../util_namespace.cuh\"\n\n#if (THRUST_VERSION >= 100700)\n // This iterator is compatible with Thrust API 1.7 and newer\n #include \n #include \n#endif // THRUST_VERSION\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIterator\n * @{\n */\n\n\n/**\n * \\brief A random-access input wrapper for transforming dereferenced values.\n *\n * \\par Overview\n * - TransformInputIteratorTwraps a unary conversion functor of type \\p\n * ConversionOp and a random-access input iterator of type InputIteratorT,\n * using the former to produce references of type \\p ValueType from the latter.\n * - Can be used with any data type.\n * - Can be constructed, manipulated, and exchanged within and between host and device\n * functions. Wrapped host memory can only be dereferenced on the host, and wrapped\n * device memory can only be dereferenced on the device.\n * - Compatible with Thrust API v1.7 or newer.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of \\p TransformInputIteratorTto\n * dereference an array of integers, tripling the values and converting them to doubles.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * // Functor for tripling integer values and converting to doubles\n * struct TripleDoubler\n * {\n * __host__ __device__ __forceinline__\n * double operator()(const int &a) const {\n * return double(a * 3);\n * }\n * };\n *\n * // Declare, allocate, and initialize a device array\n * int *d_in; // e.g., [8, 6, 7, 5, 3, 0, 9]\n * TripleDoubler conversion_op;\n *\n * // Create an iterator wrapper\n * cub::TransformInputIterator itr(d_in, conversion_op);\n *\n * // Within device code:\n * printf(\"%f\\n\", itr[0]); // 24.0\n * printf(\"%f\\n\", itr[1]); // 18.0\n * printf(\"%f\\n\", itr[6]); // 27.0\n *\n * \\endcode\n *\n * \\tparam ValueType The value type of this iterator\n * \\tparam ConversionOp Unary functor type for mapping objects of type \\p InputType to type \\p ValueType. Must have member ValueType operator()(const InputType &datum).\n * \\tparam InputIteratorT The type of the wrapped input iterator\n * \\tparam OffsetT The difference type of this iterator (Default: \\p ptrdiff_t)\n *\n */\ntemplate <\n typename ValueType,\n typename ConversionOp,\n typename InputIteratorT,\n typename OffsetT = ptrdiff_t>\nclass TransformInputIterator\n{\npublic:\n\n // Required iterator traits\n typedef TransformInputIterator self_type; ///< My own type\n typedef OffsetT difference_type; ///< Type to express the result of subtracting one iterator from another\n typedef ValueType value_type; ///< The type of the element the iterator can point to\n typedef ValueType* pointer; ///< The type of a pointer to an element the iterator can point to\n typedef ValueType reference; ///< The type of a reference to an element the iterator can point to\n\n#if (THRUST_VERSION >= 100700)\n // Use Thrust's iterator categories so we can use these iterators in Thrust 1.7 (or newer) methods\n typedef typename thrust::detail::iterator_facade_category<\n thrust::any_system_tag,\n thrust::random_access_traversal_tag,\n value_type,\n reference\n >::type iterator_category; ///< The iterator category\n#else\n typedef std::random_access_iterator_tag iterator_category; ///< The iterator category\n#endif // THRUST_VERSION\n\nprivate:\n\n ConversionOp conversion_op;\n InputIteratorT input_itr;\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__ TransformInputIterator(\n InputIteratorT input_itr, ///< Input iterator to wrap\n ConversionOp conversion_op) ///< Conversion functor to wrap\n :\n conversion_op(conversion_op),\n input_itr(input_itr)\n {}\n\n /// Postfix increment\n __host__ __device__ __forceinline__ self_type operator++(int)\n {\n self_type retval = *this;\n input_itr++;\n return retval;\n }\n\n /// Prefix increment\n __host__ __device__ __forceinline__ self_type operator++()\n {\n input_itr++;\n return *this;\n }\n\n /// Indirection\n __host__ __device__ __forceinline__ reference operator*() const\n {\n return conversion_op(*input_itr);\n }\n\n /// Addition\n template \n __host__ __device__ __forceinline__ self_type operator+(Distance n) const\n {\n self_type retval(input_itr + n, conversion_op);\n return retval;\n }\n\n /// Addition assignment\n template \n __host__ __device__ __forceinline__ self_type& operator+=(Distance n)\n {\n input_itr += n;\n return *this;\n }\n\n /// Subtraction\n template \n __host__ __device__ __forceinline__ self_type operator-(Distance n) const\n {\n self_type retval(input_itr - n, conversion_op);\n return retval;\n }\n\n /// Subtraction assignment\n template \n __host__ __device__ __forceinline__ self_type& operator-=(Distance n)\n {\n input_itr -= n;\n return *this;\n }\n\n /// Distance\n __host__ __device__ __forceinline__ difference_type operator-(self_type other) const\n {\n return input_itr - other.input_itr;\n }\n\n /// Array subscript\n template \n __host__ __device__ __forceinline__ reference operator[](Distance n) const\n {\n return conversion_op(input_itr[n]);\n }\n\n /// Structure dereference\n __host__ __device__ __forceinline__ pointer operator->()\n {\n return &conversion_op(*input_itr);\n }\n\n /// Equal to\n __host__ __device__ __forceinline__ bool operator==(const self_type& rhs)\n {\n return (input_itr == rhs.input_itr);\n }\n\n /// Not equal to\n __host__ __device__ __forceinline__ bool operator!=(const self_type& rhs)\n {\n return (input_itr != rhs.input_itr);\n }\n\n /// ostream operator\n friend std::ostream& operator<<(std::ostream& os, const self_type& itr)\n {\n return os;\n }\n};\n\n\n\n/** @} */ // end group UtilIterator\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_load.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Thread utilities for reading memory using PTX cache modifiers.\n */\n\n#pragma once\n\n#include \n\n#include \n\n#include \"../util_ptx.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIo\n * @{\n */\n\n//-----------------------------------------------------------------------------\n// Tags and constants\n//-----------------------------------------------------------------------------\n\n/**\n * \\brief Enumeration of cache modifiers for memory load operations.\n */\nenum CacheLoadModifier\n{\n LOAD_DEFAULT, ///< Default (no modifier)\n LOAD_CA, ///< Cache at all levels\n LOAD_CG, ///< Cache at global level\n LOAD_CS, ///< Cache streaming (likely to be accessed once)\n LOAD_CV, ///< Cache as volatile (including cached system lines)\n LOAD_LDG, ///< Cache as texture\n LOAD_VOLATILE, ///< Volatile (any memory space)\n};\n\n\n/**\n * \\name Thread I/O (cache modified)\n * @{\n */\n\n/**\n * \\brief Thread utility for reading memory using cub::CacheLoadModifier cache modifiers. Can be used to load any data type.\n *\n * \\par Example\n * \\code\n * #include // or equivalently \n *\n * // 32-bit load using cache-global modifier:\n * int *d_in;\n * int val = cub::ThreadLoad(d_in + threadIdx.x);\n *\n * // 16-bit load using default modifier\n * short *d_in;\n * short val = cub::ThreadLoad(d_in + threadIdx.x);\n *\n * // 256-bit load using cache-volatile modifier\n * double4 *d_in;\n * double4 val = cub::ThreadLoad(d_in + threadIdx.x);\n *\n * // 96-bit load using cache-streaming modifier\n * struct TestFoo { bool a; short b; };\n * TestFoo *d_struct;\n * TestFoo val = cub::ThreadLoad(d_in + threadIdx.x);\n * \\endcode\n *\n * \\tparam MODIFIER [inferred] CacheLoadModifier enumeration\n * \\tparam InputIteratorT [inferred] Input iterator type \\iterator\n */\ntemplate <\n CacheLoadModifier MODIFIER,\n typename InputIteratorT>\n__device__ __forceinline__ typename std::iterator_traits::value_type ThreadLoad(InputIteratorT itr);\n\n\n//@} end member group\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/// Helper structure for templated load iteration (inductive case)\ntemplate \nstruct IterateThreadLoad\n{\n template \n static __device__ __forceinline__ void Load(T const *ptr, T *vals)\n {\n vals[COUNT] = ThreadLoad(ptr + COUNT);\n IterateThreadLoad::template Load(ptr, vals);\n }\n\n template \n static __device__ __forceinline__ void Dereference(InputIteratorT itr, T *vals)\n {\n vals[COUNT] = itr[COUNT];\n IterateThreadLoad::Dereference(itr, vals);\n }\n};\n\n\n/// Helper structure for templated load iteration (termination case)\ntemplate \nstruct IterateThreadLoad\n{\n template \n static __device__ __forceinline__ void Load(T const * /*ptr*/, T * /*vals*/) {}\n\n template \n static __device__ __forceinline__ void Dereference(InputIteratorT /*itr*/, T * /*vals*/) {}\n};\n\n\n/**\n * Define a uint4 (16B) ThreadLoad specialization for the given Cache load modifier\n */\n#define _CUB_LOAD_16(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ uint4 ThreadLoad(uint4 const *ptr) \\\n { \\\n uint4 retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".v4.u32 {%0, %1, %2, %3}, [%4];\" : \\\n \"=r\"(retval.x), \\\n \"=r\"(retval.y), \\\n \"=r\"(retval.z), \\\n \"=r\"(retval.w) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n } \\\n template<> \\\n __device__ __forceinline__ ulonglong2 ThreadLoad(ulonglong2 const *ptr) \\\n { \\\n ulonglong2 retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".v2.u64 {%0, %1}, [%2];\" : \\\n \"=l\"(retval.x), \\\n \"=l\"(retval.y) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n }\n\n/**\n * Define a uint2 (8B) ThreadLoad specialization for the given Cache load modifier\n */\n#define _CUB_LOAD_8(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ ushort4 ThreadLoad(ushort4 const *ptr) \\\n { \\\n ushort4 retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".v4.u16 {%0, %1, %2, %3}, [%4];\" : \\\n \"=h\"(retval.x), \\\n \"=h\"(retval.y), \\\n \"=h\"(retval.z), \\\n \"=h\"(retval.w) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n } \\\n template<> \\\n __device__ __forceinline__ uint2 ThreadLoad(uint2 const *ptr) \\\n { \\\n uint2 retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".v2.u32 {%0, %1}, [%2];\" : \\\n \"=r\"(retval.x), \\\n \"=r\"(retval.y) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n } \\\n template<> \\\n __device__ __forceinline__ unsigned long long ThreadLoad(unsigned long long const *ptr) \\\n { \\\n unsigned long long retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".u64 %0, [%1];\" : \\\n \"=l\"(retval) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n }\n\n/**\n * Define a uint (4B) ThreadLoad specialization for the given Cache load modifier\n */\n#define _CUB_LOAD_4(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ unsigned int ThreadLoad(unsigned int const *ptr) \\\n { \\\n unsigned int retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".u32 %0, [%1];\" : \\\n \"=r\"(retval) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n }\n\n\n/**\n * Define a unsigned short (2B) ThreadLoad specialization for the given Cache load modifier\n */\n#define _CUB_LOAD_2(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ unsigned short ThreadLoad(unsigned short const *ptr) \\\n { \\\n unsigned short retval; \\\n asm volatile (\"ld.\"#ptx_modifier\".u16 %0, [%1];\" : \\\n \"=h\"(retval) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return retval; \\\n }\n\n\n/**\n * Define an unsigned char (1B) ThreadLoad specialization for the given Cache load modifier\n */\n#define _CUB_LOAD_1(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ unsigned char ThreadLoad(unsigned char const *ptr) \\\n { \\\n unsigned short retval; \\\n asm volatile ( \\\n \"{\" \\\n \" .reg .u8 datum;\" \\\n \" ld.\"#ptx_modifier\".u8 datum, [%1];\" \\\n \" cvt.u16.u8 %0, datum;\" \\\n \"}\" : \\\n \"=h\"(retval) : \\\n _CUB_ASM_PTR_(ptr)); \\\n return (unsigned char) retval; \\\n }\n\n\n/**\n * Define powers-of-two ThreadLoad specializations for the given Cache load modifier\n */\n#define _CUB_LOAD_ALL(cub_modifier, ptx_modifier) \\\n _CUB_LOAD_16(cub_modifier, ptx_modifier) \\\n _CUB_LOAD_8(cub_modifier, ptx_modifier) \\\n _CUB_LOAD_4(cub_modifier, ptx_modifier) \\\n _CUB_LOAD_2(cub_modifier, ptx_modifier) \\\n _CUB_LOAD_1(cub_modifier, ptx_modifier) \\\n\n\n/**\n * Define powers-of-two ThreadLoad specializations for the various Cache load modifiers\n */\n#if CUB_PTX_ARCH >= 200\n _CUB_LOAD_ALL(LOAD_CA, ca)\n _CUB_LOAD_ALL(LOAD_CG, cg)\n _CUB_LOAD_ALL(LOAD_CS, cs)\n _CUB_LOAD_ALL(LOAD_CV, cv)\n#else\n _CUB_LOAD_ALL(LOAD_CA, global)\n // Use volatile to ensure coherent reads when this PTX is JIT'd to run on newer architectures with L1\n _CUB_LOAD_ALL(LOAD_CG, volatile.global)\n _CUB_LOAD_ALL(LOAD_CS, global)\n _CUB_LOAD_ALL(LOAD_CV, volatile.global)\n#endif\n\n#if CUB_PTX_ARCH >= 350\n _CUB_LOAD_ALL(LOAD_LDG, global.nc)\n#else\n _CUB_LOAD_ALL(LOAD_LDG, global)\n#endif\n\n\n// Macro cleanup\n#undef _CUB_LOAD_ALL\n#undef _CUB_LOAD_1\n#undef _CUB_LOAD_2\n#undef _CUB_LOAD_4\n#undef _CUB_LOAD_8\n#undef _CUB_LOAD_16\n\n\n\n/**\n * ThreadLoad definition for LOAD_DEFAULT modifier on iterator types\n */\ntemplate \n__device__ __forceinline__ typename std::iterator_traits::value_type ThreadLoad(\n InputIteratorT itr,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n return *itr;\n}\n\n\n/**\n * ThreadLoad definition for LOAD_DEFAULT modifier on pointer types\n */\ntemplate \n__device__ __forceinline__ T ThreadLoad(\n T *ptr,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n return *ptr;\n}\n\n\n/**\n * ThreadLoad definition for LOAD_VOLATILE modifier on primitive pointer types\n */\ntemplate \n__device__ __forceinline__ T ThreadLoadVolatilePointer(\n T *ptr,\n Int2Type /*is_primitive*/)\n{\n T retval = *reinterpret_cast(ptr);\n return retval;\n}\n\n\n/**\n * ThreadLoad definition for LOAD_VOLATILE modifier on non-primitive pointer types\n */\ntemplate \n__device__ __forceinline__ T ThreadLoadVolatilePointer(\n T *ptr,\n Int2Type /*is_primitive*/)\n{\n typedef typename UnitWord::VolatileWord VolatileWord; // Word type for memcopying\n\n const int VOLATILE_MULTIPLE = sizeof(T) / sizeof(VolatileWord);\n/*\n VolatileWord words[VOLATILE_MULTIPLE];\n\n IterateThreadLoad<0, VOLATILE_MULTIPLE>::Dereference(\n reinterpret_cast(ptr),\n words);\n\n return *reinterpret_cast(words);\n*/\n\n T retval;\n VolatileWord *words = reinterpret_cast(&retval);\n IterateThreadLoad<0, VOLATILE_MULTIPLE>::Dereference(\n reinterpret_cast(ptr),\n words);\n return retval;\n}\n\n\n/**\n * ThreadLoad definition for LOAD_VOLATILE modifier on pointer types\n */\ntemplate \n__device__ __forceinline__ T ThreadLoad(\n T *ptr,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n // Apply tags for partial-specialization\n return ThreadLoadVolatilePointer(ptr, Int2Type::PRIMITIVE>());\n}\n\n\n/**\n * ThreadLoad definition for generic modifiers on pointer types\n */\ntemplate \n__device__ __forceinline__ T ThreadLoad(\n T const *ptr,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n typedef typename UnitWord::DeviceWord DeviceWord;\n\n const int DEVICE_MULTIPLE = sizeof(T) / sizeof(DeviceWord);\n\n DeviceWord words[DEVICE_MULTIPLE];\n\n IterateThreadLoad<0, DEVICE_MULTIPLE>::template Load(\n reinterpret_cast(const_cast(ptr)),\n words);\n\n return *reinterpret_cast(words);\n}\n\n\n/**\n * ThreadLoad definition for generic modifiers\n */\ntemplate <\n CacheLoadModifier MODIFIER,\n typename InputIteratorT>\n__device__ __forceinline__ typename std::iterator_traits::value_type ThreadLoad(InputIteratorT itr)\n{\n // Apply tags for partial-specialization\n return ThreadLoad(\n itr,\n Int2Type(),\n Int2Type::VALUE>());\n}\n\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/** @} */ // end group UtilIo\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_operators.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Simple binary operator functor types\n */\n\n/******************************************************************************\n * Simple functor operators\n ******************************************************************************/\n\n#pragma once\n\n#include \"../util_macro.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilModule\n * @{\n */\n\n/**\n * \\brief Default equality functor\n */\nstruct Equality\n{\n /// Boolean equality operator, returns (a == b)\n template \n __host__ __device__ __forceinline__ bool operator()(const T &a, const T &b) const\n {\n return a == b;\n }\n};\n\n\n/**\n * \\brief Default inequality functor\n */\nstruct Inequality\n{\n /// Boolean inequality operator, returns (a != b)\n template \n __host__ __device__ __forceinline__ bool operator()(const T &a, const T &b) const\n {\n return a != b;\n }\n};\n\n\n/**\n * \\brief Inequality functor (wraps equality functor)\n */\ntemplate \nstruct InequalityWrapper\n{\n /// Wrapped equality operator\n EqualityOp op;\n\n /// Constructor\n __host__ __device__ __forceinline__\n InequalityWrapper(EqualityOp op) : op(op) {}\n\n /// Boolean inequality operator, returns (a != b)\n template \n __host__ __device__ __forceinline__ bool operator()(const T &a, const T &b)\n {\n return !op(a, b);\n }\n};\n\n\n/**\n * \\brief Default sum functor\n */\nstruct Sum\n{\n /// Boolean sum operator, returns a + b\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return a + b;\n }\n};\n\n\n/**\n * \\brief Default max functor\n */\nstruct Max\n{\n /// Boolean max operator, returns (a > b) ? a : b\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return CUB_MAX(a, b);\n }\n};\n\n\n/**\n * \\brief Arg max functor (keeps the value and offset of the first occurrence of the larger item)\n */\nstruct ArgMax\n{\n /// Boolean max operator, preferring the item having the smaller offset in case of ties\n template \n __host__ __device__ __forceinline__ KeyValuePair operator()(\n const KeyValuePair &a,\n const KeyValuePair &b) const\n {\n// Mooch BUG (device reduce argmax gk110 3.2 million random fp32)\n// return ((b.value > a.value) || ((a.value == b.value) && (b.key < a.key))) ? b : a;\n\n if ((b.value > a.value) || ((a.value == b.value) && (b.key < a.key)))\n return b;\n return a;\n }\n};\n\n\n/**\n * \\brief Default min functor\n */\nstruct Min\n{\n /// Boolean min operator, returns (a < b) ? a : b\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return CUB_MIN(a, b);\n }\n};\n\n\n/**\n * \\brief Arg min functor (keeps the value and offset of the first occurrence of the smallest item)\n */\nstruct ArgMin\n{\n /// Boolean min operator, preferring the item having the smaller offset in case of ties\n template \n __host__ __device__ __forceinline__ KeyValuePair operator()(\n const KeyValuePair &a,\n const KeyValuePair &b) const\n {\n// Mooch BUG (device reduce argmax gk110 3.2 million random fp32)\n// return ((b.value < a.value) || ((a.value == b.value) && (b.key < a.key))) ? b : a;\n\n if ((b.value < a.value) || ((a.value == b.value) && (b.key < a.key)))\n return b;\n return a;\n }\n};\n\n\n/**\n * \\brief Default cast functor\n */\ntemplate \nstruct CastOp\n{\n /// Cast operator, returns (B) a\n template \n __host__ __device__ __forceinline__ B operator()(const A &a) const\n {\n return (B) a;\n }\n};\n\n\n/**\n * \\brief Binary operator wrapper for switching non-commutative scan arguments\n */\ntemplate \nclass SwizzleScanOp\n{\nprivate:\n\n /// Wrapped scan operator\n ScanOp scan_op;\n\npublic:\n\n /// Constructor\n __host__ __device__ __forceinline__\n SwizzleScanOp(ScanOp scan_op) : scan_op(scan_op) {}\n\n /// Switch the scan arguments\n template \n __host__ __device__ __forceinline__\n T operator()(const T &a, const T &b)\n {\n T _a(a);\n T _b(b);\n\n return scan_op(_b, _a);\n }\n};\n\n\n/**\n * \\brief Reduce-by-segment functor.\n *\n * Given two cub::KeyValuePair inputs \\p a and \\p b and a\n * binary associative combining operator \\p f(const T &x, const T &y),\n * an instance of this functor returns a cub::KeyValuePair whose \\p key\n * field is a.key + b.key, and whose \\p value field\n * is either b.value if b.key is non-zero, or f(a.value, b.value) otherwise.\n *\n * ReduceBySegmentOp is an associative, non-commutative binary combining operator\n * for input sequences of cub::KeyValuePair pairings. Such\n * sequences are typically used to represent a segmented set of values to be reduced\n * and a corresponding set of {0,1}-valued integer \"head flags\" demarcating the\n * first value of each segment.\n *\n */\ntemplate ///< Binary reduction operator to apply to values\nstruct ReduceBySegmentOp\n{\n /// Wrapped reduction operator\n ReductionOpT op;\n\n /// Constructor\n __host__ __device__ __forceinline__ ReduceBySegmentOp() {}\n\n /// Constructor\n __host__ __device__ __forceinline__ ReduceBySegmentOp(ReductionOpT op) : op(op) {}\n\n /// Scan operator\n template ///< KeyValuePair pairing of T (value) and OffsetT (head flag)\n __host__ __device__ __forceinline__ KeyValuePairT operator()(\n const KeyValuePairT &first, ///< First partial reduction\n const KeyValuePairT &second) ///< Second partial reduction\n {\n KeyValuePairT retval;\n retval.key = first.key + second.key;\n retval.value = (second.key) ?\n second.value : // The second partial reduction spans a segment reset, so it's value aggregate becomes the running aggregate\n op(first.value, second.value); // The second partial reduction does not span a reset, so accumulate both into the running aggregate\n return retval;\n }\n};\n\n\n\ntemplate ///< Binary reduction operator to apply to values\nstruct ReduceByKeyOp\n{\n /// Wrapped reduction operator\n ReductionOpT op;\n\n /// Constructor\n __host__ __device__ __forceinline__ ReduceByKeyOp() {}\n\n /// Constructor\n __host__ __device__ __forceinline__ ReduceByKeyOp(ReductionOpT op) : op(op) {}\n\n /// Scan operator\n template \n __host__ __device__ __forceinline__ KeyValuePairT operator()(\n const KeyValuePairT &first, ///< First partial reduction\n const KeyValuePairT &second) ///< Second partial reduction\n {\n KeyValuePairT retval = second;\n\n if (first.key == second.key)\n retval.value = op(first.value, retval.value);\n\n return retval;\n }\n};\n\n\n\n\n\n\n\n/** @} */ // end group UtilModule\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_reduce.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Thread utilities for sequential reduction over statically-sized array types\n */\n\n#pragma once\n\n#include \"../thread/thread_operators.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/// Internal namespace (to prevent ADL mishaps between static functions when mixing different CUB installations)\nnamespace internal {\n\n/**\n * Sequential reduction over statically-sized array types\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ReductionOp>\n__device__ __forceinline__ T ThreadReduce(\n T* input, ///< [in] Input array\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n T prefix, ///< [in] Prefix to seed reduction with\n Int2Type /*length*/)\n{\n T retval = prefix;\n\n #pragma unroll\n for (int i = 0; i < LENGTH; ++i)\n retval = reduction_op(retval, input[i]);\n\n return retval;\n}\n\n\n/**\n * \\brief Perform a sequential reduction over \\p LENGTH elements of the \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH LengthT of input array\n * \\tparam T [inferred] The data type to be reduced.\n * \\tparam ScanOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ReductionOp>\n__device__ __forceinline__ T ThreadReduce(\n T* input, ///< [in] Input array\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n T prefix) ///< [in] Prefix to seed reduction with\n{\n return ThreadReduce(input, reduction_op, prefix, Int2Type());\n}\n\n\n/**\n * \\brief Perform a sequential reduction over \\p LENGTH elements of the \\p input array. The aggregate is returned.\n *\n * \\tparam LENGTH LengthT of input array\n * \\tparam T [inferred] The data type to be reduced.\n * \\tparam ScanOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ReductionOp>\n__device__ __forceinline__ T ThreadReduce(\n T* input, ///< [in] Input array\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n{\n T prefix = input[0];\n return ThreadReduce(input + 1, reduction_op, prefix);\n}\n\n\n/**\n * \\brief Perform a sequential reduction over the statically-sized \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH [inferred] LengthT of \\p input array\n * \\tparam T [inferred] The data type to be reduced.\n * \\tparam ScanOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ReductionOp>\n__device__ __forceinline__ T ThreadReduce(\n T (&input)[LENGTH], ///< [in] Input array\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n T prefix) ///< [in] Prefix to seed reduction with\n{\n return ThreadReduce(input, reduction_op, prefix, Int2Type());\n}\n\n\n/**\n * \\brief Serial reduction with the specified operator\n *\n * \\tparam LENGTH [inferred] LengthT of \\p input array\n * \\tparam T [inferred] The data type to be reduced.\n * \\tparam ScanOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ReductionOp>\n__device__ __forceinline__ T ThreadReduce(\n T (&input)[LENGTH], ///< [in] Input array\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n{\n return ThreadReduce((T*) input, reduction_op);\n}\n\n\n} // internal namespace\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_scan.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Thread utilities for sequential prefix scan over statically-sized array types\n */\n\n#pragma once\n\n#include \"../thread/thread_operators.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/// Internal namespace (to prevent ADL mishaps between static functions when mixing different CUB installations)\nnamespace internal {\n\n\n/**\n * \\addtogroup UtilModule\n * @{\n */\n\n/**\n * \\name Sequential prefix scan over statically-sized array types\n * @{\n */\n\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanExclusive(\n T inclusive,\n T exclusive,\n T *input, ///< [in] Input array\n T *output, ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n Int2Type /*length*/)\n{\n #pragma unroll\n for (int i = 0; i < LENGTH; ++i)\n {\n inclusive = scan_op(exclusive, input[i]);\n output[i] = exclusive;\n exclusive = inclusive;\n }\n\n return inclusive;\n}\n\n\n\n/**\n * \\brief Perform a sequential exclusive prefix scan over \\p LENGTH elements of the \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanExclusive(\n T *input, ///< [in] Input array\n T *output, ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T prefix, ///< [in] Prefix to seed scan with\n bool apply_prefix = true) ///< [in] Whether or not the calling thread should apply its prefix. If not, the first output element is undefined. (Handy for preventing thread-0 from applying a prefix.)\n{\n T inclusive = input[0];\n if (apply_prefix)\n {\n inclusive = scan_op(prefix, inclusive);\n }\n output[0] = prefix;\n T exclusive = inclusive;\n\n return ThreadScanExclusive(inclusive, exclusive, input + 1, output + 1, scan_op, Int2Type());\n}\n\n\n/**\n * \\brief Perform a sequential exclusive prefix scan over the statically-sized \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH [inferred] LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanExclusive(\n T (&input)[LENGTH], ///< [in] Input array\n T (&output)[LENGTH], ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T prefix, ///< [in] Prefix to seed scan with\n bool apply_prefix = true) ///< [in] Whether or not the calling thread should apply its prefix. (Handy for preventing thread-0 from applying a prefix.)\n{\n return ThreadScanExclusive((T*) input, (T*) output, scan_op, prefix, apply_prefix);\n}\n\n\n\n\n\n\n\n\n\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanInclusive(\n T inclusive,\n T *input, ///< [in] Input array\n T *output, ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n Int2Type /*length*/)\n{\n #pragma unroll\n for (int i = 0; i < LENGTH; ++i)\n {\n inclusive = scan_op(inclusive, input[i]);\n output[i] = inclusive;\n }\n\n return inclusive;\n}\n\n\n/**\n * \\brief Perform a sequential inclusive prefix scan over \\p LENGTH elements of the \\p input array. The aggregate is returned.\n *\n * \\tparam LENGTH LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanInclusive(\n T *input, ///< [in] Input array\n T *output, ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n{\n T inclusive = input[0];\n output[0] = inclusive;\n\n // Continue scan\n return ThreadScanInclusive(inclusive, input + 1, output + 1, scan_op, Int2Type());\n}\n\n\n/**\n * \\brief Perform a sequential inclusive prefix scan over the statically-sized \\p input array. The aggregate is returned.\n *\n * \\tparam LENGTH [inferred] LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanInclusive(\n T (&input)[LENGTH], ///< [in] Input array\n T (&output)[LENGTH], ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op) ///< [in] Binary scan operator\n{\n return ThreadScanInclusive((T*) input, (T*) output, scan_op);\n}\n\n\n/**\n * \\brief Perform a sequential inclusive prefix scan over \\p LENGTH elements of the \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanInclusive(\n T *input, ///< [in] Input array\n T *output, ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T prefix, ///< [in] Prefix to seed scan with\n bool apply_prefix = true) ///< [in] Whether or not the calling thread should apply its prefix. (Handy for preventing thread-0 from applying a prefix.)\n{\n T inclusive = input[0];\n if (apply_prefix)\n {\n inclusive = scan_op(prefix, inclusive);\n }\n output[0] = inclusive;\n\n // Continue scan\n return ThreadScanInclusive(inclusive, input + 1, output + 1, scan_op, Int2Type());\n}\n\n\n/**\n * \\brief Perform a sequential inclusive prefix scan over the statically-sized \\p input array, seeded with the specified \\p prefix. The aggregate is returned.\n *\n * \\tparam LENGTH [inferred] LengthT of \\p input and \\p output arrays\n * \\tparam T [inferred] The data type to be scanned.\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\ntemplate <\n int LENGTH,\n typename T,\n typename ScanOp>\n__device__ __forceinline__ T ThreadScanInclusive(\n T (&input)[LENGTH], ///< [in] Input array\n T (&output)[LENGTH], ///< [out] Output array (may be aliased to \\p input)\n ScanOp scan_op, ///< [in] Binary scan operator\n T prefix, ///< [in] Prefix to seed scan with\n bool apply_prefix = true) ///< [in] Whether or not the calling thread should apply its prefix. (Handy for preventing thread-0 from applying a prefix.)\n{\n return ThreadScanInclusive((T*) input, (T*) output, scan_op, prefix, apply_prefix);\n}\n\n\n//@} end member group\n\n/** @} */ // end group UtilModule\n\n\n} // internal namespace\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_search.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Thread utilities for sequential search\n */\n\n#pragma once\n\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * Computes the begin offsets into A and B for the specific diagonal\n */\ntemplate <\n typename AIteratorT,\n typename BIteratorT,\n typename OffsetT,\n typename CoordinateT>\n__host__ __device__ __forceinline__ void MergePathSearch(\n OffsetT diagonal,\n AIteratorT a,\n BIteratorT b,\n OffsetT a_len,\n OffsetT b_len,\n CoordinateT& path_coordinate)\n{\n /// The value type of the input iterator\n typedef typename std::iterator_traits::value_type T;\n\n OffsetT split_min = CUB_MAX(diagonal - b_len, 0);\n OffsetT split_max = CUB_MIN(diagonal, a_len);\n\n while (split_min < split_max)\n {\n OffsetT split_pivot = (split_min + split_max) >> 1;\n if (a[split_pivot] <= b[diagonal - split_pivot - 1])\n {\n // Move candidate split range up A, down B\n split_min = split_pivot + 1;\n }\n else\n {\n // Move candidate split range up B, down A\n split_max = split_pivot;\n }\n }\n\n path_coordinate.x = CUB_MIN(split_min, a_len);\n path_coordinate.y = diagonal - split_min;\n}\n\n\n\n/**\n * \\brief Returns the offset of the first value within \\p input which does not compare less than \\p val\n */\ntemplate <\n typename InputIteratorT,\n typename OffsetT,\n typename T>\n__device__ __forceinline__ OffsetT LowerBound(\n InputIteratorT input, ///< [in] Input sequence\n OffsetT num_items, ///< [in] Input sequence length\n T val) ///< [in] Search key\n{\n OffsetT retval = 0;\n while (num_items > 0)\n {\n OffsetT half = num_items >> 1;\n if (input[retval + half] < val)\n {\n retval = retval + (half + 1);\n num_items = num_items - (half + 1);\n }\n else\n {\n num_items = half;\n }\n }\n\n return retval;\n}\n\n\n/**\n * \\brief Returns the offset of the first value within \\p input which compares greater than \\p val\n */\ntemplate <\n typename InputIteratorT,\n typename OffsetT,\n typename T>\n__device__ __forceinline__ OffsetT UpperBound(\n InputIteratorT input, ///< [in] Input sequence\n OffsetT num_items, ///< [in] Input sequence length\n T val) ///< [in] Search key\n{\n OffsetT retval = 0;\n while (num_items > 0)\n {\n OffsetT half = num_items >> 1;\n if (val < input[retval + half])\n {\n num_items = half;\n }\n else\n {\n retval = retval + (half + 1);\n num_items = num_items - (half + 1);\n }\n }\n\n return retval;\n}\n\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/thread/thread_store.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Thread utilities for writing memory using PTX cache modifiers.\n */\n\n#pragma once\n\n#include \n\n#include \"../util_ptx.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup UtilIo\n * @{\n */\n\n\n//-----------------------------------------------------------------------------\n// Tags and constants\n//-----------------------------------------------------------------------------\n\n/**\n * \\brief Enumeration of cache modifiers for memory store operations.\n */\nenum CacheStoreModifier\n{\n STORE_DEFAULT, ///< Default (no modifier)\n STORE_WB, ///< Cache write-back all coherent levels\n STORE_CG, ///< Cache at global level\n STORE_CS, ///< Cache streaming (likely to be accessed once)\n STORE_WT, ///< Cache write-through (to system memory)\n STORE_VOLATILE, ///< Volatile shared (any memory space)\n};\n\n\n/**\n * \\name Thread I/O (cache modified)\n * @{\n */\n\n/**\n * \\brief Thread utility for writing memory using cub::CacheStoreModifier cache modifiers. Can be used to store any data type.\n *\n * \\par Example\n * \\code\n * #include // or equivalently \n *\n * // 32-bit store using cache-global modifier:\n * int *d_out;\n * int val;\n * cub::ThreadStore(d_out + threadIdx.x, val);\n *\n * // 16-bit store using default modifier\n * short *d_out;\n * short val;\n * cub::ThreadStore(d_out + threadIdx.x, val);\n *\n * // 256-bit store using write-through modifier\n * double4 *d_out;\n * double4 val;\n * cub::ThreadStore(d_out + threadIdx.x, val);\n *\n * // 96-bit store using cache-streaming cache modifier\n * struct TestFoo { bool a; short b; };\n * TestFoo *d_struct;\n * TestFoo val;\n * cub::ThreadStore(d_out + threadIdx.x, val);\n * \\endcode\n *\n * \\tparam MODIFIER [inferred] CacheStoreModifier enumeration\n * \\tparam InputIteratorT [inferred] Output iterator type \\iterator\n * \\tparam T [inferred] Data type of output value\n */\ntemplate <\n CacheStoreModifier MODIFIER,\n typename OutputIteratorT,\n typename T>\n__device__ __forceinline__ void ThreadStore(OutputIteratorT itr, T val);\n\n\n//@} end member group\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/// Helper structure for templated store iteration (inductive case)\ntemplate \nstruct IterateThreadStore\n{\n template \n static __device__ __forceinline__ void Store(T *ptr, T *vals)\n {\n ThreadStore(ptr + COUNT, vals[COUNT]);\n IterateThreadStore::template Store(ptr, vals);\n }\n\n template \n static __device__ __forceinline__ void Dereference(OutputIteratorT ptr, T *vals)\n {\n ptr[COUNT] = vals[COUNT];\n IterateThreadStore::Dereference(ptr, vals);\n }\n\n};\n\n/// Helper structure for templated store iteration (termination case)\ntemplate \nstruct IterateThreadStore\n{\n template \n static __device__ __forceinline__ void Store(T * /*ptr*/, T * /*vals*/) {}\n\n template \n static __device__ __forceinline__ void Dereference(OutputIteratorT /*ptr*/, T * /*vals*/) {}\n};\n\n\n/**\n * Define a uint4 (16B) ThreadStore specialization for the given Cache load modifier\n */\n#define _CUB_STORE_16(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(uint4* ptr, uint4 val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".v4.u32 [%0], {%1, %2, %3, %4};\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"r\"(val.x), \\\n \"r\"(val.y), \\\n \"r\"(val.z), \\\n \"r\"(val.w)); \\\n } \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(ulonglong2* ptr, ulonglong2 val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".v2.u64 [%0], {%1, %2};\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"l\"(val.x), \\\n \"l\"(val.y)); \\\n }\n\n\n/**\n * Define a uint2 (8B) ThreadStore specialization for the given Cache load modifier\n */\n#define _CUB_STORE_8(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(ushort4* ptr, ushort4 val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".v4.u16 [%0], {%1, %2, %3, %4};\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"h\"(val.x), \\\n \"h\"(val.y), \\\n \"h\"(val.z), \\\n \"h\"(val.w)); \\\n } \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(uint2* ptr, uint2 val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".v2.u32 [%0], {%1, %2};\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"r\"(val.x), \\\n \"r\"(val.y)); \\\n } \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(unsigned long long* ptr, unsigned long long val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".u64 [%0], %1;\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"l\"(val)); \\\n }\n\n/**\n * Define a unsigned int (4B) ThreadStore specialization for the given Cache load modifier\n */\n#define _CUB_STORE_4(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(unsigned int* ptr, unsigned int val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".u32 [%0], %1;\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"r\"(val)); \\\n }\n\n\n/**\n * Define a unsigned short (2B) ThreadStore specialization for the given Cache load modifier\n */\n#define _CUB_STORE_2(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(unsigned short* ptr, unsigned short val) \\\n { \\\n asm volatile (\"st.\"#ptx_modifier\".u16 [%0], %1;\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"h\"(val)); \\\n }\n\n\n/**\n * Define a unsigned char (1B) ThreadStore specialization for the given Cache load modifier\n */\n#define _CUB_STORE_1(cub_modifier, ptx_modifier) \\\n template<> \\\n __device__ __forceinline__ void ThreadStore(unsigned char* ptr, unsigned char val) \\\n { \\\n asm volatile ( \\\n \"{\" \\\n \" .reg .u8 datum;\" \\\n \" cvt.u8.u16 datum, %1;\" \\\n \" st.\"#ptx_modifier\".u8 [%0], datum;\" \\\n \"}\" : : \\\n _CUB_ASM_PTR_(ptr), \\\n \"h\"((unsigned short) val)); \\\n }\n\n/**\n * Define powers-of-two ThreadStore specializations for the given Cache load modifier\n */\n#define _CUB_STORE_ALL(cub_modifier, ptx_modifier) \\\n _CUB_STORE_16(cub_modifier, ptx_modifier) \\\n _CUB_STORE_8(cub_modifier, ptx_modifier) \\\n _CUB_STORE_4(cub_modifier, ptx_modifier) \\\n _CUB_STORE_2(cub_modifier, ptx_modifier) \\\n _CUB_STORE_1(cub_modifier, ptx_modifier) \\\n\n\n/**\n * Define ThreadStore specializations for the various Cache load modifiers\n */\n#if CUB_PTX_ARCH >= 200\n _CUB_STORE_ALL(STORE_WB, wb)\n _CUB_STORE_ALL(STORE_CG, cg)\n _CUB_STORE_ALL(STORE_CS, cs)\n _CUB_STORE_ALL(STORE_WT, wt)\n#else\n _CUB_STORE_ALL(STORE_WB, global)\n _CUB_STORE_ALL(STORE_CG, global)\n _CUB_STORE_ALL(STORE_CS, global)\n _CUB_STORE_ALL(STORE_WT, volatile.global)\n#endif\n\n\n// Macro cleanup\n#undef _CUB_STORE_ALL\n#undef _CUB_STORE_1\n#undef _CUB_STORE_2\n#undef _CUB_STORE_4\n#undef _CUB_STORE_8\n#undef _CUB_STORE_16\n\n\n/**\n * ThreadStore definition for STORE_DEFAULT modifier on iterator types\n */\ntemplate \n__device__ __forceinline__ void ThreadStore(\n OutputIteratorT itr,\n T val,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n *itr = val;\n}\n\n\n/**\n * ThreadStore definition for STORE_DEFAULT modifier on pointer types\n */\ntemplate \n__device__ __forceinline__ void ThreadStore(\n T *ptr,\n T val,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n *ptr = val;\n}\n\n\n/**\n * ThreadStore definition for STORE_VOLATILE modifier on primitive pointer types\n */\ntemplate \n__device__ __forceinline__ void ThreadStoreVolatilePtr(\n T *ptr,\n T val,\n Int2Type /*is_primitive*/)\n{\n *reinterpret_cast(ptr) = val;\n}\n\n\n/**\n * ThreadStore definition for STORE_VOLATILE modifier on non-primitive pointer types\n */\ntemplate \n__device__ __forceinline__ void ThreadStoreVolatilePtr(\n T *ptr,\n T val,\n Int2Type /*is_primitive*/)\n{\n // Create a temporary using shuffle-words, then store using volatile-words\n typedef typename UnitWord::VolatileWord VolatileWord; \n typedef typename UnitWord::ShuffleWord ShuffleWord;\n\n const int VOLATILE_MULTIPLE = sizeof(T) / sizeof(VolatileWord);\n const int SHUFFLE_MULTIPLE = sizeof(T) / sizeof(ShuffleWord);\n \n VolatileWord words[VOLATILE_MULTIPLE];\n\n #pragma unroll\n for (int i = 0; i < SHUFFLE_MULTIPLE; ++i)\n reinterpret_cast(words)[i] = reinterpret_cast(&val)[i];\n\n IterateThreadStore<0, VOLATILE_MULTIPLE>::template Dereference(\n reinterpret_cast(ptr),\n words);\n}\n\n\n/**\n * ThreadStore definition for STORE_VOLATILE modifier on pointer types\n */\ntemplate \n__device__ __forceinline__ void ThreadStore(\n T *ptr,\n T val,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n ThreadStoreVolatilePtr(ptr, val, Int2Type::PRIMITIVE>());\n}\n\n\n/**\n * ThreadStore definition for generic modifiers on pointer types\n */\ntemplate \n__device__ __forceinline__ void ThreadStore(\n T *ptr,\n T val,\n Int2Type /*modifier*/,\n Int2Type /*is_pointer*/)\n{\n // Create a temporary using shuffle-words, then store using device-words\n typedef typename UnitWord::DeviceWord DeviceWord; \n typedef typename UnitWord::ShuffleWord ShuffleWord;\n\n const int DEVICE_MULTIPLE = sizeof(T) / sizeof(DeviceWord);\n const int SHUFFLE_MULTIPLE = sizeof(T) / sizeof(ShuffleWord);\n \n DeviceWord words[DEVICE_MULTIPLE];\n\n #pragma unroll\n for (int i = 0; i < SHUFFLE_MULTIPLE; ++i)\n reinterpret_cast(words)[i] = reinterpret_cast(&val)[i];\n\n IterateThreadStore<0, DEVICE_MULTIPLE>::template Store(\n reinterpret_cast(ptr),\n words);\n}\n\n\n/**\n * ThreadStore definition for generic modifiers\n */\ntemplate \n__device__ __forceinline__ void ThreadStore(OutputIteratorT itr, T val)\n{\n ThreadStore(\n itr,\n val,\n Int2Type(),\n Int2Type::VALUE>());\n}\n\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/** @} */ // end group UtilIo\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_allocator.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple caching allocator for device memory allocations. The allocator is\n * thread-safe and capable of managing device allocations on multiple devices.\n ******************************************************************************/\n\n#pragma once\n\n#include \"util_namespace.cuh\"\n#include \"util_debug.cuh\"\n\n#include \n#include \n\n#include \"host/mutex.cuh\"\n#include \n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilMgmt\n * @{\n */\n\n\n/******************************************************************************\n * CachingDeviceAllocator (host use)\n ******************************************************************************/\n\n/**\n * \\brief A simple caching allocator for device memory allocations.\n *\n * \\par Overview\n * The allocator is thread-safe and stream-safe and is capable of managing cached\n * device allocations on multiple devices. It behaves as follows:\n *\n * \\par\n * - Allocations from the allocator are associated with an \\p active_stream. Once freed,\n * the allocation becomes available immediately for reuse within the \\p active_stream\n * with which it was associated with during allocation, and it becomes available for\n * reuse within other streams when all prior work submitted to \\p active_stream has completed.\n * - Allocations are categorized and cached by bin size. A new allocation request of\n * a given size will only consider cached allocations within the corresponding bin.\n * - Bin limits progress geometrically in accordance with the growth factor\n * \\p bin_growth provided during construction. Unused device allocations within\n * a larger bin cache are not reused for allocation requests that categorize to\n * smaller bin sizes.\n * - Allocation requests below (\\p bin_growth ^ \\p min_bin) are rounded up to\n * (\\p bin_growth ^ \\p min_bin).\n * - Allocations above (\\p bin_growth ^ \\p max_bin) are not rounded up to the nearest\n * bin and are simply freed when they are deallocated instead of being returned\n * to a bin-cache.\n * - %If the total storage of cached allocations on a given device will exceed\n * \\p max_cached_bytes, allocations for that device are simply freed when they are\n * deallocated instead of being returned to their bin-cache.\n *\n * \\par\n * For example, the default-constructed CachingDeviceAllocator is configured with:\n * - \\p bin_growth = 8\n * - \\p min_bin = 3\n * - \\p max_bin = 7\n * - \\p max_cached_bytes = 6MB - 1B\n *\n * \\par\n * which delineates five bin-sizes: 512B, 4KB, 32KB, 256KB, and 2MB\n * and sets a maximum of 6,291,455 cached bytes per device\n *\n */\nstruct CachingDeviceAllocator\n{\n\n //---------------------------------------------------------------------\n // Constants\n //---------------------------------------------------------------------\n\n /// Out-of-bounds bin\n static const unsigned int INVALID_BIN = (unsigned int) -1;\n\n /// Invalid size\n static const size_t INVALID_SIZE = (size_t) -1;\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n /// Invalid device ordinal\n static const int INVALID_DEVICE_ORDINAL = -1;\n\n //---------------------------------------------------------------------\n // Type definitions and helper types\n //---------------------------------------------------------------------\n\n /**\n * Descriptor for device memory allocations\n */\n struct BlockDescriptor\n {\n void* d_ptr; // Device pointer\n size_t bytes; // Size of allocation in bytes\n unsigned int bin; // Bin enumeration\n int device; // device ordinal\n cudaStream_t associated_stream; // Associated associated_stream\n cudaEvent_t ready_event; // Signal when associated stream has run to the point at which this block was freed\n\n // Constructor (suitable for searching maps for a specific block, given its pointer and device)\n BlockDescriptor(void *d_ptr, int device) :\n d_ptr(d_ptr),\n bytes(0),\n bin(INVALID_BIN),\n device(device),\n associated_stream(0),\n ready_event(0)\n {}\n\n // Constructor (suitable for searching maps for a range of suitable blocks, given a device)\n BlockDescriptor(int device) :\n d_ptr(NULL),\n bytes(0),\n bin(INVALID_BIN),\n device(device),\n associated_stream(0),\n ready_event(0)\n {}\n\n // Comparison functor for comparing device pointers\n static bool PtrCompare(const BlockDescriptor &a, const BlockDescriptor &b)\n {\n if (a.device == b.device)\n return (a.d_ptr < b.d_ptr);\n else\n return (a.device < b.device);\n }\n\n // Comparison functor for comparing allocation sizes\n static bool SizeCompare(const BlockDescriptor &a, const BlockDescriptor &b)\n {\n if (a.device == b.device)\n return (a.bytes < b.bytes);\n else\n return (a.device < b.device);\n }\n };\n\n /// BlockDescriptor comparator function interface\n typedef bool (*Compare)(const BlockDescriptor &, const BlockDescriptor &);\n\n class TotalBytes {\n public:\n size_t free;\n size_t live;\n TotalBytes() { free = live = 0; }\n };\n\n /// Set type for cached blocks (ordered by size)\n typedef std::multiset CachedBlocks;\n\n /// Set type for live blocks (ordered by ptr)\n typedef std::multiset BusyBlocks;\n\n /// Map type of device ordinals to the number of cached bytes cached by each device\n typedef std::map GpuCachedBytes;\n\n\n //---------------------------------------------------------------------\n // Utility functions\n //---------------------------------------------------------------------\n\n /**\n * Integer pow function for unsigned base and exponent\n */\n static unsigned int IntPow(\n unsigned int base,\n unsigned int exp)\n {\n unsigned int retval = 1;\n while (exp > 0)\n {\n if (exp & 1) {\n retval = retval * base; // multiply the result by the current base\n }\n base = base * base; // square the base\n exp = exp >> 1; // divide the exponent in half\n }\n return retval;\n }\n\n\n /**\n * Round up to the nearest power-of\n */\n void NearestPowerOf(\n unsigned int &power,\n size_t &rounded_bytes,\n unsigned int base,\n size_t value)\n {\n power = 0;\n rounded_bytes = 1;\n\n if (value * base < value)\n {\n // Overflow\n power = sizeof(size_t) * 8;\n rounded_bytes = size_t(0) - 1;\n return;\n }\n\n while (rounded_bytes < value)\n {\n rounded_bytes *= base;\n power++;\n }\n }\n\n\n //---------------------------------------------------------------------\n // Fields\n //---------------------------------------------------------------------\n\n cub::Mutex mutex; /// Mutex for thread-safety\n\n unsigned int bin_growth; /// Geometric growth factor for bin-sizes\n unsigned int min_bin; /// Minimum bin enumeration\n unsigned int max_bin; /// Maximum bin enumeration\n\n size_t min_bin_bytes; /// Minimum bin size\n size_t max_bin_bytes; /// Maximum bin size\n size_t max_cached_bytes; /// Maximum aggregate cached bytes per device\n\n const bool skip_cleanup; /// Whether or not to skip a call to FreeAllCached() when destructor is called. (The CUDA runtime may have already shut down for statically declared allocators)\n bool debug; /// Whether or not to print (de)allocation events to stdout\n\n GpuCachedBytes cached_bytes; /// Map of device ordinal to aggregate cached bytes on that device\n CachedBlocks cached_blocks; /// Set of cached device allocations available for reuse\n BusyBlocks live_blocks; /// Set of live device allocations currently in use\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n //---------------------------------------------------------------------\n // Methods\n //---------------------------------------------------------------------\n\n /**\n * \\brief Constructor.\n */\n CachingDeviceAllocator(\n unsigned int bin_growth, ///< Geometric growth factor for bin-sizes\n unsigned int min_bin = 1, ///< Minimum bin (default is bin_growth ^ 1)\n unsigned int max_bin = INVALID_BIN, ///< Maximum bin (default is no max bin)\n size_t max_cached_bytes = INVALID_SIZE, ///< Maximum aggregate cached bytes per device (default is no limit)\n bool skip_cleanup = false, ///< Whether or not to skip a call to \\p FreeAllCached() when the destructor is called (default is to deallocate)\n bool debug = false) ///< Whether or not to print (de)allocation events to stdout (default is no stderr output)\n :\n bin_growth(bin_growth),\n min_bin(min_bin),\n max_bin(max_bin),\n min_bin_bytes(IntPow(bin_growth, min_bin)),\n max_bin_bytes(IntPow(bin_growth, max_bin)),\n max_cached_bytes(max_cached_bytes),\n skip_cleanup(skip_cleanup),\n debug(debug),\n cached_blocks(BlockDescriptor::SizeCompare),\n live_blocks(BlockDescriptor::PtrCompare)\n {}\n\n\n /**\n * \\brief Default constructor.\n *\n * Configured with:\n * \\par\n * - \\p bin_growth = 8\n * - \\p min_bin = 3\n * - \\p max_bin = 7\n * - \\p max_cached_bytes = (\\p bin_growth ^ \\p max_bin) * 3) - 1 = 6,291,455 bytes\n *\n * which delineates five bin-sizes: 512B, 4KB, 32KB, 256KB, and 2MB and\n * sets a maximum of 6,291,455 cached bytes per device\n */\n CachingDeviceAllocator(\n bool skip_cleanup = false,\n bool debug = false)\n :\n bin_growth(8),\n min_bin(3),\n max_bin(7),\n min_bin_bytes(IntPow(bin_growth, min_bin)),\n max_bin_bytes(IntPow(bin_growth, max_bin)),\n max_cached_bytes((max_bin_bytes * 3) - 1),\n skip_cleanup(skip_cleanup),\n debug(debug),\n cached_blocks(BlockDescriptor::SizeCompare),\n live_blocks(BlockDescriptor::PtrCompare)\n {}\n\n\n /**\n * \\brief Sets the limit on the number bytes this allocator is allowed to cache per device.\n *\n * Changing the ceiling of cached bytes does not cause any allocations (in-use or\n * cached-in-reserve) to be freed. See \\p FreeAllCached().\n */\n cudaError_t SetMaxCachedBytes(\n size_t max_cached_bytes)\n {\n // Lock\n mutex.Lock();\n\n if (debug) _CubLog(\"Changing max_cached_bytes (%lld -> %lld)\\n\", (long long) this->max_cached_bytes, (long long) max_cached_bytes);\n\n this->max_cached_bytes = max_cached_bytes;\n\n // Unlock\n mutex.Unlock();\n\n return cudaSuccess;\n }\n\n\n /**\n * \\brief Provides a suitable allocation of device memory for the given size on the specified device.\n *\n * Once freed, the allocation becomes available immediately for reuse within the \\p active_stream\n * with which it was associated with during allocation, and it becomes available for reuse within other\n * streams when all prior work submitted to \\p active_stream has completed.\n */\n cudaError_t DeviceAllocate(\n int device, ///< [in] Device on which to place the allocation\n void **d_ptr, ///< [out] Reference to pointer to the allocation\n size_t bytes, ///< [in] Minimum number of bytes for the allocation\n cudaStream_t active_stream = 0) ///< [in] The stream to be associated with this allocation\n {\n *d_ptr = NULL;\n int entrypoint_device = INVALID_DEVICE_ORDINAL;\n cudaError_t error = cudaSuccess;\n\n if (device == INVALID_DEVICE_ORDINAL)\n {\n if (CubDebug(error = cudaGetDevice(&entrypoint_device))) return error;\n device = entrypoint_device;\n }\n\n // Create a block descriptor for the requested allocation\n bool found = false;\n BlockDescriptor search_key(device);\n search_key.associated_stream = active_stream;\n NearestPowerOf(search_key.bin, search_key.bytes, bin_growth, bytes);\n\n if (search_key.bin > max_bin)\n {\n // Bin is greater than our maximum bin: allocate the request\n // exactly and give out-of-bounds bin. It will not be cached\n // for reuse when returned.\n search_key.bin = INVALID_BIN;\n search_key.bytes = bytes;\n }\n else\n {\n // Search for a suitable cached allocation: lock\n mutex.Lock();\n\n if (search_key.bin < min_bin)\n {\n // Bin is less than minimum bin: round up\n search_key.bin = min_bin;\n search_key.bytes = min_bin_bytes;\n }\n\n // Iterate through the range of cached blocks on the same device in the same bin\n CachedBlocks::iterator block_itr = cached_blocks.lower_bound(search_key);\n while ((block_itr != cached_blocks.end())\n && (block_itr->device == device)\n && (block_itr->bin == search_key.bin))\n {\n // To prevent races with reusing blocks returned by the host but still\n // in use by the device, only consider cached blocks that are\n // either (from the active stream) or (from an idle stream)\n if ((active_stream == block_itr->associated_stream) ||\n (cudaEventQuery(block_itr->ready_event) != cudaErrorNotReady))\n {\n // Reuse existing cache block. Insert into live blocks.\n found = true;\n search_key = *block_itr;\n search_key.associated_stream = active_stream;\n live_blocks.insert(search_key);\n\n // Remove from free blocks\n cached_bytes[device].free -= search_key.bytes;\n cached_bytes[device].live += search_key.bytes;\n\n if (debug) _CubLog(\"\\tDevice %d reused cached block at %p (%lld bytes) for stream %lld (previously associated with stream %lld).\\n\",\n device, search_key.d_ptr, (long long) search_key.bytes, (long long) search_key.associated_stream, (long long) block_itr->associated_stream);\n\n cached_blocks.erase(block_itr);\n\n break;\n }\n block_itr++;\n }\n\n // Done searching: unlock\n mutex.Unlock();\n }\n\n // Allocate the block if necessary\n if (!found)\n {\n // Set runtime's current device to specified device (entrypoint may not be set)\n if (device != entrypoint_device)\n {\n if (CubDebug(error = cudaGetDevice(&entrypoint_device))) return error;\n if (CubDebug(error = cudaSetDevice(device))) return error;\n }\n\n // Attempt to allocate\n if (CubDebug(error = cudaMalloc(&search_key.d_ptr, search_key.bytes)) == cudaErrorMemoryAllocation)\n {\n // The allocation attempt failed: free all cached blocks on device and retry\n if (debug) _CubLog(\"\\tDevice %d failed to allocate %lld bytes for stream %lld, retrying after freeing cached allocations\",\n device, (long long) search_key.bytes, (long long) search_key.associated_stream);\n\n error = cudaSuccess; // Reset the error we will return\n cudaGetLastError(); // Reset CUDART's error\n\n // Lock\n mutex.Lock();\n\n // Iterate the range of free blocks on the same device\n BlockDescriptor free_key(device);\n CachedBlocks::iterator block_itr = cached_blocks.lower_bound(free_key);\n\n while ((block_itr != cached_blocks.end()) && (block_itr->device == device))\n {\n // No need to worry about synchronization with the device: cudaFree is\n // blocking and will synchronize across all kernels executing\n // on the current device\n\n // Free device memory and destroy stream event.\n if (CubDebug(error = cudaFree(block_itr->d_ptr))) break;\n if (CubDebug(error = cudaEventDestroy(block_itr->ready_event))) break;\n\n // Reduce balance and erase entry\n cached_bytes[device].free -= block_itr->bytes;\n\n if (debug) _CubLog(\"\\tDevice %d freed %lld bytes.\\n\\t\\t %lld available blocks cached (%lld bytes), %lld live blocks (%lld bytes) outstanding.\\n\",\n device, (long long) block_itr->bytes, (long long) cached_blocks.size(), (long long) cached_bytes[device].free, (long long) live_blocks.size(), (long long) cached_bytes[device].live);\n\n cached_blocks.erase(block_itr);\n\n block_itr++;\n }\n\n // Unlock\n mutex.Unlock();\n\n // Return under error\n if (error) return error;\n\n // Try to allocate again\n if (CubDebug(error = cudaMalloc(&search_key.d_ptr, search_key.bytes))) return error;\n }\n\n // Create ready event\n if (CubDebug(error = cudaEventCreateWithFlags(&search_key.ready_event, cudaEventDisableTiming)))\n return error;\n\n // Insert into live blocks\n mutex.Lock();\n live_blocks.insert(search_key);\n cached_bytes[device].live += search_key.bytes;\n mutex.Unlock();\n\n if (debug) _CubLog(\"\\tDevice %d allocated new device block at %p (%lld bytes associated with stream %lld).\\n\",\n device, search_key.d_ptr, (long long) search_key.bytes, (long long) search_key.associated_stream);\n\n // Attempt to revert back to previous device if necessary\n if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device))\n {\n if (CubDebug(error = cudaSetDevice(entrypoint_device))) return error;\n }\n }\n\n // Copy device pointer to output parameter\n *d_ptr = search_key.d_ptr;\n\n if (debug) _CubLog(\"\\t\\t%lld available blocks cached (%lld bytes), %lld live blocks outstanding(%lld bytes).\\n\",\n (long long) cached_blocks.size(), (long long) cached_bytes[device].free, (long long) live_blocks.size(), (long long) cached_bytes[device].live);\n\n return error;\n }\n\n\n /**\n * \\brief Provides a suitable allocation of device memory for the given size on the current device.\n *\n * Once freed, the allocation becomes available immediately for reuse within the \\p active_stream\n * with which it was associated with during allocation, and it becomes available for reuse within other\n * streams when all prior work submitted to \\p active_stream has completed.\n */\n cudaError_t DeviceAllocate(\n void **d_ptr, ///< [out] Reference to pointer to the allocation\n size_t bytes, ///< [in] Minimum number of bytes for the allocation\n cudaStream_t active_stream = 0) ///< [in] The stream to be associated with this allocation\n {\n return DeviceAllocate(INVALID_DEVICE_ORDINAL, d_ptr, bytes, active_stream);\n }\n\n\n /**\n * \\brief Frees a live allocation of device memory on the specified device, returning it to the allocator.\n *\n * Once freed, the allocation becomes available immediately for reuse within the \\p active_stream\n * with which it was associated with during allocation, and it becomes available for reuse within other\n * streams when all prior work submitted to \\p active_stream has completed.\n */\n cudaError_t DeviceFree(\n int device,\n void* d_ptr)\n {\n int entrypoint_device = INVALID_DEVICE_ORDINAL;\n cudaError_t error = cudaSuccess;\n\n if (device == INVALID_DEVICE_ORDINAL)\n {\n if (CubDebug(error = cudaGetDevice(&entrypoint_device)))\n return error;\n device = entrypoint_device;\n }\n\n // Lock\n mutex.Lock();\n\n // Find corresponding block descriptor\n bool recached = false;\n BlockDescriptor search_key(d_ptr, device);\n BusyBlocks::iterator block_itr = live_blocks.find(search_key);\n if (block_itr != live_blocks.end())\n {\n // Remove from live blocks\n search_key = *block_itr;\n live_blocks.erase(block_itr);\n cached_bytes[device].live -= search_key.bytes;\n\n // Keep the returned allocation if bin is valid and we won't exceed the max cached threshold\n if ((search_key.bin != INVALID_BIN) && (cached_bytes[device].free + search_key.bytes <= max_cached_bytes))\n {\n // Insert returned allocation into free blocks\n recached = true;\n cached_blocks.insert(search_key);\n cached_bytes[device].free += search_key.bytes;\n\n if (debug) _CubLog(\"\\tDevice %d returned %lld bytes from associated stream %lld.\\n\\t\\t %lld available blocks cached (%lld bytes), %lld live blocks outstanding. (%lld bytes)\\n\",\n device, (long long) search_key.bytes, (long long) search_key.associated_stream, (long long) cached_blocks.size(),\n (long long) cached_bytes[device].free, (long long) live_blocks.size(), (long long) cached_bytes[device].live);\n }\n }\n\n // Unlock\n mutex.Unlock();\n\n // First set to specified device (entrypoint may not be set)\n if (device != entrypoint_device)\n {\n if (CubDebug(error = cudaGetDevice(&entrypoint_device))) return error;\n if (CubDebug(error = cudaSetDevice(device))) return error;\n }\n\n if (recached)\n {\n // Insert the ready event in the associated stream (must have current device set properly)\n if (CubDebug(error = cudaEventRecord(search_key.ready_event, search_key.associated_stream))) return error;\n }\n else\n {\n // Free the allocation from the runtime and cleanup the event.\n if (CubDebug(error = cudaFree(d_ptr))) return error;\n if (CubDebug(error = cudaEventDestroy(search_key.ready_event))) return error;\n\n if (debug) _CubLog(\"\\tDevice %d freed %lld bytes from associated stream %lld.\\n\\t\\t %lld available blocks cached (%lld bytes), %lld live blocks (%lld bytes) outstanding.\\n\",\n device, (long long) search_key.bytes, (long long) search_key.associated_stream, (long long) cached_blocks.size(), (long long) cached_bytes[device].free, (long long) live_blocks.size(), (long long) cached_bytes[device].live);\n }\n\n // Reset device\n if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device))\n {\n if (CubDebug(error = cudaSetDevice(entrypoint_device))) return error;\n }\n\n return error;\n }\n\n\n /**\n * \\brief Frees a live allocation of device memory on the current device, returning it to the allocator.\n *\n * Once freed, the allocation becomes available immediately for reuse within the \\p active_stream\n * with which it was associated with during allocation, and it becomes available for reuse within other\n * streams when all prior work submitted to \\p active_stream has completed.\n */\n cudaError_t DeviceFree(\n void* d_ptr)\n {\n return DeviceFree(INVALID_DEVICE_ORDINAL, d_ptr);\n }\n\n\n /**\n * \\brief Frees all cached device allocations on all devices\n */\n cudaError_t FreeAllCached()\n {\n cudaError_t error = cudaSuccess;\n int entrypoint_device = INVALID_DEVICE_ORDINAL;\n int current_device = INVALID_DEVICE_ORDINAL;\n\n mutex.Lock();\n\n while (!cached_blocks.empty())\n {\n // Get first block\n CachedBlocks::iterator begin = cached_blocks.begin();\n\n // Get entry-point device ordinal if necessary\n if (entrypoint_device == INVALID_DEVICE_ORDINAL)\n {\n if (CubDebug(error = cudaGetDevice(&entrypoint_device))) break;\n }\n\n // Set current device ordinal if necessary\n if (begin->device != current_device)\n {\n if (CubDebug(error = cudaSetDevice(begin->device))) break;\n current_device = begin->device;\n }\n\n // Free device memory\n if (CubDebug(error = cudaFree(begin->d_ptr))) break;\n if (CubDebug(error = cudaEventDestroy(begin->ready_event))) break;\n\n // Reduce balance and erase entry\n cached_bytes[current_device].free -= begin->bytes;\n\n if (debug) _CubLog(\"\\tDevice %d freed %lld bytes.\\n\\t\\t %lld available blocks cached (%lld bytes), %lld live blocks (%lld bytes) outstanding.\\n\",\n current_device, (long long) begin->bytes, (long long) cached_blocks.size(), (long long) cached_bytes[current_device].free, (long long) live_blocks.size(), (long long) cached_bytes[current_device].live);\n\n cached_blocks.erase(begin);\n }\n\n mutex.Unlock();\n\n // Attempt to revert back to entry-point device if necessary\n if (entrypoint_device != INVALID_DEVICE_ORDINAL)\n {\n if (CubDebug(error = cudaSetDevice(entrypoint_device))) return error;\n }\n\n return error;\n }\n\n\n /**\n * \\brief Destructor\n */\n virtual ~CachingDeviceAllocator()\n {\n if (!skip_cleanup)\n FreeAllCached();\n }\n\n};\n\n\n\n\n/** @} */ // end group UtilMgmt\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_arch.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Static architectural properties by SM version.\n */\n\n#pragma once\n\n#include \"util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n#if (__CUDACC_VER_MAJOR__ >= 9) && !defined(CUB_USE_COOPERATIVE_GROUPS)\n #define CUB_USE_COOPERATIVE_GROUPS\n#endif\n\n/// CUB_PTX_ARCH reflects the PTX version targeted by the active compiler pass (or zero during the host pass).\n#ifndef CUB_PTX_ARCH\n #ifndef __CUDA_ARCH__\n #define CUB_PTX_ARCH 0\n #else\n #define CUB_PTX_ARCH __CUDA_ARCH__\n #endif\n#endif\n\n\n/// Whether or not the source targeted by the active compiler pass is allowed to invoke device kernels or methods from the CUDA runtime API.\n#ifndef CUB_RUNTIME_FUNCTION\n #if !defined(__CUDA_ARCH__) || (__CUDA_ARCH__>= 350 && defined(__CUDACC_RDC__))\n #define CUB_RUNTIME_ENABLED\n #define CUB_RUNTIME_FUNCTION __host__ __device__\n #else\n #define CUB_RUNTIME_FUNCTION __host__\n #endif\n#endif\n\n\n/// Number of threads per warp\n#ifndef CUB_LOG_WARP_THREADS\n #define CUB_LOG_WARP_THREADS(arch) \\\n (5)\n #define CUB_WARP_THREADS(arch) \\\n (1 << CUB_LOG_WARP_THREADS(arch))\n\n #define CUB_PTX_WARP_THREADS CUB_WARP_THREADS(CUB_PTX_ARCH)\n #define CUB_PTX_LOG_WARP_THREADS CUB_LOG_WARP_THREADS(CUB_PTX_ARCH)\n#endif\n\n\n/// Number of smem banks\n#ifndef CUB_LOG_SMEM_BANKS\n #define CUB_LOG_SMEM_BANKS(arch) \\\n ((arch >= 200) ? \\\n (5) : \\\n (4))\n #define CUB_SMEM_BANKS(arch) \\\n (1 << CUB_LOG_SMEM_BANKS(arch))\n\n #define CUB_PTX_LOG_SMEM_BANKS CUB_LOG_SMEM_BANKS(CUB_PTX_ARCH)\n #define CUB_PTX_SMEM_BANKS CUB_SMEM_BANKS(CUB_PTX_ARCH)\n#endif\n\n\n/// Oversubscription factor\n#ifndef CUB_SUBSCRIPTION_FACTOR\n #define CUB_SUBSCRIPTION_FACTOR(arch) \\\n ((arch >= 300) ? \\\n (5) : \\\n ((arch >= 200) ? \\\n (3) : \\\n (10)))\n #define CUB_PTX_SUBSCRIPTION_FACTOR CUB_SUBSCRIPTION_FACTOR(CUB_PTX_ARCH)\n#endif\n\n\n/// Prefer padding overhead vs X-way conflicts greater than this threshold\n#ifndef CUB_PREFER_CONFLICT_OVER_PADDING\n #define CUB_PREFER_CONFLICT_OVER_PADDING(arch) \\\n ((arch >= 300) ? \\\n (1) : \\\n (4))\n #define CUB_PTX_PREFER_CONFLICT_OVER_PADDING CUB_PREFER_CONFLICT_OVER_PADDING(CUB_PTX_ARCH)\n#endif\n\n\n/// Scale down the number of warps to keep same amount of \"tile\" storage as the nominal configuration for 4B data. Minimum of two warps.\n#ifndef CUB_BLOCK_THREADS\n #define CUB_BLOCK_THREADS(NOMINAL_4B_BLOCK_THREADS, T, PTX_ARCH) \\\n (CUB_MIN( \\\n NOMINAL_4B_BLOCK_THREADS * 2, \\\n CUB_WARP_THREADS(PTX_ARCH) * CUB_MAX( \\\n (NOMINAL_4B_BLOCK_THREADS / CUB_WARP_THREADS(PTX_ARCH)) * 3 / 4, \\\n (NOMINAL_4B_BLOCK_THREADS / CUB_WARP_THREADS(PTX_ARCH)) * 4 / sizeof(T))))\n#endif\n\n/// Scale up/down number of items per thread to keep the same amount of \"tile\" storage as the nominal configuration for 4B data. Minimum 1 item per thread\n#ifndef CUB_ITEMS_PER_THREAD\n #define CUB_ITEMS_PER_THREAD(NOMINAL_4B_ITEMS_PER_THREAD, NOMINAL_4B_BLOCK_THREADS, T, PTX_ARCH) \\\n\t (CUB_MIN( \\\n\t NOMINAL_4B_ITEMS_PER_THREAD * 2, \\\n\t CUB_MAX( \\\n\t 1, \\\n\t (NOMINAL_4B_ITEMS_PER_THREAD * NOMINAL_4B_BLOCK_THREADS * 4 / sizeof(T)) / CUB_BLOCK_THREADS(NOMINAL_4B_BLOCK_THREADS, T, PTX_ARCH))))\n#endif\n\n/// Define both nominal threads-per-block and items-per-thread\n#ifndef CUB_NOMINAL_CONFIG\n #define CUB_NOMINAL_CONFIG(NOMINAL_4B_BLOCK_THREADS, NOMINAL_4B_ITEMS_PER_THREAD, T) \\\n CUB_BLOCK_THREADS(NOMINAL_4B_BLOCK_THREADS, T, 200), \\\n CUB_ITEMS_PER_THREAD(NOMINAL_4B_ITEMS_PER_THREAD, NOMINAL_4B_BLOCK_THREADS, T, 200)\n#endif\n\n\n\n#endif // Do not document\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_debug.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Error and event logging routines.\n *\n * The following macros definitions are supported:\n * - \\p CUB_LOG. Simple event messages are printed to \\p stdout.\n */\n\n#pragma once\n\n#include \n#include \"util_namespace.cuh\"\n#include \"util_arch.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilMgmt\n * @{\n */\n\n\n/// CUB error reporting macro (prints error messages to stderr)\n#if (defined(DEBUG) || defined(_DEBUG)) && !defined(CUB_STDERR)\n #define CUB_STDERR\n#endif\n\n\n\n/**\n * \\brief %If \\p CUB_STDERR is defined and \\p error is not \\p cudaSuccess, the corresponding error message is printed to \\p stderr (or \\p stdout in device code) along with the supplied source context.\n *\n * \\return The CUDA error.\n */\n__host__ __device__ __forceinline__ cudaError_t Debug(\n cudaError_t error,\n const char* filename,\n int line)\n{\n (void)filename;\n (void)line;\n#ifdef CUB_STDERR\n if (error)\n {\n #if (CUB_PTX_ARCH == 0)\n fprintf(stderr, \"CUDA error %d [%s, %d]: %s\\n\", error, filename, line, cudaGetErrorString(error));\n fflush(stderr);\n #elif (CUB_PTX_ARCH >= 200)\n printf(\"CUDA error %d [block (%d,%d,%d) thread (%d,%d,%d), %s, %d]\\n\", error, blockIdx.z, blockIdx.y, blockIdx.x, threadIdx.z, threadIdx.y, threadIdx.x, filename, line);\n #endif\n }\n#endif\n return error;\n}\n\n\n/**\n * \\brief Debug macro\n */\n#ifndef CubDebug\n #define CubDebug(e) cub::Debug((cudaError_t) (e), __FILE__, __LINE__)\n#endif\n\n\n/**\n * \\brief Debug macro with exit\n */\n#ifndef CubDebugExit\n #define CubDebugExit(e) if (cub::Debug((cudaError_t) (e), __FILE__, __LINE__)) { exit(1); }\n#endif\n\n\n/**\n * \\brief Log macro for printf statements.\n */\n#if !defined(_CubLog)\n #if !(defined(__clang__) && defined(__CUDA__))\n #if (CUB_PTX_ARCH == 0)\n #define _CubLog(format, ...) printf(format,__VA_ARGS__);\n #elif (CUB_PTX_ARCH >= 200)\n #define _CubLog(format, ...) printf(\"[block (%d,%d,%d), thread (%d,%d,%d)]: \" format, blockIdx.z, blockIdx.y, blockIdx.x, threadIdx.z, threadIdx.y, threadIdx.x, __VA_ARGS__);\n #endif\n #else\n // XXX shameless hack for clang around variadic printf...\n // Compilies w/o supplying -std=c++11 but shows warning,\n // so we sielence them :)\n #pragma clang diagnostic ignored \"-Wc++11-extensions\"\n #pragma clang diagnostic ignored \"-Wunnamed-type-template-args\"\n template \n inline __host__ __device__ void va_printf(char const* format, Args const&... args)\n {\n #ifdef __CUDA_ARCH__\n printf(format, blockIdx.z, blockIdx.y, blockIdx.x, threadIdx.z, threadIdx.y, threadIdx.x, args...);\n #else\n printf(format, args...);\n #endif\n }\n #ifndef __CUDA_ARCH__\n #define _CubLog(format, ...) va_printf(format,__VA_ARGS__);\n #else\n #define _CubLog(format, ...) va_printf(\"[block (%d,%d,%d), thread (%d,%d,%d)]: \" format, __VA_ARGS__);\n #endif\n #endif\n#endif\n\n\n\n\n/** @} */ // end group UtilMgmt\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_device.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Properties of a given CUDA device and the corresponding PTX bundle\n */\n\n#pragma once\n\n#include \"util_type.cuh\"\n#include \"util_arch.cuh\"\n#include \"util_debug.cuh\"\n#include \"util_namespace.cuh\"\n#include \"util_macro.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilMgmt\n * @{\n */\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/**\n * Alias temporaries to externally-allocated device storage (or simply return the amount of storage needed).\n */\ntemplate \n__host__ __device__ __forceinline__\ncudaError_t AliasTemporaries(\n void *d_temp_storage, ///< [in] %Device-accessible allocation of temporary storage. When NULL, the required allocation size is written to \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Size in bytes of \\t d_temp_storage allocation\n void* (&allocations)[ALLOCATIONS], ///< [in,out] Pointers to device allocations needed\n size_t (&allocation_sizes)[ALLOCATIONS]) ///< [in] Sizes in bytes of device allocations needed\n{\n const int ALIGN_BYTES = 256;\n const int ALIGN_MASK = ~(ALIGN_BYTES - 1);\n\n // Compute exclusive prefix sum over allocation requests\n size_t allocation_offsets[ALLOCATIONS];\n size_t bytes_needed = 0;\n for (int i = 0; i < ALLOCATIONS; ++i)\n {\n size_t allocation_bytes = (allocation_sizes[i] + ALIGN_BYTES - 1) & ALIGN_MASK;\n allocation_offsets[i] = bytes_needed;\n bytes_needed += allocation_bytes;\n }\n bytes_needed += ALIGN_BYTES - 1;\n\n // Check if the caller is simply requesting the size of the storage allocation\n if (!d_temp_storage)\n {\n temp_storage_bytes = bytes_needed;\n return cudaSuccess;\n }\n\n // Check if enough storage provided\n if (temp_storage_bytes < bytes_needed)\n {\n return CubDebug(cudaErrorInvalidValue);\n }\n\n // Alias\n d_temp_storage = (void *) ((size_t(d_temp_storage) + ALIGN_BYTES - 1) & ALIGN_MASK);\n for (int i = 0; i < ALLOCATIONS; ++i)\n {\n allocations[i] = static_cast(d_temp_storage) + allocation_offsets[i];\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Empty kernel for querying PTX manifest metadata (e.g., version) for the current device\n */\ntemplate \n__global__ void EmptyKernel(void) { }\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n/**\n * \\brief Retrieves the PTX version that will be used on the current device (major * 100 + minor * 10)\n */\nCUB_RUNTIME_FUNCTION __forceinline__ cudaError_t PtxVersion(int &ptx_version)\n{\n struct Dummy\n {\n /// Type definition of the EmptyKernel kernel entry point\n typedef void (*EmptyKernelPtr)();\n\n /// Force EmptyKernel to be generated if this class is used\n CUB_RUNTIME_FUNCTION __forceinline__\n EmptyKernelPtr Empty()\n {\n return EmptyKernel;\n }\n };\n\n\n#ifndef CUB_RUNTIME_ENABLED\n (void)ptx_version;\n\n // CUDA API calls not supported from this device\n return cudaErrorInvalidConfiguration;\n\n#elif (CUB_PTX_ARCH > 0)\n\n ptx_version = CUB_PTX_ARCH;\n return cudaSuccess;\n\n#else\n\n cudaError_t error = cudaSuccess;\n do\n {\n cudaFuncAttributes empty_kernel_attrs;\n if (CubDebug(error = cudaFuncGetAttributes(&empty_kernel_attrs, EmptyKernel))) break;\n ptx_version = empty_kernel_attrs.ptxVersion * 10;\n }\n while (0);\n\n return error;\n\n#endif\n}\n\n\n/**\n * \\brief Retrieves the SM version (major * 100 + minor * 10)\n */\nCUB_RUNTIME_FUNCTION __forceinline__ cudaError_t SmVersion(int &sm_version, int device_ordinal)\n{\n#ifndef CUB_RUNTIME_ENABLED\n (void)sm_version;\n (void)device_ordinal;\n\n // CUDA API calls not supported from this device\n return cudaErrorInvalidConfiguration;\n\n#else\n\n cudaError_t error = cudaSuccess;\n do\n {\n // Fill in SM version\n int major, minor;\n if (CubDebug(error = cudaDeviceGetAttribute(&major, cudaDevAttrComputeCapabilityMajor, device_ordinal))) break;\n if (CubDebug(error = cudaDeviceGetAttribute(&minor, cudaDevAttrComputeCapabilityMinor, device_ordinal))) break;\n sm_version = major * 100 + minor * 10;\n }\n while (0);\n\n return error;\n\n#endif\n}\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n/**\n * Synchronize the stream if specified\n */\nCUB_RUNTIME_FUNCTION __forceinline__\nstatic cudaError_t SyncStream(cudaStream_t stream)\n{\n#if (CUB_PTX_ARCH == 0)\n return cudaStreamSynchronize(stream);\n#else\n (void)stream;\n // Device can't yet sync on a specific stream\n return cudaDeviceSynchronize();\n#endif\n}\n\n\n/**\n * \\brief Computes maximum SM occupancy in thread blocks for executing the given kernel function pointer \\p kernel_ptr on the current device with \\p block_threads per thread block.\n *\n * \\par Snippet\n * The code snippet below illustrates the use of the MaxSmOccupancy function.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * template \n * __global__ void ExampleKernel()\n * {\n * // Allocate shared memory for BlockScan\n * __shared__ volatile T buffer[4096];\n *\n * ...\n * }\n *\n * ...\n *\n * // Determine SM occupancy for ExampleKernel specialized for unsigned char\n * int max_sm_occupancy;\n * MaxSmOccupancy(max_sm_occupancy, ExampleKernel, 64);\n *\n * // max_sm_occupancy <-- 4 on SM10\n * // max_sm_occupancy <-- 8 on SM20\n * // max_sm_occupancy <-- 12 on SM35\n *\n * \\endcode\n *\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t MaxSmOccupancy(\n int &max_sm_occupancy, ///< [out] maximum number of thread blocks that can reside on a single SM\n KernelPtr kernel_ptr, ///< [in] Kernel pointer for which to compute SM occupancy\n int block_threads, ///< [in] Number of threads per thread block\n int dynamic_smem_bytes = 0)\n{\n#ifndef CUB_RUNTIME_ENABLED\n (void)dynamic_smem_bytes;\n (void)block_threads;\n (void)kernel_ptr;\n (void)max_sm_occupancy;\n\n // CUDA API calls not supported from this device\n return CubDebug(cudaErrorInvalidConfiguration);\n\n#else\n\n return cudaOccupancyMaxActiveBlocksPerMultiprocessor (\n &max_sm_occupancy,\n kernel_ptr,\n block_threads,\n dynamic_smem_bytes);\n\n#endif // CUB_RUNTIME_ENABLED\n}\n\n\n/******************************************************************************\n * Policy management\n ******************************************************************************/\n\n/**\n * Kernel dispatch configuration\n */\nstruct KernelConfig\n{\n int block_threads;\n int items_per_thread;\n int tile_size;\n int sm_occupancy;\n\n CUB_RUNTIME_FUNCTION __forceinline__\n KernelConfig() : block_threads(0), items_per_thread(0), tile_size(0), sm_occupancy(0) {}\n\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n cudaError_t Init(KernelPtrT kernel_ptr)\n {\n block_threads = AgentPolicyT::BLOCK_THREADS;\n items_per_thread = AgentPolicyT::ITEMS_PER_THREAD;\n tile_size = block_threads * items_per_thread;\n cudaError_t retval = MaxSmOccupancy(sm_occupancy, kernel_ptr, block_threads);\n return retval;\n }\n};\n\n\n\n/// Helper for dispatching into a policy chain\ntemplate \nstruct ChainedPolicy\n{\n /// The policy for the active compiler pass\n typedef typename If<(CUB_PTX_ARCH < PTX_VERSION), typename PrevPolicyT::ActivePolicy, PolicyT>::Type ActivePolicy;\n\n /// Specializes and dispatches op in accordance to the first policy in the chain of adequate PTX version\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Invoke(int ptx_version, FunctorT &op)\n {\n if (ptx_version < PTX_VERSION) {\n return PrevPolicyT::Invoke(ptx_version, op);\n }\n return op.template Invoke();\n }\n};\n\n/// Helper for dispatching into a policy chain (end-of-chain specialization)\ntemplate \nstruct ChainedPolicy\n{\n /// The policy for the active compiler pass\n typedef PolicyT ActivePolicy;\n\n /// Specializes and dispatches op in accordance to the first policy in the chain of adequate PTX version\n template \n CUB_RUNTIME_FUNCTION __forceinline__\n static cudaError_t Invoke(int /*ptx_version*/, FunctorT &op) {\n return op.template Invoke();\n }\n};\n\n\n\n\n#endif // Do not document\n\n\n\n\n/** @} */ // end group UtilMgmt\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_macro.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Common C/C++ macro utilities\n ******************************************************************************/\n\n#pragma once\n\n#include \"util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilModule\n * @{\n */\n\n#ifndef CUB_ALIGN\n #if defined(_WIN32) || defined(_WIN64)\n /// Align struct\n #define CUB_ALIGN(bytes) __declspec(align(32))\n #else\n /// Align struct\n #define CUB_ALIGN(bytes) __attribute__((aligned(bytes)))\n #endif\n#endif\n\n#ifndef CUB_MAX\n /// Select maximum(a, b)\n #define CUB_MAX(a, b) (((b) > (a)) ? (b) : (a))\n#endif\n\n#ifndef CUB_MIN\n /// Select minimum(a, b)\n #define CUB_MIN(a, b) (((b) < (a)) ? (b) : (a))\n#endif\n\n#ifndef CUB_QUOTIENT_FLOOR\n /// Quotient of x/y rounded down to nearest integer\n #define CUB_QUOTIENT_FLOOR(x, y) ((x) / (y))\n#endif\n\n#ifndef CUB_QUOTIENT_CEILING\n /// Quotient of x/y rounded up to nearest integer\n #define CUB_QUOTIENT_CEILING(x, y) (((x) + (y) - 1) / (y))\n#endif\n\n#ifndef CUB_ROUND_UP_NEAREST\n /// x rounded up to the nearest multiple of y\n #define CUB_ROUND_UP_NEAREST(x, y) ((((x) + (y) - 1) / (y)) * y)\n#endif\n\n#ifndef CUB_ROUND_DOWN_NEAREST\n /// x rounded down to the nearest multiple of y\n #define CUB_ROUND_DOWN_NEAREST(x, y) (((x) / (y)) * y)\n#endif\n\n\n#ifndef CUB_STATIC_ASSERT\n #ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n #define CUB_CAT_(a, b) a ## b\n #define CUB_CAT(a, b) CUB_CAT_(a, b)\n #endif // DOXYGEN_SHOULD_SKIP_THIS\n\n /// Static assert\n #define CUB_STATIC_ASSERT(cond, msg) typedef int CUB_CAT(cub_static_assert, __LINE__)[(cond) ? 1 : -1]\n#endif\n\n/** @} */ // end group UtilModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_namespace.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Place-holder for prefixing the cub namespace\n */\n\n#pragma once\n\n// For example:\n//#define CUB_NS_PREFIX namespace thrust{ namespace detail {\n//#define CUB_NS_POSTFIX } }\n\n#ifndef CUB_NS_PREFIX\n#define CUB_NS_PREFIX\n#endif\n\n#ifndef CUB_NS_POSTFIX\n#define CUB_NS_POSTFIX\n#endif\n" }, { "path": "external/cub/cub/util_ptx.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * PTX intrinsics\n */\n\n\n#pragma once\n\n#include \"util_type.cuh\"\n#include \"util_arch.cuh\"\n#include \"util_namespace.cuh\"\n#include \"util_debug.cuh\"\n\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilPtx\n * @{\n */\n\n\n/******************************************************************************\n * PTX helper macros\n ******************************************************************************/\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n/**\n * Register modifier for pointer-types (for inlining PTX assembly)\n */\n#if defined(_WIN64) || defined(__LP64__)\n #define __CUB_LP64__ 1\n // 64-bit register modifier for inlined asm\n #define _CUB_ASM_PTR_ \"l\"\n #define _CUB_ASM_PTR_SIZE_ \"u64\"\n#else\n #define __CUB_LP64__ 0\n // 32-bit register modifier for inlined asm\n #define _CUB_ASM_PTR_ \"r\"\n #define _CUB_ASM_PTR_SIZE_ \"u32\"\n#endif\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/******************************************************************************\n * Inlined PTX intrinsics\n ******************************************************************************/\n\n/**\n * \\brief Shift-right then add. Returns (\\p x >> \\p shift) + \\p addend.\n */\n__device__ __forceinline__ unsigned int SHR_ADD(\n unsigned int x,\n unsigned int shift,\n unsigned int addend)\n{\n unsigned int ret;\n#if CUB_PTX_ARCH >= 200\n asm (\"vshr.u32.u32.u32.clamp.add %0, %1, %2, %3;\" :\n \"=r\"(ret) : \"r\"(x), \"r\"(shift), \"r\"(addend));\n#else\n ret = (x >> shift) + addend;\n#endif\n return ret;\n}\n\n\n/**\n * \\brief Shift-left then add. Returns (\\p x << \\p shift) + \\p addend.\n */\n__device__ __forceinline__ unsigned int SHL_ADD(\n unsigned int x,\n unsigned int shift,\n unsigned int addend)\n{\n unsigned int ret;\n#if CUB_PTX_ARCH >= 200\n asm (\"vshl.u32.u32.u32.clamp.add %0, %1, %2, %3;\" :\n \"=r\"(ret) : \"r\"(x), \"r\"(shift), \"r\"(addend));\n#else\n ret = (x << shift) + addend;\n#endif\n return ret;\n}\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n/**\n * Bitfield-extract.\n */\ntemplate \n__device__ __forceinline__ unsigned int BFE(\n UnsignedBits source,\n unsigned int bit_start,\n unsigned int num_bits,\n Int2Type /*byte_len*/)\n{\n unsigned int bits;\n#if CUB_PTX_ARCH >= 200\n asm (\"bfe.u32 %0, %1, %2, %3;\" : \"=r\"(bits) : \"r\"((unsigned int) source), \"r\"(bit_start), \"r\"(num_bits));\n#else\n const unsigned int MASK = (1 << num_bits) - 1;\n bits = (source >> bit_start) & MASK;\n#endif\n return bits;\n}\n\n\n/**\n * Bitfield-extract for 64-bit types.\n */\ntemplate \n__device__ __forceinline__ unsigned int BFE(\n UnsignedBits source,\n unsigned int bit_start,\n unsigned int num_bits,\n Int2Type<8> /*byte_len*/)\n{\n const unsigned long long MASK = (1ull << num_bits) - 1;\n return (source >> bit_start) & MASK;\n}\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n/**\n * \\brief Bitfield-extract. Extracts \\p num_bits from \\p source starting at bit-offset \\p bit_start. The input \\p source may be an 8b, 16b, 32b, or 64b unsigned integer type.\n */\ntemplate \n__device__ __forceinline__ unsigned int BFE(\n UnsignedBits source,\n unsigned int bit_start,\n unsigned int num_bits)\n{\n return BFE(source, bit_start, num_bits, Int2Type());\n}\n\n\n/**\n * \\brief Bitfield insert. Inserts the \\p num_bits least significant bits of \\p y into \\p x at bit-offset \\p bit_start.\n */\n__device__ __forceinline__ void BFI(\n unsigned int &ret,\n unsigned int x,\n unsigned int y,\n unsigned int bit_start,\n unsigned int num_bits)\n{\n#if CUB_PTX_ARCH >= 200\n asm (\"bfi.b32 %0, %1, %2, %3, %4;\" :\n \"=r\"(ret) : \"r\"(y), \"r\"(x), \"r\"(bit_start), \"r\"(num_bits));\n#else\n x <<= bit_start;\n unsigned int MASK_X = ((1 << num_bits) - 1) << bit_start;\n unsigned int MASK_Y = ~MASK_X;\n ret = (y & MASK_Y) | (x & MASK_X);\n#endif\n}\n\n\n/**\n * \\brief Three-operand add. Returns \\p x + \\p y + \\p z.\n */\n__device__ __forceinline__ unsigned int IADD3(unsigned int x, unsigned int y, unsigned int z)\n{\n#if CUB_PTX_ARCH >= 200\n asm (\"vadd.u32.u32.u32.add %0, %1, %2, %3;\" : \"=r\"(x) : \"r\"(x), \"r\"(y), \"r\"(z));\n#else\n x = x + y + z;\n#endif\n return x;\n}\n\n\n/**\n * \\brief Byte-permute. Pick four arbitrary bytes from two 32-bit registers, and reassemble them into a 32-bit destination register. For SM2.0 or later.\n *\n * \\par\n * The bytes in the two source registers \\p a and \\p b are numbered from 0 to 7:\n * {\\p b, \\p a} = {{b7, b6, b5, b4}, {b3, b2, b1, b0}}. For each of the four bytes\n * {b3, b2, b1, b0} selected in the return value, a 4-bit selector is defined within\n * the four lower \"nibbles\" of \\p index: {\\p index } = {n7, n6, n5, n4, n3, n2, n1, n0}\n *\n * \\par Snippet\n * The code snippet below illustrates byte-permute.\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * int a = 0x03020100;\n * int b = 0x07060504;\n * int index = 0x00007531;\n *\n * int selected = PRMT(a, b, index); // 0x07050301\n *\n * \\endcode\n *\n */\n__device__ __forceinline__ int PRMT(unsigned int a, unsigned int b, unsigned int index)\n{\n int ret;\n asm (\"prmt.b32 %0, %1, %2, %3;\" : \"=r\"(ret) : \"r\"(a), \"r\"(b), \"r\"(index));\n return ret;\n}\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n/**\n * Sync-threads barrier.\n */\n__device__ __forceinline__ void BAR(int count)\n{\n asm volatile(\"bar.sync 1, %0;\" : : \"r\"(count));\n}\n\n/**\n * CTA barrier\n */\n__device__ __forceinline__ void CTA_SYNC()\n{\n __syncthreads();\n}\n\n\n/**\n * CTA barrier with predicate\n */\n__device__ __forceinline__ int CTA_SYNC_AND(int p)\n{\n return __syncthreads_and(p);\n}\n\n\n/**\n * Warp barrier\n */\n__device__ __forceinline__ void WARP_SYNC(unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n __syncwarp(member_mask);\n#endif\n}\n\n\n/**\n * Warp any\n */\n__device__ __forceinline__ int WARP_ANY(int predicate, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n return __any_sync(member_mask, predicate);\n#else\n return ::__any(predicate);\n#endif\n}\n\n\n/**\n * Warp any\n */\n__device__ __forceinline__ int WARP_ALL(int predicate, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n return __all_sync(member_mask, predicate);\n#else\n return ::__all(predicate);\n#endif\n}\n\n\n/**\n * Warp ballot\n */\n__device__ __forceinline__ int WARP_BALLOT(int predicate, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n return __ballot_sync(member_mask, predicate);\n#else\n return __ballot(predicate);\n#endif\n}\n\n/**\n * Warp synchronous shfl_up\n */\n__device__ __forceinline__ \nunsigned int SHFL_UP_SYNC(unsigned int word, int src_offset, int first_lane, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\"shfl.sync.up.b32 %0, %1, %2, %3, %4;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_offset), \"r\"(first_lane), \"r\"(member_mask));\n#else\n asm volatile(\"shfl.up.b32 %0, %1, %2, %3;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_offset), \"r\"(first_lane));\n#endif\n return word;\n}\n\n/**\n * Warp synchronous shfl_down\n */\n__device__ __forceinline__ \nunsigned int SHFL_DOWN_SYNC(unsigned int word, int src_offset, int last_lane, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\"shfl.sync.down.b32 %0, %1, %2, %3, %4;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_offset), \"r\"(last_lane), \"r\"(member_mask));\n#else\n asm volatile(\"shfl.down.b32 %0, %1, %2, %3;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_offset), \"r\"(last_lane));\n#endif\n return word;\n}\n\n/**\n * Warp synchronous shfl_idx\n */\n__device__ __forceinline__ \nunsigned int SHFL_IDX_SYNC(unsigned int word, int src_lane, int last_lane, unsigned int member_mask)\n{\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\"shfl.sync.idx.b32 %0, %1, %2, %3, %4;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_lane), \"r\"(last_lane), \"r\"(member_mask));\n#else\n asm volatile(\"shfl.idx.b32 %0, %1, %2, %3;\"\n : \"=r\"(word) : \"r\"(word), \"r\"(src_lane), \"r\"(last_lane));\n#endif\n return word;\n}\n\n/**\n * Floating point multiply. (Mantissa LSB rounds towards zero.)\n */\n__device__ __forceinline__ float FMUL_RZ(float a, float b)\n{\n float d;\n asm (\"mul.rz.f32 %0, %1, %2;\" : \"=f\"(d) : \"f\"(a), \"f\"(b));\n return d;\n}\n\n\n/**\n * Floating point multiply-add. (Mantissa LSB rounds towards zero.)\n */\n__device__ __forceinline__ float FFMA_RZ(float a, float b, float c)\n{\n float d;\n asm (\"fma.rz.f32 %0, %1, %2, %3;\" : \"=f\"(d) : \"f\"(a), \"f\"(b), \"f\"(c));\n return d;\n}\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n/**\n * \\brief Terminates the calling thread\n */\n__device__ __forceinline__ void ThreadExit() {\n asm volatile(\"exit;\");\n} \n\n\n/**\n * \\brief Abort execution and generate an interrupt to the host CPU\n */\n__device__ __forceinline__ void ThreadTrap() {\n asm volatile(\"trap;\");\n}\n\n\n/**\n * \\brief Returns the row-major linear thread identifier for a multidimensional thread block\n */\n__device__ __forceinline__ int RowMajorTid(int block_dim_x, int block_dim_y, int block_dim_z)\n{\n return ((block_dim_z == 1) ? 0 : (threadIdx.z * block_dim_x * block_dim_y)) +\n ((block_dim_y == 1) ? 0 : (threadIdx.y * block_dim_x)) +\n threadIdx.x;\n}\n\n\n/**\n * \\brief Returns the warp lane ID of the calling thread\n */\n__device__ __forceinline__ unsigned int LaneId()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%laneid;\" : \"=r\"(ret) );\n return ret;\n}\n\n\n/**\n * \\brief Returns the warp ID of the calling thread. Warp ID is guaranteed to be unique among warps, but may not correspond to a zero-based ranking within the thread block.\n */\n__device__ __forceinline__ unsigned int WarpId()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%warpid;\" : \"=r\"(ret) );\n return ret;\n}\n\n/**\n * \\brief Returns the warp lane mask of all lanes less than the calling thread\n */\n__device__ __forceinline__ unsigned int LaneMaskLt()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%lanemask_lt;\" : \"=r\"(ret) );\n return ret;\n}\n\n/**\n * \\brief Returns the warp lane mask of all lanes less than or equal to the calling thread\n */\n__device__ __forceinline__ unsigned int LaneMaskLe()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%lanemask_le;\" : \"=r\"(ret) );\n return ret;\n}\n\n/**\n * \\brief Returns the warp lane mask of all lanes greater than the calling thread\n */\n__device__ __forceinline__ unsigned int LaneMaskGt()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%lanemask_gt;\" : \"=r\"(ret) );\n return ret;\n}\n\n/**\n * \\brief Returns the warp lane mask of all lanes greater than or equal to the calling thread\n */\n__device__ __forceinline__ unsigned int LaneMaskGe()\n{\n unsigned int ret;\n asm (\"mov.u32 %0, %%lanemask_ge;\" : \"=r\"(ret) );\n return ret;\n}\n\n/** @} */ // end group UtilPtx\n\n\n\n\n/**\n * \\brief Shuffle-up for any data type. Each warp-lanei obtains the value \\p input contributed by warp-lanei-src_offset. For thread lanes \\e i < src_offset, the thread's own \\p input is returned to the thread. ![](shfl_up_logo.png)\n * \\ingroup WarpModule\n *\n * \\par\n * - Available only for SM3.0 or newer\n *\n * \\par Snippet\n * The code snippet below illustrates each thread obtaining a \\p double value from the\n * predecessor of its predecessor.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Obtain one input item per thread\n * double thread_data = ...\n *\n * // Obtain item from two ranks below\n * double peer_data = ShuffleUp(thread_data, 2, 0, 0xffffffff);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the first warp of threads is {1.0, 2.0, 3.0, 4.0, 5.0, ..., 32.0}.\n * The corresponding output \\p peer_data will be {1.0, 2.0, 1.0, 2.0, 3.0, ..., 30.0}.\n *\n */\ntemplate \n__device__ __forceinline__ T ShuffleUp(\n T input, ///< [in] The value to broadcast\n int src_offset, ///< [in] The relative down-offset of the peer to read from\n int first_lane, ///< [in] Index of first lane in segment (typically 0)\n unsigned int member_mask) ///< [in] 32-bit mask of participating warp lanes\n{\n typedef typename UnitWord::ShuffleWord ShuffleWord;\n\n const int WORDS = (sizeof(T) + sizeof(ShuffleWord) - 1) / sizeof(ShuffleWord);\n \n T output;\n ShuffleWord *output_alias = reinterpret_cast(&output);\n ShuffleWord *input_alias = reinterpret_cast(&input);\n\n unsigned int shuffle_word;\n shuffle_word = SHFL_UP_SYNC((unsigned int)input_alias[0], src_offset, first_lane, member_mask);\n output_alias[0] = shuffle_word;\n\n #pragma unroll\n for (int WORD = 1; WORD < WORDS; ++WORD)\n {\n shuffle_word = SHFL_UP_SYNC((unsigned int)input_alias[WORD], src_offset, first_lane, member_mask);\n output_alias[WORD] = shuffle_word;\n }\n\n return output;\n}\n\n\n/**\n * \\brief Shuffle-down for any data type. Each warp-lanei obtains the value \\p input contributed by warp-lanei+src_offset. For thread lanes \\e i >= WARP_THREADS, the thread's own \\p input is returned to the thread. ![](shfl_down_logo.png)\n * \\ingroup WarpModule\n *\n * \\par\n * - Available only for SM3.0 or newer\n *\n * \\par Snippet\n * The code snippet below illustrates each thread obtaining a \\p double value from the\n * successor of its successor.\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Obtain one input item per thread\n * double thread_data = ...\n *\n * // Obtain item from two ranks below\n * double peer_data = ShuffleDown(thread_data, 2, 31, 0xffffffff);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the first warp of threads is {1.0, 2.0, 3.0, 4.0, 5.0, ..., 32.0}.\n * The corresponding output \\p peer_data will be {3.0, 4.0, 5.0, 6.0, 7.0, ..., 32.0}.\n *\n */\ntemplate \n__device__ __forceinline__ T ShuffleDown(\n T input, ///< [in] The value to broadcast\n int src_offset, ///< [in] The relative up-offset of the peer to read from\n int last_lane, ///< [in] Index of first lane in segment (typically 31)\n unsigned int member_mask) ///< [in] 32-bit mask of participating warp lanes\n{\n typedef typename UnitWord::ShuffleWord ShuffleWord;\n\n const int WORDS = (sizeof(T) + sizeof(ShuffleWord) - 1) / sizeof(ShuffleWord);\n\n T output;\n ShuffleWord *output_alias = reinterpret_cast(&output);\n ShuffleWord *input_alias = reinterpret_cast(&input);\n\n unsigned int shuffle_word;\n shuffle_word = SHFL_DOWN_SYNC((unsigned int)input_alias[0], src_offset, last_lane, member_mask);\n output_alias[0] = shuffle_word;\n\n #pragma unroll\n for (int WORD = 1; WORD < WORDS; ++WORD)\n {\n shuffle_word = SHFL_DOWN_SYNC((unsigned int)input_alias[WORD], src_offset, last_lane, member_mask);\n output_alias[WORD] = shuffle_word;\n }\n\n return output;\n}\n\n\n/**\n * \\brief Shuffle-broadcast for any data type. Each warp-lanei obtains the value \\p input\n * contributed by warp-lanesrc_lane. For \\p src_lane < 0 or \\p src_lane >= WARP_THREADS,\n * then the thread's own \\p input is returned to the thread. ![](shfl_broadcast_logo.png)\n *\n * \\ingroup WarpModule\n *\n * \\par\n * - Available only for SM3.0 or newer\n *\n * \\par Snippet\n * The code snippet below illustrates each thread obtaining a \\p double value from warp-lane0.\n *\n * \\par\n * \\code\n * #include // or equivalently \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Obtain one input item per thread\n * double thread_data = ...\n *\n * // Obtain item from thread 0\n * double peer_data = ShuffleIndex(thread_data, 0, 32, 0xffffffff);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the first warp of threads is {1.0, 2.0, 3.0, 4.0, 5.0, ..., 32.0}.\n * The corresponding output \\p peer_data will be {1.0, 1.0, 1.0, 1.0, 1.0, ..., 1.0}.\n *\n */\ntemplate \n__device__ __forceinline__ T ShuffleIndex(\n T input, ///< [in] The value to broadcast\n int src_lane, ///< [in] Which warp lane is to do the broadcasting\n int logical_warp_threads, ///< [in] Number of threads per logical warp\n unsigned int member_mask) ///< [in] 32-bit mask of participating warp lanes\n{\n typedef typename UnitWord::ShuffleWord ShuffleWord;\n\n const int WORDS = (sizeof(T) + sizeof(ShuffleWord) - 1) / sizeof(ShuffleWord);\n\n T output;\n ShuffleWord *output_alias = reinterpret_cast(&output);\n ShuffleWord *input_alias = reinterpret_cast(&input);\n\n unsigned int shuffle_word;\n shuffle_word = SHFL_IDX_SYNC((unsigned int)input_alias[0],\n src_lane,\n logical_warp_threads - 1,\n member_mask);\n\n output_alias[0] = shuffle_word;\n\n #pragma unroll\n for (int WORD = 1; WORD < WORDS; ++WORD)\n {\n shuffle_word = SHFL_IDX_SYNC((unsigned int)input_alias[WORD],\n src_lane,\n logical_warp_threads - 1,\n member_mask);\n\n output_alias[WORD] = shuffle_word;\n }\n\n return output;\n}\n\n\n\n/**\n * Compute a 32b mask of threads having the same least-significant\n * LABEL_BITS of \\p label as the calling thread.\n */\ntemplate \ninline __device__ unsigned int MatchAny(unsigned int label)\n{\n unsigned int retval;\n\n // Extract masks of common threads for each bit\n #pragma unroll\n for (int BIT = 0; BIT < LABEL_BITS; ++BIT)\n {\n unsigned int mask;\n unsigned int current_bit = 1 << BIT;\n asm (\"{\\n\"\n \" .reg .pred p;\\n\"\n \" and.b32 %0, %1, %2;\"\n \" setp.eq.u32 p, %0, %2;\\n\"\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n \" vote.ballot.sync.b32 %0, p, 0xffffffff;\\n\"\n#else\n \" vote.ballot.b32 %0, p;\\n\"\n#endif\n \" @!p not.b32 %0, %0;\\n\"\n \"}\\n\" : \"=r\"(mask) : \"r\"(label), \"r\"(current_bit));\n\n // Remove peers who differ\n retval = (BIT == 0) ? mask : retval & mask;\n }\n\n return retval;\n\n// // VOLTA match\n// unsigned int retval;\n// asm (\"{\\n\"\n// \" match.any.sync.b32 %0, %1, 0xffffffff;\\n\"\n// \"}\\n\" : \"=r\"(retval) : \"r\"(label));\n// return retval;\n\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/util_type.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * Common type manipulation (metaprogramming) utilities\n */\n\n#pragma once\n\n#include \n#include \n#include \n\n#include \"util_macro.cuh\"\n#include \"util_arch.cuh\"\n#include \"util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup UtilModule\n * @{\n */\n\n\n\n/******************************************************************************\n * Type equality\n ******************************************************************************/\n\n/**\n * \\brief Type selection (IF ? ThenType : ElseType)\n */\ntemplate \nstruct If\n{\n /// Conditional type result\n typedef ThenType Type; // true\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct If\n{\n typedef ElseType Type; // false\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n/******************************************************************************\n * Conditional types\n ******************************************************************************/\n\n/**\n * \\brief Type equality test\n */\ntemplate \nstruct Equals\n{\n enum {\n VALUE = 0,\n NEGATE = 1\n };\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct Equals \n{\n enum {\n VALUE = 1,\n NEGATE = 0\n };\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/******************************************************************************\n * Static math\n ******************************************************************************/\n\n/**\n * \\brief Statically determine log2(N), rounded up.\n *\n * For example:\n * Log2<8>::VALUE // 3\n * Log2<3>::VALUE // 2\n */\ntemplate \nstruct Log2\n{\n /// Static logarithm value\n enum { VALUE = Log2> 1), COUNT + 1>::VALUE }; // Inductive case\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct Log2\n{\n enum {VALUE = (1 << (COUNT - 1) < N) ? // Base case\n COUNT :\n COUNT - 1 };\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/**\n * \\brief Statically determine if N is a power-of-two\n */\ntemplate \nstruct PowerOfTwo\n{\n enum { VALUE = ((N & (N - 1)) == 0) };\n};\n\n\n\n/******************************************************************************\n * Pointer vs. iterator detection\n ******************************************************************************/\n\n/**\n * \\brief Pointer vs. iterator\n */\ntemplate \nstruct IsPointer\n{\n enum { VALUE = 0 };\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct IsPointer\n{\n enum { VALUE = 1 };\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n/******************************************************************************\n * Qualifier detection\n ******************************************************************************/\n\n/**\n * \\brief Volatile modifier test\n */\ntemplate \nstruct IsVolatile\n{\n enum { VALUE = 0 };\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct IsVolatile\n{\n enum { VALUE = 1 };\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/******************************************************************************\n * Qualifier removal\n ******************************************************************************/\n\n/**\n * \\brief Removes \\p const and \\p volatile qualifiers from type \\p Tp.\n *\n * For example:\n * typename RemoveQualifiers::Type // int;\n */\ntemplate \nstruct RemoveQualifiers\n{\n /// Type without \\p const and \\p volatile qualifiers\n typedef Up Type;\n};\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\ntemplate \nstruct RemoveQualifiers\n{\n typedef Up Type;\n};\n\ntemplate \nstruct RemoveQualifiers\n{\n typedef Up Type;\n};\n\ntemplate \nstruct RemoveQualifiers\n{\n typedef Up Type;\n};\n\n\n/******************************************************************************\n * Marker types\n ******************************************************************************/\n\n/**\n * \\brief A simple \"NULL\" marker type\n */\nstruct NullType\n{\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n template \n __host__ __device__ __forceinline__ NullType& operator =(const T&) { return *this; }\n\n __host__ __device__ __forceinline__ bool operator ==(const NullType&) { return true; }\n\n __host__ __device__ __forceinline__ bool operator !=(const NullType&) { return false; }\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n};\n\n\n/**\n * \\brief Allows for the treatment of an integral constant as a type at compile-time (e.g., to achieve static call dispatch based on constant integral values)\n */\ntemplate \nstruct Int2Type\n{\n enum {VALUE = A};\n};\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/******************************************************************************\n * Size and alignment\n ******************************************************************************/\n\n/// Structure alignment\ntemplate \nstruct AlignBytes\n{\n struct Pad\n {\n T val;\n char byte;\n };\n\n enum\n {\n /// The \"true CUDA\" alignment of T in bytes\n ALIGN_BYTES = sizeof(Pad) - sizeof(T)\n };\n\n /// The \"truly aligned\" type\n typedef T Type;\n};\n\n// Specializations where host C++ compilers (e.g., 32-bit Windows) may disagree\n// with device C++ compilers (EDG) on types passed as template parameters through\n// kernel functions\n\n#define __CUB_ALIGN_BYTES(t, b) \\\n template <> struct AlignBytes \\\n { enum { ALIGN_BYTES = b }; typedef __align__(b) t Type; };\n\n__CUB_ALIGN_BYTES(short4, 8)\n__CUB_ALIGN_BYTES(ushort4, 8)\n__CUB_ALIGN_BYTES(int2, 8)\n__CUB_ALIGN_BYTES(uint2, 8)\n__CUB_ALIGN_BYTES(long long, 8)\n__CUB_ALIGN_BYTES(unsigned long long, 8)\n__CUB_ALIGN_BYTES(float2, 8)\n__CUB_ALIGN_BYTES(double, 8)\n#ifdef _WIN32\n __CUB_ALIGN_BYTES(long2, 8)\n __CUB_ALIGN_BYTES(ulong2, 8)\n#else\n __CUB_ALIGN_BYTES(long2, 16)\n __CUB_ALIGN_BYTES(ulong2, 16)\n#endif\n__CUB_ALIGN_BYTES(int4, 16)\n__CUB_ALIGN_BYTES(uint4, 16)\n__CUB_ALIGN_BYTES(float4, 16)\n__CUB_ALIGN_BYTES(long4, 16)\n__CUB_ALIGN_BYTES(ulong4, 16)\n__CUB_ALIGN_BYTES(longlong2, 16)\n__CUB_ALIGN_BYTES(ulonglong2, 16)\n__CUB_ALIGN_BYTES(double2, 16)\n__CUB_ALIGN_BYTES(longlong4, 16)\n__CUB_ALIGN_BYTES(ulonglong4, 16)\n__CUB_ALIGN_BYTES(double4, 16)\n\ntemplate struct AlignBytes : AlignBytes {};\ntemplate struct AlignBytes : AlignBytes {};\ntemplate struct AlignBytes : AlignBytes {};\n\n\n/// Unit-words of data movement\ntemplate \nstruct UnitWord\n{\n enum {\n ALIGN_BYTES = AlignBytes::ALIGN_BYTES\n };\n\n template \n struct IsMultiple\n {\n enum {\n UNIT_ALIGN_BYTES = AlignBytes::ALIGN_BYTES,\n IS_MULTIPLE = (sizeof(T) % sizeof(Unit) == 0) && (ALIGN_BYTES % UNIT_ALIGN_BYTES == 0)\n };\n };\n\n /// Biggest shuffle word that T is a whole multiple of and is not larger than the alignment of T\n typedef typename If::IS_MULTIPLE,\n unsigned int,\n typename If::IS_MULTIPLE,\n unsigned short,\n unsigned char>::Type>::Type ShuffleWord;\n\n /// Biggest volatile word that T is a whole multiple of and is not larger than the alignment of T\n typedef typename If::IS_MULTIPLE,\n unsigned long long,\n ShuffleWord>::Type VolatileWord;\n\n /// Biggest memory-access word that T is a whole multiple of and is not larger than the alignment of T\n typedef typename If::IS_MULTIPLE,\n ulonglong2,\n VolatileWord>::Type DeviceWord;\n\n /// Biggest texture reference word that T is a whole multiple of and is not larger than the alignment of T\n typedef typename If::IS_MULTIPLE,\n uint4,\n typename If::IS_MULTIPLE,\n uint2,\n ShuffleWord>::Type>::Type TextureWord;\n};\n\n\n// float2 specialization workaround (for SM10-SM13)\ntemplate <>\nstruct UnitWord \n{\n typedef int ShuffleWord;\n#if (CUB_PTX_ARCH > 0) && (CUB_PTX_ARCH <= 130)\n typedef float VolatileWord;\n typedef uint2 DeviceWord;\n#else\n typedef unsigned long long VolatileWord;\n typedef unsigned long long DeviceWord;\n#endif\n typedef float2 TextureWord;\n};\n\n// float4 specialization workaround (for SM10-SM13)\ntemplate <>\nstruct UnitWord \n{\n typedef int ShuffleWord;\n#if (CUB_PTX_ARCH > 0) && (CUB_PTX_ARCH <= 130)\n typedef float VolatileWord;\n typedef uint4 DeviceWord;\n#else\n typedef unsigned long long VolatileWord;\n typedef ulonglong2 DeviceWord;\n#endif\n typedef float4 TextureWord;\n};\n\n\n// char2 specialization workaround (for SM10-SM13)\ntemplate <>\nstruct UnitWord \n{\n typedef unsigned short ShuffleWord;\n#if (CUB_PTX_ARCH > 0) && (CUB_PTX_ARCH <= 130)\n typedef unsigned short VolatileWord;\n typedef short DeviceWord;\n#else\n typedef unsigned short VolatileWord;\n typedef unsigned short DeviceWord;\n#endif\n typedef unsigned short TextureWord;\n};\n\n\ntemplate struct UnitWord : UnitWord {};\ntemplate struct UnitWord : UnitWord {};\ntemplate struct UnitWord : UnitWord {};\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n/******************************************************************************\n * Vector type inference utilities.\n ******************************************************************************/\n\n/**\n * \\brief Exposes a member typedef \\p Type that names the corresponding CUDA vector type if one exists. Otherwise \\p Type refers to the CubVector structure itself, which will wrap the corresponding \\p x, \\p y, etc. vector fields.\n */\ntemplate struct CubVector;\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\nenum\n{\n /// The maximum number of elements in CUDA vector types\n MAX_VEC_ELEMENTS = 4,\n};\n\n\n/**\n * Generic vector-1 type\n */\ntemplate \nstruct CubVector\n{\n T x;\n\n typedef T BaseType;\n typedef CubVector Type;\n};\n\n/**\n * Generic vector-2 type\n */\ntemplate \nstruct CubVector\n{\n T x;\n T y;\n\n typedef T BaseType;\n typedef CubVector Type;\n};\n\n/**\n * Generic vector-3 type\n */\ntemplate \nstruct CubVector\n{\n T x;\n T y;\n T z;\n\n typedef T BaseType;\n typedef CubVector Type;\n};\n\n/**\n * Generic vector-4 type\n */\ntemplate \nstruct CubVector\n{\n T x;\n T y;\n T z;\n T w;\n\n typedef T BaseType;\n typedef CubVector Type;\n};\n\n\n/**\n * Macro for expanding partially-specialized built-in vector types\n */\n#define CUB_DEFINE_VECTOR_TYPE(base_type,short_type) \\\n \\\n template<> struct CubVector : short_type##1 \\\n { \\\n typedef base_type BaseType; \\\n typedef short_type##1 Type; \\\n __host__ __device__ __forceinline__ CubVector operator+(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x + other.x; \\\n return retval; \\\n } \\\n __host__ __device__ __forceinline__ CubVector operator-(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x - other.x; \\\n return retval; \\\n } \\\n }; \\\n \\\n template<> struct CubVector : short_type##2 \\\n { \\\n typedef base_type BaseType; \\\n typedef short_type##2 Type; \\\n __host__ __device__ __forceinline__ CubVector operator+(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x + other.x; \\\n retval.y = y + other.y; \\\n return retval; \\\n } \\\n __host__ __device__ __forceinline__ CubVector operator-(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x - other.x; \\\n retval.y = y - other.y; \\\n return retval; \\\n } \\\n }; \\\n \\\n template<> struct CubVector : short_type##3 \\\n { \\\n typedef base_type BaseType; \\\n typedef short_type##3 Type; \\\n __host__ __device__ __forceinline__ CubVector operator+(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x + other.x; \\\n retval.y = y + other.y; \\\n retval.z = z + other.z; \\\n return retval; \\\n } \\\n __host__ __device__ __forceinline__ CubVector operator-(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x - other.x; \\\n retval.y = y - other.y; \\\n retval.z = z - other.z; \\\n return retval; \\\n } \\\n }; \\\n \\\n template<> struct CubVector : short_type##4 \\\n { \\\n typedef base_type BaseType; \\\n typedef short_type##4 Type; \\\n __host__ __device__ __forceinline__ CubVector operator+(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x + other.x; \\\n retval.y = y + other.y; \\\n retval.z = z + other.z; \\\n retval.w = w + other.w; \\\n return retval; \\\n } \\\n __host__ __device__ __forceinline__ CubVector operator-(const CubVector &other) const { \\\n CubVector retval; \\\n retval.x = x - other.x; \\\n retval.y = y - other.y; \\\n retval.z = z - other.z; \\\n retval.w = w - other.w; \\\n return retval; \\\n } \\\n };\n\n\n\n// Expand CUDA vector types for built-in primitives\nCUB_DEFINE_VECTOR_TYPE(char, char)\nCUB_DEFINE_VECTOR_TYPE(signed char, char)\nCUB_DEFINE_VECTOR_TYPE(short, short)\nCUB_DEFINE_VECTOR_TYPE(int, int)\nCUB_DEFINE_VECTOR_TYPE(long, long)\nCUB_DEFINE_VECTOR_TYPE(long long, longlong)\nCUB_DEFINE_VECTOR_TYPE(unsigned char, uchar)\nCUB_DEFINE_VECTOR_TYPE(unsigned short, ushort)\nCUB_DEFINE_VECTOR_TYPE(unsigned int, uint)\nCUB_DEFINE_VECTOR_TYPE(unsigned long, ulong)\nCUB_DEFINE_VECTOR_TYPE(unsigned long long, ulonglong)\nCUB_DEFINE_VECTOR_TYPE(float, float)\nCUB_DEFINE_VECTOR_TYPE(double, double)\nCUB_DEFINE_VECTOR_TYPE(bool, uchar)\n\n// Undefine macros\n#undef CUB_DEFINE_VECTOR_TYPE\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n\n/******************************************************************************\n * Wrapper types\n ******************************************************************************/\n\n/**\n * \\brief A storage-backing wrapper that allows types with non-trivial constructors to be aliased in unions\n */\ntemplate \nstruct Uninitialized\n{\n /// Biggest memory-access word that T is a whole multiple of and is not larger than the alignment of T\n typedef typename UnitWord::DeviceWord DeviceWord;\n\n enum\n {\n WORDS = sizeof(T) / sizeof(DeviceWord)\n };\n\n /// Backing storage\n DeviceWord storage[WORDS];\n\n /// Alias\n __host__ __device__ __forceinline__ T& Alias()\n {\n return reinterpret_cast(*this);\n }\n};\n\n\n/**\n * \\brief A key identifier paired with a corresponding value\n */\ntemplate <\n typename _Key,\n typename _Value\n#if defined(_WIN32) && !defined(_WIN64)\n , bool KeyIsLT = (AlignBytes<_Key>::ALIGN_BYTES < AlignBytes<_Value>::ALIGN_BYTES)\n , bool ValIsLT = (AlignBytes<_Value>::ALIGN_BYTES < AlignBytes<_Key>::ALIGN_BYTES)\n#endif // #if defined(_WIN32) && !defined(_WIN64)\n >\nstruct KeyValuePair\n{\n typedef _Key Key; ///< Key data type\n typedef _Value Value; ///< Value data type\n\n Key key; ///< Item key\n Value value; ///< Item value\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair() {}\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair(Key const& key, Value const& value) : key(key), value(value) {}\n\n /// Inequality operator\n __host__ __device__ __forceinline__ bool operator !=(const KeyValuePair &b)\n {\n return (value != b.value) || (key != b.key);\n }\n};\n\n#if defined(_WIN32) && !defined(_WIN64)\n\n/**\n * Win32 won't do 16B alignment. This can present two problems for\n * should-be-16B-aligned (but actually 8B aligned) built-in and intrinsics members:\n * 1) If a smaller-aligned item were to be listed first, the host compiler places the\n * should-be-16B item at too early an offset (and disagrees with device compiler)\n * 2) Or, if a smaller-aligned item lists second, the host compiler gets the size\n * of the struct wrong (and disagrees with device compiler)\n *\n * So we put the larger-should-be-aligned item first, and explicitly pad the\n * end of the struct\n */\n\n/// Smaller key specialization\ntemplate \nstruct KeyValuePair\n{\n typedef K Key;\n typedef V Value;\n\n typedef char Pad[AlignBytes::ALIGN_BYTES - AlignBytes::ALIGN_BYTES];\n\n Value value; // Value has larger would-be alignment and goes first\n Key key;\n Pad pad;\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair() {}\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair(Key const& key, Value const& value) : key(key), value(value) {}\n\n /// Inequality operator\n __host__ __device__ __forceinline__ bool operator !=(const KeyValuePair &b)\n {\n return (value != b.value) || (key != b.key);\n }\n};\n\n\n/// Smaller value specialization\ntemplate \nstruct KeyValuePair\n{\n typedef K Key;\n typedef V Value;\n\n typedef char Pad[AlignBytes::ALIGN_BYTES - AlignBytes::ALIGN_BYTES];\n\n Key key; // Key has larger would-be alignment and goes first\n Value value;\n Pad pad;\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair() {}\n\n /// Constructor\n __host__ __device__ __forceinline__\n KeyValuePair(Key const& key, Value const& value) : key(key), value(value) {}\n\n /// Inequality operator\n __host__ __device__ __forceinline__ bool operator !=(const KeyValuePair &b)\n {\n return (value != b.value) || (key != b.key);\n }\n};\n\n#endif // #if defined(_WIN32) && !defined(_WIN64)\n\n\n#ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n\n/**\n * \\brief A wrapper for passing simple static arrays as kernel parameters\n */\ntemplate \nstruct ArrayWrapper\n{\n\n /// Statically-sized array of type \\p T\n T array[COUNT];\n\n /// Constructor\n __host__ __device__ __forceinline__ ArrayWrapper() {}\n};\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n/**\n * \\brief Double-buffer storage wrapper for multi-pass stream transformations that require more than one storage array for streaming intermediate results back and forth.\n *\n * Many multi-pass computations require a pair of \"ping-pong\" storage\n * buffers (e.g., one for reading from and the other for writing to, and then\n * vice-versa for the subsequent pass). This structure wraps a set of device\n * buffers and a \"selector\" member to track which is \"current\".\n */\ntemplate \nstruct DoubleBuffer\n{\n /// Pair of device buffer pointers\n T *d_buffers[2];\n\n /// Selector into \\p d_buffers (i.e., the active/valid buffer)\n int selector;\n\n /// \\brief Constructor\n __host__ __device__ __forceinline__ DoubleBuffer()\n {\n selector = 0;\n d_buffers[0] = NULL;\n d_buffers[1] = NULL;\n }\n\n /// \\brief Constructor\n __host__ __device__ __forceinline__ DoubleBuffer(\n T *d_current, ///< The currently valid buffer\n T *d_alternate) ///< Alternate storage buffer of the same size as \\p d_current\n {\n selector = 0;\n d_buffers[0] = d_current;\n d_buffers[1] = d_alternate;\n }\n\n /// \\brief Return pointer to the currently valid buffer\n __host__ __device__ __forceinline__ T* Current() { return d_buffers[selector]; }\n\n /// \\brief Return pointer to the currently invalid buffer\n __host__ __device__ __forceinline__ T* Alternate() { return d_buffers[selector ^ 1]; }\n\n};\n\n\n\n/******************************************************************************\n * Typedef-detection\n ******************************************************************************/\n\n\n/**\n * \\brief Defines a structure \\p detector_name that is templated on type \\p T. The \\p detector_name struct exposes a constant member \\p VALUE indicating whether or not parameter \\p T exposes a nested type \\p nested_type_name\n */\n#define CUB_DEFINE_DETECT_NESTED_TYPE(detector_name, nested_type_name) \\\n template \\\n struct detector_name \\\n { \\\n template \\\n static char& test(typename C::nested_type_name*); \\\n template \\\n static int& test(...); \\\n enum \\\n { \\\n VALUE = sizeof(test(0)) < sizeof(int) \\\n }; \\\n };\n\n\n\n/******************************************************************************\n * Simple enable-if (similar to Boost)\n ******************************************************************************/\n\n/**\n * \\brief Simple enable-if (similar to Boost)\n */\ntemplate \nstruct EnableIf\n{\n /// Enable-if type for SFINAE dummy variables\n typedef T Type;\n};\n\n\ntemplate \nstruct EnableIf {};\n\n\n\n/******************************************************************************\n * Typedef-detection\n ******************************************************************************/\n\n/**\n * \\brief Determine whether or not BinaryOp's functor is of the form bool operator()(const T& a, const T&b) or bool operator()(const T& a, const T&b, unsigned int idx)\n */\ntemplate \nstruct BinaryOpHasIdxParam\n{\nprivate:\n/*\n template struct SFINAE1 {};\n template struct SFINAE2 {};\n template struct SFINAE3 {};\n template struct SFINAE4 {};\n*/\n template struct SFINAE5 {};\n template struct SFINAE6 {};\n template struct SFINAE7 {};\n template struct SFINAE8 {};\n/*\n template static char Test(SFINAE1 *);\n template static char Test(SFINAE2 *);\n template static char Test(SFINAE3 *);\n template static char Test(SFINAE4 *);\n*/\n template static char Test(SFINAE5 *);\n template static char Test(SFINAE6 *);\n template static char Test(SFINAE7 *);\n template static char Test(SFINAE8 *);\n\n template static int Test(...);\n\npublic:\n\n /// Whether the functor BinaryOp has a third unsigned int index param\n static const bool HAS_PARAM = sizeof(Test(NULL)) == sizeof(char);\n};\n\n\n\n\n/******************************************************************************\n * Simple type traits utilities.\n *\n * For example:\n * Traits::CATEGORY // SIGNED_INTEGER\n * Traits::NULL_TYPE // true\n * Traits::CATEGORY // NOT_A_NUMBER\n * Traits::PRIMITIVE; // false\n *\n ******************************************************************************/\n\n/**\n * \\brief Basic type traits categories\n */\nenum Category\n{\n NOT_A_NUMBER,\n SIGNED_INTEGER,\n UNSIGNED_INTEGER,\n FLOATING_POINT\n};\n\n\n/**\n * \\brief Basic type traits\n */\ntemplate \nstruct BaseTraits\n{\n /// Category\n static const Category CATEGORY = _CATEGORY;\n enum\n {\n PRIMITIVE = _PRIMITIVE,\n NULL_TYPE = _NULL_TYPE,\n };\n};\n\n\n/**\n * Basic type traits (unsigned primitive specialization)\n */\ntemplate \nstruct BaseTraits\n{\n typedef _UnsignedBits UnsignedBits;\n\n static const Category CATEGORY = UNSIGNED_INTEGER;\n static const UnsignedBits LOWEST_KEY = UnsignedBits(0);\n static const UnsignedBits MAX_KEY = UnsignedBits(-1);\n\n enum\n {\n PRIMITIVE = true,\n NULL_TYPE = false,\n };\n\n\n static __device__ __forceinline__ UnsignedBits TwiddleIn(UnsignedBits key)\n {\n return key;\n }\n\n static __device__ __forceinline__ UnsignedBits TwiddleOut(UnsignedBits key)\n {\n return key;\n }\n\n static __host__ __device__ __forceinline__ T Max()\n {\n UnsignedBits retval = MAX_KEY;\n return reinterpret_cast(retval);\n }\n\n static __host__ __device__ __forceinline__ T Lowest()\n {\n UnsignedBits retval = LOWEST_KEY;\n return reinterpret_cast(retval);\n }\n};\n\n\n/**\n * Basic type traits (signed primitive specialization)\n */\ntemplate \nstruct BaseTraits\n{\n typedef _UnsignedBits UnsignedBits;\n\n static const Category CATEGORY = SIGNED_INTEGER;\n static const UnsignedBits HIGH_BIT = UnsignedBits(1) << ((sizeof(UnsignedBits) * 8) - 1);\n static const UnsignedBits LOWEST_KEY = HIGH_BIT;\n static const UnsignedBits MAX_KEY = UnsignedBits(-1) ^ HIGH_BIT;\n\n enum\n {\n PRIMITIVE = true,\n NULL_TYPE = false,\n };\n\n static __device__ __forceinline__ UnsignedBits TwiddleIn(UnsignedBits key)\n {\n return key ^ HIGH_BIT;\n };\n\n static __device__ __forceinline__ UnsignedBits TwiddleOut(UnsignedBits key)\n {\n return key ^ HIGH_BIT;\n };\n\n static __host__ __device__ __forceinline__ T Max()\n {\n UnsignedBits retval = MAX_KEY;\n return reinterpret_cast(retval);\n }\n\n static __host__ __device__ __forceinline__ T Lowest()\n {\n UnsignedBits retval = LOWEST_KEY;\n return reinterpret_cast(retval);\n }\n};\n\ntemplate \nstruct FpLimits;\n\ntemplate <>\nstruct FpLimits\n{\n static __host__ __device__ __forceinline__ float Max() {\n return FLT_MAX;\n }\n\n static __host__ __device__ __forceinline__ float Lowest() {\n return FLT_MAX * float(-1);\n }\n};\n\ntemplate <>\nstruct FpLimits\n{\n static __host__ __device__ __forceinline__ double Max() {\n return DBL_MAX;\n }\n\n static __host__ __device__ __forceinline__ double Lowest() {\n return DBL_MAX * double(-1);\n }\n};\n\n\n/**\n * Basic type traits (fp primitive specialization)\n */\ntemplate \nstruct BaseTraits\n{\n typedef _UnsignedBits UnsignedBits;\n\n static const Category CATEGORY = FLOATING_POINT;\n static const UnsignedBits HIGH_BIT = UnsignedBits(1) << ((sizeof(UnsignedBits) * 8) - 1);\n static const UnsignedBits LOWEST_KEY = UnsignedBits(-1);\n static const UnsignedBits MAX_KEY = UnsignedBits(-1) ^ HIGH_BIT;\n\n enum\n {\n PRIMITIVE = true,\n NULL_TYPE = false,\n };\n\n static __device__ __forceinline__ UnsignedBits TwiddleIn(UnsignedBits key)\n {\n UnsignedBits mask = (key & HIGH_BIT) ? UnsignedBits(-1) : HIGH_BIT;\n return key ^ mask;\n };\n\n static __device__ __forceinline__ UnsignedBits TwiddleOut(UnsignedBits key)\n {\n UnsignedBits mask = (key & HIGH_BIT) ? HIGH_BIT : UnsignedBits(-1);\n return key ^ mask;\n };\n\n static __host__ __device__ __forceinline__ T Max() {\n return FpLimits::Max();\n }\n\n static __host__ __device__ __forceinline__ T Lowest() {\n return FpLimits::Lowest();\n }\n};\n\n\n/**\n * \\brief Numeric type traits\n */\ntemplate struct NumericTraits : BaseTraits {};\n\ntemplate <> struct NumericTraits : BaseTraits {};\n\ntemplate <> struct NumericTraits : BaseTraits<(std::numeric_limits::is_signed) ? SIGNED_INTEGER : UNSIGNED_INTEGER, true, false, unsigned char, char> {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\n\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\n\ntemplate <> struct NumericTraits : BaseTraits {};\ntemplate <> struct NumericTraits : BaseTraits {};\n\ntemplate <> struct NumericTraits : BaseTraits::VolatileWord, bool> {};\n\n\n\n/**\n * \\brief Type traits\n */\ntemplate \nstruct Traits : NumericTraits::Type> {};\n\n\n#endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\n/** @} */ // end group UtilModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/specializations/warp_reduce_shfl.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::WarpReduceShfl provides SHFL-based variants of parallel reduction of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_macro.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\brief WarpReduceShfl provides SHFL-based variants of parallel reduction of items partitioned across a CUDA thread warp.\n *\n * LOGICAL_WARP_THREADS must be a power-of-two\n */\ntemplate <\n typename T, ///< Data type being reduced\n int LOGICAL_WARP_THREADS, ///< Number of threads per logical warp\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct WarpReduceShfl\n{\n //---------------------------------------------------------------------\n // Constants and type definitions\n //---------------------------------------------------------------------\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// The number of warp reduction steps\n STEPS = Log2::VALUE,\n\n /// Number of logical warps in a PTX warp\n LOGICAL_WARPS = CUB_WARP_THREADS(PTX_ARCH) / LOGICAL_WARP_THREADS,\n };\n\n template \n struct IsInteger\n {\n enum {\n ///Whether the data type is a small (32b or less) integer for which we can use a single SFHL instruction per exchange\n IS_SMALL_UNSIGNED = (Traits::CATEGORY == UNSIGNED_INTEGER) && (sizeof(S) <= sizeof(unsigned int))\n };\n };\n\n\n // Creates a mask where the last thread in each logical warp is set\n template \n struct LastLaneMask\n {\n enum {\n BASE_MASK = 1 << (LOGICAL_WARP_THREADS - 1),\n MASK = (LastLaneMask::MASK << LOGICAL_WARP_THREADS) | BASE_MASK,\n };\n };\n\n // Creates a mask where the last thread in each logical warp is set\n template \n struct LastLaneMask\n {\n enum {\n MASK = 1 << (LOGICAL_WARP_THREADS - 1),\n };\n };\n\n\n\n /// Shared memory storage layout type\n typedef NullType TempStorage;\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n\n unsigned int lane_id;\n\n unsigned int member_mask;\n\n //---------------------------------------------------------------------\n // Construction\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ WarpReduceShfl(\n TempStorage &/*temp_storage*/)\n :\n lane_id(LaneId()),\n\n member_mask((0xffffffff >> (32 - LOGICAL_WARP_THREADS)) << ((IS_ARCH_WARP) ?\n 0 : // arch-width subwarps need not be tiled within the arch-warp\n ((lane_id / LOGICAL_WARP_THREADS) * LOGICAL_WARP_THREADS)))\n {}\n\n\n //---------------------------------------------------------------------\n // Reduction steps\n //---------------------------------------------------------------------\n\n /// Reduction (specialized for summation across uint32 types)\n __device__ __forceinline__ unsigned int ReduceStep(\n unsigned int input, ///< [in] Calling thread's input item.\n cub::Sum /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n unsigned int output;\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 r0;\"\n \" .reg .pred p;\"\n \" shfl.sync.down.b32 r0|p, %1, %2, %3, %5;\"\n \" @p add.u32 r0, r0, %4;\"\n \" mov.u32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(last_lane), \"r\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 r0;\"\n \" .reg .pred p;\"\n \" shfl.down.b32 r0|p, %1, %2, %3;\"\n \" @p add.u32 r0, r0, %4;\"\n \" mov.u32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(last_lane), \"r\"(input));\n#endif\n\n return output;\n }\n\n\n /// Reduction (specialized for summation across fp32 types)\n __device__ __forceinline__ float ReduceStep(\n float input, ///< [in] Calling thread's input item.\n cub::Sum /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n float output;\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .f32 r0;\"\n \" .reg .pred p;\"\n \" shfl.sync.down.b32 r0|p, %1, %2, %3, %5;\"\n \" @p add.f32 r0, r0, %4;\"\n \" mov.f32 %0, r0;\"\n \"}\"\n : \"=f\"(output) : \"f\"(input), \"r\"(offset), \"r\"(last_lane), \"f\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .f32 r0;\"\n \" .reg .pred p;\"\n \" shfl.down.b32 r0|p, %1, %2, %3;\"\n \" @p add.f32 r0, r0, %4;\"\n \" mov.f32 %0, r0;\"\n \"}\"\n : \"=f\"(output) : \"f\"(input), \"r\"(offset), \"r\"(last_lane), \"f\"(input));\n#endif\n\n return output;\n }\n\n\n /// Reduction (specialized for summation across unsigned long long types)\n __device__ __forceinline__ unsigned long long ReduceStep(\n unsigned long long input, ///< [in] Calling thread's input item.\n cub::Sum /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n unsigned long long output;\n\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.down.b32 lo|p, lo, %2, %3, %4;\"\n \" shfl.sync.down.b32 hi|p, hi, %2, %3, %4;\"\n \" mov.b64 %0, {lo, hi};\"\n \" @p add.u64 %0, %0, %1;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(last_lane), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.down.b32 lo|p, lo, %2, %3;\"\n \" shfl.down.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 %0, {lo, hi};\"\n \" @p add.u64 %0, %0, %1;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(last_lane));\n#endif\n\n return output;\n }\n\n\n /// Reduction (specialized for summation across long long types)\n __device__ __forceinline__ long long ReduceStep(\n long long input, ///< [in] Calling thread's input item.\n cub::Sum /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n long long output;\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.down.b32 lo|p, lo, %2, %3, %4;\"\n \" shfl.sync.down.b32 hi|p, hi, %2, %3, %4;\"\n \" mov.b64 %0, {lo, hi};\"\n \" @p add.s64 %0, %0, %1;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(last_lane), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.down.b32 lo|p, lo, %2, %3;\"\n \" shfl.down.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 %0, {lo, hi};\"\n \" @p add.s64 %0, %0, %1;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(last_lane));\n#endif\n\n return output;\n }\n\n\n /// Reduction (specialized for summation across double types)\n __device__ __forceinline__ double ReduceStep(\n double input, ///< [in] Calling thread's input item.\n cub::Sum /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n double output;\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" .reg .f64 r0;\"\n \" mov.b64 %0, %1;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.down.b32 lo|p, lo, %2, %3, %4;\"\n \" shfl.sync.down.b32 hi|p, hi, %2, %3, %4;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.f64 %0, %0, r0;\"\n \"}\"\n : \"=d\"(output) : \"d\"(input), \"r\"(offset), \"r\"(last_lane), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" .reg .f64 r0;\"\n \" mov.b64 %0, %1;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.down.b32 lo|p, lo, %2, %3;\"\n \" shfl.down.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.f64 %0, %0, r0;\"\n \"}\"\n : \"=d\"(output) : \"d\"(input), \"r\"(offset), \"r\"(last_lane));\n#endif\n\n return output;\n }\n\n\n /// Reduction (specialized for swizzled ReduceByKeyOp across KeyValuePair types)\n template \n __device__ __forceinline__ KeyValuePair ReduceStep(\n KeyValuePair input, ///< [in] Calling thread's input item.\n SwizzleScanOp > /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n KeyValuePair output;\n\n KeyT other_key = ShuffleDown(input.key, offset, last_lane, member_mask);\n \n output.key = input.key;\n output.value = ReduceStep(\n input.value, \n cub::Sum(), \n last_lane, \n offset, \n Int2Type::IS_SMALL_UNSIGNED>());\n\n if (input.key != other_key)\n output.value = input.value;\n\n return output;\n }\n\n\n\n /// Reduction (specialized for swizzled ReduceBySegmentOp across KeyValuePair types)\n template \n __device__ __forceinline__ KeyValuePair ReduceStep(\n KeyValuePair input, ///< [in] Calling thread's input item.\n SwizzleScanOp > /*reduction_op*/, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n KeyValuePair output;\n\n output.value = ReduceStep(input.value, cub::Sum(), last_lane, offset, Int2Type::IS_SMALL_UNSIGNED>());\n output.key = ReduceStep(input.key, cub::Sum(), last_lane, offset, Int2Type::IS_SMALL_UNSIGNED>());\n\n if (input.key > 0)\n output.value = input.value;\n\n return output;\n }\n\n\n /// Reduction step (generic)\n template \n __device__ __forceinline__ _T ReduceStep(\n _T input, ///< [in] Calling thread's input item.\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n _T output = input;\n\n _T temp = ShuffleDown(output, offset, last_lane, member_mask);\n\n // Perform reduction op if valid\n if (offset + lane_id <= last_lane)\n output = reduction_op(input, temp);\n\n return output;\n }\n\n\n /// Reduction step (specialized for small unsigned integers size 32b or less)\n template \n __device__ __forceinline__ _T ReduceStep(\n _T input, ///< [in] Calling thread's input item.\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset, ///< [in] Up-offset to pull from\n Int2Type /*is_small_unsigned*/) ///< [in] Marker type indicating whether T is a small unsigned integer\n {\n return ReduceStep(input, reduction_op, last_lane, offset);\n }\n\n\n /// Reduction step (specialized for types other than small unsigned integers size 32b or less)\n template \n __device__ __forceinline__ _T ReduceStep(\n _T input, ///< [in] Calling thread's input item.\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n int offset, ///< [in] Up-offset to pull from\n Int2Type /*is_small_unsigned*/) ///< [in] Marker type indicating whether T is a small unsigned integer\n {\n return ReduceStep(input, reduction_op, last_lane, offset);\n }\n\n\n //---------------------------------------------------------------------\n // Templated inclusive scan iteration\n //---------------------------------------------------------------------\n\n template \n __device__ __forceinline__ void ReduceStep(\n T& input, ///< [in] Calling thread's input item.\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n int last_lane, ///< [in] Index of last lane in segment\n Int2Type /*step*/)\n {\n input = ReduceStep(input, reduction_op, last_lane, 1 << STEP, Int2Type::IS_SMALL_UNSIGNED>());\n\n ReduceStep(input, reduction_op, last_lane, Int2Type());\n }\n\n template \n __device__ __forceinline__ void ReduceStep(\n T& /*input*/, ///< [in] Calling thread's input item.\n ReductionOp /*reduction_op*/, ///< [in] Binary reduction operator\n int /*last_lane*/, ///< [in] Index of last lane in segment\n Int2Type /*step*/)\n {}\n\n\n //---------------------------------------------------------------------\n // Reduction operations\n //---------------------------------------------------------------------\n\n /// Reduction\n template <\n bool ALL_LANES_VALID, ///< Whether all lanes in each warp are contributing a valid fold of items\n int FOLDED_ITEMS_PER_LANE, ///< Number of items folded into each lane\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n int folded_items_per_warp, ///< [in] Total number of valid items folded into each logical warp\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n // Get the lane of the first and last thread in the logical warp\n int first_thread = 0;\n int last_thread = LOGICAL_WARP_THREADS - 1;\n if (!IS_ARCH_WARP)\n {\n first_thread = lane_id & (~(LOGICAL_WARP_THREADS - 1));\n last_thread |= lane_id;\n }\n\n // Common case is FOLDED_ITEMS_PER_LANE = 1 (or a multiple of 32)\n int lanes_with_valid_data = (folded_items_per_warp - 1) / FOLDED_ITEMS_PER_LANE;\n\n // Get the last valid lane\n int last_lane = (ALL_LANES_VALID) ?\n last_thread :\n CUB_MIN(last_thread, first_thread + lanes_with_valid_data);\n\n T output = input;\n\n// // Iterate reduction steps\n// #pragma unroll\n// for (int STEP = 0; STEP < STEPS; STEP++)\n// {\n// output = ReduceStep(output, reduction_op, last_lane, 1 << STEP, Int2Type::IS_SMALL_UNSIGNED>());\n// }\n\n // Template-iterate reduction steps\n ReduceStep(output, reduction_op, last_lane, Int2Type<0>());\n\n return output;\n }\n\n\n /// Segmented reduction\n template <\n bool HEAD_SEGMENTED, ///< Whether flags indicate a segment-head or a segment-tail\n typename FlagT,\n typename ReductionOp>\n __device__ __forceinline__ T SegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT flag, ///< [in] Whether or not the current lane is a segment head/tail\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n // Get the start flags for each thread in the warp.\n int warp_flags = WARP_BALLOT(flag, member_mask);\n\n // Convert to tail-segmented\n if (HEAD_SEGMENTED)\n warp_flags >>= 1;\n\n // Mask in the last lanes of each logical warp\n warp_flags |= LastLaneMask<1, LOGICAL_WARPS>::MASK;\n\n // Mask out the bits below the current thread\n warp_flags &= LaneMaskGe();\n\n // Find the next set flag\n int last_lane = __clz(__brev(warp_flags));\n\n T output = input;\n\n// // Iterate reduction steps\n// #pragma unroll\n// for (int STEP = 0; STEP < STEPS; STEP++)\n// {\n// output = ReduceStep(output, reduction_op, last_lane, 1 << STEP, Int2Type::IS_SMALL_UNSIGNED>());\n// }\n\n // Template-iterate reduction steps\n ReduceStep(output, reduction_op, last_lane, Int2Type<0>());\n\n return output;\n }\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/specializations/warp_reduce_smem.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::WarpReduceSmem provides smem-based variants of parallel reduction of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../thread/thread_load.cuh\"\n#include \"../../thread/thread_store.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief WarpReduceSmem provides smem-based variants of parallel reduction of items partitioned across a CUDA thread warp.\n */\ntemplate <\n typename T, ///< Data type being reduced\n int LOGICAL_WARP_THREADS, ///< Number of threads per logical warp\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct WarpReduceSmem\n{\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// Whether the logical warp size is a power-of-two\n IS_POW_OF_TWO = PowerOfTwo::VALUE,\n\n /// The number of warp scan steps\n STEPS = Log2::VALUE,\n\n /// The number of threads in half a warp\n HALF_WARP_THREADS = 1 << (STEPS - 1),\n\n /// The number of shared memory elements per warp\n WARP_SMEM_ELEMENTS = LOGICAL_WARP_THREADS + HALF_WARP_THREADS,\n\n /// FlagT status (when not using ballot)\n UNSET = 0x0, // Is initially unset\n SET = 0x1, // Is initially set\n SEEN = 0x2, // Has seen another head flag from a successor peer\n };\n\n /// Shared memory flag type\n typedef unsigned char SmemFlag;\n\n /// Shared memory storage layout type (1.5 warps-worth of elements for each warp)\n struct _TempStorage\n {\n T reduce[WARP_SMEM_ELEMENTS];\n SmemFlag flags[WARP_SMEM_ELEMENTS];\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n _TempStorage &temp_storage;\n unsigned int lane_id;\n unsigned int member_mask;\n\n\n /******************************************************************************\n * Construction\n ******************************************************************************/\n\n /// Constructor\n __device__ __forceinline__ WarpReduceSmem(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n\n lane_id(IS_ARCH_WARP ?\n LaneId() :\n LaneId() % LOGICAL_WARP_THREADS),\n\n member_mask((0xffffffff >> (32 - LOGICAL_WARP_THREADS)) << ((IS_ARCH_WARP || !IS_POW_OF_TWO ) ?\n 0 : // arch-width and non-power-of-two subwarps cannot be tiled with the arch-warp\n ((LaneId() / LOGICAL_WARP_THREADS) * LOGICAL_WARP_THREADS)))\n {}\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n //---------------------------------------------------------------------\n // Regular reduction\n //---------------------------------------------------------------------\n\n /**\n * Reduction step\n */\n template <\n bool ALL_LANES_VALID, ///< Whether all lanes in each warp are contributing a valid fold of items\n int FOLDED_ITEMS_PER_LANE, ///< Number of items folded into each lane\n typename ReductionOp,\n int STEP>\n __device__ __forceinline__ T ReduceStep(\n T input, ///< [in] Calling thread's input\n int folded_items_per_warp, ///< [in] Total number of valid items folded into each logical warp\n ReductionOp reduction_op, ///< [in] Reduction operator\n Int2Type /*step*/)\n {\n const int OFFSET = 1 << STEP;\n\n // Share input through buffer\n ThreadStore(&temp_storage.reduce[lane_id], input);\n\n WARP_SYNC(member_mask);\n\n // Update input if peer_addend is in range\n if ((ALL_LANES_VALID && IS_POW_OF_TWO) || ((lane_id + OFFSET) * FOLDED_ITEMS_PER_LANE < folded_items_per_warp))\n {\n T peer_addend = ThreadLoad(&temp_storage.reduce[lane_id + OFFSET]);\n input = reduction_op(input, peer_addend);\n }\n\n WARP_SYNC(member_mask);\n\n return ReduceStep(input, folded_items_per_warp, reduction_op, Int2Type());\n }\n\n\n /**\n * Reduction step (terminate)\n */\n template <\n bool ALL_LANES_VALID, ///< Whether all lanes in each warp are contributing a valid fold of items\n int FOLDED_ITEMS_PER_LANE, ///< Number of items folded into each lane\n typename ReductionOp>\n __device__ __forceinline__ T ReduceStep(\n T input, ///< [in] Calling thread's input\n int /*folded_items_per_warp*/, ///< [in] Total number of valid items folded into each logical warp\n ReductionOp /*reduction_op*/, ///< [in] Reduction operator\n Int2Type /*step*/)\n {\n return input;\n }\n\n\n //---------------------------------------------------------------------\n // Segmented reduction\n //---------------------------------------------------------------------\n\n\n /**\n * Ballot-based segmented reduce\n */\n template <\n bool HEAD_SEGMENTED, ///< Whether flags indicate a segment-head or a segment-tail\n typename FlagT,\n typename ReductionOp>\n __device__ __forceinline__ T SegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT flag, ///< [in] Whether or not the current lane is a segment head/tail\n ReductionOp reduction_op, ///< [in] Reduction operator\n Int2Type /*has_ballot*/) ///< [in] Marker type for whether the target arch has ballot functionality\n {\n // Get the start flags for each thread in the warp.\n int warp_flags = WARP_BALLOT(flag, member_mask);\n\n if (!HEAD_SEGMENTED)\n warp_flags <<= 1;\n\n // Keep bits above the current thread.\n warp_flags &= LaneMaskGt();\n\n // Accommodate packing of multiple logical warps in a single physical warp\n if (!IS_ARCH_WARP)\n {\n warp_flags >>= (LaneId() / LOGICAL_WARP_THREADS) * LOGICAL_WARP_THREADS;\n }\n\n // Find next flag\n int next_flag = __clz(__brev(warp_flags));\n\n // Clip the next segment at the warp boundary if necessary\n if (LOGICAL_WARP_THREADS != 32)\n next_flag = CUB_MIN(next_flag, LOGICAL_WARP_THREADS);\n\n #pragma unroll\n for (int STEP = 0; STEP < STEPS; STEP++)\n {\n const int OFFSET = 1 << STEP;\n\n // Share input into buffer\n ThreadStore(&temp_storage.reduce[lane_id], input);\n\n WARP_SYNC(member_mask);\n\n // Update input if peer_addend is in range\n if (OFFSET + lane_id < next_flag)\n {\n T peer_addend = ThreadLoad(&temp_storage.reduce[lane_id + OFFSET]);\n input = reduction_op(input, peer_addend);\n }\n\n WARP_SYNC(member_mask);\n }\n\n return input;\n }\n\n\n /**\n * Smem-based segmented reduce\n */\n template <\n bool HEAD_SEGMENTED, ///< Whether flags indicate a segment-head or a segment-tail\n typename FlagT,\n typename ReductionOp>\n __device__ __forceinline__ T SegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT flag, ///< [in] Whether or not the current lane is a segment head/tail\n ReductionOp reduction_op, ///< [in] Reduction operator\n Int2Type /*has_ballot*/) ///< [in] Marker type for whether the target arch has ballot functionality\n {\n enum\n {\n UNSET = 0x0, // Is initially unset\n SET = 0x1, // Is initially set\n SEEN = 0x2, // Has seen another head flag from a successor peer\n };\n\n // Alias flags onto shared data storage\n volatile SmemFlag *flag_storage = temp_storage.flags;\n\n SmemFlag flag_status = (flag) ? SET : UNSET;\n\n for (int STEP = 0; STEP < STEPS; STEP++)\n {\n const int OFFSET = 1 << STEP;\n\n // Share input through buffer\n ThreadStore(&temp_storage.reduce[lane_id], input);\n\n WARP_SYNC(member_mask);\n\n // Get peer from buffer\n T peer_addend = ThreadLoad(&temp_storage.reduce[lane_id + OFFSET]);\n\n WARP_SYNC(member_mask);\n\n // Share flag through buffer\n flag_storage[lane_id] = flag_status;\n\n // Get peer flag from buffer\n SmemFlag peer_flag_status = flag_storage[lane_id + OFFSET];\n\n // Update input if peer was in range\n if (lane_id < LOGICAL_WARP_THREADS - OFFSET)\n {\n if (HEAD_SEGMENTED)\n {\n // Head-segmented\n if ((flag_status & SEEN) == 0)\n {\n // Has not seen a more distant head flag\n if (peer_flag_status & SET)\n {\n // Has now seen a head flag\n flag_status |= SEEN;\n }\n else\n {\n // Peer is not a head flag: grab its count\n input = reduction_op(input, peer_addend);\n }\n\n // Update seen status to include that of peer\n flag_status |= (peer_flag_status & SEEN);\n }\n }\n else\n {\n // Tail-segmented. Simply propagate flag status\n if (!flag_status)\n {\n input = reduction_op(input, peer_addend);\n flag_status |= peer_flag_status;\n }\n\n }\n }\n }\n\n return input;\n }\n\n\n /******************************************************************************\n * Interface\n ******************************************************************************/\n\n /**\n * Reduction\n */\n template <\n bool ALL_LANES_VALID, ///< Whether all lanes in each warp are contributing a valid fold of items\n int FOLDED_ITEMS_PER_LANE, ///< Number of items folded into each lane\n typename ReductionOp>\n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n int folded_items_per_warp, ///< [in] Total number of valid items folded into each logical warp\n ReductionOp reduction_op) ///< [in] Reduction operator\n {\n return ReduceStep(input, folded_items_per_warp, reduction_op, Int2Type<0>());\n }\n\n\n /**\n * Segmented reduction\n */\n template <\n bool HEAD_SEGMENTED, ///< Whether flags indicate a segment-head or a segment-tail\n typename FlagT,\n typename ReductionOp>\n __device__ __forceinline__ T SegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT flag, ///< [in] Whether or not the current lane is a segment head/tail\n ReductionOp reduction_op) ///< [in] Reduction operator\n {\n return SegmentedReduce(input, flag, reduction_op, Int2Type<(PTX_ARCH >= 200)>());\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/specializations/warp_scan_shfl.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::WarpScanShfl provides SHFL-based variants of parallel prefix scan of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_ptx.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief WarpScanShfl provides SHFL-based variants of parallel prefix scan of items partitioned across a CUDA thread warp.\n *\n * LOGICAL_WARP_THREADS must be a power-of-two\n */\ntemplate <\n typename T, ///< Data type being scanned\n int LOGICAL_WARP_THREADS, ///< Number of threads per logical warp\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct WarpScanShfl\n{\n //---------------------------------------------------------------------\n // Constants and type definitions\n //---------------------------------------------------------------------\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// The number of warp scan steps\n STEPS = Log2::VALUE,\n\n /// The 5-bit SHFL mask for logically splitting warps into sub-segments starts 8-bits up\n SHFL_C = ((0xFFFFFFFFU << STEPS) & 31) << 8,\n };\n\n template \n struct IntegerTraits\n {\n enum {\n ///Whether the data type is a small (32b or less) integer for which we can use a single SFHL instruction per exchange\n IS_SMALL_UNSIGNED = (Traits::CATEGORY == UNSIGNED_INTEGER) && (sizeof(S) <= sizeof(unsigned int))\n };\n };\n\n /// Shared memory storage layout type\n struct TempStorage {};\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n unsigned int lane_id;\n\n unsigned int member_mask;\n\n //---------------------------------------------------------------------\n // Construction\n //---------------------------------------------------------------------\n\n /// Constructor\n __device__ __forceinline__ WarpScanShfl(\n TempStorage &/*temp_storage*/)\n :\n lane_id(LaneId()),\n\n member_mask((0xffffffff >> (32 - LOGICAL_WARP_THREADS)) << ((IS_ARCH_WARP) ?\n 0 : // arch-width subwarps need not be tiled within the arch-warp\n ((lane_id / LOGICAL_WARP_THREADS) * LOGICAL_WARP_THREADS)))\n {}\n\n\n //---------------------------------------------------------------------\n // Inclusive scan steps\n //---------------------------------------------------------------------\n\n /// Inclusive prefix scan step (specialized for summation across int32 types)\n __device__ __forceinline__ int InclusiveScanStep(\n int input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n int output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .s32 r0;\"\n \" .reg .pred p;\"\n \" shfl.sync.up.b32 r0|p, %1, %2, %3, %5;\"\n \" @p add.s32 r0, r0, %4;\"\n \" mov.s32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(shfl_c), \"r\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .s32 r0;\"\n \" .reg .pred p;\"\n \" shfl.up.b32 r0|p, %1, %2, %3;\"\n \" @p add.s32 r0, r0, %4;\"\n \" mov.s32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(shfl_c), \"r\"(input));\n#endif\n\n return output;\n }\n\n /// Inclusive prefix scan step (specialized for summation across uint32 types)\n __device__ __forceinline__ unsigned int InclusiveScanStep(\n unsigned int input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n unsigned int output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 r0;\"\n \" .reg .pred p;\"\n \" shfl.sync.up.b32 r0|p, %1, %2, %3, %5;\"\n \" @p add.u32 r0, r0, %4;\"\n \" mov.u32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(shfl_c), \"r\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 r0;\"\n \" .reg .pred p;\"\n \" shfl.up.b32 r0|p, %1, %2, %3;\"\n \" @p add.u32 r0, r0, %4;\"\n \" mov.u32 %0, r0;\"\n \"}\"\n : \"=r\"(output) : \"r\"(input), \"r\"(offset), \"r\"(shfl_c), \"r\"(input));\n#endif\n\n return output;\n }\n\n\n /// Inclusive prefix scan step (specialized for summation across fp32 types)\n __device__ __forceinline__ float InclusiveScanStep(\n float input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n float output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .f32 r0;\"\n \" .reg .pred p;\"\n \" shfl.sync.up.b32 r0|p, %1, %2, %3, %5;\"\n \" @p add.f32 r0, r0, %4;\"\n \" mov.f32 %0, r0;\"\n \"}\"\n : \"=f\"(output) : \"f\"(input), \"r\"(offset), \"r\"(shfl_c), \"f\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .f32 r0;\"\n \" .reg .pred p;\"\n \" shfl.up.b32 r0|p, %1, %2, %3;\"\n \" @p add.f32 r0, r0, %4;\"\n \" mov.f32 %0, r0;\"\n \"}\"\n : \"=f\"(output) : \"f\"(input), \"r\"(offset), \"r\"(shfl_c), \"f\"(input));\n#endif\n\n return output;\n }\n\n\n /// Inclusive prefix scan step (specialized for summation across unsigned long long types)\n __device__ __forceinline__ unsigned long long InclusiveScanStep(\n unsigned long long input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n unsigned long long output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u64 r0;\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.up.b32 lo|p, lo, %2, %3, %5;\"\n \" shfl.sync.up.b32 hi|p, hi, %2, %3, %5;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.u64 r0, r0, %4;\"\n \" mov.u64 %0, r0;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(shfl_c), \"l\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u64 r0;\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.up.b32 lo|p, lo, %2, %3;\"\n \" shfl.up.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.u64 r0, r0, %4;\"\n \" mov.u64 %0, r0;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(shfl_c), \"l\"(input));\n#endif\n\n return output;\n }\n\n\n /// Inclusive prefix scan step (specialized for summation across long long types)\n __device__ __forceinline__ long long InclusiveScanStep(\n long long input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n long long output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .s64 r0;\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.up.b32 lo|p, lo, %2, %3, %5;\"\n \" shfl.sync.up.b32 hi|p, hi, %2, %3, %5;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.s64 r0, r0, %4;\"\n \" mov.s64 %0, r0;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(shfl_c), \"l\"(input), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .s64 r0;\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.up.b32 lo|p, lo, %2, %3;\"\n \" shfl.up.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.s64 r0, r0, %4;\"\n \" mov.s64 %0, r0;\"\n \"}\"\n : \"=l\"(output) : \"l\"(input), \"r\"(offset), \"r\"(shfl_c), \"l\"(input));\n#endif\n\n return output;\n }\n\n\n /// Inclusive prefix scan step (specialized for summation across fp64 types)\n __device__ __forceinline__ double InclusiveScanStep(\n double input, ///< [in] Calling thread's input item.\n cub::Sum /*scan_op*/, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n double output;\n int shfl_c = first_lane | SHFL_C; // Shuffle control (mask and first-lane)\n\n // Use predicate set from SHFL to guard against invalid peers\n#ifdef CUB_USE_COOPERATIVE_GROUPS\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" .reg .f64 r0;\"\n \" mov.b64 %0, %1;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.sync.up.b32 lo|p, lo, %2, %3, %4;\"\n \" shfl.sync.up.b32 hi|p, hi, %2, %3, %4;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.f64 %0, %0, r0;\"\n \"}\"\n : \"=d\"(output) : \"d\"(input), \"r\"(offset), \"r\"(shfl_c), \"r\"(member_mask));\n#else\n asm volatile(\n \"{\"\n \" .reg .u32 lo;\"\n \" .reg .u32 hi;\"\n \" .reg .pred p;\"\n \" .reg .f64 r0;\"\n \" mov.b64 %0, %1;\"\n \" mov.b64 {lo, hi}, %1;\"\n \" shfl.up.b32 lo|p, lo, %2, %3;\"\n \" shfl.up.b32 hi|p, hi, %2, %3;\"\n \" mov.b64 r0, {lo, hi};\"\n \" @p add.f64 %0, %0, r0;\"\n \"}\"\n : \"=d\"(output) : \"d\"(input), \"r\"(offset), \"r\"(shfl_c));\n#endif\n\n return output;\n }\n\n\n/*\n /// Inclusive prefix scan (specialized for ReduceBySegmentOp across KeyValuePair types)\n template \n __device__ __forceinline__ KeyValuePairInclusiveScanStep(\n KeyValuePair input, ///< [in] Calling thread's input item.\n ReduceBySegmentOp scan_op, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n KeyValuePair output;\n\n output.value = InclusiveScanStep(input.value, cub::Sum(), first_lane, offset, Int2Type::IS_SMALL_UNSIGNED>());\n output.key = InclusiveScanStep(input.key, cub::Sum(), first_lane, offset, Int2Type::IS_SMALL_UNSIGNED>());\n\n if (input.key > 0)\n output.value = input.value;\n\n return output;\n }\n*/\n\n /// Inclusive prefix scan step (generic)\n template \n __device__ __forceinline__ _T InclusiveScanStep(\n _T input, ///< [in] Calling thread's input item.\n ScanOpT scan_op, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset) ///< [in] Up-offset to pull from\n {\n _T temp = ShuffleUp(input, offset, first_lane, member_mask);\n\n // Perform scan op if from a valid peer\n _T output = scan_op(temp, input);\n if (static_cast(lane_id) < first_lane + offset)\n output = input;\n\n return output;\n }\n\n\n /// Inclusive prefix scan step (specialized for small integers size 32b or less)\n template \n __device__ __forceinline__ _T InclusiveScanStep(\n _T input, ///< [in] Calling thread's input item.\n ScanOpT scan_op, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset, ///< [in] Up-offset to pull from\n Int2Type /*is_small_unsigned*/) ///< [in] Marker type indicating whether T is a small integer\n {\n return InclusiveScanStep(input, scan_op, first_lane, offset);\n }\n\n\n /// Inclusive prefix scan step (specialized for types other than small integers size 32b or less)\n template \n __device__ __forceinline__ _T InclusiveScanStep(\n _T input, ///< [in] Calling thread's input item.\n ScanOpT scan_op, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n int offset, ///< [in] Up-offset to pull from\n Int2Type /*is_small_unsigned*/) ///< [in] Marker type indicating whether T is a small integer\n {\n return InclusiveScanStep(input, scan_op, first_lane, offset);\n }\n\n //---------------------------------------------------------------------\n // Templated inclusive scan iteration\n //---------------------------------------------------------------------\n\n template \n __device__ __forceinline__ void InclusiveScanStep(\n _T& input, ///< [in] Calling thread's input item.\n ScanOp scan_op, ///< [in] Binary scan operator\n int first_lane, ///< [in] Index of first lane in segment\n Int2Type /*step*/) ///< [in] Marker type indicating scan step\n {\n input = InclusiveScanStep(input, scan_op, first_lane, 1 << STEP, Int2Type::IS_SMALL_UNSIGNED>());\n\n InclusiveScanStep(input, scan_op, first_lane, Int2Type());\n }\n\n template \n __device__ __forceinline__ void InclusiveScanStep(\n _T& /*input*/, ///< [in] Calling thread's input item.\n ScanOp /*scan_op*/, ///< [in] Binary scan operator\n int /*first_lane*/, ///< [in] Index of first lane in segment\n Int2Type /*step*/) ///< [in] Marker type indicating scan step\n {}\n\n\n /******************************************************************************\n * Interface\n ******************************************************************************/\n\n //---------------------------------------------------------------------\n // Broadcast\n //---------------------------------------------------------------------\n\n /// Broadcast\n __device__ __forceinline__ T Broadcast(\n T input, ///< [in] The value to broadcast\n int src_lane) ///< [in] Which warp lane is to do the broadcasting\n {\n return ShuffleIndex(input, src_lane, LOGICAL_WARP_THREADS, member_mask);\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive operations\n //---------------------------------------------------------------------\n\n /// Inclusive scan\n template \n __device__ __forceinline__ void InclusiveScan(\n _T input, ///< [in] Calling thread's input item.\n _T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOpT scan_op) ///< [in] Binary scan operator\n {\n inclusive_output = input;\n\n // Iterate scan steps\n int segment_first_lane = 0;\n\n // Iterate scan steps\n// InclusiveScanStep(inclusive_output, scan_op, segment_first_lane, Int2Type<0>());\n\n // Iterate scan steps\n #pragma unroll\n for (int STEP = 0; STEP < STEPS; STEP++)\n {\n inclusive_output = InclusiveScanStep(\n inclusive_output,\n scan_op,\n segment_first_lane,\n (1 << STEP),\n Int2Type::IS_SMALL_UNSIGNED>());\n }\n\n }\n\n /// Inclusive scan, specialized for reduce-value-by-key\n template \n __device__ __forceinline__ void InclusiveScan(\n KeyValuePair input, ///< [in] Calling thread's input item.\n KeyValuePair &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ReduceByKeyOp scan_op) ///< [in] Binary scan operator\n {\n inclusive_output = input;\n\n KeyT pred_key = ShuffleUp(inclusive_output.key, 1, 0, member_mask);\n\n unsigned int ballot = WARP_BALLOT((pred_key != inclusive_output.key), member_mask);\n\n // Mask away all lanes greater than ours\n ballot = ballot & LaneMaskLe();\n\n // Find index of first set bit\n int segment_first_lane = CUB_MAX(0, 31 - __clz(ballot));\n\n // Iterate scan steps\n// InclusiveScanStep(inclusive_output.value, scan_op.op, segment_first_lane, Int2Type<0>());\n\n // Iterate scan steps\n #pragma unroll\n for (int STEP = 0; STEP < STEPS; STEP++)\n {\n inclusive_output.value = InclusiveScanStep(\n inclusive_output.value,\n scan_op.op,\n segment_first_lane,\n (1 << STEP),\n Int2Type::IS_SMALL_UNSIGNED>());\n }\n }\n\n\n /// Inclusive scan with aggregate\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOpT scan_op, ///< [in] Binary scan operator\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InclusiveScan(input, inclusive_output, scan_op);\n\n // Grab aggregate from last warp lane\n warp_aggregate = ShuffleIndex(inclusive_output, LOGICAL_WARP_THREADS - 1, LOGICAL_WARP_THREADS, member_mask);\n }\n\n\n //---------------------------------------------------------------------\n // Get exclusive from inclusive\n //---------------------------------------------------------------------\n\n /// Update inclusive and exclusive using input and inclusive\n template \n __device__ __forceinline__ void Update(\n T /*input*/, ///< [in]\n T &inclusive, ///< [in, out]\n T &exclusive, ///< [out]\n ScanOpT /*scan_op*/, ///< [in]\n IsIntegerT /*is_integer*/) ///< [in]\n {\n // initial value unknown\n exclusive = ShuffleUp(inclusive, 1, 0, member_mask);\n }\n\n /// Update inclusive and exclusive using input and inclusive (specialized for summation of integer types)\n __device__ __forceinline__ void Update(\n T input,\n T &inclusive,\n T &exclusive,\n cub::Sum /*scan_op*/,\n Int2Type /*is_integer*/)\n {\n // initial value presumed 0\n exclusive = inclusive - input;\n }\n\n /// Update inclusive and exclusive using initial value using input, inclusive, and initial value\n template \n __device__ __forceinline__ void Update (\n T /*input*/,\n T &inclusive,\n T &exclusive,\n ScanOpT scan_op,\n T initial_value,\n IsIntegerT /*is_integer*/)\n {\n inclusive = scan_op(initial_value, inclusive);\n exclusive = ShuffleUp(inclusive, 1, 0, member_mask);\n\n unsigned int segment_id = (IS_ARCH_WARP) ?\n lane_id :\n lane_id % LOGICAL_WARP_THREADS;\n\n if (segment_id == 0)\n exclusive = initial_value;\n }\n\n /// Update inclusive and exclusive using initial value using input and inclusive (specialized for summation of integer types)\n __device__ __forceinline__ void Update (\n T input,\n T &inclusive,\n T &exclusive,\n cub::Sum scan_op,\n T initial_value,\n Int2Type /*is_integer*/)\n {\n inclusive = scan_op(initial_value, inclusive);\n exclusive = inclusive - input;\n }\n\n\n /// Update inclusive, exclusive, and warp aggregate using input and inclusive\n template \n __device__ __forceinline__ void Update (\n T input,\n T &inclusive,\n T &exclusive,\n T &warp_aggregate,\n ScanOpT scan_op,\n IsIntegerT is_integer)\n {\n warp_aggregate = ShuffleIndex(inclusive, LOGICAL_WARP_THREADS - 1, LOGICAL_WARP_THREADS, member_mask);\n Update(input, inclusive, exclusive, scan_op, is_integer);\n }\n\n /// Update inclusive, exclusive, and warp aggregate using input, inclusive, and initial value\n template \n __device__ __forceinline__ void Update (\n T input,\n T &inclusive,\n T &exclusive,\n T &warp_aggregate,\n ScanOpT scan_op,\n T initial_value,\n IsIntegerT is_integer)\n {\n warp_aggregate = ShuffleIndex(inclusive, LOGICAL_WARP_THREADS - 1, LOGICAL_WARP_THREADS, member_mask);\n Update(input, inclusive, exclusive, scan_op, initial_value, is_integer);\n }\n\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/specializations/warp_scan_smem.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * cub::WarpScanSmem provides smem-based variants of parallel prefix scan of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"../../thread/thread_operators.cuh\"\n#include \"../../thread/thread_load.cuh\"\n#include \"../../thread/thread_store.cuh\"\n#include \"../../util_type.cuh\"\n#include \"../../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\brief WarpScanSmem provides smem-based variants of parallel prefix scan of items partitioned across a CUDA thread warp.\n */\ntemplate <\n typename T, ///< Data type being scanned\n int LOGICAL_WARP_THREADS, ///< Number of threads per logical warp\n int PTX_ARCH> ///< The PTX compute capability for which to to specialize this collective\nstruct WarpScanSmem\n{\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// Whether the logical warp size is a power-of-two\n IS_POW_OF_TWO = PowerOfTwo::VALUE,\n\n /// The number of warp scan steps\n STEPS = Log2::VALUE,\n\n /// The number of threads in half a warp\n HALF_WARP_THREADS = 1 << (STEPS - 1),\n\n /// The number of shared memory elements per warp\n WARP_SMEM_ELEMENTS = LOGICAL_WARP_THREADS + HALF_WARP_THREADS,\n };\n\n /// Storage cell type (workaround for SM1x compiler bugs with custom-ops like Max() on signed chars)\n typedef typename If<((Equals::VALUE || Equals::VALUE) && (PTX_ARCH < 200)), int, T>::Type CellT;\n\n /// Shared memory storage layout type (1.5 warps-worth of elements for each warp)\n typedef CellT _TempStorage[WARP_SMEM_ELEMENTS];\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n _TempStorage &temp_storage;\n unsigned int lane_id;\n unsigned int member_mask;\n\n\n /******************************************************************************\n * Construction\n ******************************************************************************/\n\n /// Constructor\n __device__ __forceinline__ WarpScanSmem(\n TempStorage &temp_storage)\n :\n temp_storage(temp_storage.Alias()),\n\n lane_id(IS_ARCH_WARP ?\n LaneId() :\n LaneId() % LOGICAL_WARP_THREADS),\n\n member_mask((0xffffffff >> (32 - LOGICAL_WARP_THREADS)) << ((IS_ARCH_WARP || !IS_POW_OF_TWO ) ?\n 0 : // arch-width and non-power-of-two subwarps cannot be tiled with the arch-warp\n ((LaneId() / LOGICAL_WARP_THREADS) * LOGICAL_WARP_THREADS)))\n {}\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\n /// Basic inclusive scan iteration (template unrolled, inductive-case specialization)\n template <\n bool HAS_IDENTITY,\n int STEP,\n typename ScanOp>\n __device__ __forceinline__ void ScanStep(\n T &partial,\n ScanOp scan_op,\n Int2Type /*step*/)\n {\n const int OFFSET = 1 << STEP;\n\n // Share partial into buffer\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) partial);\n\n WARP_SYNC(member_mask);\n\n // Update partial if addend is in range\n if (HAS_IDENTITY || (lane_id >= OFFSET))\n {\n T addend = (T) ThreadLoad(&temp_storage[HALF_WARP_THREADS + lane_id - OFFSET]);\n partial = scan_op(addend, partial);\n }\n WARP_SYNC(member_mask);\n\n ScanStep(partial, scan_op, Int2Type());\n }\n\n\n /// Basic inclusive scan iteration(template unrolled, base-case specialization)\n template <\n bool HAS_IDENTITY,\n typename ScanOp>\n __device__ __forceinline__ void ScanStep(\n T &/*partial*/,\n ScanOp /*scan_op*/,\n Int2Type /*step*/)\n {}\n\n\n /// Inclusive prefix scan (specialized for summation across primitive types)\n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n Sum scan_op, ///< [in] Binary scan operator\n Int2Type /*is_primitive*/) ///< [in] Marker type indicating whether T is primitive type\n {\n T identity = 0;\n ThreadStore(&temp_storage[lane_id], (CellT) identity);\n\n WARP_SYNC(member_mask);\n\n // Iterate scan steps\n output = input;\n ScanStep(output, scan_op, Int2Type<0>());\n }\n\n\n /// Inclusive prefix scan\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op, ///< [in] Binary scan operator\n Int2Type /*is_primitive*/) ///< [in] Marker type indicating whether T is primitive type\n {\n // Iterate scan steps\n output = input;\n ScanStep(output, scan_op, Int2Type<0>());\n }\n\n\n /******************************************************************************\n * Interface\n ******************************************************************************/\n\n //---------------------------------------------------------------------\n // Broadcast\n //---------------------------------------------------------------------\n\n /// Broadcast\n __device__ __forceinline__ T Broadcast(\n T input, ///< [in] The value to broadcast\n unsigned int src_lane) ///< [in] Which warp lane is to do the broadcasting\n {\n if (lane_id == src_lane)\n {\n ThreadStore(temp_storage, (CellT) input);\n }\n\n WARP_SYNC(member_mask);\n\n return (T)ThreadLoad(temp_storage);\n }\n\n\n //---------------------------------------------------------------------\n // Inclusive operations\n //---------------------------------------------------------------------\n\n /// Inclusive scan\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InclusiveScan(input, inclusive_output, scan_op, Int2Type::PRIMITIVE>());\n }\n\n\n /// Inclusive scan with aggregate\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op, ///< [in] Binary scan operator\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InclusiveScan(input, inclusive_output, scan_op);\n\n // Retrieve aggregate\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive_output);\n\n WARP_SYNC(member_mask);\n\n warp_aggregate = (T) ThreadLoad(&temp_storage[WARP_SMEM_ELEMENTS - 1]);\n\n WARP_SYNC(member_mask);\n }\n\n\n //---------------------------------------------------------------------\n // Get exclusive from inclusive\n //---------------------------------------------------------------------\n\n /// Update inclusive and exclusive using input and inclusive\n template \n __device__ __forceinline__ void Update(\n T /*input*/, ///< [in]\n T &inclusive, ///< [in, out]\n T &exclusive, ///< [out]\n ScanOpT /*scan_op*/, ///< [in]\n IsIntegerT /*is_integer*/) ///< [in]\n {\n // initial value unknown\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n exclusive = (T) ThreadLoad(&temp_storage[HALF_WARP_THREADS + lane_id - 1]);\n }\n\n /// Update inclusive and exclusive using input and inclusive (specialized for summation of integer types)\n __device__ __forceinline__ void Update(\n T input,\n T &inclusive,\n T &exclusive,\n cub::Sum /*scan_op*/,\n Int2Type /*is_integer*/)\n {\n // initial value presumed 0\n exclusive = inclusive - input;\n }\n\n /// Update inclusive and exclusive using initial value using input, inclusive, and initial value\n template \n __device__ __forceinline__ void Update (\n T /*input*/,\n T &inclusive,\n T &exclusive,\n ScanOpT scan_op,\n T initial_value,\n IsIntegerT /*is_integer*/)\n {\n inclusive = scan_op(initial_value, inclusive);\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n exclusive = (T) ThreadLoad(&temp_storage[HALF_WARP_THREADS + lane_id - 1]);\n if (lane_id == 0)\n exclusive = initial_value;\n }\n\n /// Update inclusive and exclusive using initial value using input and inclusive (specialized for summation of integer types)\n __device__ __forceinline__ void Update (\n T input,\n T &inclusive,\n T &exclusive,\n cub::Sum scan_op,\n T initial_value,\n Int2Type /*is_integer*/)\n {\n inclusive = scan_op(initial_value, inclusive);\n exclusive = inclusive - input;\n }\n\n\n /// Update inclusive, exclusive, and warp aggregate using input and inclusive\n template \n __device__ __forceinline__ void Update (\n T /*input*/,\n T &inclusive,\n T &exclusive,\n T &warp_aggregate,\n ScanOpT /*scan_op*/,\n IsIntegerT /*is_integer*/)\n {\n // Initial value presumed to be unknown or identity (either way our padding is correct)\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n exclusive = (T) ThreadLoad(&temp_storage[HALF_WARP_THREADS + lane_id - 1]);\n warp_aggregate = (T) ThreadLoad(&temp_storage[WARP_SMEM_ELEMENTS - 1]);\n }\n\n /// Update inclusive, exclusive, and warp aggregate using input and inclusive (specialized for summation of integer types)\n __device__ __forceinline__ void Update (\n T input,\n T &inclusive,\n T &exclusive,\n T &warp_aggregate,\n cub::Sum /*scan_o*/,\n Int2Type /*is_integer*/)\n {\n // Initial value presumed to be unknown or identity (either way our padding is correct)\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n warp_aggregate = (T) ThreadLoad(&temp_storage[WARP_SMEM_ELEMENTS - 1]);\n exclusive = inclusive - input;\n }\n\n /// Update inclusive, exclusive, and warp aggregate using input, inclusive, and initial value\n template \n __device__ __forceinline__ void Update (\n T /*input*/,\n T &inclusive,\n T &exclusive,\n T &warp_aggregate,\n ScanOpT scan_op,\n T initial_value,\n IsIntegerT /*is_integer*/)\n {\n // Broadcast warp aggregate\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n warp_aggregate = (T) ThreadLoad(&temp_storage[WARP_SMEM_ELEMENTS - 1]);\n\n WARP_SYNC(member_mask);\n\n // Update inclusive with initial value\n inclusive = scan_op(initial_value, inclusive);\n\n // Get exclusive from exclusive\n ThreadStore(&temp_storage[HALF_WARP_THREADS + lane_id - 1], (CellT) inclusive);\n\n WARP_SYNC(member_mask);\n\n exclusive = (T) ThreadLoad(&temp_storage[HALF_WARP_THREADS + lane_id - 2]);\n\n if (lane_id == 0)\n exclusive = initial_value;\n }\n\n\n};\n\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/warp_reduce.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::WarpReduce class provides [collective](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"specializations/warp_reduce_shfl.cuh\"\n#include \"specializations/warp_reduce_smem.cuh\"\n#include \"../thread/thread_operators.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n\n/**\n * \\addtogroup WarpModule\n * @{\n */\n\n/**\n * \\brief The WarpReduce class provides [collective](index.html#sec0) methods for computing a parallel reduction of items partitioned across a CUDA thread warp. ![](warp_reduce_logo.png)\n *\n * \\tparam T The reduction input/output element type\n * \\tparam LOGICAL_WARP_THREADS [optional] The number of threads per \"logical\" warp (may be less than the number of hardware warp threads). Default is the warp size of the targeted CUDA compute-capability (e.g., 32 threads for SM20).\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - A reduction (or fold)\n * uses a binary combining operator to compute a single aggregate from a list of input elements.\n * - Supports \"logical\" warps smaller than the physical warp size (e.g., logical warps of 8 threads)\n * - The number of entrant threads must be an multiple of \\p LOGICAL_WARP_THREADS\n *\n * \\par Performance Considerations\n * - Uses special instructions when applicable (e.g., warp \\p SHFL instructions)\n * - Uses synchronization-free communication between warp lanes when applicable\n * - Incurs zero bank conflicts for most types\n * - Computation is slightly more efficient (i.e., having lower instruction overhead) for:\n * - Summation (vs. generic reduction)\n * - The architecture's warp size is a whole multiple of \\p LOGICAL_WARP_THREADS\n *\n * \\par Simple Examples\n * \\warpcollective{WarpReduce}\n * \\par\n * The code snippet below illustrates four concurrent warp sum reductions within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for 4 warps\n * __shared__ typename WarpReduce::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Return the warp-wide sums to each lane0 (threads 0, 32, 64, and 96)\n * int warp_id = threadIdx.x / 32;\n * int aggregate = WarpReduce(temp_storage[warp_id]).Sum(thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, 1, 2, 3, ..., 127}.\n * The corresponding output \\p aggregate in threads 0, 32, 64, and 96 will \\p 496, \\p 1520,\n * \\p 2544, and \\p 3568, respectively (and is undefined in other threads).\n *\n * \\par\n * The code snippet below illustrates a single warp sum reduction within a block of\n * 128 threads.\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n * ...\n *\n * // Only the first warp performs a reduction\n * if (threadIdx.x < 32)\n * {\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Return the warp-wide sum to lane0\n * int aggregate = WarpReduce(temp_storage).Sum(thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the warp of threads is {0, 1, 2, 3, ..., 31}.\n * The corresponding output \\p aggregate in thread0 will be \\p 496 (and is undefined in other threads).\n *\n */\ntemplate <\n typename T,\n int LOGICAL_WARP_THREADS = CUB_PTX_WARP_THREADS,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass WarpReduce\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// Whether the logical warp size is a power-of-two\n IS_POW_OF_TWO = PowerOfTwo::VALUE,\n };\n\npublic:\n\n #ifndef DOXYGEN_SHOULD_SKIP_THIS // Do not document\n\n /// Internal specialization. Use SHFL-based reduction if (architecture is >= SM30) and (LOGICAL_WARP_THREADS is a power-of-two)\n typedef typename If<(PTX_ARCH >= 300) && (IS_POW_OF_TWO),\n WarpReduceShfl,\n WarpReduceSmem >::Type InternalWarpReduce;\n\n #endif // DOXYGEN_SHOULD_SKIP_THIS\n\n\nprivate:\n\n /// Shared memory storage layout type for WarpReduce\n typedef typename InternalWarpReduce::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n\n\n /******************************************************************************\n * Utility methods\n ******************************************************************************/\n\npublic:\n\n /// \\smemstorage{WarpReduce}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage. Logical warp and lane identifiers are constructed from threadIdx.x.\n */\n __device__ __forceinline__ WarpReduce(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias())\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Summation reductions\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes a warp-wide sum in the calling warp. The output is valid in warp lane0.\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp sum reductions within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for 4 warps\n * __shared__ typename WarpReduce::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Return the warp-wide sums to each lane0\n * int warp_id = threadIdx.x / 32;\n * int aggregate = WarpReduce(temp_storage[warp_id]).Sum(thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, 1, 2, 3, ..., 127}.\n * The corresponding output \\p aggregate in threads 0, 32, 64, and 96 will \\p 496, \\p 1520,\n * \\p 2544, and \\p 3568, respectively (and is undefined in other threads).\n *\n */\n __device__ __forceinline__ T Sum(\n T input) ///< [in] Calling thread's input\n {\n return InternalWarpReduce(temp_storage).template Reduce(input, LOGICAL_WARP_THREADS, cub::Sum());\n }\n\n /**\n * \\brief Computes a partially-full warp-wide sum in the calling warp. The output is valid in warp lane0.\n *\n * All threads across the calling warp must agree on the same value for \\p valid_items. Otherwise the result is undefined.\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a sum reduction within a single, partially-full\n * block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(int *d_data, int valid_items)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item per thread if in range\n * int thread_data;\n * if (threadIdx.x < valid_items)\n * thread_data = d_data[threadIdx.x];\n *\n * // Return the warp-wide sums to each lane0\n * int aggregate = WarpReduce(temp_storage).Sum(\n * thread_data, valid_items);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is {0, 1, 2, 3, 4, ... and \\p valid_items\n * is \\p 4. The corresponding output \\p aggregate in thread0 is \\p 6 (and is\n * undefined in other threads).\n *\n */\n __device__ __forceinline__ T Sum(\n T input, ///< [in] Calling thread's input\n int valid_items) ///< [in] Total number of valid items in the calling thread's logical warp (may be less than \\p LOGICAL_WARP_THREADS)\n {\n // Determine if we don't need bounds checking\n return InternalWarpReduce(temp_storage).template Reduce(input, valid_items, cub::Sum());\n }\n\n\n /**\n * \\brief Computes a segmented sum in the calling warp where segments are defined by head-flags. The sum of each segment is returned to the first lane in that segment (which always includes lane0).\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a head-segmented warp sum\n * reduction within a block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item and flag per thread\n * int thread_data = ...\n * int head_flag = ...\n *\n * // Return the warp-wide sums to each lane0\n * int aggregate = WarpReduce(temp_storage).HeadSegmentedSum(\n * thread_data, head_flag);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data and \\p head_flag across the block of threads\n * is {0, 1, 2, 3, ..., 31 and is {1, 0, 0, 0, 1, 0, 0, 0, ..., 1, 0, 0, 0,\n * respectively. The corresponding output \\p aggregate in threads 0, 4, 8, etc. will be\n * \\p 6, \\p 22, \\p 38, etc. (and is undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n *\n */\n template <\n typename FlagT>\n __device__ __forceinline__ T HeadSegmentedSum(\n T input, ///< [in] Calling thread's input\n FlagT head_flag) ///< [in] Head flag denoting whether or not \\p input is the start of a new segment\n {\n return HeadSegmentedReduce(input, head_flag, cub::Sum());\n }\n\n\n /**\n * \\brief Computes a segmented sum in the calling warp where segments are defined by tail-flags. The sum of each segment is returned to the first lane in that segment (which always includes lane0).\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a tail-segmented warp sum\n * reduction within a block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item and flag per thread\n * int thread_data = ...\n * int tail_flag = ...\n *\n * // Return the warp-wide sums to each lane0\n * int aggregate = WarpReduce(temp_storage).TailSegmentedSum(\n * thread_data, tail_flag);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data and \\p tail_flag across the block of threads\n * is {0, 1, 2, 3, ..., 31 and is {0, 0, 0, 1, 0, 0, 0, 1, ..., 0, 0, 0, 1,\n * respectively. The corresponding output \\p aggregate in threads 0, 4, 8, etc. will be\n * \\p 6, \\p 22, \\p 38, etc. (and is undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template <\n typename FlagT>\n __device__ __forceinline__ T TailSegmentedSum(\n T input, ///< [in] Calling thread's input\n FlagT tail_flag) ///< [in] Head flag denoting whether or not \\p input is the start of a new segment\n {\n return TailSegmentedReduce(input, tail_flag, cub::Sum());\n }\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Generic reductions\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes a warp-wide reduction in the calling warp using the specified binary reduction functor. The output is valid in warp lane0.\n *\n * Supports non-commutative reduction operators\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp max reductions within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for 4 warps\n * __shared__ typename WarpReduce::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Return the warp-wide reductions to each lane0\n * int warp_id = threadIdx.x / 32;\n * int aggregate = WarpReduce(temp_storage[warp_id]).Reduce(\n * thread_data, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, 1, 2, 3, ..., 127}.\n * The corresponding output \\p aggregate in threads 0, 32, 64, and 96 will \\p 31, \\p 63,\n * \\p 95, and \\p 127, respectively (and is undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n ReductionOp reduction_op) ///< [in] Binary reduction operator\n {\n return InternalWarpReduce(temp_storage).template Reduce(input, LOGICAL_WARP_THREADS, reduction_op);\n }\n\n /**\n * \\brief Computes a partially-full warp-wide reduction in the calling warp using the specified binary reduction functor. The output is valid in warp lane0.\n *\n * All threads across the calling warp must agree on the same value for \\p valid_items. Otherwise the result is undefined.\n *\n * Supports non-commutative reduction operators\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a max reduction within a single, partially-full\n * block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(int *d_data, int valid_items)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item per thread if in range\n * int thread_data;\n * if (threadIdx.x < valid_items)\n * thread_data = d_data[threadIdx.x];\n *\n * // Return the warp-wide reductions to each lane0\n * int aggregate = WarpReduce(temp_storage).Reduce(\n * thread_data, cub::Max(), valid_items);\n *\n * \\endcode\n * \\par\n * Suppose the input \\p d_data is {0, 1, 2, 3, 4, ... and \\p valid_items\n * is \\p 4. The corresponding output \\p aggregate in thread0 is \\p 3 (and is\n * undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ T Reduce(\n T input, ///< [in] Calling thread's input\n ReductionOp reduction_op, ///< [in] Binary reduction operator\n int valid_items) ///< [in] Total number of valid items in the calling thread's logical warp (may be less than \\p LOGICAL_WARP_THREADS)\n {\n return InternalWarpReduce(temp_storage).template Reduce(input, valid_items, reduction_op);\n }\n\n\n /**\n * \\brief Computes a segmented reduction in the calling warp where segments are defined by head-flags. The reduction of each segment is returned to the first lane in that segment (which always includes lane0).\n *\n * Supports non-commutative reduction operators\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a head-segmented warp max\n * reduction within a block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item and flag per thread\n * int thread_data = ...\n * int head_flag = ...\n *\n * // Return the warp-wide reductions to each lane0\n * int aggregate = WarpReduce(temp_storage).HeadSegmentedReduce(\n * thread_data, head_flag, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data and \\p head_flag across the block of threads\n * is {0, 1, 2, 3, ..., 31 and is {1, 0, 0, 0, 1, 0, 0, 0, ..., 1, 0, 0, 0,\n * respectively. The corresponding output \\p aggregate in threads 0, 4, 8, etc. will be\n * \\p 3, \\p 7, \\p 11, etc. (and is undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template <\n typename ReductionOp,\n typename FlagT>\n __device__ __forceinline__ T HeadSegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT head_flag, ///< [in] Head flag denoting whether or not \\p input is the start of a new segment\n ReductionOp reduction_op) ///< [in] Reduction operator\n {\n return InternalWarpReduce(temp_storage).template SegmentedReduce(input, head_flag, reduction_op);\n }\n\n\n /**\n * \\brief Computes a segmented reduction in the calling warp where segments are defined by tail-flags. The reduction of each segment is returned to the first lane in that segment (which always includes lane0).\n *\n * Supports non-commutative reduction operators\n *\n * \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates a tail-segmented warp max\n * reduction within a block of 32 threads (one warp).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpReduce for type int\n * typedef cub::WarpReduce WarpReduce;\n *\n * // Allocate WarpReduce shared memory for one warp\n * __shared__ typename WarpReduce::TempStorage temp_storage;\n *\n * // Obtain one input item and flag per thread\n * int thread_data = ...\n * int tail_flag = ...\n *\n * // Return the warp-wide reductions to each lane0\n * int aggregate = WarpReduce(temp_storage).TailSegmentedReduce(\n * thread_data, tail_flag, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data and \\p tail_flag across the block of threads\n * is {0, 1, 2, 3, ..., 31 and is {0, 0, 0, 1, 0, 0, 0, 1, ..., 0, 0, 0, 1,\n * respectively. The corresponding output \\p aggregate in threads 0, 4, 8, etc. will be\n * \\p 3, \\p 7, \\p 11, etc. (and is undefined in other threads).\n *\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template <\n typename ReductionOp,\n typename FlagT>\n __device__ __forceinline__ T TailSegmentedReduce(\n T input, ///< [in] Calling thread's input\n FlagT tail_flag, ///< [in] Tail flag denoting whether or not \\p input is the end of the current segment\n ReductionOp reduction_op) ///< [in] Reduction operator\n {\n return InternalWarpReduce(temp_storage).template SegmentedReduce(input, tail_flag, reduction_op);\n }\n\n\n\n //@} end member group\n};\n\n/** @} */ // end group WarpModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/cub/warp/warp_scan.cuh", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/**\n * \\file\n * The cub::WarpScan class provides [collective](index.html#sec0) methods for computing a parallel prefix scan of items partitioned across a CUDA thread warp.\n */\n\n#pragma once\n\n#include \"specializations/warp_scan_shfl.cuh\"\n#include \"specializations/warp_scan_smem.cuh\"\n#include \"../thread/thread_operators.cuh\"\n#include \"../util_arch.cuh\"\n#include \"../util_type.cuh\"\n#include \"../util_namespace.cuh\"\n\n/// Optional outer namespace(s)\nCUB_NS_PREFIX\n\n/// CUB namespace\nnamespace cub {\n\n/**\n * \\addtogroup WarpModule\n * @{\n */\n\n/**\n * \\brief The WarpScan class provides [collective](index.html#sec0) methods for computing a parallel prefix scan of items partitioned across a CUDA thread warp. ![](warp_scan_logo.png)\n *\n * \\tparam T The scan input/output element type\n * \\tparam LOGICAL_WARP_THREADS [optional] The number of threads per \"logical\" warp (may be less than the number of hardware warp threads). Default is the warp size associated with the CUDA Compute Capability targeted by the compiler (e.g., 32 threads for SM20).\n * \\tparam PTX_ARCH [optional] \\ptxversion\n *\n * \\par Overview\n * - Given a list of input elements and a binary reduction operator, a [prefix scan](http://en.wikipedia.org/wiki/Prefix_sum)\n * produces an output list where each element is computed to be the reduction\n * of the elements occurring earlier in the input list. Prefix sum\n * connotes a prefix scan with the addition operator. The term \\em inclusive indicates\n * that the ith output reduction incorporates the ith input.\n * The term \\em exclusive indicates the ith input is not incorporated into\n * the ith output reduction.\n * - Supports non-commutative scan operators\n * - Supports \"logical\" warps smaller than the physical warp size (e.g., a logical warp of 8 threads)\n * - The number of entrant threads must be an multiple of \\p LOGICAL_WARP_THREADS\n *\n * \\par Performance Considerations\n * - Uses special instructions when applicable (e.g., warp \\p SHFL)\n * - Uses synchronization-free communication between warp lanes when applicable\n * - Incurs zero bank conflicts for most types\n * - Computation is slightly more efficient (i.e., having lower instruction overhead) for:\n * - Summation (vs. generic scan)\n * - The architecture's warp size is a whole multiple of \\p LOGICAL_WARP_THREADS\n *\n * \\par Simple Examples\n * \\warpcollective{WarpScan}\n * \\par\n * The code snippet below illustrates four concurrent warp prefix sums within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute warp-wide prefix sums\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data in each of the four warps of threads will be\n * 0, 1, 2, 3, ..., 31}.\n *\n * \\par\n * The code snippet below illustrates a single warp prefix sum within a block of\n * 128 threads.\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for one warp\n * __shared__ typename WarpScan::TempStorage temp_storage;\n * ...\n *\n * // Only the first warp performs a prefix sum\n * if (threadIdx.x < 32)\n * {\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute warp-wide prefix sums\n * WarpScan(temp_storage).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the warp of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data will be {0, 1, 2, 3, ..., 31}.\n *\n */\ntemplate <\n typename T,\n int LOGICAL_WARP_THREADS = CUB_PTX_WARP_THREADS,\n int PTX_ARCH = CUB_PTX_ARCH>\nclass WarpScan\n{\nprivate:\n\n /******************************************************************************\n * Constants and type definitions\n ******************************************************************************/\n\n enum\n {\n /// Whether the logical warp size and the PTX warp size coincide\n IS_ARCH_WARP = (LOGICAL_WARP_THREADS == CUB_WARP_THREADS(PTX_ARCH)),\n\n /// Whether the logical warp size is a power-of-two\n IS_POW_OF_TWO = ((LOGICAL_WARP_THREADS & (LOGICAL_WARP_THREADS - 1)) == 0),\n\n /// Whether the data type is an integer (which has fully-associative addition)\n IS_INTEGER = ((Traits::CATEGORY == SIGNED_INTEGER) || (Traits::CATEGORY == UNSIGNED_INTEGER))\n };\n\n /// Internal specialization. Use SHFL-based scan if (architecture is >= SM30) and (LOGICAL_WARP_THREADS is a power-of-two)\n typedef typename If<(PTX_ARCH >= 300) && (IS_POW_OF_TWO),\n WarpScanShfl,\n WarpScanSmem >::Type InternalWarpScan;\n\n /// Shared memory storage layout type for WarpScan\n typedef typename InternalWarpScan::TempStorage _TempStorage;\n\n\n /******************************************************************************\n * Thread fields\n ******************************************************************************/\n\n /// Shared storage reference\n _TempStorage &temp_storage;\n unsigned int lane_id;\n\n\n\n /******************************************************************************\n * Public types\n ******************************************************************************/\n\npublic:\n\n /// \\smemstorage{WarpScan}\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n /******************************************************************//**\n * \\name Collective constructors\n *********************************************************************/\n //@{\n\n /**\n * \\brief Collective constructor using the specified memory allocation as temporary storage. Logical warp and lane identifiers are constructed from threadIdx.x.\n */\n __device__ __forceinline__ WarpScan(\n TempStorage &temp_storage) ///< [in] Reference to memory allocation having layout type TempStorage\n :\n temp_storage(temp_storage.Alias()),\n lane_id(IS_ARCH_WARP ?\n LaneId() :\n LaneId() % LOGICAL_WARP_THREADS)\n {}\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive prefix sums\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an inclusive prefix sum across the calling warp.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide inclusive prefix sums within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute inclusive warp-wide prefix sums\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).InclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data in each of the four warps of threads will be\n * 1, 2, 3, ..., 32}.\n */\n __device__ __forceinline__ void InclusiveSum(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output) ///< [out] Calling thread's output item. May be aliased with \\p input.\n {\n InclusiveScan(input, inclusive_output, cub::Sum());\n }\n\n\n /**\n * \\brief Computes an inclusive prefix sum across the calling warp. Also provides every thread with the warp-wide \\p warp_aggregate of all inputs.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide inclusive prefix sums within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute inclusive warp-wide prefix sums\n * int warp_aggregate;\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).InclusiveSum(thread_data, thread_data, warp_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data in each of the four warps of threads will be\n * 1, 2, 3, ..., 32}. Furthermore, \\p warp_aggregate for all threads in all warps will be \\p 32.\n */\n __device__ __forceinline__ void InclusiveSum(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InclusiveScan(input, inclusive_output, cub::Sum(), warp_aggregate);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Exclusive prefix sums\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes an exclusive prefix sum across the calling warp. The value of 0 is applied as the initial value, and is assigned to \\p exclusive_output in thread0.\n *\n * \\par\n * - \\identityzero\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix sums within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix sums\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveSum(thread_data, thread_data);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data in each of the four warps of threads will be\n * 0, 1, 2, ..., 31}.\n *\n */\n __device__ __forceinline__ void ExclusiveSum(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output) ///< [out] Calling thread's output item. May be aliased with \\p input.\n {\n T initial_value = 0;\n ExclusiveScan(input, exclusive_output, initial_value, cub::Sum());\n }\n\n\n /**\n * \\brief Computes an exclusive prefix sum across the calling warp. The value of 0 is applied as the initial value, and is assigned to \\p exclusive_output in thread0. Also provides every thread with the warp-wide \\p warp_aggregate of all inputs.\n *\n * \\par\n * - \\identityzero\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix sums within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix sums\n * int warp_aggregate;\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveSum(thread_data, thread_data, warp_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {1, 1, 1, 1, ...}.\n * The corresponding output \\p thread_data in each of the four warps of threads will be\n * 0, 1, 2, ..., 31}. Furthermore, \\p warp_aggregate for all threads in all warps will be \\p 32.\n */\n __device__ __forceinline__ void ExclusiveSum(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n T initial_value = 0;\n ExclusiveScan(input, exclusive_output, initial_value, cub::Sum(), warp_aggregate);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Inclusive prefix scans\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes an inclusive prefix scan using the specified binary scan functor across the calling warp.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide inclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute inclusive warp-wide prefix max scans\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).InclusiveScan(thread_data, thread_data, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * 0, 0, 2, 2, ..., 30, 30, the output for the second warp would be 32, 32, 34, 34, ..., 62, 62, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InternalWarpScan(temp_storage).InclusiveScan(input, inclusive_output, scan_op);\n }\n\n\n /**\n * \\brief Computes an inclusive prefix scan using the specified binary scan functor across the calling warp. Also provides every thread with the warp-wide \\p warp_aggregate of all inputs.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide inclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute inclusive warp-wide prefix max scans\n * int warp_aggregate;\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).InclusiveScan(\n * thread_data, thread_data, cub::Max(), warp_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * 0, 0, 2, 2, ..., 30, 30, the output for the second warp would be 32, 32, 34, 34, ..., 62, 62, etc.\n * Furthermore, \\p warp_aggregate would be assigned \\p 30 for threads in the first warp, \\p 62 for threads\n * in the second warp, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void InclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op, ///< [in] Binary scan operator\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InternalWarpScan(temp_storage).InclusiveScan(input, inclusive_output, scan_op, warp_aggregate);\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Exclusive prefix scans\n *********************************************************************/\n //@{\n\n /**\n * \\brief Computes an exclusive prefix scan using the specified binary scan functor across the calling warp. Because no initial value is supplied, the \\p output computed for warp-lane0 is undefined.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix max scans\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveScan(thread_data, thread_data, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * ?, 0, 0, 2, ..., 28, 30, the output for the second warp would be ?, 32, 32, 34, ..., 60, 62, etc.\n * (The output \\p thread_data in warp lane0 is undefined.)\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InternalWarpScan internal(temp_storage);\n\n T inclusive_output;\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n scan_op,\n Int2Type());\n }\n\n\n /**\n * \\brief Computes an exclusive prefix scan using the specified binary scan functor across the calling warp.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix max scans\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * INT_MIN, 0, 0, 2, ..., 28, 30, the output for the second warp would be 30, 32, 32, 34, ..., 60, 62, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n T initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InternalWarpScan internal(temp_storage);\n\n T inclusive_output;\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n scan_op,\n initial_value,\n Int2Type());\n }\n\n\n /**\n * \\brief Computes an exclusive prefix scan using the specified binary scan functor across the calling warp. Because no initial value is supplied, the \\p output computed for warp-lane0 is undefined. Also provides every thread with the warp-wide \\p warp_aggregate of all inputs.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix max scans\n * int warp_aggregate;\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveScan(thread_data, thread_data, cub::Max(), warp_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * ?, 0, 0, 2, ..., 28, 30, the output for the second warp would be ?, 32, 32, 34, ..., 60, 62, etc.\n * (The output \\p thread_data in warp lane0 is undefined.) Furthermore, \\p warp_aggregate would be assigned \\p 30 for threads in the first warp, \\p 62 for threads\n * in the second warp, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n ScanOp scan_op, ///< [in] Binary scan operator\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InternalWarpScan internal(temp_storage);\n\n T inclusive_output;\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n warp_aggregate,\n scan_op,\n Int2Type());\n }\n\n\n /**\n * \\brief Computes an exclusive prefix scan using the specified binary scan functor across the calling warp. Also provides every thread with the warp-wide \\p warp_aggregate of all inputs.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix max scans\n * int warp_aggregate;\n * int warp_id = threadIdx.x / 32;\n * WarpScan(temp_storage[warp_id]).ExclusiveScan(thread_data, thread_data, INT_MIN, cub::Max(), warp_aggregate);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p thread_data in the first warp would be\n * INT_MIN, 0, 0, 2, ..., 28, 30, the output for the second warp would be 30, 32, 32, 34, ..., 60, 62, etc.\n * Furthermore, \\p warp_aggregate would be assigned \\p 30 for threads in the first warp, \\p 62 for threads\n * in the second warp, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void ExclusiveScan(\n T input, ///< [in] Calling thread's input item.\n T &exclusive_output, ///< [out] Calling thread's output item. May be aliased with \\p input.\n T initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op, ///< [in] Binary scan operator\n T &warp_aggregate) ///< [out] Warp-wide aggregate reduction of input items.\n {\n InternalWarpScan internal(temp_storage);\n\n T inclusive_output;\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n warp_aggregate,\n scan_op,\n initial_value,\n Int2Type());\n }\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Combination (inclusive & exclusive) prefix scans\n *********************************************************************/\n //@{\n\n\n /**\n * \\brief Computes both inclusive and exclusive prefix scans using the specified binary scan functor across the calling warp. Because no initial value is supplied, the \\p exclusive_output computed for warp-lane0 is undefined.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide exclusive prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute exclusive warp-wide prefix max scans\n * int inclusive_partial, exclusive_partial;\n * WarpScan(temp_storage[warp_id]).Scan(thread_data, inclusive_partial, exclusive_partial, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p inclusive_partial in the first warp would be\n * 0, 0, 2, 2, ..., 30, 30, the output for the second warp would be 32, 32, 34, 34, ..., 62, 62, etc.\n * The corresponding output \\p exclusive_partial in the first warp would be\n * ?, 0, 0, 2, ..., 28, 30, the output for the second warp would be ?, 32, 32, 34, ..., 60, 62, etc.\n * (The output \\p thread_data in warp lane0 is undefined.)\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void Scan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's inclusive-scan output item.\n T &exclusive_output, ///< [out] Calling thread's exclusive-scan output item.\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InternalWarpScan internal(temp_storage);\n\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n scan_op,\n Int2Type());\n }\n\n\n /**\n * \\brief Computes both inclusive and exclusive prefix scans using the specified binary scan functor across the calling warp.\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates four concurrent warp-wide prefix max scans within a block of\n * 128 threads (one per each of the 32-thread warps).\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Compute inclusive warp-wide prefix max scans\n * int warp_id = threadIdx.x / 32;\n * int inclusive_partial, exclusive_partial;\n * WarpScan(temp_storage[warp_id]).Scan(thread_data, inclusive_partial, exclusive_partial, INT_MIN, cub::Max());\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, -1, 2, -3, ..., 126, -127}.\n * The corresponding output \\p inclusive_partial in the first warp would be\n * 0, 0, 2, 2, ..., 30, 30, the output for the second warp would be 32, 32, 34, 34, ..., 62, 62, etc.\n * The corresponding output \\p exclusive_partial in the first warp would be\n * INT_MIN, 0, 0, 2, ..., 28, 30, the output for the second warp would be 30, 32, 32, 34, ..., 60, 62, etc.\n *\n * \\tparam ScanOp [inferred] Binary scan operator type having member T operator()(const T &a, const T &b)\n */\n template \n __device__ __forceinline__ void Scan(\n T input, ///< [in] Calling thread's input item.\n T &inclusive_output, ///< [out] Calling thread's inclusive-scan output item.\n T &exclusive_output, ///< [out] Calling thread's exclusive-scan output item.\n T initial_value, ///< [in] Initial value to seed the exclusive scan\n ScanOp scan_op) ///< [in] Binary scan operator\n {\n InternalWarpScan internal(temp_storage);\n\n internal.InclusiveScan(input, inclusive_output, scan_op);\n\n internal.Update(\n input,\n inclusive_output,\n exclusive_output,\n scan_op,\n initial_value,\n Int2Type());\n }\n\n\n\n //@} end member group\n /******************************************************************//**\n * \\name Data exchange\n *********************************************************************/\n //@{\n\n /**\n * \\brief Broadcast the value \\p input from warp-lanesrc_lane to all lanes in the warp\n *\n * \\par\n * - \\smemreuse\n *\n * \\par Snippet\n * The code snippet below illustrates the warp-wide broadcasts of values from\n * lanes0 in each of four warps to all other threads in those warps.\n * \\par\n * \\code\n * #include \n *\n * __global__ void ExampleKernel(...)\n * {\n * // Specialize WarpScan for type int\n * typedef cub::WarpScan WarpScan;\n *\n * // Allocate WarpScan shared memory for 4 warps\n * __shared__ typename WarpScan::TempStorage temp_storage[4];\n *\n * // Obtain one input item per thread\n * int thread_data = ...\n *\n * // Broadcast from lane0 in each warp to all other threads in the warp\n * int warp_id = threadIdx.x / 32;\n * thread_data = WarpScan(temp_storage[warp_id]).Broadcast(thread_data, 0);\n *\n * \\endcode\n * \\par\n * Suppose the set of input \\p thread_data across the block of threads is {0, 1, 2, 3, ..., 127}.\n * The corresponding output \\p thread_data will be\n * {0, 0, ..., 0} in warp0,\n * {32, 32, ..., 32} in warp1,\n * {64, 64, ..., 64} in warp2, etc.\n */\n __device__ __forceinline__ T Broadcast(\n T input, ///< [in] The value to broadcast\n unsigned int src_lane) ///< [in] Which warp lane is to do the broadcasting\n {\n return InternalWarpScan(temp_storage).Broadcast(input, src_lane);\n }\n\n //@} end member group\n\n};\n\n/** @} */ // end group WarpModule\n\n} // CUB namespace\nCUB_NS_POSTFIX // Optional outer namespace(s)\n" }, { "path": "external/cub/eclipse code style profile.xml", "content": "\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n" }, { "path": "external/cub/examples/block/.gitignore", "content": "/bin\n/Debug\n/Release\n/cuda55.sdf\n/cuda55.suo\n/cuda60.sdf\n/cuda60.suo\n" }, { "path": "external/cub/examples/block/Makefile", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n\n#-------------------------------------------------------------------------------\n#\n# Makefile usage\n#\n# make [sm=] [cdp=<0|1>] [force32=<0|1>] [abi=<0|1>] [open64=<0|1>] [verbose=<0|1>] [keep=<0|1>]\n#\n#-------------------------------------------------------------------------------\n \ninclude ../../common.mk \n \n \n#-------------------------------------------------------------------------------\n# Includes\n#-------------------------------------------------------------------------------\n\nINC += -I$(CUB_DIR) -I$(CUB_DIR)test \n\n\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nrwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nDEPS =\t\t\t\t$(CUB_DEPS) \\\n\t\t\t\t\t$(CUB_DIR)test/Makefile \\\n\t\t\t\t\t$(CUB_DIR)test/test_util.h \\\n\t\t\t\t\t$(CUB_DIR)test/mersenne.h \\\n\t\t\nALL = \texample_block_radix_sort \\\n\t \texample_block_reduce \\\n\t \texample_block_scan\n\t\t\n\n\n#-------------------------------------------------------------------------------\n# make default\n#-------------------------------------------------------------------------------\n\ndefault:\n\n\n#-------------------------------------------------------------------------------\n# make clean\n#-------------------------------------------------------------------------------\n\nclean :\n\trm -f bin/*$(CPU_ARCH_SUFFIX)* \n\trm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o\n\n\n#-------------------------------------------------------------------------------\n# make all\n#-------------------------------------------------------------------------------\n\nall : $(ALL)\n\n#-------------------------------------------------------------------------------\n# make run\n#-------------------------------------------------------------------------------\n\nrun : \n\tfor i in $(ALL); do ./bin/$${i}_$(BIN_SUFFIX) --device=$(device) || exit 1; done\n\n\n\n\n#-------------------------------------------------------------------------------\n# make example_block_reduce\n#-------------------------------------------------------------------------------\n\nexample_block_reduce: bin/example_block_reduce_$(BIN_SUFFIX)\n\nbin/example_block_reduce_$(BIN_SUFFIX) : example_block_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_block_reduce_$(BIN_SUFFIX) example_block_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_block_scan\n#-------------------------------------------------------------------------------\n\nexample_block_scan: bin/example_block_scan_$(BIN_SUFFIX)\n\nbin/example_block_scan_$(BIN_SUFFIX) : example_block_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_block_scan_$(BIN_SUFFIX) example_block_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_block_radix_sort\n#-------------------------------------------------------------------------------\n\nexample_block_radix_sort: bin/example_block_radix_sort_$(BIN_SUFFIX)\n\nbin/example_block_radix_sort_$(BIN_SUFFIX) : example_block_radix_sort.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_block_radix_sort_$(BIN_SUFFIX) example_block_radix_sort.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\t\n" }, { "path": "external/cub/examples/block/example_block_radix_sort.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple demonstration of cub::BlockRadixSort\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_block_radix_sort.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console (define before including cub.h)\n#define CUB_STDERR\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\n/// Verbose output\nbool g_verbose = false;\n\n/// Timing iterations\nint g_timing_iterations = 100;\n\n/// Default grid size\nint g_grid_size = 1;\n\n/// Uniform key samples\nbool g_uniform_keys;\n\n\n//---------------------------------------------------------------------\n// Kernels\n//---------------------------------------------------------------------\n\n/**\n * Simple kernel for performing a block-wide sorting over integers\n */\ntemplate <\n typename Key,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\n__launch_bounds__ (BLOCK_THREADS)\n__global__ void BlockSortKernel(\n Key *d_in, // Tile of input\n Key *d_out, // Tile of output\n clock_t *d_elapsed) // Elapsed cycle count of block scan\n{\n enum { TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD };\n\n // Specialize BlockLoad type for our thread block (uses warp-striped loads for coalescing, then transposes in shared memory to a blocked arrangement)\n typedef BlockLoad BlockLoadT;\n\n // Specialize BlockRadixSort type for our thread block\n typedef BlockRadixSort BlockRadixSortT;\n\n // Shared memory\n __shared__ union TempStorage\n {\n typename BlockLoadT::TempStorage load;\n typename BlockRadixSortT::TempStorage sort;\n } temp_storage;\n\n // Per-thread tile items\n Key items[ITEMS_PER_THREAD];\n\n // Our current block's offset\n int block_offset = blockIdx.x * TILE_SIZE;\n\n // Load items into a blocked arrangement\n BlockLoadT(temp_storage.load).Load(d_in + block_offset, items);\n\n // Barrier for smem reuse\n __syncthreads();\n\n // Start cycle timer\n clock_t start = clock();\n\n // Sort keys\n BlockRadixSortT(temp_storage.sort).SortBlockedToStriped(items);\n\n // Stop cycle timer\n clock_t stop = clock();\n\n // Store output in striped fashion\n StoreDirectStriped(threadIdx.x, d_out + block_offset, items);\n\n // Store elapsed clocks\n if (threadIdx.x == 0)\n {\n d_elapsed[blockIdx.x] = (start > stop) ? start - stop : stop - start;\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Host utilities\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize sorting problem (and solution).\n */\ntemplate \nvoid Initialize(\n Key *h_in,\n Key *h_reference,\n int num_items,\n int tile_size)\n{\n for (int i = 0; i < num_items; ++i)\n {\n if (g_uniform_keys)\n {\n h_in[i] = 0;\n }\n else\n {\n RandomBits(h_in[i]);\n }\n h_reference[i] = h_in[i];\n }\n\n // Only sort the first tile\n std::sort(h_reference, h_reference + tile_size);\n}\n\n\n/**\n * Test BlockScan\n */\ntemplate <\n typename Key,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid Test()\n{\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Allocate host arrays\n Key *h_in = new Key[TILE_SIZE * g_grid_size];\n Key *h_reference = new Key[TILE_SIZE * g_grid_size];\n clock_t *h_elapsed = new clock_t[g_grid_size];\n\n // Initialize problem and reference output on host\n Initialize(h_in, h_reference, TILE_SIZE * g_grid_size, TILE_SIZE);\n\n // Initialize device arrays\n Key *d_in = NULL;\n Key *d_out = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(cudaMalloc((void**)&d_in, sizeof(Key) * TILE_SIZE * g_grid_size));\n CubDebugExit(cudaMalloc((void**)&d_out, sizeof(Key) * TILE_SIZE * g_grid_size));\n CubDebugExit(cudaMalloc((void**)&d_elapsed, sizeof(clock_t) * g_grid_size));\n\n // Display input problem data\n if (g_verbose)\n {\n printf(\"Input data: \");\n for (int i = 0; i < TILE_SIZE; i++)\n std::cout << h_in[i] << \", \";\n printf(\"\\n\\n\");\n }\n\n // Kernel props\n int max_sm_occupancy;\n CubDebugExit(MaxSmOccupancy(max_sm_occupancy, BlockSortKernel, BLOCK_THREADS));\n\n // Copy problem to device\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(Key) * TILE_SIZE * g_grid_size, cudaMemcpyHostToDevice));\n\n printf(\"BlockRadixSort %d items (%d timing iterations, %d blocks, %d threads, %d items per thread, %d SM occupancy):\\n\",\n TILE_SIZE * g_grid_size, g_timing_iterations, g_grid_size, BLOCK_THREADS, ITEMS_PER_THREAD, max_sm_occupancy);\n fflush(stdout);\n\n // Run kernel once to prime caches and check result\n BlockSortKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n // Check for kernel errors and STDIO from the kernel, if any\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Check results\n printf(\"\\tOutput items: \");\n int compare = CompareDeviceResults(h_reference, d_out, TILE_SIZE, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n fflush(stdout);\n\n // Run this several times and average the performance results\n GpuTimer timer;\n float elapsed_millis = 0.0;\n unsigned long long elapsed_clocks = 0;\n\n for (int i = 0; i < g_timing_iterations; ++i)\n {\n timer.Start();\n\n // Run kernel\n BlockSortKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n // Copy clocks from device\n CubDebugExit(cudaMemcpy(h_elapsed, d_elapsed, sizeof(clock_t) * g_grid_size, cudaMemcpyDeviceToHost));\n for (int i = 0; i < g_grid_size; i++)\n elapsed_clocks += h_elapsed[i];\n }\n\n // Check for kernel errors and STDIO from the kernel, if any\n CubDebugExit(cudaDeviceSynchronize());\n\n // Display timing results\n float avg_millis = elapsed_millis / g_timing_iterations;\n float avg_items_per_sec = float(TILE_SIZE * g_grid_size) / avg_millis / 1000.0f;\n double avg_clocks = double(elapsed_clocks) / g_timing_iterations / g_grid_size;\n double avg_clocks_per_item = avg_clocks / TILE_SIZE;\n\n printf(\"\\tAverage BlockRadixSort::SortBlocked clocks: %.3f\\n\", avg_clocks);\n printf(\"\\tAverage BlockRadixSort::SortBlocked clocks per item: %.3f\\n\", avg_clocks_per_item);\n printf(\"\\tAverage kernel millis: %.4f\\n\", avg_millis);\n printf(\"\\tAverage million items / sec: %.4f\\n\", avg_items_per_sec);\n fflush(stdout);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (h_elapsed) delete[] h_elapsed;\n if (d_in) CubDebugExit(cudaFree(d_in));\n if (d_out) CubDebugExit(cudaFree(d_out));\n if (d_elapsed) CubDebugExit(cudaFree(d_elapsed));\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_uniform_keys = args.CheckCmdLineFlag(\"uniform\");\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"grid-size\", g_grid_size);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--i=]\"\n \"[--grid-size=]\"\n \"[--v] \"\n \"\\n\", argv[0], g_timing_iterations, g_grid_size);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n fflush(stdout);\n\n // Run tests\n printf(\"\\nuint32:\\n\"); fflush(stdout);\n Test();\n printf(\"\\n\"); fflush(stdout);\n\n printf(\"\\nfp32:\\n\"); fflush(stdout);\n Test();\n printf(\"\\n\"); fflush(stdout);\n\n printf(\"\\nuint8:\\n\"); fflush(stdout);\n Test();\n printf(\"\\n\"); fflush(stdout);\n\n return 0;\n}\n\n" }, { "path": "external/cub/examples/block/example_block_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple demonstration of cub::BlockReduce\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_block_reduce.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console (define before including cub.h)\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\n/// Verbose output\nbool g_verbose = false;\n\n/// Timing iterations\nint g_timing_iterations = 100;\n\n/// Default grid size\nint g_grid_size = 1;\n\n\n\n//---------------------------------------------------------------------\n// Kernels\n//---------------------------------------------------------------------\n\n/**\n * Simple kernel for performing a block-wide exclusive prefix sum over integers\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockReduceAlgorithm ALGORITHM>\n__global__ void BlockSumKernel(\n int *d_in, // Tile of input\n int *d_out, // Tile aggregate\n clock_t *d_elapsed) // Elapsed cycle count of block reduction\n{\n // Specialize BlockReduce type for our thread block\n typedef BlockReduce BlockReduceT;\n\n // Shared memory\n __shared__ typename BlockReduceT::TempStorage temp_storage;\n\n // Per-thread tile data\n int data[ITEMS_PER_THREAD];\n LoadDirectStriped(threadIdx.x, d_in, data);\n\n // Start cycle timer\n clock_t start = clock();\n\n // Compute sum\n int aggregate = BlockReduceT(temp_storage).Sum(data);\n\n // Stop cycle timer\n clock_t stop = clock();\n\n // Store aggregate and elapsed clocks\n if (threadIdx.x == 0)\n {\n *d_elapsed = (start > stop) ? start - stop : stop - start;\n *d_out = aggregate;\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Host utilities\n//---------------------------------------------------------------------\n\n/**\n * Initialize reduction problem (and solution).\n * Returns the aggregate\n */\nint Initialize(int *h_in, int num_items)\n{\n int inclusive = 0;\n\n for (int i = 0; i < num_items; ++i)\n {\n h_in[i] = i % 17;\n inclusive += h_in[i];\n }\n\n return inclusive;\n}\n\n\n/**\n * Test thread block reduction\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockReduceAlgorithm ALGORITHM>\nvoid Test()\n{\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Allocate host arrays\n int *h_in = new int[TILE_SIZE];\n int *h_gpu = new int[TILE_SIZE + 1];\n\n // Initialize problem and reference output on host\n int h_aggregate = Initialize(h_in, TILE_SIZE);\n\n // Initialize device arrays\n int *d_in = NULL;\n int *d_out = NULL;\n clock_t *d_elapsed = NULL;\n cudaMalloc((void**)&d_in, sizeof(int) * TILE_SIZE);\n cudaMalloc((void**)&d_out, sizeof(int) * 1);\n cudaMalloc((void**)&d_elapsed, sizeof(clock_t));\n\n // Display input problem data\n if (g_verbose)\n {\n printf(\"Input data: \");\n for (int i = 0; i < TILE_SIZE; i++)\n printf(\"%d, \", h_in[i]);\n printf(\"\\n\\n\");\n }\n\n // Kernel props\n int max_sm_occupancy;\n CubDebugExit(MaxSmOccupancy(max_sm_occupancy, BlockSumKernel, BLOCK_THREADS));\n\n // Copy problem to device\n cudaMemcpy(d_in, h_in, sizeof(int) * TILE_SIZE, cudaMemcpyHostToDevice);\n\n printf(\"BlockReduce algorithm %s on %d items (%d timing iterations, %d blocks, %d threads, %d items per thread, %d SM occupancy):\\n\",\n (ALGORITHM == BLOCK_REDUCE_RAKING) ? \"BLOCK_REDUCE_RAKING\" : \"BLOCK_REDUCE_WARP_REDUCTIONS\",\n TILE_SIZE, g_timing_iterations, g_grid_size, BLOCK_THREADS, ITEMS_PER_THREAD, max_sm_occupancy);\n\n // Run aggregate/prefix kernel\n BlockSumKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n // Check total aggregate\n printf(\"\\tAggregate: \");\n int compare = CompareDeviceResults(&h_aggregate, d_out, 1, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Run this several times and average the performance results\n GpuTimer timer;\n float elapsed_millis = 0.0;\n clock_t elapsed_clocks = 0;\n\n for (int i = 0; i < g_timing_iterations; ++i)\n {\n // Copy problem to device\n cudaMemcpy(d_in, h_in, sizeof(int) * TILE_SIZE, cudaMemcpyHostToDevice);\n\n timer.Start();\n\n // Run aggregate/prefix kernel\n BlockSumKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n // Copy clocks from device\n clock_t clocks;\n CubDebugExit(cudaMemcpy(&clocks, d_elapsed, sizeof(clock_t), cudaMemcpyDeviceToHost));\n elapsed_clocks += clocks;\n\n }\n\n // Check for kernel errors and STDIO from the kernel, if any\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Display timing results\n float avg_millis = elapsed_millis / g_timing_iterations;\n float avg_items_per_sec = float(TILE_SIZE * g_grid_size) / avg_millis / 1000.0f;\n float avg_clocks = float(elapsed_clocks) / g_timing_iterations;\n float avg_clocks_per_item = avg_clocks / TILE_SIZE;\n\n printf(\"\\tAverage BlockReduce::Sum clocks: %.3f\\n\", avg_clocks);\n printf(\"\\tAverage BlockReduce::Sum clocks per item: %.3f\\n\", avg_clocks_per_item);\n printf(\"\\tAverage kernel millis: %.4f\\n\", avg_millis);\n printf(\"\\tAverage million items / sec: %.4f\\n\", avg_items_per_sec);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_gpu) delete[] h_gpu;\n if (d_in) cudaFree(d_in);\n if (d_out) cudaFree(d_out);\n if (d_elapsed) cudaFree(d_elapsed);\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"grid-size\", g_grid_size);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--i=] \"\n \"[--grid-size=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Run tests\n Test<1024, 1, BLOCK_REDUCE_RAKING>();\n Test<512, 2, BLOCK_REDUCE_RAKING>();\n Test<256, 4, BLOCK_REDUCE_RAKING>();\n Test<128, 8, BLOCK_REDUCE_RAKING>();\n Test<64, 16, BLOCK_REDUCE_RAKING>();\n Test<32, 32, BLOCK_REDUCE_RAKING>();\n Test<16, 64, BLOCK_REDUCE_RAKING>();\n\n printf(\"-------------\\n\");\n\n Test<1024, 1, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<512, 2, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<256, 4, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<128, 8, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<64, 16, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<32, 32, BLOCK_REDUCE_WARP_REDUCTIONS>();\n Test<16, 64, BLOCK_REDUCE_WARP_REDUCTIONS>();\n\n return 0;\n}\n\n" }, { "path": "external/cub/examples/block/example_block_scan.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple demonstration of cub::BlockScan\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_block_scan.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console (define before including cub.h)\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\n/// Verbose output\nbool g_verbose = false;\n\n/// Timing iterations\nint g_timing_iterations = 100;\n\n/// Default grid size\nint g_grid_size = 1;\n\n\n\n//---------------------------------------------------------------------\n// Kernels\n//---------------------------------------------------------------------\n\n/**\n * Simple kernel for performing a block-wide exclusive prefix sum over integers\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockScanAlgorithm ALGORITHM>\n__global__ void BlockPrefixSumKernel(\n int *d_in, // Tile of input\n int *d_out, // Tile of output\n clock_t *d_elapsed) // Elapsed cycle count of block scan\n{\n // Specialize BlockLoad type for our thread block (uses warp-striped loads for coalescing, then transposes in shared memory to a blocked arrangement)\n typedef BlockLoad BlockLoadT;\n\n // Specialize BlockStore type for our thread block (uses warp-striped loads for coalescing, then transposes in shared memory to a blocked arrangement)\n typedef BlockStore BlockStoreT;\n\n // Specialize BlockScan type for our thread block\n typedef BlockScan BlockScanT;\n\n // Shared memory\n __shared__ union TempStorage\n {\n typename BlockLoadT::TempStorage load;\n typename BlockStoreT::TempStorage store;\n typename BlockScanT::TempStorage scan;\n } temp_storage;\n\n // Per-thread tile data\n int data[ITEMS_PER_THREAD];\n\n // Load items into a blocked arrangement\n BlockLoadT(temp_storage.load).Load(d_in, data);\n\n // Barrier for smem reuse\n __syncthreads();\n\n // Start cycle timer\n clock_t start = clock();\n\n // Compute exclusive prefix sum\n int aggregate;\n BlockScanT(temp_storage.scan).ExclusiveSum(data, data, aggregate);\n\n // Stop cycle timer\n clock_t stop = clock();\n\n // Barrier for smem reuse\n __syncthreads();\n\n // Store items from a blocked arrangement\n BlockStoreT(temp_storage.store).Store(d_out, data);\n\n // Store aggregate and elapsed clocks\n if (threadIdx.x == 0)\n {\n *d_elapsed = (start > stop) ? start - stop : stop - start;\n d_out[BLOCK_THREADS * ITEMS_PER_THREAD] = aggregate;\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Host utilities\n//---------------------------------------------------------------------\n\n/**\n * Initialize exclusive prefix sum problem (and solution).\n * Returns the aggregate\n */\nint Initialize(\n int *h_in,\n int *h_reference,\n int num_items)\n{\n int inclusive = 0;\n\n for (int i = 0; i < num_items; ++i)\n {\n h_in[i] = i % 17;\n\n h_reference[i] = inclusive;\n inclusive += h_in[i];\n }\n\n return inclusive;\n}\n\n\n/**\n * Test thread block scan\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockScanAlgorithm ALGORITHM>\nvoid Test()\n{\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Allocate host arrays\n int *h_in = new int[TILE_SIZE];\n int *h_reference = new int[TILE_SIZE];\n int *h_gpu = new int[TILE_SIZE + 1];\n\n // Initialize problem and reference output on host\n int h_aggregate = Initialize(h_in, h_reference, TILE_SIZE);\n\n // Initialize device arrays\n int *d_in = NULL;\n int *d_out = NULL;\n clock_t *d_elapsed = NULL;\n cudaMalloc((void**)&d_in, sizeof(int) * TILE_SIZE);\n cudaMalloc((void**)&d_out, sizeof(int) * (TILE_SIZE + 1));\n cudaMalloc((void**)&d_elapsed, sizeof(clock_t));\n\n // Display input problem data\n if (g_verbose)\n {\n printf(\"Input data: \");\n for (int i = 0; i < TILE_SIZE; i++)\n printf(\"%d, \", h_in[i]);\n printf(\"\\n\\n\");\n }\n\n // Kernel props\n int max_sm_occupancy;\n CubDebugExit(MaxSmOccupancy(max_sm_occupancy, BlockPrefixSumKernel, BLOCK_THREADS));\n\n // Copy problem to device\n cudaMemcpy(d_in, h_in, sizeof(int) * TILE_SIZE, cudaMemcpyHostToDevice);\n\n printf(\"BlockScan algorithm %s on %d items (%d timing iterations, %d blocks, %d threads, %d items per thread, %d SM occupancy):\\n\",\n (ALGORITHM == BLOCK_SCAN_RAKING) ? \"BLOCK_SCAN_RAKING\" : (ALGORITHM == BLOCK_SCAN_RAKING_MEMOIZE) ? \"BLOCK_SCAN_RAKING_MEMOIZE\" : \"BLOCK_SCAN_WARP_SCANS\",\n TILE_SIZE, g_timing_iterations, g_grid_size, BLOCK_THREADS, ITEMS_PER_THREAD, max_sm_occupancy);\n\n // Run aggregate/prefix kernel\n BlockPrefixSumKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n // Check results\n printf(\"\\tOutput items: \");\n int compare = CompareDeviceResults(h_reference, d_out, TILE_SIZE, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Check total aggregate\n printf(\"\\tAggregate: \");\n compare = CompareDeviceResults(&h_aggregate, d_out + TILE_SIZE, 1, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Run this several times and average the performance results\n GpuTimer timer;\n float elapsed_millis = 0.0;\n clock_t elapsed_clocks = 0;\n\n for (int i = 0; i < g_timing_iterations; ++i)\n {\n // Copy problem to device\n cudaMemcpy(d_in, h_in, sizeof(int) * TILE_SIZE, cudaMemcpyHostToDevice);\n\n timer.Start();\n\n // Run aggregate/prefix kernel\n BlockPrefixSumKernel<<>>(\n d_in,\n d_out,\n d_elapsed);\n\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n // Copy clocks from device\n clock_t clocks;\n CubDebugExit(cudaMemcpy(&clocks, d_elapsed, sizeof(clock_t), cudaMemcpyDeviceToHost));\n elapsed_clocks += clocks;\n\n }\n\n // Check for kernel errors and STDIO from the kernel, if any\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Display timing results\n float avg_millis = elapsed_millis / g_timing_iterations;\n float avg_items_per_sec = float(TILE_SIZE * g_grid_size) / avg_millis / 1000.0f;\n float avg_clocks = float(elapsed_clocks) / g_timing_iterations;\n float avg_clocks_per_item = avg_clocks / TILE_SIZE;\n\n printf(\"\\tAverage BlockScan::Sum clocks: %.3f\\n\", avg_clocks);\n printf(\"\\tAverage BlockScan::Sum clocks per item: %.3f\\n\", avg_clocks_per_item);\n printf(\"\\tAverage kernel millis: %.4f\\n\", avg_millis);\n printf(\"\\tAverage million items / sec: %.4f\\n\", avg_items_per_sec);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (h_gpu) delete[] h_gpu;\n if (d_in) cudaFree(d_in);\n if (d_out) cudaFree(d_out);\n if (d_elapsed) cudaFree(d_elapsed);\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"grid-size\", g_grid_size);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--i=]\"\n \"[--grid-size=]\"\n \"[--v] \"\n \"\\n\", argv[0], g_timing_iterations, g_grid_size);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Run tests\n Test<1024, 1, BLOCK_SCAN_RAKING>();\n Test<512, 2, BLOCK_SCAN_RAKING>();\n Test<256, 4, BLOCK_SCAN_RAKING>();\n Test<128, 8, BLOCK_SCAN_RAKING>();\n Test<64, 16, BLOCK_SCAN_RAKING>();\n Test<32, 32, BLOCK_SCAN_RAKING>();\n\n printf(\"-------------\\n\");\n\n Test<1024, 1, BLOCK_SCAN_RAKING_MEMOIZE>();\n Test<512, 2, BLOCK_SCAN_RAKING_MEMOIZE>();\n Test<256, 4, BLOCK_SCAN_RAKING_MEMOIZE>();\n Test<128, 8, BLOCK_SCAN_RAKING_MEMOIZE>();\n Test<64, 16, BLOCK_SCAN_RAKING_MEMOIZE>();\n Test<32, 32, BLOCK_SCAN_RAKING_MEMOIZE>();\n\n printf(\"-------------\\n\");\n\n Test<1024, 1, BLOCK_SCAN_WARP_SCANS>();\n Test<512, 2, BLOCK_SCAN_WARP_SCANS>();\n Test<256, 4, BLOCK_SCAN_WARP_SCANS>();\n Test<128, 8, BLOCK_SCAN_WARP_SCANS>();\n Test<64, 16, BLOCK_SCAN_WARP_SCANS>();\n Test<32, 32, BLOCK_SCAN_WARP_SCANS>();\n\n\n return 0;\n}\n\n" }, { "path": "external/cub/examples/block/reduce_by_key.cu", "content": "\n\n#include \n\n\ntemplate <\n int BLOCK_THREADS, ///< Number of CTA threads\n typename KeyT, ///< Key type\n typename ValueT> ///< Value type\n__global__ void Kernel()\n{\n // Tuple type for scanning (pairs accumulated segment-value with segment-index)\n typedef cub::KeyValuePair OffsetValuePairT;\n\n // Reduce-value-by-segment scan operator\n typedef cub::ReduceBySegmentOp ReduceBySegmentOpT;\n\n // Parameterized BlockDiscontinuity type for setting head flags\n typedef cub::BlockDiscontinuity<\n KeyT,\n BLOCK_THREADS>\n BlockDiscontinuityKeysT;\n\n // Parameterized BlockScan type\n typedef cub::BlockScan<\n OffsetValuePairT,\n BLOCK_THREADS,\n cub::BLOCK_SCAN_WARP_SCANS>\n BlockScanT;\n\n // Shared memory\n __shared__ union TempStorage\n {\n typename BlockScanT::TempStorage scan; // Scan storage\n typename BlockDiscontinuityKeysT::TempStorage discontinuity; // Discontinuity storage\n } temp_storage;\n\n\n // Read data (each thread gets 3 items each, every 9 items is a segment)\n KeyT my_keys[3] = {threadIdx.x / 3, threadIdx.x / 3, threadIdx.x / 3};\n ValueT my_values[3] = {1, 1, 1};\n\n // Set head segment head flags\n int my_flags[3];\n BlockDiscontinuityKeysT(temp_storage.discontinuity).FlagHeads(\n my_flags,\n my_keys,\n cub::Inequality());\n\n __syncthreads();\n\n\n\n\n\n\n}\n" }, { "path": "external/cub/examples/device/.gitignore", "content": "/bin\n/Debug\n/ipch\n/Release\n/cuda55.sdf\n/cuda55.suo\n/cuda60.sdf\n/cuda60.suo\n" }, { "path": "external/cub/examples/device/Makefile", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n\n#-------------------------------------------------------------------------------\n#\n# Makefile usage\n#\n# make [sm=] [cdp=<0|1>] [force32=<0|1>] [abi=<0|1>] [open64=<0|1>] [verbose=<0|1>] [keep=<0|1>]\n#\n#-------------------------------------------------------------------------------\n \ninclude ../../common.mk \n \n \n#-------------------------------------------------------------------------------\n# Includes\n#-------------------------------------------------------------------------------\n\nINC += -I$(CUB_DIR) -I$(CUB_DIR)test \n\n\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nrwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nDEPS =\t\t\t\t$(CUB_DEPS) \\\n\t\t\t\t\t$(CUB_DIR)test/Makefile \\\n\t\t\t\t\t$(CUB_DIR)test/test_util.h \\\n\t\t\t\t\t$(CUB_DIR)test/mersenne.h \\\n\t\t\nALL = \texample_device_partition_flagged \\\n\t\texample_device_partition_if \\\n\t \texample_device_radix_sort \\\n\t\texample_device_reduce \\\n\t \texample_device_scan \\\n\t \texample_device_select_unique \\\n\t\texample_device_select_flagged \\\n\t\texample_device_select_if \\\n\t\texample_device_sort_find_non_trivial_runs\n\t\t\n\n\n#-------------------------------------------------------------------------------\n# make default\n#-------------------------------------------------------------------------------\n\ndefault:\n\n\n#-------------------------------------------------------------------------------\n# make clean\n#-------------------------------------------------------------------------------\n\nclean :\n\trm -f bin/*$(CPU_ARCH_SUFFIX)* \n\trm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o\n\n\n#-------------------------------------------------------------------------------\n# make all\n#-------------------------------------------------------------------------------\n\nall : $(ALL)\n\n#-------------------------------------------------------------------------------\n# make run\n#-------------------------------------------------------------------------------\n\nrun : \n\tfor i in $(ALL); do ./bin/$${i}_$(BIN_SUFFIX) --device=$(device) || exit 1; done\n\n\n#-------------------------------------------------------------------------------\n# make example_device_reduce\n#-------------------------------------------------------------------------------\n\nexample_device_reduce: bin/example_device_reduce_$(BIN_SUFFIX)\n\nbin/example_device_reduce_$(BIN_SUFFIX) : example_device_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_reduce_$(BIN_SUFFIX) example_device_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_device_partition_flagged\n#-------------------------------------------------------------------------------\n\nexample_device_partition_flagged: bin/example_device_partition_flagged_$(BIN_SUFFIX)\n\nbin/example_device_partition_flagged_$(BIN_SUFFIX) : example_device_partition_flagged.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_partition_flagged_$(BIN_SUFFIX) example_device_partition_flagged.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make example_device_partition_if\n#-------------------------------------------------------------------------------\n\nexample_device_partition_if: bin/example_device_partition_if_$(BIN_SUFFIX)\n\nbin/example_device_partition_if_$(BIN_SUFFIX) : example_device_partition_if.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_partition_if_$(BIN_SUFFIX) example_device_partition_if.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make example_device_scan\n#-------------------------------------------------------------------------------\n\nexample_device_scan: bin/example_device_scan_$(BIN_SUFFIX)\n\nbin/example_device_scan_$(BIN_SUFFIX) : example_device_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_scan_$(BIN_SUFFIX) example_device_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_device_radix_sort\n#-------------------------------------------------------------------------------\n\nexample_device_radix_sort: bin/example_device_radix_sort_$(BIN_SUFFIX)\n\nbin/example_device_radix_sort_$(BIN_SUFFIX) : example_device_radix_sort.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_radix_sort_$(BIN_SUFFIX) example_device_radix_sort.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_device_select_unique\n#-------------------------------------------------------------------------------\n\nexample_device_select_unique: bin/example_device_select_unique_$(BIN_SUFFIX)\n\nbin/example_device_select_unique_$(BIN_SUFFIX) : example_device_select_unique.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_select_unique_$(BIN_SUFFIX) example_device_select_unique.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_device_select_flagged\n#-------------------------------------------------------------------------------\n\nexample_device_select_flagged: bin/example_device_select_flagged_$(BIN_SUFFIX)\n\nbin/example_device_select_flagged_$(BIN_SUFFIX) : example_device_select_flagged.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_select_flagged_$(BIN_SUFFIX) example_device_select_flagged.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make example_device_select_if\n#-------------------------------------------------------------------------------\n\nexample_device_select_if: bin/example_device_select_if_$(BIN_SUFFIX)\n\nbin/example_device_select_if_$(BIN_SUFFIX) : example_device_select_if.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_select_if_$(BIN_SUFFIX) example_device_select_if.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make example_device_sort_find_non_trivial_runs\n#-------------------------------------------------------------------------------\n\nexample_device_sort_find_non_trivial_runs: bin/example_device_sort_find_non_trivial_runs_$(BIN_SUFFIX)\n\nbin/example_device_sort_find_non_trivial_runs_$(BIN_SUFFIX) : example_device_sort_find_non_trivial_runs.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/example_device_sort_find_non_trivial_runs_$(BIN_SUFFIX) example_device_sort_find_non_trivial_runs.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_partition_flagged.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DevicePartition::Flagged().\n *\n * Partition flagged items from from a sequence of int keys using a\n * corresponding sequence of unsigned char flags.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_partition_flagged.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem, setting flags at distances of random length\n * chosen from [1..max_segment]\n */\nvoid Initialize(\n int *h_in,\n unsigned char *h_flags,\n int num_items,\n int max_segment)\n{\n unsigned short max_short = (unsigned short) -1;\n\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Select number of repeating occurrences\n unsigned short repeat;\n RandomBits(repeat);\n repeat = (unsigned short) ((float(repeat) * (float(max_segment) / float(max_short))));\n repeat = CUB_MAX(1, repeat);\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n h_flags[j] = 0;\n h_in[j] = key;\n j++;\n }\n\n h_flags[i] = 1;\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"Flags:\\n\");\n DisplayResults(h_flags, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\nint Solve(\n int *h_in,\n unsigned char *h_flags,\n int *h_reference,\n int num_items)\n{\n int num_selected = 0;\n for (int i = 0; i < num_items; ++i)\n {\n if (h_flags[i])\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n else\n {\n h_reference[num_items - (i - num_selected) - 1] = h_in[i];\n }\n }\n\n return num_selected;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n int max_segment = 40; // Maximum segment length\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--maxseg=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays\n int *h_in = new int[num_items];\n int *h_reference = new int[num_items];\n unsigned char *h_flags = new unsigned char[num_items];\n\n // Initialize problem and solution\n Initialize(h_in, h_flags, num_items, max_segment);\n int num_selected = Solve(h_in, h_flags, h_reference, num_items);\n\n printf(\"cub::DevicePartition::Flagged %d items, %d selected (avg distance %d), %d-byte elements\\n\",\n num_items, num_selected, (num_selected > 0) ? num_items / num_selected : 0, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n unsigned char *d_flags = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_flags, sizeof(unsigned char) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_flags, h_flags, sizeof(unsigned char) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array and num selected\n int *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_items, true, g_verbose);\n printf(\"\\t Data %s \", compare ? \"FAIL\" : \"PASS\");\n compare |= CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_flags) CubDebugExit(g_allocator.DeviceFree(d_flags));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_partition_if.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DevicePartition::If().\n *\n * Partitions items from from a sequence of int keys using a\n * section functor (greater-than)\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_select_if.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n/// Selection functor type\nstruct GreaterThan\n{\n int compare;\n\n __host__ __device__ __forceinline__\n GreaterThan(int compare) : compare(compare) {}\n\n __host__ __device__ __forceinline__\n bool operator()(const int &a) const {\n return (a > compare);\n }\n};\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem, setting runs of random length chosen from [1..max_segment]\n */\nvoid Initialize(\n int *h_in,\n int num_items,\n int max_segment)\n{\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Randomly select number of repeating occurrences uniformly from [1..max_segment]\n unsigned short max_short = (unsigned short) -1;\n unsigned short repeat;\n RandomBits(repeat);\n repeat = (unsigned short) ((float(repeat) * (float(max_segment) / float(max_short))));\n repeat = CUB_MAX(1, repeat);\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n h_in[j] = key;\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\ntemplate \nint Solve(\n int *h_in,\n SelectOp select_op,\n int *h_reference,\n int num_items)\n{\n int num_selected = 0;\n for (int i = 0; i < num_items; ++i)\n {\n if (select_op(h_in[i]))\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n else\n {\n h_reference[num_items - (i - num_selected) - 1] = h_in[i];\n }\n }\n\n return num_selected;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n int max_segment = 40; // Maximum segment length\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays\n int *h_in = new int[num_items];\n int *h_reference = new int[num_items];\n\n // DevicePartition a pivot index\n unsigned int pivot_index;\n unsigned int max_int = (unsigned int) -1;\n RandomBits(pivot_index);\n pivot_index = (unsigned int) ((float(pivot_index) * (float(num_items - 1) / float(max_int))));\n printf(\"Pivot idx: %d\\n\", pivot_index); fflush(stdout);\n\n // Initialize problem and solution\n Initialize(h_in, num_items, max_segment);\n GreaterThan select_op(h_in[pivot_index]);\n\n int num_selected = Solve(h_in, select_op, h_reference, num_items);\n\n printf(\"cub::DevicePartition::If %d items, %d selected (avg run length %d), %d-byte elements\\n\",\n num_items, num_selected, (num_selected > 0) ? num_items / num_selected : 0, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array and num selected\n int *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DevicePartition::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DevicePartition::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_items, true, g_verbose);\n printf(\"\\t Data %s \", compare ? \"FAIL\" : \"PASS\");\n compare = compare | CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_radix_sort.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceRadixSort::SortPairs().\n *\n * Sorts an array of float keys paired with a corresponding array of int values.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_radix_sort.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Simple key-value pairing for floating point types. Distinguishes\n * between positive and negative zero.\n */\nstruct Pair\n{\n float key;\n int value;\n\n bool operator<(const Pair &b) const\n {\n if (key < b.key)\n return true;\n\n if (key > b.key)\n return false;\n\n // Return true if key is negative zero and b.key is positive zero\n unsigned int key_bits = *reinterpret_cast(const_cast(&key));\n unsigned int b_key_bits = *reinterpret_cast(const_cast(&b.key));\n unsigned int HIGH_BIT = 1u << 31;\n\n return ((key_bits & HIGH_BIT) != 0) && ((b_key_bits & HIGH_BIT) == 0);\n }\n};\n\n\n/**\n * Initialize key-value sorting problem.\n */\nvoid Initialize(\n float *h_keys,\n int *h_values,\n float *h_reference_keys,\n int *h_reference_values,\n int num_items)\n{\n Pair *h_pairs = new Pair[num_items];\n\n for (int i = 0; i < num_items; ++i)\n {\n RandomBits(h_keys[i]);\n RandomBits(h_values[i]);\n h_pairs[i].key = h_keys[i];\n h_pairs[i].value = h_values[i];\n }\n\n if (g_verbose)\n {\n printf(\"Input keys:\\n\");\n DisplayResults(h_keys, num_items);\n printf(\"\\n\\n\");\n\n printf(\"Input values:\\n\");\n DisplayResults(h_values, num_items);\n printf(\"\\n\\n\");\n }\n\n std::stable_sort(h_pairs, h_pairs + num_items);\n\n for (int i = 0; i < num_items; ++i)\n {\n h_reference_keys[i] = h_pairs[i].key;\n h_reference_values[i] = h_pairs[i].value;\n }\n\n delete[] h_pairs;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n printf(\"cub::DeviceRadixSort::SortPairs() %d items (%d-byte keys %d-byte values)\\n\",\n num_items, int(sizeof(float)), int(sizeof(int)));\n fflush(stdout);\n\n // Allocate host arrays\n float *h_keys = new float[num_items];\n float *h_reference_keys = new float[num_items];\n int *h_values = new int[num_items];\n int *h_reference_values = new int[num_items];\n\n // Initialize problem and solution on host\n Initialize(h_keys, h_values, h_reference_keys, h_reference_values, num_items);\n\n // Allocate device arrays\n DoubleBuffer d_keys;\n DoubleBuffer d_values;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[0], sizeof(float) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(float) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[0], sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[1], sizeof(int) * num_items));\n\n // Allocate temporary storage\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n\n CubDebugExit(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Initialize device arrays\n CubDebugExit(cudaMemcpy(d_keys.d_buffers[d_keys.selector], h_keys, sizeof(float) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_values.d_buffers[d_values.selector], h_values, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Run\n CubDebugExit(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference_keys, d_keys.Current(), num_items, true, g_verbose);\n printf(\"\\t Compare keys (selector %d): %s\\n\", d_keys.selector, compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n compare = CompareDeviceResults(h_reference_values, d_values.Current(), num_items, true, g_verbose);\n printf(\"\\t Compare values (selector %d): %s\\n\", d_values.selector, compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_keys) delete[] h_keys;\n if (h_reference_keys) delete[] h_reference_keys;\n if (h_values) delete[] h_values;\n if (h_reference_values) delete[] h_reference_values;\n\n if (d_keys.d_buffers[0]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[0]));\n if (d_keys.d_buffers[1]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[1]));\n if (d_values.d_buffers[0]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[0]));\n if (d_values.d_buffers[1]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[1]));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceReduce::Sum().\n *\n * Sums an array of int keys.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_reduce.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem\n */\nvoid Initialize(\n int *h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n h_in[i] = i;\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Compute solution\n */\nvoid Solve(\n int *h_in,\n int &h_reference,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n if (i == 0)\n h_reference = h_in[0];\n else\n h_reference += h_in[i];\n }\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n printf(\"cub::DeviceReduce::Sum() %d items (%d-byte elements)\\n\",\n num_items, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate host arrays\n int* h_in = new int[num_items];\n int h_reference;\n\n // Initialize problem and solution\n Initialize(h_in, num_items);\n Solve(h_in, h_reference, num_items);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array\n int *d_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * 1));\n\n // Request and allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DeviceReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(&h_reference, d_out, 1, g_verbose, g_verbose);\n printf(\"\\t%s\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_scan.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceScan::ExclusiveSum().\n *\n * Computes an exclusive sum of int keys.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_scan.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\nvoid Initialize(\n int *h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n h_in[i] = i;\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n/**\n * Solve exclusive-scan problem\n */\nint Solve(\n int *h_in,\n int *h_reference,\n int num_items)\n{\n int inclusive = 0;\n int aggregate = 0;\n\n for (int i = 0; i < num_items; ++i)\n {\n h_reference[i] = inclusive;\n inclusive += h_in[i];\n aggregate += h_in[i];\n }\n\n return aggregate;\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n printf(\"cub::DeviceScan::ExclusiveSum %d items (%d-byte elements)\\n\",\n num_items, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate host arrays\n int* h_in = new int[num_items];\n int* h_reference = new int[num_items];\n\n // Initialize problem and solution\n Initialize(h_in, num_items);\n Solve(h_in, h_reference, num_items);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array\n int *d_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_items, true, g_verbose);\n printf(\"\\t%s\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_select_flagged.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceSelect::Flagged().\n *\n * Selects flagged items from from a sequence of int keys using a\n * corresponding sequence of unsigned char flags.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_select_flagged.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem, setting flags at distances of random length\n * chosen from [1..max_segment]\n */\nvoid Initialize(\n int *h_in,\n unsigned char *h_flags,\n int num_items,\n int max_segment)\n{\n unsigned short max_short = (unsigned short) -1;\n\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Select number of repeating occurrences\n unsigned short repeat;\n RandomBits(repeat);\n repeat = (unsigned short) ((float(repeat) * (float(max_segment) / float(max_short))));\n repeat = CUB_MAX(1, repeat);\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n h_flags[j] = 0;\n h_in[j] = key;\n j++;\n }\n\n h_flags[i] = 1;\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"Flags:\\n\");\n DisplayResults(h_flags, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\nint Solve(\n int *h_in,\n unsigned char *h_flags,\n int *h_reference,\n int num_items)\n{\n int num_selected = 0;\n for (int i = 0; i < num_items; ++i)\n {\n if (h_flags[i])\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n else\n {\n h_reference[num_items - (i - num_selected) - 1] = h_in[i];\n }\n }\n\n return num_selected;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n int max_segment = 40; // Maximum segment length\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--maxseg=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays\n int *h_in = new int[num_items];\n int *h_reference = new int[num_items];\n unsigned char *h_flags = new unsigned char[num_items];\n\n // Initialize problem and solution\n Initialize(h_in, h_flags, num_items, max_segment);\n int num_selected = Solve(h_in, h_flags, h_reference, num_items);\n\n printf(\"cub::DeviceSelect::Flagged %d items, %d selected (avg distance %d), %d-byte elements\\n\",\n num_items, num_selected, (num_selected > 0) ? num_items / num_selected : 0, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n unsigned char *d_flags = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_flags, sizeof(unsigned char) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_flags, h_flags, sizeof(unsigned char) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array and num selected\n int *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceSelect::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DeviceSelect::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_selected, true, g_verbose);\n printf(\"\\t Data %s \", compare ? \"FAIL\" : \"PASS\");\n compare |= CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_flags) CubDebugExit(g_allocator.DeviceFree(d_flags));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_select_if.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceSelect::If().\n *\n * Selects items from from a sequence of int keys using a\n * section functor (greater-than)\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_select_if.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n/// Selection functor type\nstruct GreaterThan\n{\n int compare;\n\n __host__ __device__ __forceinline__\n GreaterThan(int compare) : compare(compare) {}\n\n __host__ __device__ __forceinline__\n bool operator()(const int &a) const {\n return (a > compare);\n }\n};\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem, setting runs of random length chosen from [1..max_segment]\n */\nvoid Initialize(\n int *h_in,\n int num_items,\n int max_segment)\n{\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Randomly select number of repeating occurrences uniformly from [1..max_segment]\n unsigned short max_short = (unsigned short) -1;\n unsigned short repeat;\n RandomBits(repeat);\n repeat = (unsigned short) ((float(repeat) * (float(max_segment) / float(max_short))));\n repeat = CUB_MAX(1, repeat);\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n h_in[j] = key;\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\ntemplate \nint Solve(\n int *h_in,\n SelectOp select_op,\n int *h_reference,\n int num_items)\n{\n int num_selected = 0;\n for (int i = 0; i < num_items; ++i)\n {\n if (select_op(h_in[i]))\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n else\n {\n h_reference[num_items - (i - num_selected) - 1] = h_in[i];\n }\n }\n\n return num_selected;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n int max_segment = 40; // Maximum segment length\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays\n int *h_in = new int[num_items];\n int *h_reference = new int[num_items];\n\n // Select a pivot index\n unsigned int pivot_index;\n unsigned int max_int = (unsigned int) -1;\n RandomBits(pivot_index);\n pivot_index = (unsigned int) ((float(pivot_index) * (float(num_items - 1) / float(max_int))));\n printf(\"Pivot idx: %d\\n\", pivot_index); fflush(stdout);\n\n // Initialize problem and solution\n Initialize(h_in, num_items, max_segment);\n GreaterThan select_op(h_in[pivot_index]);\n\n int num_selected = Solve(h_in, select_op, h_reference, num_items);\n\n printf(\"cub::DeviceSelect::If %d items, %d selected (avg run length %d), %d-byte elements\\n\",\n num_items, num_selected, (num_selected > 0) ? num_items / num_selected : 0, (int) sizeof(int));\n fflush(stdout);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array and num selected\n int *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_selected, true, g_verbose);\n printf(\"\\t Data %s \", compare ? \"FAIL\" : \"PASS\");\n compare = compare | CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n" }, { "path": "external/cub/examples/device/example_device_select_unique.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of DeviceSelect::Unique().\n *\n * Selects the first element from each run of identical values from a sequence\n * of int keys.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_select_unique.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem, setting runs of random length chosen from [1..max_segment]\n */\nvoid Initialize(\n int *h_in,\n int num_items,\n int max_segment)\n{\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Randomly select number of repeating occurrences uniformly from [1..max_segment]\n unsigned short max_short = (unsigned short) -1;\n unsigned short repeat;\n RandomBits(repeat);\n repeat = (unsigned short) ((float(repeat) * (float(max_segment) / float(max_short))));\n repeat = CUB_MAX(1, repeat);\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n h_in[j] = key;\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\nint Solve(\n int *h_in,\n int *h_reference,\n int num_items)\n{\n int num_selected = 0;\n if (num_items > 0)\n {\n h_reference[num_selected] = h_in[0];\n num_selected++;\n }\n\n for (int i = 1; i < num_items; ++i)\n {\n if (h_in[i] != h_in[i - 1])\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n }\n\n return num_selected;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = 150;\n int max_segment = 40; // Maximum segment length\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays\n int* h_in = new int[num_items];\n int* h_reference = new int[num_items];\n\n // Initialize problem and solution\n Initialize(h_in, num_items, max_segment);\n int num_selected = Solve(h_in, h_reference, num_items);\n\n printf(\"cub::DeviceSelect::Unique %d items (%d-byte elements), %d selected (avg run length %d)\\n\",\n num_items, (int) sizeof(int), num_selected, num_items / num_selected);\n fflush(stdout);\n\n // Allocate problem device arrays\n int *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(int) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Allocate device output array and num selected\n int *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceSelect::Unique(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Run\n CubDebugExit(DeviceSelect::Unique(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_selected, true, g_verbose);\n printf(\"\\t Data %s \", compare ? \"FAIL\" : \"PASS\");\n compare = compare | CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n printf(\"\\n\\n\");\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/examples/device/example_device_sort_find_non_trivial_runs.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Simple example of sorting a sequence of keys and values (each pair is a\n * randomly-selected int32 paired with its original offset in the unsorted sequence), and then\n * isolating all maximal, non-trivial (having length > 1) \"runs\" of duplicates.\n *\n * To compile using the command line:\n * nvcc -arch=sm_XX example_device_sort_find_non_trivial_runs.cu -I../.. -lcudart -O3\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \"../../test/test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false; // Whether to display input/output to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Simple key-value pairing for using std::sort on key-value pairs.\n */\ntemplate \nstruct Pair\n{\n Key key;\n Value value;\n\n bool operator<(const Pair &b) const\n {\n return (key < b.key);\n }\n};\n\n\n/**\n * Pair ostream operator\n */\ntemplate \nstd::ostream& operator<<(std::ostream& os, const Pair& val)\n{\n os << '<' << val.key << ',' << val.value << '>';\n return os;\n}\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n Key *h_keys,\n Value *h_values,\n int num_items,\n int max_key)\n{\n float scale = float(max_key) / float(UINT_MAX);\n for (int i = 0; i < num_items; ++i)\n {\n Key sample;\n RandomBits(sample);\n h_keys[i] = (max_key == -1) ? i : (Key) (scale * sample);\n h_values[i] = i;\n }\n\n if (g_verbose)\n {\n printf(\"Keys:\\n\");\n DisplayResults(h_keys, num_items);\n printf(\"\\n\\n\");\n\n printf(\"Values:\\n\");\n DisplayResults(h_values, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve sorted non-trivial subrange problem. Returns the number\n * of non-trivial runs found.\n */\ntemplate \nint Solve(\n Key *h_keys,\n Value *h_values,\n int num_items,\n int *h_offsets_reference,\n int *h_lengths_reference)\n{\n // Sort\n\n Pair *h_pairs = new Pair[num_items];\n for (int i = 0; i < num_items; ++i)\n {\n h_pairs[i].key = h_keys[i];\n h_pairs[i].value = h_values[i];\n }\n\n std::stable_sort(h_pairs, h_pairs + num_items);\n\n if (g_verbose)\n {\n printf(\"Sorted pairs:\\n\");\n DisplayResults(h_pairs, num_items);\n printf(\"\\n\\n\");\n }\n\n // Find non-trivial runs\n\n Key previous = h_pairs[0].key;\n int length = 1;\n int num_runs = 0;\n int run_begin = 0;\n\n for (int i = 1; i < num_items; ++i)\n {\n if (previous != h_pairs[i].key)\n {\n if (length > 1)\n {\n h_offsets_reference[num_runs] = run_begin;\n h_lengths_reference[num_runs] = length;\n num_runs++;\n }\n length = 1;\n run_begin = i;\n }\n else\n {\n length++;\n }\n previous = h_pairs[i].key;\n }\n\n if (length > 1)\n {\n h_offsets_reference[num_runs] = run_begin;\n h_lengths_reference[num_runs] = length;\n num_runs++;\n }\n\n delete[] h_pairs;\n\n return num_runs;\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n typedef unsigned int Key;\n typedef int Value;\n\n int timing_iterations = 0;\n int num_items = 40;\n Key max_key = 20; // Max item\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"maxkey\", max_key);\n args.GetCmdLineArgument(\"i\", timing_iterations);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--i= \"\n \"[--n= \"\n \"[--maxkey=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Allocate host arrays (problem and reference solution)\n\n Key *h_keys = new Key[num_items];\n Value *h_values = new Value[num_items];\n int *h_offsets_reference = new int[num_items];\n int *h_lengths_reference = new int[num_items];\n\n // Initialize key-value pairs and compute reference solution (sort them, and identify non-trivial runs)\n printf(\"Computing reference solution on CPU for %d items (max key %d)\\n\", num_items, max_key);\n fflush(stdout);\n\n Initialize(h_keys, h_values, num_items, max_key);\n int num_runs = Solve(h_keys, h_values, num_items, h_offsets_reference, h_lengths_reference);\n\n printf(\"%d non-trivial runs\\n\", num_runs);\n fflush(stdout);\n\n // Repeat for performance timing\n GpuTimer gpu_timer;\n GpuTimer gpu_rle_timer;\n float elapsed_millis = 0.0;\n float elapsed_rle_millis = 0.0;\n for (int i = 0; i <= timing_iterations; ++i)\n {\n\n // Allocate and initialize device arrays for sorting\n DoubleBuffer d_keys;\n DoubleBuffer d_values;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[0], sizeof(Key) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(Key) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[0], sizeof(Value) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[1], sizeof(Value) * num_items));\n\n CubDebugExit(cudaMemcpy(d_keys.d_buffers[d_keys.selector], h_keys, sizeof(float) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_values.d_buffers[d_values.selector], h_values, sizeof(int) * num_items, cudaMemcpyHostToDevice));\n\n // Start timer\n gpu_timer.Start();\n\n // Allocate temporary storage for sorting\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n CubDebugExit(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Do the sort\n CubDebugExit(DeviceRadixSort::SortPairs(d_temp_storage, temp_storage_bytes, d_keys, d_values, num_items));\n\n // Free unused buffers and sorting temporary storage\n if (d_keys.d_buffers[d_keys.selector ^ 1]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[d_keys.selector ^ 1]));\n if (d_values.d_buffers[d_values.selector ^ 1]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[d_values.selector ^ 1]));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Start timer\n gpu_rle_timer.Start();\n\n // Allocate device arrays for enumerating non-trivial runs\n int *d_offests_out = NULL;\n int *d_lengths_out = NULL;\n int *d_num_runs = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_offests_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_lengths_out, sizeof(int) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_runs, sizeof(int) * 1));\n\n // Allocate temporary storage for isolating non-trivial runs\n d_temp_storage = NULL;\n CubDebugExit(DeviceRunLengthEncode::NonTrivialRuns(\n d_temp_storage,\n temp_storage_bytes,\n d_keys.d_buffers[d_keys.selector],\n d_offests_out,\n d_lengths_out,\n d_num_runs,\n num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Do the isolation\n CubDebugExit(DeviceRunLengthEncode::NonTrivialRuns(\n d_temp_storage,\n temp_storage_bytes,\n d_keys.d_buffers[d_keys.selector],\n d_offests_out,\n d_lengths_out,\n d_num_runs,\n num_items));\n\n // Free keys buffer\n if (d_keys.d_buffers[d_keys.selector]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[d_keys.selector]));\n\n //\n // Hypothetically do stuff with the original key-indices corresponding to non-trivial runs of identical keys\n //\n\n // Stop sort timer\n gpu_timer.Stop();\n gpu_rle_timer.Stop();\n\n if (i == 0)\n {\n // First iteration is a warmup: // Check for correctness (and display results, if specified)\n\n printf(\"\\nRUN OFFSETS: \\n\");\n int compare = CompareDeviceResults(h_offsets_reference, d_offests_out, num_runs, true, g_verbose);\n printf(\"\\t\\t %s \", compare ? \"FAIL\" : \"PASS\");\n\n printf(\"\\nRUN LENGTHS: \\n\");\n compare |= CompareDeviceResults(h_lengths_reference, d_lengths_out, num_runs, true, g_verbose);\n printf(\"\\t\\t %s \", compare ? \"FAIL\" : \"PASS\");\n\n printf(\"\\nNUM RUNS: \\n\");\n compare |= CompareDeviceResults(&num_runs, d_num_runs, 1, true, g_verbose);\n printf(\"\\t\\t %s \", compare ? \"FAIL\" : \"PASS\");\n\n AssertEquals(0, compare);\n }\n else\n {\n elapsed_millis += gpu_timer.ElapsedMillis();\n elapsed_rle_millis += gpu_rle_timer.ElapsedMillis();\n }\n\n // GPU cleanup\n\n if (d_values.d_buffers[d_values.selector]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[d_values.selector]));\n if (d_offests_out) CubDebugExit(g_allocator.DeviceFree(d_offests_out));\n if (d_lengths_out) CubDebugExit(g_allocator.DeviceFree(d_lengths_out));\n if (d_num_runs) CubDebugExit(g_allocator.DeviceFree(d_num_runs));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n }\n\n // Host cleanup\n if (h_keys) delete[] h_keys;\n if (h_values) delete[] h_values;\n if (h_offsets_reference) delete[] h_offsets_reference;\n if (h_lengths_reference) delete[] h_lengths_reference;\n\n printf(\"\\n\\n\");\n\n if (timing_iterations > 0)\n {\n printf(\"%d timing iterations, average time to sort and isolate non-trivial duplicates: %.3f ms (%.3f ms spent in RLE isolation)\\n\",\n timing_iterations,\n elapsed_millis / timing_iterations,\n elapsed_rle_millis / timing_iterations);\n }\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/experimental/.gitignore", "content": "/bin\n" }, { "path": "external/cub/experimental/Makefile", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n\n#-------------------------------------------------------------------------------\n#\n# Makefile usage\n#\n# make [sm=] [cdp=<0|1>] [force32=<0|1>] [abi=<0|1>] [open64=<0|1>] [verbose=<0|1>] [keep=<0|1>] [quicktest=<0|1>]\n#\n#-------------------------------------------------------------------------------\n \ninclude ../common.mk \n\n#-------------------------------------------------------------------------------\n# Commandline Options\n#-------------------------------------------------------------------------------\n\n# [mkl=<0|1>] compile against Intel MKL\nifeq ($(mkl), 1)\n\tDEFINES \t+= -DCUB_MKL\n\nifeq (WIN_NT, $(findstring WIN_NT, $(OSUPPER)))\n\tLIBS \t\t+=\tmkl_intel_lp64.lib mkl_intel_thread.lib mkl_core.lib libiomp5md.lib\n\tNVCCFLAGS \t+= -Xcompiler /openmp\nelse\n\tLIBS\t\t+= -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -liomp5 -lpthread -lm\n\tNVCCFLAGS \t+= -Xcompiler -fopenmp\n\t\nendif\t\n\nendif\n\n\n#-------------------------------------------------------------------------------\n# Compiler and compilation platform\n#-------------------------------------------------------------------------------\n\n# Includes\nINC += -I$(CUB_DIR) -I$(CUB_DIR)test \n\n# detect OS\nOSUPPER = $(shell uname -s 2>/dev/null | tr [:lower:] [:upper:])\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nexp_rwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nEXP_DEPS = \t$(call rwildcard, ./,*.cuh) \\\n\t\t\t$(call rwildcard, ./,*.h)\n\nDEPS =\t\t\t\t$(CUB_DEPS) \\\n\t\t\t\t\t$(EXP_DEPS) \\\n\t\t\t\t\t$(CUB_DIR)test/Makefile \\\n\t\t\t\t\t$(CUB_DIR)test/test_util.h \\\n\t\t\t\t\t$(CUB_DIR)test/mersenne.h \\\n\n\t\t\n\n#-------------------------------------------------------------------------------\n# make default\n#-------------------------------------------------------------------------------\n\ndefault:\n\n\n#-------------------------------------------------------------------------------\n# make clean\n#-------------------------------------------------------------------------------\n\nclean :\n\trm -f bin/*$(CPU_ARCH_SUFFIX)* \n\trm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o\n\n\n\n#-------------------------------------------------------------------------------\n# make histogram_compare\n#-------------------------------------------------------------------------------\n\nhistogram_compare: bin/histogram_compare_$(BIN_SUFFIX)\n\nbin/histogram_compare_$(BIN_SUFFIX) : histogram_compare.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/histogram_compare_$(BIN_SUFFIX) histogram_compare.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\t\n\n\n#-------------------------------------------------------------------------------\n# make spmv_compare\n#-------------------------------------------------------------------------------\n\nspmv_compare: bin/spmv_compare_$(BIN_SUFFIX)\n\nbin/spmv_compare_$(BIN_SUFFIX) : spmv_compare.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/spmv_compare_$(BIN_SUFFIX) spmv_compare.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -lcusparse $(MKL_LIBS) -O3\n\t\n\n" }, { "path": "external/cub/experimental/defunct/example_coo_spmv.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * An implementation of COO SpMV using prefix scan to implement a\n * reduce-value-by-row strategy\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n#include \n\n#include \n\n#include \"coo_graph.cuh\"\n#include \"../test/test_util.h\"\n\nusing namespace cub;\nusing namespace std;\n\n\n/******************************************************************************\n * Globals, constants, and typedefs\n ******************************************************************************/\n\ntypedef int VertexId; // uint32s as vertex ids\ntypedef double Value; // double-precision floating point values\n\nbool g_verbose = false;\nint g_timing_iterations = 1;\nCachingDeviceAllocator g_allocator;\n\n\n/******************************************************************************\n * Texture referencing\n ******************************************************************************/\n\n/**\n * Templated texture reference type for multiplicand vector\n */\ntemplate \nstruct TexVector\n{\n // Texture type to actually use (e.g., because CUDA doesn't load doubles as texture items)\n typedef typename If<(Equals::VALUE), uint2, Value>::Type CastType;\n\n // Texture reference type\n typedef texture TexRef;\n\n static TexRef ref;\n\n /**\n * Bind textures\n */\n static void BindTexture(void *d_in, int elements)\n {\n cudaChannelFormatDesc tex_desc = cudaCreateChannelDesc();\n if (d_in)\n {\n size_t offset;\n size_t bytes = sizeof(CastType) * elements;\n CubDebugExit(cudaBindTexture(&offset, ref, d_in, tex_desc, bytes));\n }\n }\n\n /**\n * Unbind textures\n */\n static void UnbindTexture()\n {\n CubDebugExit(cudaUnbindTexture(ref));\n }\n\n /**\n * Load\n */\n static __device__ __forceinline__ Value Load(int offset)\n {\n Value output;\n reinterpret_cast::CastType &>(output) = tex1Dfetch(TexVector::ref, offset);\n return output;\n }\n};\n\n// Texture reference definitions\ntemplate \ntypename TexVector::TexRef TexVector::ref = 0;\n\n\n/******************************************************************************\n * Utility types\n ******************************************************************************/\n\n\n/**\n * A partial dot-product sum paired with a corresponding row-id\n */\ntemplate \nstruct PartialProduct\n{\n VertexId row; /// Row-id\n Value partial; /// PartialProduct sum\n};\n\n\n/**\n * A partial dot-product sum paired with a corresponding row-id (specialized for double-int pairings)\n */\ntemplate <>\nstruct PartialProduct\n{\n long long row; /// Row-id\n double partial; /// PartialProduct sum\n};\n\n\n/**\n * Reduce-value-by-row scan operator\n */\nstruct ReduceByKeyOp\n{\n template \n __device__ __forceinline__ PartialProduct operator()(\n const PartialProduct &first,\n const PartialProduct &second)\n {\n PartialProduct retval;\n\n retval.partial = (second.row != first.row) ?\n second.partial :\n first.partial + second.partial;\n\n retval.row = second.row;\n return retval;\n }\n};\n\n\n/**\n * Stateful block-wide prefix operator for BlockScan\n */\ntemplate \nstruct BlockPrefixCallbackOp\n{\n // Running block-wide prefix\n PartialProduct running_prefix;\n\n /**\n * Returns the block-wide running_prefix in thread-0\n */\n __device__ __forceinline__ PartialProduct operator()(\n const PartialProduct &block_aggregate) ///< The aggregate sum of the BlockScan inputs\n {\n ReduceByKeyOp scan_op;\n\n PartialProduct retval = running_prefix;\n running_prefix = scan_op(running_prefix, block_aggregate);\n return retval;\n }\n};\n\n\n/**\n * Operator for detecting discontinuities in a list of row identifiers.\n */\nstruct NewRowOp\n{\n /// Returns true if row_b is the start of a new row\n template \n __device__ __forceinline__ bool operator()(\n const VertexId& row_a,\n const VertexId& row_b)\n {\n return (row_a != row_b);\n }\n};\n\n\n\n/******************************************************************************\n * Persistent thread block types\n ******************************************************************************/\n\n/**\n * SpMV thread block abstraction for processing a contiguous segment of\n * sparse COO tiles.\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename VertexId,\n typename Value>\nstruct PersistentBlockSpmv\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // Constants\n enum\n {\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n // Head flag type\n typedef int HeadFlag;\n\n // Partial dot product type\n typedef PartialProduct PartialProduct;\n\n // Parameterized BlockScan type for reduce-value-by-row scan\n typedef BlockScan BlockScan;\n\n // Parameterized BlockExchange type for exchanging rows between warp-striped -> blocked arrangements\n typedef BlockExchange BlockExchangeRows;\n\n // Parameterized BlockExchange type for exchanging values between warp-striped -> blocked arrangements\n typedef BlockExchange BlockExchangeValues;\n\n // Parameterized BlockDiscontinuity type for setting head-flags for each new row segment\n typedef BlockDiscontinuity BlockDiscontinuity;\n\n // Shared memory type for this thread block\n struct TempStorage\n {\n union\n {\n typename BlockExchangeRows::TempStorage exchange_rows; // Smem needed for BlockExchangeRows\n typename BlockExchangeValues::TempStorage exchange_values; // Smem needed for BlockExchangeValues\n struct\n {\n typename BlockScan::TempStorage scan; // Smem needed for BlockScan\n typename BlockDiscontinuity::TempStorage discontinuity; // Smem needed for BlockDiscontinuity\n };\n };\n\n VertexId first_block_row; ///< The first row-ID seen by this thread block\n VertexId last_block_row; ///< The last row-ID seen by this thread block\n Value first_product; ///< The first dot-product written by this thread block\n };\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n TempStorage &temp_storage;\n BlockPrefixCallbackOp prefix_op;\n VertexId *d_rows;\n VertexId *d_columns;\n Value *d_values;\n Value *d_vector;\n Value *d_result;\n PartialProduct *d_block_partials;\n int block_offset;\n int block_end;\n\n\n //---------------------------------------------------------------------\n // Operations\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__\n PersistentBlockSpmv(\n TempStorage &temp_storage,\n VertexId *d_rows,\n VertexId *d_columns,\n Value *d_values,\n Value *d_vector,\n Value *d_result,\n PartialProduct *d_block_partials,\n int block_offset,\n int block_end)\n :\n temp_storage(temp_storage),\n d_rows(d_rows),\n d_columns(d_columns),\n d_values(d_values),\n d_vector(d_vector),\n d_result(d_result),\n d_block_partials(d_block_partials),\n block_offset(block_offset),\n block_end(block_end)\n {\n // Initialize scalar shared memory values\n if (threadIdx.x == 0)\n {\n VertexId first_block_row = d_rows[block_offset];\n VertexId last_block_row = d_rows[block_end - 1];\n\n temp_storage.first_block_row = first_block_row;\n temp_storage.last_block_row = last_block_row;\n temp_storage.first_product = Value(0);\n\n // Initialize prefix_op to identity\n prefix_op.running_prefix.row = first_block_row;\n prefix_op.running_prefix.partial = Value(0);\n }\n\n __syncthreads();\n }\n\n\n /**\n * Processes a COO input tile of edges, outputting dot products for each row\n */\n template \n __device__ __forceinline__ void ProcessTile(\n int block_offset,\n int guarded_items = 0)\n {\n VertexId columns[ITEMS_PER_THREAD];\n VertexId rows[ITEMS_PER_THREAD];\n Value values[ITEMS_PER_THREAD];\n PartialProduct partial_sums[ITEMS_PER_THREAD];\n HeadFlag head_flags[ITEMS_PER_THREAD];\n\n // Load a thread block-striped tile of A (sparse row-ids, column-ids, and values)\n if (FULL_TILE)\n {\n // Unguarded loads\n LoadDirectWarpStriped(threadIdx.x, d_columns + block_offset, columns);\n LoadDirectWarpStriped(threadIdx.x, d_values + block_offset, values);\n LoadDirectWarpStriped(threadIdx.x, d_rows + block_offset, rows);\n }\n else\n {\n // This is a partial-tile (e.g., the last tile of input). Extend the coordinates of the last\n // vertex for out-of-bound items, but zero-valued\n LoadDirectWarpStriped(threadIdx.x, d_columns + block_offset, columns, guarded_items, VertexId(0));\n LoadDirectWarpStriped(threadIdx.x, d_values + block_offset, values, guarded_items, Value(0));\n LoadDirectWarpStriped(threadIdx.x, d_rows + block_offset, rows, guarded_items, temp_storage.last_block_row);\n }\n\n // Load the referenced values from x and compute the dot product partials sums\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n#if CUB_PTX_ARCH >= 350\n values[ITEM] *= ThreadLoad(d_vector + columns[ITEM]);\n#else\n values[ITEM] *= TexVector::Load(columns[ITEM]);\n#endif\n }\n\n // Transpose from warp-striped to blocked arrangement\n BlockExchangeValues(temp_storage.exchange_values).WarpStripedToBlocked(values);\n\n __syncthreads();\n\n // Transpose from warp-striped to blocked arrangement\n BlockExchangeRows(temp_storage.exchange_rows).WarpStripedToBlocked(rows);\n\n // Barrier for smem reuse and coherence\n __syncthreads();\n\n // FlagT row heads by looking for discontinuities\n BlockDiscontinuity(temp_storage.discontinuity).FlagHeads(\n head_flags, // (Out) Head flags\n rows, // Original row ids\n NewRowOp(), // Functor for detecting start of new rows\n prefix_op.running_prefix.row); // Last row ID from previous tile to compare with first row ID in this tile\n\n // Assemble partial product structures\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n partial_sums[ITEM].partial = values[ITEM];\n partial_sums[ITEM].row = rows[ITEM];\n }\n\n // Reduce reduce-value-by-row across partial_sums using exclusive prefix scan\n PartialProduct block_aggregate;\n BlockScan(temp_storage.scan).ExclusiveScan(\n partial_sums, // Scan input\n partial_sums, // Scan output\n ReduceByKeyOp(), // Scan operator\n block_aggregate, // Block-wide total (unused)\n prefix_op); // Prefix operator for seeding the block-wide scan with the running total\n\n // Barrier for smem reuse and coherence\n __syncthreads();\n\n // Scatter an accumulated dot product if it is the head of a valid row\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (head_flags[ITEM])\n {\n d_result[partial_sums[ITEM].row] = partial_sums[ITEM].partial;\n\n // Save off the first partial product that this thread block will scatter\n if (partial_sums[ITEM].row == temp_storage.first_block_row)\n {\n temp_storage.first_product = partial_sums[ITEM].partial;\n }\n }\n }\n }\n\n\n /**\n * Iterate over input tiles belonging to this thread block\n */\n __device__ __forceinline__\n void ProcessTiles()\n {\n // Process full tiles\n while (block_offset <= block_end - TILE_ITEMS)\n {\n ProcessTile(block_offset);\n block_offset += TILE_ITEMS;\n }\n\n // Process the last, partially-full tile (if present)\n int guarded_items = block_end - block_offset;\n if (guarded_items)\n {\n ProcessTile(block_offset, guarded_items);\n }\n\n if (threadIdx.x == 0)\n {\n if (gridDim.x == 1)\n {\n // Scatter the final aggregate (this kernel contains only 1 thread block)\n d_result[prefix_op.running_prefix.row] = prefix_op.running_prefix.partial;\n }\n else\n {\n // Write the first and last partial products from this thread block so\n // that they can be subsequently \"fixed up\" in the next kernel.\n\n PartialProduct first_product;\n first_product.row = temp_storage.first_block_row;\n first_product.partial = temp_storage.first_product;\n\n d_block_partials[blockIdx.x * 2] = first_product;\n d_block_partials[(blockIdx.x * 2) + 1] = prefix_op.running_prefix;\n }\n }\n }\n};\n\n\n/**\n * Threadblock abstraction for \"fixing up\" an array of interblock SpMV partial products.\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename VertexId,\n typename Value>\nstruct FinalizeSpmvBlock\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // Constants\n enum\n {\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n // Head flag type\n typedef int HeadFlag;\n\n // Partial dot product type\n typedef PartialProduct PartialProduct;\n\n // Parameterized BlockScan type for reduce-value-by-row scan\n typedef BlockScan BlockScan;\n\n // Parameterized BlockDiscontinuity type for setting head-flags for each new row segment\n typedef BlockDiscontinuity BlockDiscontinuity;\n\n // Shared memory type for this thread block\n struct TempStorage\n {\n typename BlockScan::TempStorage scan; // Smem needed for reduce-value-by-row scan\n typename BlockDiscontinuity::TempStorage discontinuity; // Smem needed for head-flagging\n\n VertexId last_block_row;\n };\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n TempStorage &temp_storage;\n BlockPrefixCallbackOp prefix_op;\n Value *d_result;\n PartialProduct *d_block_partials;\n int num_partials;\n\n\n //---------------------------------------------------------------------\n // Operations\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__\n FinalizeSpmvBlock(\n TempStorage &temp_storage,\n Value *d_result,\n PartialProduct *d_block_partials,\n int num_partials)\n :\n temp_storage(temp_storage),\n d_result(d_result),\n d_block_partials(d_block_partials),\n num_partials(num_partials)\n {\n // Initialize scalar shared memory values\n if (threadIdx.x == 0)\n {\n VertexId first_block_row = d_block_partials[0].row;\n VertexId last_block_row = d_block_partials[num_partials - 1].row;\n temp_storage.last_block_row = last_block_row;\n\n // Initialize prefix_op to identity\n prefix_op.running_prefix.row = first_block_row;\n prefix_op.running_prefix.partial = Value(0);\n }\n\n __syncthreads();\n }\n\n\n /**\n * Processes a COO input tile of edges, outputting dot products for each row\n */\n template \n __device__ __forceinline__\n void ProcessTile(\n int block_offset,\n int guarded_items = 0)\n {\n VertexId rows[ITEMS_PER_THREAD];\n PartialProduct partial_sums[ITEMS_PER_THREAD];\n HeadFlag head_flags[ITEMS_PER_THREAD];\n\n // Load a tile of block partials from previous kernel\n if (FULL_TILE)\n {\n // Full tile\n#if CUB_PTX_ARCH >= 350\n LoadDirectBlocked(threadIdx.x, d_block_partials + block_offset, partial_sums);\n#else\n LoadDirectBlocked(threadIdx.x, d_block_partials + block_offset, partial_sums);\n#endif\n }\n else\n {\n // Partial tile (extend zero-valued coordinates of the last partial-product for out-of-bounds items)\n PartialProduct default_sum;\n default_sum.row = temp_storage.last_block_row;\n default_sum.partial = Value(0);\n\n#if CUB_PTX_ARCH >= 350\n LoadDirectBlocked(threadIdx.x, d_block_partials + block_offset, partial_sums, guarded_items, default_sum);\n#else\n LoadDirectBlocked(threadIdx.x, d_block_partials + block_offset, partial_sums, guarded_items, default_sum);\n#endif\n }\n\n // Copy out row IDs for row-head flagging\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n rows[ITEM] = partial_sums[ITEM].row;\n }\n\n // FlagT row heads by looking for discontinuities\n BlockDiscontinuity(temp_storage.discontinuity).FlagHeads(\n rows, // Original row ids\n head_flags, // (Out) Head flags\n NewRowOp(), // Functor for detecting start of new rows\n prefix_op.running_prefix.row); // Last row ID from previous tile to compare with first row ID in this tile\n\n // Reduce reduce-value-by-row across partial_sums using exclusive prefix scan\n PartialProduct block_aggregate;\n BlockScan(temp_storage.scan).ExclusiveScan(\n partial_sums, // Scan input\n partial_sums, // Scan output\n ReduceByKeyOp(), // Scan operator\n block_aggregate, // Block-wide total (unused)\n prefix_op); // Prefix operator for seeding the block-wide scan with the running total\n\n // Scatter an accumulated dot product if it is the head of a valid row\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (head_flags[ITEM])\n {\n d_result[partial_sums[ITEM].row] = partial_sums[ITEM].partial;\n }\n }\n }\n\n\n /**\n * Iterate over input tiles belonging to this thread block\n */\n __device__ __forceinline__\n void ProcessTiles()\n {\n // Process full tiles\n int block_offset = 0;\n while (block_offset <= num_partials - TILE_ITEMS)\n {\n ProcessTile(block_offset);\n block_offset += TILE_ITEMS;\n }\n\n // Process final partial tile (if present)\n int guarded_items = num_partials - block_offset;\n if (guarded_items)\n {\n ProcessTile(block_offset, guarded_items);\n }\n\n // Scatter the final aggregate (this kernel contains only 1 thread block)\n if (threadIdx.x == 0)\n {\n d_result[prefix_op.running_prefix.row] = prefix_op.running_prefix.partial;\n }\n }\n};\n\n\n/******************************************************************************\n * Kernel entrypoints\n ******************************************************************************/\n\n\n\n/**\n * SpMV kernel whose thread blocks each process a contiguous segment of sparse COO tiles.\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename VertexId,\n typename Value>\n__launch_bounds__ (BLOCK_THREADS)\n__global__ void CooKernel(\n GridEvenShare even_share,\n PartialProduct *d_block_partials,\n VertexId *d_rows,\n VertexId *d_columns,\n Value *d_values,\n Value *d_vector,\n Value *d_result)\n{\n // Specialize SpMV thread block abstraction type\n typedef PersistentBlockSpmv PersistentBlockSpmv;\n\n // Shared memory allocation\n __shared__ typename PersistentBlockSpmv::TempStorage temp_storage;\n\n // Initialize thread block even-share to tell us where to start and stop our tile-processing\n even_share.BlockInit();\n\n // Construct persistent thread block\n PersistentBlockSpmv persistent_block(\n temp_storage,\n d_rows,\n d_columns,\n d_values,\n d_vector,\n d_result,\n d_block_partials,\n even_share.block_offset,\n even_share.block_end);\n\n // Process input tiles\n persistent_block.ProcessTiles();\n}\n\n\n/**\n * Kernel for \"fixing up\" an array of interblock SpMV partial products.\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename VertexId,\n typename Value>\n__launch_bounds__ (BLOCK_THREADS, 1)\n__global__ void CooFinalizeKernel(\n PartialProduct *d_block_partials,\n int num_partials,\n Value *d_result)\n{\n // Specialize \"fix-up\" thread block abstraction type\n typedef FinalizeSpmvBlock FinalizeSpmvBlock;\n\n // Shared memory allocation\n __shared__ typename FinalizeSpmvBlock::TempStorage temp_storage;\n\n // Construct persistent thread block\n FinalizeSpmvBlock persistent_block(temp_storage, d_result, d_block_partials, num_partials);\n\n // Process input tiles\n persistent_block.ProcessTiles();\n}\n\n\n\n//---------------------------------------------------------------------\n// Host subroutines\n//---------------------------------------------------------------------\n\n\n/**\n * Simple test of device\n */\ntemplate <\n int COO_BLOCK_THREADS,\n int COO_ITEMS_PER_THREAD,\n int COO_SUBSCRIPTION_FACTOR,\n int FINALIZE_BLOCK_THREADS,\n int FINALIZE_ITEMS_PER_THREAD,\n typename VertexId,\n typename Value>\nvoid TestDevice(\n CooGraph& coo_graph,\n Value* h_vector,\n Value* h_reference)\n{\n typedef PartialProduct PartialProduct;\n\n const int COO_TILE_SIZE = COO_BLOCK_THREADS * COO_ITEMS_PER_THREAD;\n\n // SOA device storage\n VertexId *d_rows; // SOA graph row coordinates\n VertexId *d_columns; // SOA graph col coordinates\n Value *d_values; // SOA graph values\n Value *d_vector; // Vector multiplicand\n Value *d_result; // Output row\n PartialProduct *d_block_partials; // Temporary storage for communicating dot product partials between thread blocks\n\n // Create SOA version of coo_graph on host\n int num_edges = coo_graph.coo_tuples.size();\n VertexId *h_rows = new VertexId[num_edges];\n VertexId *h_columns = new VertexId[num_edges];\n Value *h_values = new Value[num_edges];\n for (int i = 0; i < num_edges; i++)\n {\n h_rows[i] = coo_graph.coo_tuples[i].row;\n h_columns[i] = coo_graph.coo_tuples[i].col;\n h_values[i] = coo_graph.coo_tuples[i].val;\n }\n\n // Get CUDA properties\n Device device_props;\n CubDebugExit(device_props.Init());\n\n // Determine launch configuration from kernel properties\n int coo_sm_occupancy;\n CubDebugExit(device_props.MaxSmOccupancy(\n coo_sm_occupancy,\n CooKernel,\n COO_BLOCK_THREADS));\n int max_coo_grid_size = device_props.sm_count * coo_sm_occupancy * COO_SUBSCRIPTION_FACTOR;\n\n // Construct an even-share work distribution\n GridEvenShare even_share(num_edges, max_coo_grid_size, COO_TILE_SIZE);\n int coo_grid_size = even_share.grid_size;\n int num_partials = coo_grid_size * 2;\n\n // Allocate COO device arrays\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_rows, sizeof(VertexId) * num_edges));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_columns, sizeof(VertexId) * num_edges));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values, sizeof(Value) * num_edges));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_vector, sizeof(Value) * coo_graph.col_dim));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_result, sizeof(Value) * coo_graph.row_dim));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_block_partials, sizeof(PartialProduct) * num_partials));\n\n // Copy host arrays to device\n CubDebugExit(cudaMemcpy(d_rows, h_rows, sizeof(VertexId) * num_edges, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_columns, h_columns, sizeof(VertexId) * num_edges, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_values, h_values, sizeof(Value) * num_edges, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_vector, h_vector, sizeof(Value) * coo_graph.col_dim, cudaMemcpyHostToDevice));\n\n // Bind textures\n TexVector::BindTexture(d_vector, coo_graph.col_dim);\n\n // Print debug info\n printf(\"CooKernel<%d, %d><<<%d, %d>>>(...), Max SM occupancy: %d\\n\",\n COO_BLOCK_THREADS, COO_ITEMS_PER_THREAD, coo_grid_size, COO_BLOCK_THREADS, coo_sm_occupancy);\n if (coo_grid_size > 1)\n {\n printf(\"CooFinalizeKernel<<<1, %d>>>(...)\\n\", FINALIZE_BLOCK_THREADS);\n }\n fflush(stdout);\n\n CubDebugExit(cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte));\n\n // Run kernel (always run one iteration without timing)\n GpuTimer gpu_timer;\n float elapsed_millis = 0.0;\n for (int i = 0; i <= g_timing_iterations; i++)\n {\n gpu_timer.Start();\n\n // Initialize output\n CubDebugExit(cudaMemset(d_result, 0, coo_graph.row_dim * sizeof(Value)));\n\n // Run the COO kernel\n CooKernel<<>>(\n even_share,\n d_block_partials,\n d_rows,\n d_columns,\n d_values,\n d_vector,\n d_result);\n\n if (coo_grid_size > 1)\n {\n // Run the COO finalize kernel\n CooFinalizeKernel<<<1, FINALIZE_BLOCK_THREADS>>>(\n d_block_partials,\n num_partials,\n d_result);\n }\n\n gpu_timer.Stop();\n\n if (i > 0)\n elapsed_millis += gpu_timer.ElapsedMillis();\n }\n\n // Force any kernel stdio to screen\n CubDebugExit(cudaThreadSynchronize());\n fflush(stdout);\n\n // Display timing\n if (g_timing_iterations > 0)\n {\n float avg_elapsed = elapsed_millis / g_timing_iterations;\n int total_bytes = ((sizeof(VertexId) + sizeof(VertexId)) * 2 * num_edges) + (sizeof(Value) * coo_graph.row_dim);\n printf(\"%d iterations, average elapsed (%.3f ms), utilized bandwidth (%.3f GB/s), GFLOPS(%.3f)\\n\",\n g_timing_iterations,\n avg_elapsed,\n total_bytes / avg_elapsed / 1000.0 / 1000.0,\n num_edges * 2 / avg_elapsed / 1000.0 / 1000.0);\n }\n\n // Check results\n int compare = CompareDeviceResults(h_reference, d_result, coo_graph.row_dim, true, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n TexVector::UnbindTexture();\n CubDebugExit(g_allocator.DeviceFree(d_block_partials));\n CubDebugExit(g_allocator.DeviceFree(d_rows));\n CubDebugExit(g_allocator.DeviceFree(d_columns));\n CubDebugExit(g_allocator.DeviceFree(d_values));\n CubDebugExit(g_allocator.DeviceFree(d_vector));\n CubDebugExit(g_allocator.DeviceFree(d_result));\n delete[] h_rows;\n delete[] h_columns;\n delete[] h_values;\n}\n\n\n/**\n * Compute reference answer on CPU\n */\ntemplate \nvoid ComputeReference(\n CooGraph& coo_graph,\n Value* h_vector,\n Value* h_reference)\n{\n for (VertexId i = 0; i < coo_graph.row_dim; i++)\n {\n h_reference[i] = 0.0;\n }\n\n for (VertexId i = 0; i < coo_graph.coo_tuples.size(); i++)\n {\n h_reference[coo_graph.coo_tuples[i].row] +=\n coo_graph.coo_tuples[i].val *\n h_vector[coo_graph.coo_tuples[i].col];\n }\n}\n\n\n/**\n * Assign arbitrary values to vector items\n */\ntemplate \nvoid AssignVectorValues(Value *vector, int col_dim)\n{\n for (int i = 0; i < col_dim; i++)\n {\n vector[i] = 1.0;\n }\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s\\n [--device=] [--v] [--iterations=] [--grid-size=]\\n\"\n \"\\t--type=wheel --spokes=\\n\"\n \"\\t--type=grid2d --width= [--no-self-loops]\\n\"\n \"\\t--type=grid3d --width= [--no-self-loops]\\n\"\n \"\\t--type=market --file=\\n\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get graph type\n string type;\n args.GetCmdLineArgument(\"type\", type);\n\n // Generate graph structure\n\n CpuTimer timer;\n timer.Start();\n CooGraph coo_graph;\n if (type == string(\"grid2d\"))\n {\n VertexId width;\n args.GetCmdLineArgument(\"width\", width);\n bool self_loops = !args.CheckCmdLineFlag(\"no-self-loops\");\n printf(\"Generating %s grid2d width(%d)... \", (self_loops) ? \"5-pt\" : \"4-pt\", width); fflush(stdout);\n if (coo_graph.InitGrid2d(width, self_loops)) exit(1);\n } else if (type == string(\"grid3d\"))\n {\n VertexId width;\n args.GetCmdLineArgument(\"width\", width);\n bool self_loops = !args.CheckCmdLineFlag(\"no-self-loops\");\n printf(\"Generating %s grid3d width(%d)... \", (self_loops) ? \"7-pt\" : \"6-pt\", width); fflush(stdout);\n if (coo_graph.InitGrid3d(width, self_loops)) exit(1);\n }\n else if (type == string(\"wheel\"))\n {\n VertexId spokes;\n args.GetCmdLineArgument(\"spokes\", spokes);\n printf(\"Generating wheel spokes(%d)... \", spokes); fflush(stdout);\n if (coo_graph.InitWheel(spokes)) exit(1);\n }\n else if (type == string(\"market\"))\n {\n string filename;\n args.GetCmdLineArgument(\"file\", filename);\n printf(\"Generating MARKET for %s... \", filename.c_str()); fflush(stdout);\n if (coo_graph.InitMarket(filename)) exit(1);\n }\n else\n {\n printf(\"Unsupported graph type\\n\");\n exit(1);\n }\n timer.Stop();\n printf(\"Done (%.3fs). %d non-zeros, %d rows, %d columns\\n\",\n timer.ElapsedMillis() / 1000.0,\n coo_graph.coo_tuples.size(),\n coo_graph.row_dim,\n coo_graph.col_dim);\n fflush(stdout);\n\n if (g_verbose)\n {\n cout << coo_graph << \"\\n\";\n }\n\n // Create vector\n Value *h_vector = new Value[coo_graph.col_dim];\n AssignVectorValues(h_vector, coo_graph.col_dim);\n if (g_verbose)\n {\n printf(\"Vector[%d]: \", coo_graph.col_dim);\n DisplayResults(h_vector, coo_graph.col_dim);\n printf(\"\\n\\n\");\n }\n\n // Compute reference answer\n Value *h_reference = new Value[coo_graph.row_dim];\n ComputeReference(coo_graph, h_vector, h_reference);\n if (g_verbose)\n {\n printf(\"Results[%d]: \", coo_graph.row_dim);\n DisplayResults(h_reference, coo_graph.row_dim);\n printf(\"\\n\\n\");\n }\n\n // Parameterization for SM35\n enum\n {\n COO_BLOCK_THREADS = 64,\n COO_ITEMS_PER_THREAD = 10,\n COO_SUBSCRIPTION_FACTOR = 4,\n FINALIZE_BLOCK_THREADS = 256,\n FINALIZE_ITEMS_PER_THREAD = 4,\n };\n\n // Run GPU version\n TestDevice<\n COO_BLOCK_THREADS,\n COO_ITEMS_PER_THREAD,\n COO_SUBSCRIPTION_FACTOR,\n FINALIZE_BLOCK_THREADS,\n FINALIZE_ITEMS_PER_THREAD>(coo_graph, h_vector, h_reference);\n\n // Cleanup\n delete[] h_vector;\n delete[] h_reference;\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/experimental/defunct/test_device_seg_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * An implementation of segmented reduction using a load-balanced parallelization\n * strategy based on the MergePath decision path.\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n#include \n\n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\nusing namespace std;\n\n\n/******************************************************************************\n * Globals, constants, and typedefs\n ******************************************************************************/\n\nbool g_verbose = false;\nint g_timing_iterations = 1;\nCachingDeviceAllocator g_allocator(true);\n\n\n/******************************************************************************\n * Utility routines\n ******************************************************************************/\n\n\n/**\n * An pair of index offsets\n */\ntemplate \nstruct IndexPair\n{\n OffsetT a_idx;\n OffsetT b_idx;\n};\n\n\n/**\n * Computes the begin offsets into A and B for the specified\n * location (diagonal) along the merge decision path\n */\ntemplate <\n int BLOCK_THREADS,\n typename IteratorA,\n typename IteratorB,\n typename OffsetT>\n__device__ __forceinline__ void ParallelMergePathSearch(\n OffsetT diagonal,\n IteratorA a,\n IteratorB b,\n IndexPair begin, // Begin offsets into a and b\n IndexPair end, // End offsets into a and b\n IndexPair &intersection) // [out] Intersection offsets into a and b\n{\n OffsetT a_split_min = CUB_MAX(diagonal - end.b_idx, begin.a_idx);\n OffsetT a_split_max = CUB_MIN(diagonal, end.a_idx);\n\n while (a_split_min < a_split_max)\n {\n OffsetT a_distance = a_split_max - a_split_min;\n OffsetT a_slice = (a_distance + BLOCK_THREADS - 1) >> Log2::VALUE;\n OffsetT a_split_pivot = CUB_MIN(a_split_min + (threadIdx.x * a_slice), end.a_idx - 1);\n\n int move_up = (a[a_split_pivot] <= b[diagonal - a_split_pivot - 1]);\n int num_up = __syncthreads_count(move_up);\n/*\n _CubLog(\"a_split_min(%d), a_split_max(%d) a_distance(%d), a_slice(%d), a_split_pivot(%d), move_up(%d), num_up(%d), a_begin(%d), a_end(%d)\\n\",\n a_split_min, a_split_max, a_distance, a_slice, a_split_pivot, move_up, num_up, a_begin, a_end);\n*/\n a_split_max = CUB_MIN(num_up * a_slice, end.a_idx);\n a_split_min = CUB_MAX(a_split_max - a_slice, begin.a_idx) + 1;\n }\n\n intersection.a_idx = CUB_MIN(a_split_min, end.a_idx);\n intersection.b_idx = CUB_MIN(diagonal - a_split_min, end.b_idx);\n}\n\n/**\n * Computes the begin offsets into A and B for the specified\n * location (diagonal) along the merge decision path\n */\ntemplate <\n typename IteratorA,\n typename IteratorB,\n typename OffsetT>\n__device__ __forceinline__ void MergePathSearch(\n OffsetT diagonal,\n IteratorA a,\n IteratorB b,\n IndexPair begin, // Begin offsets into a and b\n IndexPair end, // End offsets into a and b\n IndexPair &intersection) // [out] Intersection offsets into a and b\n{\n OffsetT split_min = CUB_MAX(diagonal - end.b_idx, begin.a_idx);\n OffsetT split_max = CUB_MIN(diagonal, end.a_idx);\n\n while (split_min < split_max)\n {\n OffsetT split_pivot = (split_min + split_max) >> 1;\n if (a[split_pivot] <= b[diagonal - split_pivot - 1])\n {\n // Move candidate split range up A, down B\n split_min = split_pivot + 1;\n }\n else\n {\n // Move candidate split range up B, down A\n split_max = split_pivot;\n }\n }\n\n intersection.a_idx = CUB_MIN(split_min, end.a_idx);\n intersection.b_idx = CUB_MIN(diagonal - split_min, end.b_idx);\n}\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for BlockSegReduceRegion\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n bool _USE_SMEM_SEGMENT_CACHE, ///< Whether or not to cache incoming segment offsets in shared memory before reducing each tile\n bool _USE_SMEM_VALUE_CACHE, ///< Whether or not to cache incoming values in shared memory before reducing each tile\n CacheLoadModifier _LOAD_MODIFIER_SEGMENTS, ///< Cache load modifier for reading segment offsets\n CacheLoadModifier _LOAD_MODIFIER_VALUES, ///< Cache load modifier for reading values\n BlockReduceAlgorithm _REDUCE_ALGORITHM, ///< The BlockReduce algorithm to use\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct BlockSegReduceRegionPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n USE_SMEM_SEGMENT_CACHE = _USE_SMEM_SEGMENT_CACHE, ///< Whether or not to cache incoming segment offsets in shared memory before reducing each tile\n USE_SMEM_VALUE_CACHE = _USE_SMEM_VALUE_CACHE, ///< Whether or not to cache incoming upcoming values in shared memory before reducing each tile\n };\n\n static const CacheLoadModifier LOAD_MODIFIER_SEGMENTS = _LOAD_MODIFIER_SEGMENTS; ///< Cache load modifier for reading segment offsets\n static const CacheLoadModifier LOAD_MODIFIER_VALUES = _LOAD_MODIFIER_VALUES; ///< Cache load modifier for reading values\n static const BlockReduceAlgorithm REDUCE_ALGORITHM = _REDUCE_ALGORITHM; ///< The BlockReduce algorithm to use\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n/******************************************************************************\n * Persistent thread block types\n ******************************************************************************/\n\n/**\n * \\brief BlockSegReduceTiles implements a stateful abstraction of CUDA thread blocks for participating in device-wide segmented reduction.\n */\ntemplate <\n typename BlockSegReduceRegionPolicy, ///< Parameterized BlockSegReduceRegionPolicy tuning policy\n typename SegmentOffsetIterator, ///< Random-access input iterator type for reading segment end-offsets\n typename ValueIterator, ///< Random-access input iterator type for reading values\n typename OutputIteratorT, ///< Random-access output iterator type for writing segment reductions\n typename ReductionOp, ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\n typename OffsetT> ///< Signed integer type for global offsets\nstruct BlockSegReduceRegion\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // Constants\n enum\n {\n BLOCK_THREADS = BlockSegReduceRegionPolicy::BLOCK_THREADS,\n ITEMS_PER_THREAD = BlockSegReduceRegionPolicy::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD, /// Number of work items to be processed per tile\n\n USE_SMEM_SEGMENT_CACHE = BlockSegReduceRegionPolicy::USE_SMEM_SEGMENT_CACHE, ///< Whether or not to cache incoming segment offsets in shared memory before reducing each tile\n USE_SMEM_VALUE_CACHE = BlockSegReduceRegionPolicy::USE_SMEM_VALUE_CACHE, ///< Whether or not to cache incoming upcoming values in shared memory before reducing each tile\n\n SMEM_SEGMENT_CACHE_ITEMS = USE_SMEM_SEGMENT_CACHE ? TILE_ITEMS : 1,\n SMEM_VALUE_CACHE_ITEMS = USE_SMEM_VALUE_CACHE ? TILE_ITEMS : 1,\n };\n\n // Segment offset type\n typedef typename std::iterator_traits::value_type SegmentOffset;\n\n // Value type\n typedef typename std::iterator_traits::value_type Value;\n\n // Counting iterator type\n typedef CountingInputIterator CountingIterator;\n\n // Segment offsets iterator wrapper type\n typedef typename If<(IsPointer::VALUE),\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n SegmentOffsetIterator>::Type // Directly use the supplied input iterator type\n WrappedSegmentOffsetIterator;\n\n // Values iterator wrapper type\n typedef typename If<(IsPointer::VALUE),\n CacheModifiedInputIterator, // Wrap the native input pointer with CacheModifiedInputIterator\n ValueIterator>::Type // Directly use the supplied input iterator type\n WrappedValueIterator;\n\n // Tail flag type for marking segment discontinuities\n typedef int TailFlag;\n\n // Reduce-by-key data type tuple (segment-ID, value)\n typedef KeyValuePair KeyValuePair;\n\n // Index pair data type\n typedef IndexPair IndexPair;\n\n // BlockScan scan operator for reduction-by-segment\n typedef ReduceByKeyOp ReduceByKeyOp;\n\n // Stateful BlockScan prefix callback type for managing a running total while scanning consecutive tiles\n typedef RunningBlockPrefixCallbackOp<\n KeyValuePair,\n ReduceByKeyOp>\n RunningPrefixCallbackOp;\n\n // Parameterized BlockShift type for exchanging index pairs\n typedef BlockShift<\n IndexPair,\n BLOCK_THREADS>\n BlockShift;\n\n // Parameterized BlockReduce type for block-wide reduction\n typedef BlockReduce<\n Value,\n BLOCK_THREADS,\n BlockSegReduceRegionPolicy::REDUCE_ALGORITHM>\n BlockReduce;\n\n // Parameterized BlockScan type for block-wide reduce-value-by-key\n typedef BlockScan<\n KeyValuePair,\n BLOCK_THREADS,\n BlockSegReduceRegionPolicy::SCAN_ALGORITHM>\n BlockScan;\n\n // Shared memory type for this thread block\n struct _TempStorage\n {\n union\n {\n // Smem needed for BlockScan\n typename BlockScan::TempStorage scan;\n\n // Smem needed for BlockReduce\n typename BlockReduce::TempStorage reduce;\n\n struct\n {\n // Smem needed for communicating start/end indices between threads for a given work tile\n typename BlockShift::TempStorage shift;\n\n // Smem needed for caching segment end-offsets\n SegmentOffset cached_segment_end_offsets[SMEM_SEGMENT_CACHE_ITEMS + 1];\n };\n\n // Smem needed for caching values\n Value cached_values[SMEM_VALUE_CACHE_ITEMS];\n };\n\n IndexPair block_region_idx[2]; // The starting [0] and ending [1] pairs of segment and value indices for the thread block's region\n\n // The first partial reduction tuple scattered by this thread block\n KeyValuePair first_tuple;\n };\n\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n _TempStorage &temp_storage; ///< Reference to shared storage\n WrappedSegmentOffsetIterator d_segment_end_offsets; ///< A sequence of \\p num_segments segment end-offsets\n WrappedValueIterator d_values; ///< A sequence of \\p num_values data to reduce\n OutputIteratorT d_output; ///< A sequence of \\p num_segments segment totals\n CountingIterator d_value_offsets; ///< A sequence of \\p num_values value-offsets\n IndexPair *d_block_idx;\n OffsetT num_values; ///< Total number of values to reduce\n OffsetT num_segments; ///< Number of segments being reduced\n Value identity; ///< Identity value (for zero-length segments)\n ReductionOp reduction_op; ///< Reduction operator\n ReduceByKeyOp scan_op; ///< Reduce-by-key scan operator\n RunningPrefixCallbackOp prefix_op; ///< Stateful running total for block-wide prefix scan of partial reduction tuples\n\n\n //---------------------------------------------------------------------\n // Operations\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__\n BlockSegReduceRegion(\n TempStorage &temp_storage, ///< Reference to shared storage\n SegmentOffsetIterator d_segment_end_offsets, ///< A sequence of \\p num_segments segment end-offsets\n ValueIterator d_values, ///< A sequence of \\p num_values values\n OutputIteratorT d_output, ///< A sequence of \\p num_segments segment totals\n IndexPair *d_block_idx,\n OffsetT num_values, ///< Number of values to reduce\n OffsetT num_segments, ///< Number of segments being reduced\n Value identity, ///< Identity value (for zero-length segments)\n ReductionOp reduction_op) ///< Reduction operator\n :\n temp_storage(temp_storage.Alias()),\n d_segment_end_offsets(d_segment_end_offsets),\n d_values(d_values),\n d_value_offsets(0),\n d_output(d_output),\n d_block_idx(d_block_idx),\n num_values(num_values),\n num_segments(num_segments),\n identity(identity),\n reduction_op(reduction_op),\n scan_op(reduction_op),\n prefix_op(scan_op)\n {}\n\n\n /**\n * Fast-path single-segment tile reduction. Perform a\n * simple block-wide reduction and accumulate the result into\n * the running total.\n */\n __device__ __forceinline__ void SingleSegmentTile(\n IndexPair next_tile_idx,\n IndexPair block_idx)\n {\n OffsetT tile_values = next_tile_idx.b_idx - block_idx.b_idx;\n\n // Load a tile's worth of values (using identity for out-of-bounds items)\n Value values[ITEMS_PER_THREAD];\n LoadDirectStriped(threadIdx.x, d_values + block_idx.b_idx, values, tile_values, identity);\n\n // Barrier for smem reuse\n __syncthreads();\n\n // Reduce the tile of values and update the running total in thread-0\n KeyValuePair tile_aggregate;\n tile_aggregate.key = block_idx.a_idx;\n tile_aggregate.value = BlockReduce(temp_storage.reduce).Reduce(values, reduction_op);\n\n if (threadIdx.x == 0)\n {\n prefix_op.running_total = scan_op(prefix_op.running_total, tile_aggregate);\n }\n }\n\n /**\n * Fast-path empty-segment tile reduction. Write out a tile of identity\n * values to output.\n */\n __device__ __forceinline__ void EmptySegmentsTile(\n IndexPair next_tile_idx,\n IndexPair block_idx)\n {\n Value segment_reductions[ITEMS_PER_THREAD];\n\n if (threadIdx.x == 0)\n {\n // The first segment gets the running segment total\n segment_reductions[0] = prefix_op.running_total.value;\n\n // Update the running prefix\n prefix_op.running_total.value = identity;\n prefix_op.running_total.key = next_tile_idx.a_idx;\n }\n else\n {\n // Remainder of segments in this tile get identity\n segment_reductions[0] = identity;\n }\n\n // Remainder of segments in this tile get identity\n #pragma unroll\n for (int ITEM = 1; ITEM < ITEMS_PER_THREAD; ++ITEM)\n segment_reductions[ITEM] = identity;\n\n // Store reductions\n OffsetT tile_segments = next_tile_idx.a_idx - block_idx.a_idx;\n StoreDirectStriped(threadIdx.x, d_output + block_idx.a_idx, segment_reductions, tile_segments);\n }\n\n\n /**\n * Multi-segment tile reduction.\n */\n template \n __device__ __forceinline__ void MultiSegmentTile(\n IndexPair block_idx,\n IndexPair thread_idx,\n IndexPair next_thread_idx,\n IndexPair next_tile_idx)\n {\n IndexPair local_thread_idx;\n local_thread_idx.a_idx = thread_idx.a_idx - block_idx.a_idx;\n local_thread_idx.b_idx = thread_idx.b_idx - block_idx.b_idx;\n\n // Check if first segment end-offset is in range\n bool valid_segment = FULL_TILE || (thread_idx.a_idx < next_thread_idx.a_idx);\n\n // Check if first value offset is in range\n bool valid_value = FULL_TILE || (thread_idx.b_idx < next_thread_idx.b_idx);\n\n // Load first segment end-offset\n OffsetT segment_end_offset = (valid_segment) ?\n (USE_SMEM_SEGMENT_CACHE)?\n temp_storage.cached_segment_end_offsets[local_thread_idx.a_idx] :\n d_segment_end_offsets[thread_idx.a_idx] :\n -1;\n\n OffsetT segment_ids[ITEMS_PER_THREAD];\n OffsetT value_offsets[ITEMS_PER_THREAD];\n\n KeyValuePair first_partial;\n first_partial.key = thread_idx.a_idx;\n first_partial.value = identity;\n\n // Get segment IDs and gather-offsets for values\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n segment_ids[ITEM] = -1;\n value_offsets[ITEM] = -1;\n\n // Whether or not we slide (a) right along the segment path or (b) down the value path\n if (valid_segment && (!valid_value || (segment_end_offset <= thread_idx.b_idx)))\n {\n // Consume this segment index\n segment_ids[ITEM] = thread_idx.a_idx;\n thread_idx.a_idx++;\n local_thread_idx.a_idx++;\n\n valid_segment = FULL_TILE || (thread_idx.a_idx < next_thread_idx.a_idx);\n\n // Read next segment end-offset (if valid)\n if (valid_segment)\n {\n if (USE_SMEM_SEGMENT_CACHE)\n segment_end_offset = temp_storage.cached_segment_end_offsets[local_thread_idx.a_idx];\n else\n segment_end_offset = d_segment_end_offsets[thread_idx.a_idx];\n }\n }\n else if (valid_value)\n {\n // Consume this value index\n value_offsets[ITEM] = thread_idx.b_idx;\n thread_idx.b_idx++;\n local_thread_idx.b_idx++;\n\n valid_value = FULL_TILE || (thread_idx.b_idx < next_thread_idx.b_idx);\n }\n }\n\n // Load values\n Value values[ITEMS_PER_THREAD];\n\n if (USE_SMEM_VALUE_CACHE)\n {\n // Barrier for smem reuse\n __syncthreads();\n\n OffsetT tile_values = next_tile_idx.b_idx - block_idx.b_idx;\n\n // Load a tile's worth of values (using identity for out-of-bounds items)\n LoadDirectStriped(threadIdx.x, d_values + block_idx.b_idx, values, tile_values, identity);\n\n // Store to shared\n StoreDirectStriped(threadIdx.x, temp_storage.cached_values, values, tile_values);\n\n // Barrier for smem reuse\n __syncthreads();\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n values[ITEM] = (value_offsets[ITEM] == -1) ?\n identity :\n temp_storage.cached_values[value_offsets[ITEM] - block_idx.b_idx];\n }\n }\n else\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n values[ITEM] = (value_offsets[ITEM] == -1) ?\n identity :\n d_values[value_offsets[ITEM]];\n }\n }\n\n // Reduce within thread segments\n KeyValuePair running_total = first_partial;\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n if (segment_ids[ITEM] != -1)\n {\n // Consume this segment index\n d_output[segment_ids[ITEM]] = running_total.value;\n\n// _CubLog(\"Updating segment %d with value %lld\\n\", segment_ids[ITEM], running_total.value)\n\n if (first_partial.key == segment_ids[ITEM])\n first_partial.value = running_total.value;\n\n running_total.key = segment_ids[ITEM];\n running_total.value = identity;\n }\n\n running_total.value = reduction_op(running_total.value, values[ITEM]);\n }\n/*\n\n // Barrier for smem reuse\n __syncthreads();\n\n // Use prefix scan to reduce values by segment-id. The segment-reductions end up in items flagged as segment-tails.\n KeyValuePair block_aggregate;\n BlockScan(temp_storage.scan).InclusiveScan(\n pairs, // Scan input\n pairs, // Scan output\n scan_op, // Scan operator\n block_aggregate, // Block-wide total (unused)\n prefix_op); // Prefix operator for seeding the block-wide scan with the running total\n*/\n\n/*\n // Check if first segment end-offset is in range\n bool valid_segment = (thread_idx.a_idx < next_thread_idx.a_idx);\n\n // Check if first value offset is in range\n bool valid_value = (thread_idx.b_idx < next_thread_idx.b_idx);\n\n // Load first segment end-offset\n OffsetT segment_end_offset = (valid_segment) ?\n d_segment_end_offsets[thread_idx.a_idx] :\n num_values; // Out of range (the last segment end-offset is one-past the last value offset)\n\n // Load first value offset\n OffsetT value_offset = (valid_value) ?\n d_value_offsets[thread_idx.b_idx] :\n num_values; // Out of range (one-past the last value offset)\n\n // Assemble segment-demarcating tail flags and partial reduction tuples\n TailFlag tail_flags[ITEMS_PER_THREAD];\n KeyValuePair partial_reductions[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n // Default tuple and flag values\n partial_reductions[ITEM].key = thread_idx.a_idx;\n partial_reductions[ITEM].value = identity;\n tail_flags[ITEM] = 0;\n\n // Whether or not we slide (a) right along the segment path or (b) down the value path\n if (valid_segment && (!valid_value || (segment_end_offset <= value_offset)))\n {\n // Consume this segment index\n\n // Set tail flag noting the end of the segment\n tail_flags[ITEM] = 1;\n\n // Increment segment index\n thread_idx.a_idx++;\n\n // Read next segment end-offset (if valid)\n if ((valid_segment = (thread_idx.a_idx < next_thread_idx.a_idx)))\n segment_end_offset = d_segment_end_offsets[thread_idx.a_idx];\n }\n else if (valid_value)\n {\n // Consume this value index\n\n // Update the tuple's value with the value at this index.\n partial_reductions[ITEM].value = d_values[value_offset];\n\n // Increment value index\n thread_idx.b_idx++;\n\n // Read next value offset (if valid)\n if ((valid_value = (thread_idx.b_idx < next_thread_idx.b_idx)))\n value_offset = d_value_offsets[thread_idx.b_idx];\n }\n }\n\n // Use prefix scan to reduce values by segment-id. The segment-reductions end up in items flagged as segment-tails.\n KeyValuePair block_aggregate;\n BlockScan(temp_storage.scan).InclusiveScan(\n partial_reductions, // Scan input\n partial_reductions, // Scan output\n scan_op, // Scan operator\n block_aggregate, // Block-wide total (unused)\n prefix_op); // Prefix operator for seeding the block-wide scan with the running total\n\n // The first segment index for this region (hoist?)\n OffsetT first_segment_idx = temp_storage.block_idx.a_idx[0];\n\n // Scatter an accumulated reduction if it is the head of a valid segment\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (tail_flags[ITEM])\n {\n OffsetT segment_idx = partial_reductions[ITEM].key;\n Value value = partial_reductions[ITEM].value;\n\n // Write value reduction to corresponding segment id\n d_output[segment_idx] = value;\n\n // Save off the first value product that this thread block will scatter\n if (segment_idx == first_segment_idx)\n {\n temp_storage.first_tuple.value = value;\n }\n }\n }\n*/\n }\n\n\n\n /**\n * Have the thread block process the specified region of the MergePath decision path\n */\n __device__ __forceinline__ void ProcessRegion(\n OffsetT block_diagonal,\n OffsetT next_block_diagonal,\n KeyValuePair &first_tuple, // [Out] Valid in thread-0\n KeyValuePair &last_tuple) // [Out] Valid in thread-0\n {\n // Thread block initialization\n if (threadIdx.x < 2)\n {\n // Retrieve block starting and ending indices\n IndexPair block_idx = {0, 0};\n if (gridDim.x > 1)\n {\n block_idx = d_block_idx[blockIdx.x + threadIdx.x];\n }\n else if (threadIdx.x > 0)\n {\n block_idx.a_idx = num_segments;\n block_idx.b_idx = num_values;\n }\n\n // Share block starting and ending indices\n temp_storage.block_region_idx[threadIdx.x] = block_idx;\n\n // Initialize the block's running prefix\n if (threadIdx.x == 0)\n {\n prefix_op.running_total.key = block_idx.a_idx;\n prefix_op.running_total.value = identity;\n\n // Initialize the \"first scattered partial reduction tuple\" to the prefix tuple (in case we don't actually scatter one)\n temp_storage.first_tuple = prefix_op.running_total;\n }\n }\n\n // Ensure coherence of region indices\n __syncthreads();\n\n // Read block's starting indices\n IndexPair block_idx = temp_storage.block_region_idx[0];\n\n // Have the thread block iterate over the region\n #pragma unroll 1\n while (block_diagonal < next_block_diagonal)\n {\n // Read block's ending indices (hoist?)\n IndexPair next_block_idx = temp_storage.block_region_idx[1];\n\n // Clamp the per-thread search range to within one work-tile of block's current indices\n IndexPair next_tile_idx;\n next_tile_idx.a_idx = CUB_MIN(next_block_idx.a_idx, block_idx.a_idx + TILE_ITEMS);\n next_tile_idx.b_idx = CUB_MIN(next_block_idx.b_idx, block_idx.b_idx + TILE_ITEMS);\n\n // Have each thread search for the end-indices of its subranges within the segment and value inputs\n IndexPair next_thread_idx;\n if (USE_SMEM_SEGMENT_CACHE)\n {\n // Search in smem cache\n OffsetT num_segments = next_tile_idx.a_idx - block_idx.a_idx;\n\n // Load global\n SegmentOffset segment_offsets[ITEMS_PER_THREAD];\n LoadDirectStriped(threadIdx.x, d_segment_end_offsets + block_idx.a_idx, segment_offsets, num_segments, num_values);\n\n // Store to shared\n StoreDirectStriped(threadIdx.x, temp_storage.cached_segment_end_offsets, segment_offsets);\n\n __syncthreads();\n\n OffsetT next_thread_diagonal = block_diagonal + ((threadIdx.x + 1) * ITEMS_PER_THREAD);\n\n MergePathSearch(\n next_thread_diagonal, // Next thread diagonal\n temp_storage.cached_segment_end_offsets - block_idx.a_idx, // A (segment end-offsets)\n d_value_offsets, // B (value offsets)\n block_idx, // Start indices into A and B\n next_tile_idx, // End indices into A and B\n next_thread_idx); // [out] diagonal intersection indices into A and B\n }\n else\n {\n // Search in global\n\n OffsetT next_thread_diagonal = block_diagonal + ((threadIdx.x + 1) * ITEMS_PER_THREAD);\n\n MergePathSearch(\n next_thread_diagonal, // Next thread diagonal\n d_segment_end_offsets, // A (segment end-offsets)\n d_value_offsets, // B (value offsets)\n block_idx, // Start indices into A and B\n next_tile_idx, // End indices into A and B\n next_thread_idx); // [out] diagonal intersection indices into A and B\n }\n\n // Share thread end-indices to get thread begin-indices and tile end-indices\n IndexPair thread_idx;\n\n BlockShift(temp_storage.shift).Up(\n next_thread_idx, // Input item\n thread_idx, // [out] Output item\n block_idx, // Prefix item to be provided to thread0\n next_tile_idx); // [out] Suffix item shifted out by the threadBLOCK_THREADS-1 to be provided to all threads\n\n// if (block_idx.a_idx == next_tile_idx.a_idx)\n// {\n// // There are no segment end-offsets in this tile. Perform a\n// // simple block-wide reduction and accumulate the result into\n// // the running total.\n// SingleSegmentTile(next_tile_idx, block_idx);\n// }\n// else if (block_idx.b_idx == next_tile_idx.b_idx)\n// {\n// // There are no values in this tile (only empty segments).\n// EmptySegmentsTile(next_tile_idx.a_idx, block_idx.a_idx);\n// }\n// else\n if ((next_tile_idx.a_idx < num_segments) && (next_tile_idx.b_idx < num_values))\n {\n // Merge the tile's segment and value indices (full tile)\n MultiSegmentTile(block_idx, thread_idx, next_thread_idx, next_tile_idx);\n }\n else\n {\n // Merge the tile's segment and value indices (partially full tile)\n MultiSegmentTile(block_idx, thread_idx, next_thread_idx, next_tile_idx);\n }\n\n // Advance the block's indices in preparation for the next tile\n block_idx = next_tile_idx;\n\n // Advance to the next region in the decision path\n block_diagonal += TILE_ITEMS;\n\n // Barrier for smem reuse\n __syncthreads();\n }\n\n // Get first and last tuples for the region\n if (threadIdx.x == 0)\n {\n first_tuple = temp_storage.first_tuple;\n last_tuple = prefix_op.running_total;\n }\n\n }\n\n\n};\n\n\n\n\n\n\n\n\n/******************************************************************************\n * Tuning policy types\n ******************************************************************************/\n\n/**\n * Parameterizable tuning policy type for BlockSegReduceRegionByKey\n */\ntemplate <\n int _BLOCK_THREADS, ///< Threads per thread block\n int _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n BlockLoadAlgorithm _LOAD_ALGORITHM, ///< The BlockLoad algorithm to use\n bool _LOAD_WARP_TIME_SLICING, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage)\n CacheLoadModifier _LOAD_MODIFIER, ///< Cache load modifier for reading input elements\n BlockScanAlgorithm _SCAN_ALGORITHM> ///< The BlockScan algorithm to use\nstruct BlockSegReduceRegionByKeyPolicy\n{\n enum\n {\n BLOCK_THREADS = _BLOCK_THREADS, ///< Threads per thread block\n ITEMS_PER_THREAD = _ITEMS_PER_THREAD, ///< Items per thread (per tile of input)\n LOAD_WARP_TIME_SLICING = _LOAD_WARP_TIME_SLICING, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage) };\n };\n\n static const BlockLoadAlgorithm LOAD_ALGORITHM = _LOAD_ALGORITHM; ///< The BlockLoad algorithm to use\n static const CacheLoadModifier LOAD_MODIFIER = _LOAD_MODIFIER; ///< Cache load modifier for reading input elements\n static const BlockScanAlgorithm SCAN_ALGORITHM = _SCAN_ALGORITHM; ///< The BlockScan algorithm to use\n};\n\n\n/******************************************************************************\n * Persistent thread block types\n ******************************************************************************/\n\n/**\n * \\brief BlockSegReduceRegionByKey implements a stateful abstraction of CUDA thread blocks for participating in device-wide reduce-value-by-key.\n */\ntemplate <\n typename BlockSegReduceRegionByKeyPolicy, ///< Parameterized BlockSegReduceRegionByKeyPolicy tuning policy\n typename InputIteratorT, ///< Random-access iterator referencing key-value input tuples\n typename OutputIteratorT, ///< Random-access iterator referencing segment output totals\n typename ReductionOp> ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\nstruct BlockSegReduceRegionByKey\n{\n //---------------------------------------------------------------------\n // Types and constants\n //---------------------------------------------------------------------\n\n // Constants\n enum\n {\n BLOCK_THREADS = BlockSegReduceRegionByKeyPolicy::BLOCK_THREADS,\n ITEMS_PER_THREAD = BlockSegReduceRegionByKeyPolicy::ITEMS_PER_THREAD,\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n // KeyValuePair input type\n typedef typename std::iterator_traits::value_type KeyValuePair;\n\n // Signed integer type for global offsets\n typedef typename KeyValuePair::Key OffsetT;\n\n // Value type\n typedef typename KeyValuePair::Value Value;\n\n // Head flag type\n typedef int HeadFlag;\n\n // Input iterator wrapper type for loading KeyValuePair elements through cache\n typedef CacheModifiedInputIterator<\n BlockSegReduceRegionByKeyPolicy::LOAD_MODIFIER,\n KeyValuePair,\n OffsetT>\n WrappedInputIteratorT;\n\n // Parameterized BlockLoad type\n typedef BlockLoad<\n WrappedInputIteratorT,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n BlockSegReduceRegionByKeyPolicy::LOAD_ALGORITHM,\n BlockSegReduceRegionByKeyPolicy::LOAD_WARP_TIME_SLICING>\n BlockLoad;\n\n // BlockScan scan operator for reduction-by-segment\n typedef ReduceByKeyOp ReduceByKeyOp;\n\n // Stateful BlockScan prefix callback type for managing a running total while scanning consecutive tiles\n typedef RunningBlockPrefixCallbackOp<\n KeyValuePair,\n ReduceByKeyOp>\n RunningPrefixCallbackOp;\n\n // Parameterized BlockScan type for block-wide reduce-value-by-key\n typedef BlockScan<\n KeyValuePair,\n BLOCK_THREADS,\n BlockSegReduceRegionByKeyPolicy::SCAN_ALGORITHM>\n BlockScan;\n\n // Parameterized BlockDiscontinuity type for identifying key discontinuities\n typedef BlockDiscontinuity<\n OffsetT,\n BLOCK_THREADS>\n BlockDiscontinuity;\n\n // Operator for detecting discontinuities in a list of segment identifiers.\n struct NewSegmentOp\n {\n /// Returns true if row_b is the start of a new row\n __device__ __forceinline__ bool operator()(const OffsetT& b, const OffsetT& a)\n {\n return (a != b);\n }\n };\n\n // Shared memory type for this thread block\n struct _TempStorage\n {\n union\n {\n typename BlockLoad::TempStorage load; // Smem needed for tile loading\n struct {\n typename BlockScan::TempStorage scan; // Smem needed for reduce-value-by-segment scan\n typename BlockDiscontinuity::TempStorage discontinuity; // Smem needed for head-flagging\n };\n };\n };\n\n // Alias wrapper allowing storage to be unioned\n struct TempStorage : Uninitialized<_TempStorage> {};\n\n\n //---------------------------------------------------------------------\n // Thread fields\n //---------------------------------------------------------------------\n\n _TempStorage &temp_storage; ///< Reference to shared storage\n WrappedInputIteratorT d_tuple_partials; ///< A sequence of partial reduction tuples to scan\n OutputIteratorT d_output; ///< A sequence of segment totals\n Value identity; ///< Identity value (for zero-length segments)\n ReduceByKeyOp scan_op; ///< Reduce-by-key scan operator\n RunningPrefixCallbackOp prefix_op; ///< Stateful running total for block-wide prefix scan of partial reduction tuples\n\n\n //---------------------------------------------------------------------\n // Operations\n //---------------------------------------------------------------------\n\n /**\n * Constructor\n */\n __device__ __forceinline__\n BlockSegReduceRegionByKey(\n TempStorage &temp_storage, ///< Reference to shared storage\n InputIteratorT d_tuple_partials, ///< A sequence of partial reduction tuples to scan\n OutputIteratorT d_output, ///< A sequence of segment totals\n Value identity, ///< Identity value (for zero-length segments)\n ReductionOp reduction_op) ///< Reduction operator\n :\n temp_storage(temp_storage.Alias()),\n d_tuple_partials(d_tuple_partials),\n d_output(d_output),\n identity(identity),\n scan_op(reduction_op),\n prefix_op(scan_op)\n {}\n\n\n\n /**\n * Processes a reduce-value-by-key input tile, outputting reductions for each segment\n */\n template \n __device__ __forceinline__\n void ProcessTile(\n OffsetT block_offset,\n OffsetT first_segment_idx,\n OffsetT last_segment_idx,\n int guarded_items = TILE_ITEMS)\n {\n KeyValuePair partial_reductions[ITEMS_PER_THREAD];\n OffsetT segment_ids[ITEMS_PER_THREAD];\n HeadFlag head_flags[ITEMS_PER_THREAD];\n\n // Load a tile of block partials from previous kernel\n if (FULL_TILE)\n {\n // Full tile\n BlockLoad(temp_storage.load).Load(d_tuple_partials + block_offset, partial_reductions);\n }\n else\n {\n KeyValuePair oob_default;\n oob_default.key = last_segment_idx; // The last segment ID to be reduced\n oob_default.value = identity;\n\n // Partially-full tile\n BlockLoad(temp_storage.load).Load(d_tuple_partials + block_offset, partial_reductions, guarded_items, oob_default);\n }\n\n // Barrier for shared memory reuse\n __syncthreads();\n\n // Copy the segment IDs for head-flagging\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n segment_ids[ITEM] = partial_reductions[ITEM].key;\n }\n\n // FlagT segment heads by looking for discontinuities\n BlockDiscontinuity(temp_storage.discontinuity).FlagHeads(\n head_flags, // [out] Head flags\n segment_ids, // Segment ids\n NewSegmentOp(), // Functor for detecting start of new rows\n prefix_op.running_total.key); // Last segment ID from previous tile to compare with first segment ID in this tile\n\n // Reduce-value-by-segment across partial_reductions using exclusive prefix scan\n KeyValuePair block_aggregate;\n BlockScan(temp_storage.scan).ExclusiveScan(\n partial_reductions, // Scan input\n partial_reductions, // Scan output\n scan_op, // Scan operator\n block_aggregate, // Block-wide total (unused)\n prefix_op); // Prefix operator for seeding the block-wide scan with the running total\n\n // Scatter an accumulated reduction if it is the head of a valid segment\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ITEM++)\n {\n if (head_flags[ITEM])\n {\n d_output[partial_reductions[ITEM].key] = partial_reductions[ITEM].value;\n }\n }\n }\n\n\n /**\n * Iterate over input tiles belonging to this thread block\n */\n __device__ __forceinline__\n void ProcessRegion(\n OffsetT block_offset,\n OffsetT block_end,\n OffsetT first_segment_idx,\n OffsetT last_segment_idx)\n {\n if (threadIdx.x == 0)\n {\n // Initialize running prefix to the first segment index paired with identity\n prefix_op.running_total.key = first_segment_idx;\n prefix_op.running_total.value = identity;\n }\n\n // Process full tiles\n while (block_offset + TILE_ITEMS <= block_end)\n {\n ProcessTile(block_offset, first_segment_idx, last_segment_idx);\n __syncthreads();\n\n block_offset += TILE_ITEMS;\n }\n\n // Process final value tile (if present)\n int guarded_items = block_end - block_offset;\n if (guarded_items)\n {\n ProcessTile(block_offset, first_segment_idx, last_segment_idx, guarded_items);\n }\n }\n};\n\n\n\n/******************************************************************************\n * Kernel entrypoints\n ******************************************************************************/\n\n/**\n * Segmented reduce region kernel entry point (multi-block).\n */\n\ntemplate <\n typename SegmentOffsetIterator, ///< Random-access input iterator type for reading segment end-offsets\n typename OffsetT> ///< Signed integer type for global offsets\n__global__ void SegReducePartitionKernel(\n SegmentOffsetIterator d_segment_end_offsets, ///< [in] A sequence of \\p num_segments segment end-offsets\n IndexPair *d_block_idx,\n int num_partition_samples,\n OffsetT num_values, ///< [in] Number of values to reduce\n OffsetT num_segments, ///< [in] Number of segments being reduced\n GridEvenShare even_share) ///< [in] Even-share descriptor for mapping an equal number of tiles onto each thread block\n{\n // Segment offset type\n typedef typename std::iterator_traits::value_type SegmentOffset;\n\n // Counting iterator type\n typedef CountingInputIterator CountingIterator;\n\n // Cache-modified iterator for segment end-offsets\n CacheModifiedInputIterator d_wrapped_segment_end_offsets(d_segment_end_offsets);\n\n // Counting iterator for value offsets\n CountingIterator d_value_offsets(0);\n\n // Initialize even-share to tell us where to start and stop our tile-processing\n int partition_id = (blockDim.x * blockIdx.x) + threadIdx.x;\n even_share.Init(partition_id);\n\n // Search for block starting and ending indices\n IndexPair start_idx = {0, 0};\n IndexPair end_idx = {num_segments, num_values};\n IndexPair block_idx;\n\n MergePathSearch(\n even_share.block_offset, // Next thread diagonal\n d_wrapped_segment_end_offsets, // A (segment end-offsets)\n d_value_offsets, // B (value offsets)\n start_idx, // Start indices into A and B\n end_idx, // End indices into A and B\n block_idx); // [out] diagonal intersection indices into A and B\n\n // Write output\n if (partition_id < num_partition_samples)\n {\n d_block_idx[partition_id] = block_idx;\n }\n}\n\n\n/**\n * Segmented reduce region kernel entry point (multi-block).\n */\ntemplate <\n typename BlockSegReduceRegionPolicy, ///< Parameterized BlockSegReduceRegionPolicy tuning policy\n typename SegmentOffsetIterator, ///< Random-access input iterator type for reading segment end-offsets\n typename ValueIterator, ///< Random-access input iterator type for reading values\n typename OutputIteratorT, ///< Random-access output iterator type for writing segment reductions\n typename ReductionOp, ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\n typename OffsetT, ///< Signed integer type for global offsets\n typename Value> ///< Value type\n__launch_bounds__ (BlockSegReduceRegionPolicy::BLOCK_THREADS)\n__global__ void SegReduceRegionKernel(\n SegmentOffsetIterator d_segment_end_offsets, ///< [in] A sequence of \\p num_segments segment end-offsets\n ValueIterator d_values, ///< [in] A sequence of \\p num_values values\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n KeyValuePair *d_tuple_partials, ///< [out] A sequence of (gridDim.x * 2) partial reduction tuples\n IndexPair *d_block_idx,\n OffsetT num_values, ///< [in] Number of values to reduce\n OffsetT num_segments, ///< [in] Number of segments being reduced\n Value identity, ///< [in] Identity value (for zero-length segments)\n ReductionOp reduction_op, ///< [in] Reduction operator\n GridEvenShare even_share) ///< [in] Even-share descriptor for mapping an equal number of tiles onto each thread block\n{\n typedef KeyValuePair KeyValuePair;\n\n // Specialize thread block abstraction type for reducing a range of segmented values\n typedef BlockSegReduceRegion<\n BlockSegReduceRegionPolicy,\n SegmentOffsetIterator,\n ValueIterator,\n OutputIteratorT,\n ReductionOp,\n OffsetT>\n BlockSegReduceRegion;\n\n // Shared memory allocation\n __shared__ typename BlockSegReduceRegion::TempStorage temp_storage;\n\n // Initialize thread block even-share to tell us where to start and stop our tile-processing\n even_share.BlockInit();\n\n // Construct persistent thread block\n BlockSegReduceRegion thread_block(\n temp_storage,\n d_segment_end_offsets,\n d_values,\n d_output,\n d_block_idx,\n num_values,\n num_segments,\n identity,\n reduction_op);\n\n // First and last partial reduction tuples within the range (valid in thread-0)\n KeyValuePair first_tuple, last_tuple;\n\n // Consume block's region of work\n thread_block.ProcessRegion(\n even_share.block_offset,\n even_share.block_end,\n first_tuple,\n last_tuple);\n\n if (threadIdx.x == 0)\n {\n if (gridDim.x > 1)\n {\n // Special case where the first segment written and the carry-out are for the same segment\n if (first_tuple.key == last_tuple.key)\n {\n first_tuple.value = identity;\n }\n\n // Write the first and last partial products from this thread block so\n // that they can be subsequently \"fixed up\" in the next kernel.\n d_tuple_partials[blockIdx.x * 2] = first_tuple;\n d_tuple_partials[(blockIdx.x * 2) + 1] = last_tuple;\n }\n }\n\n}\n\n\n/**\n * Segmented reduce region kernel entry point (single-block).\n */\ntemplate <\n typename BlockSegReduceRegionByKeyPolicy, ///< Parameterized BlockSegReduceRegionByKeyPolicy tuning policy\n typename InputIteratorT, ///< Random-access iterator referencing key-value input tuples\n typename OutputIteratorT, ///< Random-access iterator referencing segment output totals\n typename ReductionOp, ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\n typename OffsetT, ///< Signed integer type for global offsets\n typename Value> ///< Value type\n__launch_bounds__ (BlockSegReduceRegionByKeyPolicy::BLOCK_THREADS, 1)\n__global__ void SegReduceRegionByKeyKernel(\n InputIteratorT d_tuple_partials, ///< [in] A sequence of partial reduction tuples\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n OffsetT num_segments, ///< [in] Number of segments in the \\p d_output sequence\n int num_tuple_partials, ///< [in] Number of partial reduction tuples being reduced\n Value identity, ///< [in] Identity value (for zero-length segments)\n ReductionOp reduction_op) ///< [in] Reduction operator\n{\n // Specialize thread block abstraction type for reducing a range of values by key\n typedef BlockSegReduceRegionByKey<\n BlockSegReduceRegionByKeyPolicy,\n InputIteratorT,\n OutputIteratorT,\n ReductionOp>\n BlockSegReduceRegionByKey;\n\n // Shared memory allocation\n __shared__ typename BlockSegReduceRegionByKey::TempStorage temp_storage;\n\n // Construct persistent thread block\n BlockSegReduceRegionByKey thread_block(\n temp_storage,\n d_tuple_partials,\n d_output,\n identity,\n reduction_op);\n\n // Process input tiles\n thread_block.ProcessRegion(\n 0, // Region start\n num_tuple_partials, // Region end\n 0, // First segment ID\n num_segments); // Last segment ID (one-past)\n}\n\n\n\n\n/******************************************************************************\n * Dispatch\n ******************************************************************************/\n\n/**\n * Utility class for dispatching the appropriately-tuned kernels for DeviceReduce\n */\ntemplate <\n typename ValueIterator, ///< Random-access input iterator type for reading values\n typename SegmentOffsetIterator, ///< Random-access input iterator type for reading segment end-offsets\n typename OutputIteratorT, ///< Random-access output iterator type for writing segment reductions\n typename ReductionOp, ///< Binary reduction operator type having member T operator()(const T &a, const T &b)\n typename OffsetT> ///< Signed integer type for global offsets\nstruct DeviceSegReduceDispatch\n{\n // Value type\n typedef typename std::iterator_traits::value_type Value;\n\n // Reduce-by-key data type tuple (segment-ID, value)\n typedef KeyValuePair KeyValuePair;\n\n // Index pair data type\n typedef IndexPairIndexPair;\n\n\n /******************************************************************************\n * Tuning policies\n ******************************************************************************/\n\n /// SM35\n struct Policy350\n {\n // ReduceRegionPolicy\n typedef BlockSegReduceRegionPolicy<\n 128, ///< Threads per thread block\n 6, ///< Items per thread (per tile of input)\n true, ///< Whether or not to cache incoming segment offsets in shared memory before reducing each tile\n false, ///< Whether or not to cache incoming values in shared memory before reducing each tile\n LOAD_DEFAULT, ///< Cache load modifier for reading segment offsets\n LOAD_LDG, ///< Cache load modifier for reading values\n BLOCK_REDUCE_RAKING, ///< The BlockReduce algorithm to use\n BLOCK_SCAN_WARP_SCANS> ///< The BlockScan algorithm to use\n SegReduceRegionPolicy;\n\n // ReduceRegionByKeyPolicy\n typedef BlockSegReduceRegionByKeyPolicy<\n 256, ///< Threads per thread block\n 9, ///< Items per thread (per tile of input)\n BLOCK_LOAD_DIRECT, ///< The BlockLoad algorithm to use\n false, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage)\n LOAD_LDG, ///< Cache load modifier for reading input elements\n BLOCK_SCAN_WARP_SCANS> ///< The BlockScan algorithm to use\n SegReduceRegionByKeyPolicy;\n };\n\n\n /// SM10\n struct Policy100\n {\n // ReduceRegionPolicy\n typedef BlockSegReduceRegionPolicy<\n 128, ///< Threads per thread block\n 3, ///< Items per thread (per tile of input)\n false, ///< Whether or not to cache incoming segment offsets in shared memory before reducing each tile\n false, ///< Whether or not to cache incoming values in shared memory before reducing each tile\n LOAD_DEFAULT, ///< Cache load modifier for reading segment offsets\n LOAD_DEFAULT, ///< Cache load modifier for reading values\n BLOCK_REDUCE_RAKING, ///< The BlockReduce algorithm to use\n BLOCK_SCAN_RAKING> ///< The BlockScan algorithm to use\n SegReduceRegionPolicy;\n\n // ReduceRegionByKeyPolicy\n typedef BlockSegReduceRegionByKeyPolicy<\n 128, ///< Threads per thread block\n 3, ///< Items per thread (per tile of input)\n BLOCK_LOAD_WARP_TRANSPOSE, ///< The BlockLoad algorithm to use\n false, ///< Whether or not only one warp's worth of shared memory should be allocated and time-sliced among block-warps during any load-related data transpositions (versus each warp having its own storage)\n LOAD_DEFAULT, ///< Cache load modifier for reading input elements\n BLOCK_SCAN_WARP_SCANS> ///< The BlockScan algorithm to use\n SegReduceRegionByKeyPolicy;\n };\n\n\n /******************************************************************************\n * Tuning policies of current PTX compiler pass\n ******************************************************************************/\n\n#if (CUB_PTX_ARCH >= 350)\n typedef Policy350 PtxPolicy;\n/*\n#elif (CUB_PTX_ARCH >= 300)\n typedef Policy300 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 200)\n typedef Policy200 PtxPolicy;\n\n#elif (CUB_PTX_ARCH >= 130)\n typedef Policy130 PtxPolicy;\n*/\n#else\n typedef Policy100 PtxPolicy;\n\n#endif\n\n // \"Opaque\" policies (whose parameterizations aren't reflected in the type signature)\n struct PtxSegReduceRegionPolicy : PtxPolicy::SegReduceRegionPolicy {};\n struct PtxSegReduceRegionByKeyPolicy : PtxPolicy::SegReduceRegionByKeyPolicy {};\n\n\n /******************************************************************************\n * Utilities\n ******************************************************************************/\n\n /**\n * Initialize kernel dispatch configurations with the policies corresponding to the PTX assembly we will use\n */\n template <\n typename SegReduceKernelConfig,\n typename SegReduceByKeyKernelConfig>\n __host__ __device__ __forceinline__\n static void InitConfigs(\n int ptx_version,\n SegReduceKernelConfig &seg_reduce_region_config,\n SegReduceByKeyKernelConfig &seg_reduce_region_by_key_config)\n {\n #if (CUB_PTX_ARCH > 0)\n\n // We're on the device, so initialize the kernel dispatch configurations with the current PTX policy\n seg_reduce_region_config.Init();\n seg_reduce_region_by_key_config.Init();\n\n #else\n\n // We're on the host, so lookup and initialize the kernel dispatch configurations with the policies that match the device's PTX version\n if (ptx_version >= 350)\n {\n seg_reduce_region_config.template Init();\n seg_reduce_region_by_key_config.template Init();\n }\n/*\n else if (ptx_version >= 300)\n {\n seg_reduce_region_config.template Init();\n seg_reduce_region_by_key_config.template Init();\n }\n else if (ptx_version >= 200)\n {\n seg_reduce_region_config.template Init();\n seg_reduce_region_by_key_config.template Init();\n }\n else if (ptx_version >= 130)\n {\n seg_reduce_region_config.template Init();\n seg_reduce_region_by_key_config.template Init();\n }\n*/\n else\n {\n seg_reduce_region_config.template Init();\n seg_reduce_region_by_key_config.template Init();\n }\n\n #endif\n }\n\n\n /**\n * SegReduceRegionKernel kernel dispatch configuration\n */\n struct SegReduceKernelConfig\n {\n int block_threads;\n int items_per_thread;\n bool use_smem_segment_cache;\n bool use_smem_value_cache;\n CacheLoadModifier load_modifier_segments;\n CacheLoadModifier load_modifier_values;\n BlockReduceAlgorithm reduce_algorithm;\n BlockScanAlgorithm scan_algorithm;\n\n template \n __host__ __device__ __forceinline__\n void Init()\n {\n block_threads = SegReduceRegionPolicy::BLOCK_THREADS;\n items_per_thread = SegReduceRegionPolicy::ITEMS_PER_THREAD;\n use_smem_segment_cache = SegReduceRegionPolicy::USE_SMEM_SEGMENT_CACHE;\n use_smem_value_cache = SegReduceRegionPolicy::USE_SMEM_VALUE_CACHE;\n load_modifier_segments = SegReduceRegionPolicy::LOAD_MODIFIER_SEGMENTS;\n load_modifier_values = SegReduceRegionPolicy::LOAD_MODIFIER_VALUES;\n reduce_algorithm = SegReduceRegionPolicy::REDUCE_ALGORITHM;\n scan_algorithm = SegReduceRegionPolicy::SCAN_ALGORITHM;\n }\n };\n\n /**\n * SegReduceRegionByKeyKernel kernel dispatch configuration\n */\n struct SegReduceByKeyKernelConfig\n {\n int block_threads;\n int items_per_thread;\n BlockLoadAlgorithm load_algorithm;\n bool load_warp_time_slicing;\n CacheLoadModifier load_modifier;\n BlockScanAlgorithm scan_algorithm;\n\n template \n __host__ __device__ __forceinline__\n void Init()\n {\n block_threads = SegReduceRegionByKeyPolicy::BLOCK_THREADS;\n items_per_thread = SegReduceRegionByKeyPolicy::ITEMS_PER_THREAD;\n load_algorithm = SegReduceRegionByKeyPolicy::LOAD_ALGORITHM;\n load_warp_time_slicing = SegReduceRegionByKeyPolicy::LOAD_WARP_TIME_SLICING;\n load_modifier = SegReduceRegionByKeyPolicy::LOAD_MODIFIER;\n scan_algorithm = SegReduceRegionByKeyPolicy::SCAN_ALGORITHM;\n }\n };\n\n\n /******************************************************************************\n * Dispatch entrypoints\n ******************************************************************************/\n\n /**\n * Internal dispatch routine for computing a device-wide segmented reduction.\n */\n template <\n typename SegReducePartitionKernelPtr,\n typename SegReduceRegionKernelPtr, ///< Function type of cub::SegReduceRegionKernel\n typename SegReduceRegionByKeyKernelPtr> ///< Function type of cub::SegReduceRegionByKeyKernel\n __host__ __device__ __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device allocation of temporary storage. When NULL, the required allocation size is returned in \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation.\n ValueIterator d_values, ///< [in] A sequence of \\p num_values data to reduce\n SegmentOffsetIterator d_segment_offsets, ///< [in] A sequence of (\\p num_segments + 1) segment offsets\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n OffsetT num_values, ///< [in] Total number of values to reduce\n OffsetT num_segments, ///< [in] Number of segments being reduced\n Value identity, ///< [in] Identity value (for zero-length segments)\n ReductionOp reduction_op, ///< [in] Reduction operator\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous, ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n int sm_version, ///< [in] SM version of target device to use when computing SM occupancy\n SegReducePartitionKernelPtr seg_reduce_partition_kernel, ///< [in] Kernel function pointer to parameterization of cub::SegReduceRegionKernel\n SegReduceRegionKernelPtr seg_reduce_region_kernel, ///< [in] Kernel function pointer to parameterization of cub::SegReduceRegionKernel\n SegReduceRegionByKeyKernelPtr seg_reduce_region_by_key_kernel, ///< [in] Kernel function pointer to parameterization of cub::SegReduceRegionByKeyKernel\n SegReduceKernelConfig &seg_reduce_region_config, ///< [in] Dispatch parameters that match the policy that \\p seg_reduce_region_kernel was compiled for\n SegReduceByKeyKernelConfig &seg_reduce_region_by_key_config) ///< [in] Dispatch parameters that match the policy that \\p seg_reduce_region_by_key_kernel was compiled for\n {\n#ifndef CUB_RUNTIME_ENABLED\n\n // Kernel launch not supported from this device\n return CubDebug(cudaErrorNotSupported );\n\n#else\n\n cudaError error = cudaSuccess;\n do\n {\n // Dispatch two kernels: (1) a multi-block segmented reduction\n // to reduce regions by block, and (2) a single-block reduce-by-key kernel\n // to \"fix up\" segments spanning more than one region.\n\n // Tile size of seg_reduce_region_kernel\n int tile_size = seg_reduce_region_config.block_threads * seg_reduce_region_config.items_per_thread;\n\n // Get device ordinal\n int device_ordinal;\n if (CubDebug(error = cudaGetDevice(&device_ordinal))) break;\n\n // Get SM count\n int sm_count;\n if (CubDebug(error = cudaDeviceGetAttribute (&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal))) break;\n\n // Get SM occupancy for histogram_region_kernel\n int seg_reduce_region_sm_occupancy;\n if (CubDebug(error = MaxSmOccupancy(\n seg_reduce_region_sm_occupancy,\n sm_version,\n seg_reduce_region_kernel,\n seg_reduce_region_config.block_threads))) break;\n\n // Get device occupancy for histogram_region_kernel\n int seg_reduce_region_occupancy = seg_reduce_region_sm_occupancy * sm_count;\n\n // Even-share work distribution\n int num_diagonals = num_values + num_segments; // Total number of work items\n int subscription_factor = seg_reduce_region_sm_occupancy; // Amount of CTAs to oversubscribe the device beyond actively-resident (heuristic)\n int max_grid_size = seg_reduce_region_occupancy * subscription_factor;\n GridEvenShareeven_share(\n num_diagonals,\n max_grid_size,\n tile_size);\n\n // Get grid size for seg_reduce_region_kernel\n int seg_reduce_region_grid_size = even_share.grid_size;\n\n // Number of \"fix-up\" reduce-by-key tuples (2 per thread block)\n int num_tuple_partials = seg_reduce_region_grid_size * 2;\n int num_partition_samples = seg_reduce_region_grid_size + 1;\n\n // Temporary storage allocation requirements\n void* allocations[2];\n size_t allocation_sizes[2] =\n {\n num_tuple_partials * sizeof(KeyValuePair), // bytes needed for \"fix-up\" reduce-by-key tuples\n num_partition_samples * sizeof(IndexPair), // bytes needed block indices\n };\n\n // Alias the temporary allocations from the single storage blob (or set the necessary size of the blob)\n if (CubDebug(error = AliasTemporaries(d_temp_storage, temp_storage_bytes, allocations, allocation_sizes))) break;\n if (d_temp_storage == NULL)\n {\n // Return if the caller is simply requesting the size of the storage allocation\n return cudaSuccess;\n }\n\n // Alias the allocations\n KeyValuePair *d_tuple_partials = (KeyValuePair*) allocations[0]; // \"fix-up\" tuples\n IndexPair *d_block_idx = (IndexPair *) allocations[1]; // block starting/ending indices\n\n // Array of segment end-offsets\n SegmentOffsetIterator d_segment_end_offsets = d_segment_offsets + 1;\n\n // Grid launch params for seg_reduce_partition_kernel\n int partition_block_size = 32;\n int partition_grid_size = (num_partition_samples + partition_block_size - 1) / partition_block_size;\n\n // Partition work among multiple thread blocks if necessary\n if (seg_reduce_region_grid_size > 1)\n {\n // Log seg_reduce_partition_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking seg_reduce_partition_kernel<<<%d, %d, 0, %lld>>>()\\n\",\n partition_grid_size, partition_block_size, (long long) stream);\n\n // Invoke seg_reduce_partition_kernel\n seg_reduce_partition_kernel<<>>(\n d_segment_end_offsets, ///< [in] A sequence of \\p num_segments segment end-offsets\n d_block_idx,\n num_partition_samples,\n num_values, ///< [in] Number of values to reduce\n num_segments, ///< [in] Number of segments being reduced\n even_share); ///< [in] Even-share descriptor for mapping an equal number of tiles onto each thread block\n\n // Sync the stream if specified\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n\n // Log seg_reduce_region_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking seg_reduce_region_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread, %d SM occupancy\\n\",\n seg_reduce_region_grid_size, seg_reduce_region_config.block_threads, (long long) stream, seg_reduce_region_config.items_per_thread, seg_reduce_region_sm_occupancy);\n\n // Mooch\n if (CubDebug(error = cudaDeviceSetSharedMemConfig(cudaSharedMemBankSizeEightByte))) break;\n\n // Invoke seg_reduce_region_kernel\n seg_reduce_region_kernel<<>>(\n d_segment_end_offsets,\n d_values,\n d_output,\n d_tuple_partials,\n d_block_idx,\n num_values,\n num_segments,\n identity,\n reduction_op,\n even_share);\n\n // Sync the stream if specified\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n/*\n // Perform \"fix-up\" of region partial reductions if grid size is greater than one thread block\n if (seg_reduce_region_grid_size > 1)\n {\n // Log seg_reduce_region_by_key_kernel configuration\n if (debug_synchronous) _CubLog(\"Invoking seg_reduce_region_by_key_kernel<<<%d, %d, 0, %lld>>>(), %d items per thread\\n\",\n 1, seg_reduce_region_by_key_config.block_threads, (long long) stream, seg_reduce_region_by_key_config.items_per_thread);\n\n // Invoke seg_reduce_region_by_key_kernel\n seg_reduce_region_by_key_kernel<<<1, seg_reduce_region_by_key_config.block_threads, 0, stream>>>(\n d_tuple_partials,\n d_output,\n num_segments,\n num_tuple_partials,\n identity,\n reduction_op);\n\n // Sync the stream if specified\n if (debug_synchronous && (CubDebug(error = SyncStream(stream)))) break;\n }\n*/\n }\n\n while (0);\n\n return error;\n\n#endif // CUB_RUNTIME_ENABLED\n }\n\n\n /**\n * Internal dispatch routine for computing a device-wide segmented reduction.\n */\n __host__ __device__ __forceinline__\n static cudaError_t Dispatch(\n void* d_temp_storage, ///< [in] %Device allocation of temporary storage. When NULL, the required allocation size is returned in \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation.\n ValueIterator d_values, ///< [in] A sequence of \\p num_values data to reduce\n SegmentOffsetIterator d_segment_offsets, ///< [in] A sequence of (\\p num_segments + 1) segment offsets\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n OffsetT num_values, ///< [in] Total number of values to reduce\n OffsetT num_segments, ///< [in] Number of segments being reduced\n Value identity, ///< [in] Identity value (for zero-length segments)\n ReductionOp reduction_op, ///< [in] Reduction operator\n cudaStream_t stream, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n cudaError error = cudaSuccess;\n do\n {\n // Get PTX version\n int ptx_version;\n #if (CUB_PTX_ARCH == 0)\n if (CubDebug(error = PtxVersion(ptx_version))) break;\n #else\n ptx_version = CUB_PTX_ARCH;\n #endif\n\n // Get kernel kernel dispatch configurations\n SegReduceKernelConfig seg_reduce_region_config;\n SegReduceByKeyKernelConfig seg_reduce_region_by_key_config;\n\n InitConfigs(ptx_version, seg_reduce_region_config, seg_reduce_region_by_key_config);\n\n // Dispatch\n if (CubDebug(error = Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_values,\n d_segment_offsets,\n d_output,\n num_values,\n num_segments,\n identity,\n reduction_op,\n stream,\n debug_synchronous,\n ptx_version, // Use PTX version instead of SM version because, as a statically known quantity, this improves device-side launch dramatically but at the risk of imprecise occupancy calculation for mismatches\n SegReducePartitionKernel,\n SegReduceRegionKernel,\n SegReduceRegionByKeyKernel,\n seg_reduce_region_config,\n seg_reduce_region_by_key_config))) break;\n }\n while (0);\n\n return error;\n\n }\n};\n\n\n\n\n/******************************************************************************\n * DeviceSegReduce\n *****************************************************************************/\n\n/**\n * \\brief DeviceSegReduce provides operations for computing a device-wide, parallel segmented reduction across a sequence of data items residing within global memory.\n * \\ingroup DeviceModule\n *\n * \\par Overview\n * A reduction (or fold)\n * uses a binary combining operator to compute a single aggregate from a list of input elements.\n *\n * \\par Usage Considerations\n * \\cdp_class{DeviceReduce}\n *\n */\nstruct DeviceSegReduce\n{\n /**\n * \\brief Computes a device-wide segmented reduction using the specified binary \\p reduction_op functor.\n *\n * \\par\n * Does not support non-commutative reduction operators.\n *\n * \\devicestorage\n *\n * \\cdp\n *\n * \\iterator\n *\n * \\tparam ValueIterator [inferred] Random-access input iterator type for reading values\n * \\tparam SegmentOffsetIterator [inferred] Random-access input iterator type for reading segment end-offsets\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing segment reductions\n * \\tparam Value [inferred] Value type\n * \\tparam ReductionOp [inferred] Binary reduction operator type having member T operator()(const T &a, const T &b)\n */\n template <\n typename ValueIterator,\n typename SegmentOffsetIterator,\n typename OutputIteratorT,\n typename Value,\n typename ReductionOp>\n __host__ __device__ __forceinline__\n static cudaError_t Reduce(\n void* d_temp_storage, ///< [in] %Device allocation of temporary storage. When NULL, the required allocation size is returned in \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation.\n ValueIterator d_values, ///< [in] A sequence of \\p num_values data to reduce\n SegmentOffsetIterator d_segment_offsets, ///< [in] A sequence of (\\p num_segments + 1) segment offsets\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n int num_values, ///< [in] Total number of values to reduce\n int num_segments, ///< [in] Number of segments being reduced\n Value identity, ///< [in] Identity value (for zero-length segments)\n ReductionOp reduction_op, ///< [in] Reduction operator\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n typedef DeviceSegReduceDispatch<\n ValueIterator,\n SegmentOffsetIterator,\n OutputIteratorT,\n ReductionOp,\n OffsetT>\n DeviceSegReduceDispatch;\n\n return DeviceSegReduceDispatch::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_values,\n d_segment_offsets,\n d_output,\n num_values,\n num_segments,\n identity,\n reduction_op,\n stream,\n debug_synchronous);\n }\n\n\n /**\n * \\brief Computes a device-wide segmented sum using the addition ('+') operator.\n *\n * \\par\n * Does not support non-commutative summation.\n *\n * \\devicestorage\n *\n * \\cdp\n *\n * \\iterator\n *\n * \\tparam ValueIterator [inferred] Random-access input iterator type for reading values\n * \\tparam SegmentOffsetIterator [inferred] Random-access input iterator type for reading segment end-offsets\n * \\tparam OutputIteratorT [inferred] Random-access output iterator type for writing segment reductions\n */\n template <\n typename ValueIterator,\n typename SegmentOffsetIterator,\n typename OutputIteratorT>\n __host__ __device__ __forceinline__\n static cudaError_t Sum(\n void* d_temp_storage, ///< [in] %Device allocation of temporary storage. When NULL, the required allocation size is returned in \\p temp_storage_bytes and no work is done.\n size_t &temp_storage_bytes, ///< [in,out] Reference to size in bytes of \\p d_temp_storage allocation.\n ValueIterator d_values, ///< [in] A sequence of \\p num_values data to reduce\n SegmentOffsetIterator d_segment_offsets, ///< [in] A sequence of (\\p num_segments + 1) segment offsets\n OutputIteratorT d_output, ///< [out] A sequence of \\p num_segments segment totals\n int num_values, ///< [in] Total number of values to reduce\n int num_segments, ///< [in] Number of segments being reduced\n cudaStream_t stream = 0, ///< [in] [optional] CUDA stream to launch kernels within. Default is stream0.\n bool debug_synchronous = false) ///< [in] [optional] Whether or not to synchronize the stream after every kernel launch to check for errors. Also causes launch configurations to be printed to the console. Default is \\p false.\n {\n // Signed integer type for global offsets\n typedef int OffsetT;\n\n // Value type\n typedef typename std::iterator_traits::value_type Value;\n\n Value identity = Value();\n cub::Sum reduction_op;\n\n typedef DeviceSegReduceDispatch<\n ValueIterator,\n SegmentOffsetIterator,\n OutputIteratorT,\n cub::Sum,\n OffsetT>\n DeviceSegReduceDispatch;\n\n return DeviceSegReduceDispatch::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n d_values,\n d_segment_offsets,\n d_output,\n num_values,\n num_segments,\n identity,\n reduction_op,\n stream,\n debug_synchronous);\n }\n};\n\n\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n GenMode gen_mode,\n Value *h_values,\n vector &segment_offsets,\n int num_values,\n int avg_segment_size)\n{\n // Initialize values\n// if (g_verbose) printf(\"Values: \");\n for (int i = 0; i < num_values; ++i)\n {\n InitValue(gen_mode, h_values[i], i);\n// if (g_verbose) std::cout << h_values[i] << \", \";\n }\n// if (g_verbose) printf(\"\\n\\n\");\n\n // Initialize segment lengths\n const unsigned int MAX_INTEGER = -1u;\n const unsigned int MAX_SEGMENT_LENGTH = avg_segment_size * 2;\n const double SCALE_FACTOR = double(MAX_SEGMENT_LENGTH) / double(MAX_INTEGER);\n\n segment_offsets.push_back(0);\n\n OffsetT consumed = 0;\n OffsetT remaining = num_values;\n while (remaining > 0)\n {\n // Randomly sample a 32-bit unsigned int\n unsigned int segment_length;\n RandomBits(segment_length);\n\n // Scale to maximum segment length\n segment_length = (unsigned int) (double(segment_length) * SCALE_FACTOR);\n segment_length = CUB_MIN(segment_length, remaining);\n\n consumed += segment_length;\n remaining -= segment_length;\n\n segment_offsets.push_back(consumed);\n }\n}\n\n\n/**\n * Compute reference answer\n */\ntemplate \nvoid ComputeReference(\n Value *h_values,\n OffsetT *h_segment_offsets,\n Value *h_reference,\n int num_segments,\n Value identity)\n{\n if (g_verbose) printf(\"%d segment reductions: \", num_segments);\n for (int segment = 0; segment < num_segments; ++segment)\n {\n h_reference[segment] = identity;\n\n for (int i = h_segment_offsets[segment]; i < h_segment_offsets[segment + 1]; ++i)\n {\n h_reference[segment] += h_values[i];\n }\n if (g_verbose) std::cout << h_reference[segment] << \", \";\n }\n if (g_verbose) printf(\"\\n\\n\");\n}\n\n\n/**\n * Simple test of device\n */\ntemplate <\n bool CDP,\n typename OffsetT,\n typename Value,\n typename ReductionOp>\nvoid Test(\n OffsetT num_values,\n int avg_segment_size,\n ReductionOp reduction_op,\n Value identity,\n char* type_string)\n{\n Value *h_values = NULL;\n Value *h_reference = NULL;\n OffsetT *h_segment_offsets = NULL;\n\n printf(\"%d\\n\", num_values);\n\n // Initialize problem on host\n h_values = new Value[num_values];\n vector segment_offsets;\n Initialize(UNIFORM, h_values, segment_offsets, num_values, avg_segment_size);\n\n // Allocate simple offsets array and copy STL vector into it\n h_segment_offsets = new OffsetT[segment_offsets.size()];\n for (int i = 0; i < segment_offsets.size(); ++i)\n h_segment_offsets[i] = segment_offsets[i];\n\n OffsetT num_segments = segment_offsets.size() - 1;\n if (g_verbose)\n {\n printf(\"%d segment offsets: \", num_segments);\n for (int i = 0; i < num_segments; ++i)\n std::cout << h_segment_offsets[i] << \"(\" << h_segment_offsets[i + 1] - h_segment_offsets[i] << \"), \";\n if (g_verbose) std::cout << std::endl << std::endl;\n }\n\n // Solve problem on host\n h_reference = new Value[num_segments];\n ComputeReference(h_values, h_segment_offsets, h_reference, num_segments, identity);\n\n printf(\"\\n\\n%s cub::DeviceSegReduce::%s %d items (%d-byte %s), %d segments (%d-byte offset indices)\\n\",\n (CDP) ? \"CDP device invoked\" : \"Host-invoked\",\n (Equals::VALUE) ? \"Sum\" : \"Reduce\",\n num_values, (int) sizeof(Value), type_string,\n num_segments, (int) sizeof(OffsetT));\n fflush(stdout);\n\n // Allocate and initialize problem on device\n Value *d_values = NULL;\n OffsetT *d_segment_offsets = NULL;\n Value *d_output = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values, sizeof(Value) * num_values));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_segment_offsets, sizeof(OffsetT) * (num_segments + 1)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_output, sizeof(Value) * num_segments));\n CubDebugExit(cudaMemcpy(d_values, h_values, sizeof(Value) * num_values, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_segment_offsets, h_segment_offsets, sizeof(OffsetT) * (num_segments + 1), cudaMemcpyHostToDevice));\n\n // Request and allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceSegReduce::Sum(d_temp_storage, temp_storage_bytes, d_values, d_segment_offsets, d_output, num_values, num_segments, 0, false));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output\n CubDebugExit(cudaMemset(d_output, 0, sizeof(Value) * num_segments));\n\n // Run warmup/correctness iteration\n CubDebugExit(DeviceSegReduce::Sum(d_temp_storage, temp_storage_bytes, d_values, d_segment_offsets, d_output, num_values, num_segments, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_output, num_segments, true, g_verbose);\n printf(\"\\t%s\", compare ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n for (int i = 0; i < g_timing_iterations; ++i)\n {\n CubDebugExit(DeviceSegReduce::Sum(d_temp_storage, temp_storage_bytes, d_values, d_segment_offsets, d_output, num_values, num_segments, 0, false));\n }\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_values) / avg_millis / 1000.0 / 1000.0;\n float giga_bandwidth = giga_rate *\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s\", avg_millis, giga_rate, giga_bandwidth);\n }\n\n // Device cleanup\n if (d_values) CubDebugExit(g_allocator.DeviceFree(d_values));\n if (d_segment_offsets) CubDebugExit(g_allocator.DeviceFree(d_segment_offsets));\n if (d_output) CubDebugExit(g_allocator.DeviceFree(d_output));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Host cleanup\n if (h_values) delete[] h_values;\n if (h_segment_offsets) delete[] h_segment_offsets;\n if (h_reference) delete[] h_reference;\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_values = 32 * 1024 * 1024;\n int avg_segment_size = 500;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_values);\n args.GetCmdLineArgument(\"ss\", avg_segment_size);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--v] \"\n \"[--i=] \"\n \"[--n=]\\n\"\n \"[--ss=]\\n\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n Test((int) num_values, avg_segment_size, Sum(), (long long) 0, CUB_TYPE_STRING(long long));\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/experimental/histogram/histogram_cub.h", "content": "/******************************************************************************\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n#include \n\nusing namespace cub;\n\ntemplate <\n int NUM_CHANNELS,\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\ndouble run_cub_histogram(\n PixelType *d_image,\n int width,\n int height,\n unsigned int *d_hist, \n bool is_warmup)\n{\n enum {\n is_float = Equals::VALUE,\n };\n\n typedef typename If::Type SampleT; // Sample type\n typedef typename If::Type LevelT; // Level type (uint32 for uchar)\n\n // Setup data structures\n unsigned int* d_histogram[ACTIVE_CHANNELS];\n int num_levels[ACTIVE_CHANNELS]; ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[ACTIVE_CHANNELS]; ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[ACTIVE_CHANNELS]; ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n {\n d_histogram[CHANNEL] = d_hist + (CHANNEL * NUM_BINS);\n num_levels[CHANNEL] = NUM_BINS + 1;\n lower_level[CHANNEL] = 0;\n upper_level[CHANNEL] = (is_float) ? 1 : 256;\n }\n\n // Allocate temporary storage\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n\n SampleT* d_image_samples = (SampleT*) d_image;\n\n // Get amount of temporary storage needed\n DeviceHistogram::MultiHistogramEven(\n d_temp_storage,\n temp_storage_bytes,\n d_image_samples,\n d_histogram,\n num_levels,\n lower_level,\n upper_level,\n width * height, \n (cudaStream_t) 0,\n is_warmup);\n\n cudaMalloc(&d_temp_storage, temp_storage_bytes);\n\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n // Compute histogram\n DeviceHistogram::MultiHistogramEven(\n d_temp_storage,\n temp_storage_bytes,\n d_image_samples,\n d_histogram,\n num_levels,\n lower_level,\n upper_level,\n width * height, \n (cudaStream_t) 0,\n is_warmup);\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n cudaFree(d_temp_storage);\n\n return elapsed_millis;\n}\n\n" }, { "path": "external/cub/experimental/histogram/histogram_gmem_atomics.h", "content": "/******************************************************************************\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n#include \n\nnamespace histogram_gmem_atomics\n{\n // Decode float4 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(float4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n float* samples = reinterpret_cast(&pixel);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL] * float(NUM_BINS));\n }\n\n // Decode uchar4 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(uchar4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n unsigned char* samples = reinterpret_cast(&pixel);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL]);\n }\n\n // Decode uchar1 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(uchar1 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n bins[0] = (unsigned int) pixel.x;\n }\n\n // First-pass histogram kernel (binning into privatized counters)\n template <\n int NUM_PARTS,\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\n __global__ void histogram_gmem_atomics(\n const PixelType *in,\n int width,\n int height,\n unsigned int *out)\n {\n // global position and size\n int x = blockIdx.x * blockDim.x + threadIdx.x;\n int y = blockIdx.y * blockDim.y + threadIdx.y;\n int nx = blockDim.x * gridDim.x;\n int ny = blockDim.y * gridDim.y;\n\n // threads in workgroup\n int t = threadIdx.x + threadIdx.y * blockDim.x; // thread index in workgroup, linear in 0..nt-1\n int nt = blockDim.x * blockDim.y; // total threads in workgroup\n\n // group index in 0..ngroups-1\n int g = blockIdx.x + blockIdx.y * gridDim.x;\n\n // initialize smem\n unsigned int *gmem = out + g * NUM_PARTS;\n for (int i = t; i < ACTIVE_CHANNELS * NUM_BINS; i += nt)\n gmem[i] = 0;\n __syncthreads();\n\n // process pixels (updates our group's partial histogram in gmem)\n for (int col = x; col < width; col += nx)\n {\n for (int row = y; row < height; row += ny)\n {\n PixelType pixel = in[row * width + col];\n\n unsigned int bins[ACTIVE_CHANNELS];\n DecodePixel(pixel, bins);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n atomicAdd(&gmem[(NUM_BINS * CHANNEL) + bins[CHANNEL]], 1);\n }\n }\n }\n\n // Second pass histogram kernel (accumulation)\n template <\n int NUM_PARTS,\n int ACTIVE_CHANNELS,\n int NUM_BINS>\n __global__ void histogram_gmem_accum(\n const unsigned int *in,\n int n,\n unsigned int *out)\n {\n int i = blockIdx.x * blockDim.x + threadIdx.x;\n if (i > ACTIVE_CHANNELS * NUM_BINS)\n return; // out of range\n\n unsigned int total = 0;\n for (int j = 0; j < n; j++)\n total += in[i + NUM_PARTS * j];\n\n out[i] = total;\n }\n\n\n} // namespace histogram_gmem_atomics\n\n\ntemplate <\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\ndouble run_gmem_atomics(\n PixelType *d_image,\n int width,\n int height,\n unsigned int *d_hist,\n bool warmup)\n{\n enum\n {\n NUM_PARTS = 1024\n };\n\n cudaDeviceProp props;\n cudaGetDeviceProperties(&props, 0);\n\n dim3 block(32, 4);\n dim3 grid(16, 16);\n int total_blocks = grid.x * grid.y;\n\n // allocate partial histogram\n unsigned int *d_part_hist;\n cudaMalloc(&d_part_hist, total_blocks * NUM_PARTS * sizeof(unsigned int));\n\n dim3 block2(128);\n dim3 grid2((3 * NUM_BINS + block.x - 1) / block.x);\n\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n histogram_gmem_atomics::histogram_gmem_atomics<<>>(\n d_image,\n width,\n height,\n d_part_hist);\n\n histogram_gmem_atomics::histogram_gmem_accum<<>>(\n d_part_hist,\n total_blocks,\n d_hist);\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n cudaFree(d_part_hist);\n\n return elapsed_millis;\n}\n\n" }, { "path": "external/cub/experimental/histogram/histogram_smem_atomics.h", "content": "/******************************************************************************\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n#include \n\nnamespace histogram_smem_atomics\n{\n // Decode float4 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(float4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n float* samples = reinterpret_cast(&pixel);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL] * float(NUM_BINS));\n }\n\n // Decode uchar4 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(uchar4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n unsigned char* samples = reinterpret_cast(&pixel);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL]);\n }\n\n // Decode uchar1 pixel into bins\n template \n __device__ __forceinline__ void DecodePixel(uchar1 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n {\n bins[0] = (unsigned int) pixel.x;\n }\n\n // First-pass histogram kernel (binning into privatized counters)\n template <\n int NUM_PARTS,\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\n __global__ void histogram_smem_atomics(\n const PixelType *in,\n int width,\n int height,\n unsigned int *out)\n {\n // global position and size\n int x = blockIdx.x * blockDim.x + threadIdx.x;\n int y = blockIdx.y * blockDim.y + threadIdx.y;\n int nx = blockDim.x * gridDim.x;\n int ny = blockDim.y * gridDim.y;\n\n // threads in workgroup\n int t = threadIdx.x + threadIdx.y * blockDim.x; // thread index in workgroup, linear in 0..nt-1\n int nt = blockDim.x * blockDim.y; // total threads in workgroup\n\n // group index in 0..ngroups-1\n int g = blockIdx.x + blockIdx.y * gridDim.x;\n\n // initialize smem\n __shared__ unsigned int smem[ACTIVE_CHANNELS * NUM_BINS + 3];\n for (int i = t; i < ACTIVE_CHANNELS * NUM_BINS + 3; i += nt)\n smem[i] = 0;\n __syncthreads();\n\n // process pixels\n // updates our group's partial histogram in smem\n for (int col = x; col < width; col += nx)\n {\n for (int row = y; row < height; row += ny)\n {\n PixelType pixel = in[row * width + col];\n\n unsigned int bins[ACTIVE_CHANNELS];\n DecodePixel(pixel, bins);\n\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n atomicAdd(&smem[(NUM_BINS * CHANNEL) + bins[CHANNEL] + CHANNEL], 1);\n }\n }\n\n __syncthreads();\n\n // move to our workgroup's slice of output\n out += g * NUM_PARTS;\n\n // store local output to global\n for (int i = t; i < NUM_BINS; i += nt)\n {\n #pragma unroll\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n out[i + NUM_BINS * CHANNEL] = smem[i + NUM_BINS * CHANNEL + CHANNEL];\n }\n }\n\n // Second pass histogram kernel (accumulation)\n template <\n int NUM_PARTS,\n int ACTIVE_CHANNELS,\n int NUM_BINS>\n __global__ void histogram_smem_accum(\n const unsigned int *in,\n int n,\n unsigned int *out)\n {\n int i = blockIdx.x * blockDim.x + threadIdx.x;\n if (i > ACTIVE_CHANNELS * NUM_BINS) return; // out of range\n unsigned int total = 0;\n for (int j = 0; j < n; j++)\n total += in[i + NUM_PARTS * j];\n out[i] = total;\n }\n\n} // namespace histogram_smem_atomics\n\n\ntemplate <\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\ndouble run_smem_atomics(\n PixelType *d_image,\n int width,\n int height,\n unsigned int *d_hist, \n bool warmup)\n{\n enum\n {\n NUM_PARTS = 1024\n };\n\n cudaDeviceProp props;\n cudaGetDeviceProperties(&props, 0);\n\n dim3 block(32, 4);\n dim3 grid(16, 16);\n int total_blocks = grid.x * grid.y;\n\n // allocate partial histogram\n unsigned int *d_part_hist;\n cudaMalloc(&d_part_hist, total_blocks * NUM_PARTS * sizeof(unsigned int));\n\n dim3 block2(128);\n dim3 grid2((ACTIVE_CHANNELS * NUM_BINS + block.x - 1) / block.x);\n\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n histogram_smem_atomics::histogram_smem_atomics<<>>(\n d_image,\n width,\n height,\n d_part_hist);\n\n histogram_smem_atomics::histogram_smem_accum<<>>(\n d_part_hist,\n total_blocks,\n d_hist);\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n cudaFree(d_part_hist);\n\n return elapsed_millis;\n}\n\n" }, { "path": "external/cub/experimental/histogram_compare.cu", "content": "/******************************************************************************\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \"histogram/histogram_gmem_atomics.h\"\n#include \"histogram/histogram_smem_atomics.h\"\n#include \"histogram/histogram_cub.h\"\n\n#include \n#include \n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants, and type declarations\n//---------------------------------------------------------------------\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\nbool g_verbose = false; // Whether to display input/output to console\nbool g_report = false; // Whether to display a full report in CSV format\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\nstruct less_than_value\n{\n inline bool operator()(\n const std::pair &a,\n const std::pair &b)\n {\n return a.second < b.second;\n }\n};\n\n\n//---------------------------------------------------------------------\n// Targa (.tga) image file parsing\n//---------------------------------------------------------------------\n\n/**\n * TGA image header info\n */\nstruct TgaHeader\n{\n char idlength;\n char colormaptype;\n char datatypecode;\n short colormaporigin;\n short colormaplength;\n char colormapdepth;\n short x_origin;\n short y_origin;\n short width;\n short height;\n char bitsperpixel;\n char imagedescriptor;\n\n void Parse (FILE *fptr)\n {\n idlength = fgetc(fptr);\n colormaptype = fgetc(fptr);\n datatypecode = fgetc(fptr);\n fread(&colormaporigin, 2, 1, fptr);\n fread(&colormaplength, 2, 1, fptr);\n colormapdepth = fgetc(fptr);\n fread(&x_origin, 2, 1, fptr);\n fread(&y_origin, 2, 1, fptr);\n fread(&width, 2, 1, fptr);\n fread(&height, 2, 1, fptr);\n bitsperpixel = fgetc(fptr);\n imagedescriptor = fgetc(fptr);\n }\n\n void Display (FILE *fptr)\n {\n fprintf(fptr, \"ID length: %d\\n\", idlength);\n fprintf(fptr, \"Color map type: %d\\n\", colormaptype);\n fprintf(fptr, \"Image type: %d\\n\", datatypecode);\n fprintf(fptr, \"Color map offset: %d\\n\", colormaporigin);\n fprintf(fptr, \"Color map length: %d\\n\", colormaplength);\n fprintf(fptr, \"Color map depth: %d\\n\", colormapdepth);\n fprintf(fptr, \"X origin: %d\\n\", x_origin);\n fprintf(fptr, \"Y origin: %d\\n\", y_origin);\n fprintf(fptr, \"Width: %d\\n\", width);\n fprintf(fptr, \"Height: %d\\n\", height);\n fprintf(fptr, \"Bits per pixel: %d\\n\", bitsperpixel);\n fprintf(fptr, \"Descriptor: %d\\n\", imagedescriptor);\n }\n};\n\n\n/**\n * Decode image byte data into pixel\n */\nvoid ParseTgaPixel(uchar4 &pixel, unsigned char *tga_pixel, int bytes)\n{\n if (bytes == 4)\n {\n pixel.x = tga_pixel[2];\n pixel.y = tga_pixel[1];\n pixel.z = tga_pixel[0];\n pixel.w = tga_pixel[3];\n }\n else if (bytes == 3)\n {\n pixel.x = tga_pixel[2];\n pixel.y = tga_pixel[1];\n pixel.z = tga_pixel[0];\n pixel.w = 0;\n }\n else if (bytes == 2)\n {\n pixel.x = (tga_pixel[1] & 0x7c) << 1;\n pixel.y = ((tga_pixel[1] & 0x03) << 6) | ((tga_pixel[0] & 0xe0) >> 2);\n pixel.z = (tga_pixel[0] & 0x1f) << 3;\n pixel.w = (tga_pixel[1] & 0x80);\n }\n}\n\n\n/**\n * Reads a .tga image file\n */\nvoid ReadTga(uchar4* &pixels, int &width, int &height, const char *filename)\n{\n // Open the file\n FILE *fptr;\n if ((fptr = fopen(filename, \"rb\")) == NULL)\n {\n fprintf(stderr, \"File open failed\\n\");\n exit(-1);\n }\n\n // Parse header\n TgaHeader header;\n header.Parse(fptr);\n// header.Display(stdout);\n width = header.width;\n height = header.height;\n\n // Verify compatibility\n if (header.datatypecode != 2 && header.datatypecode != 10)\n {\n fprintf(stderr, \"Can only handle image type 2 and 10\\n\");\n exit(-1);\n }\n if (header.bitsperpixel != 16 && header.bitsperpixel != 24 && header.bitsperpixel != 32)\n {\n fprintf(stderr, \"Can only handle pixel depths of 16, 24, and 32\\n\");\n exit(-1);\n }\n if (header.colormaptype != 0 && header.colormaptype != 1)\n {\n fprintf(stderr, \"Can only handle color map types of 0 and 1\\n\");\n exit(-1);\n }\n\n // Skip unnecessary header info\n int skip_bytes = header.idlength + (header.colormaptype * header.colormaplength);\n fseek(fptr, skip_bytes, SEEK_CUR);\n\n // Read the image\n int pixel_bytes = header.bitsperpixel / 8;\n\n // Allocate and initialize pixel data\n size_t image_bytes = width * height * sizeof(uchar4);\n if ((pixels == NULL) && ((pixels = (uchar4*) malloc(image_bytes)) == NULL))\n {\n fprintf(stderr, \"malloc of image failed\\n\");\n exit(-1);\n }\n memset(pixels, 0, image_bytes);\n\n // Parse pixels\n unsigned char tga_pixel[5];\n int current_pixel = 0;\n while (current_pixel < header.width * header.height)\n {\n if (header.datatypecode == 2)\n {\n // Uncompressed\n if (fread(tga_pixel, 1, pixel_bytes, fptr) != pixel_bytes)\n {\n fprintf(stderr, \"Unexpected end of file at pixel %d (uncompressed)\\n\", current_pixel);\n exit(-1);\n }\n ParseTgaPixel(pixels[current_pixel], tga_pixel, pixel_bytes);\n current_pixel++;\n }\n else if (header.datatypecode == 10)\n {\n // Compressed\n if (fread(tga_pixel, 1, pixel_bytes + 1, fptr) != pixel_bytes + 1)\n {\n fprintf(stderr, \"Unexpected end of file at pixel %d (compressed)\\n\", current_pixel);\n exit(-1);\n }\n int run_length = tga_pixel[0] & 0x7f;\n ParseTgaPixel(pixels[current_pixel], &(tga_pixel[1]), pixel_bytes);\n current_pixel++;\n\n if (tga_pixel[0] & 0x80)\n {\n // RLE chunk\n for (int i = 0; i < run_length; i++)\n {\n ParseTgaPixel(pixels[current_pixel], &(tga_pixel[1]), pixel_bytes);\n current_pixel++;\n }\n }\n else\n {\n // Normal chunk\n for (int i = 0; i < run_length; i++)\n {\n if (fread(tga_pixel, 1, pixel_bytes, fptr) != pixel_bytes)\n {\n fprintf(stderr, \"Unexpected end of file at pixel %d (normal)\\n\", current_pixel);\n exit(-1);\n }\n ParseTgaPixel(pixels[current_pixel], tga_pixel, pixel_bytes);\n current_pixel++;\n }\n }\n }\n }\n\n // Close file\n fclose(fptr);\n}\n\n\n\n//---------------------------------------------------------------------\n// Random image generation\n//---------------------------------------------------------------------\n\n/**\n * Generate a random image with specified entropy\n */\nvoid GenerateRandomImage(uchar4* &pixels, int width, int height, int entropy_reduction)\n{\n int num_pixels = width * height;\n size_t image_bytes = num_pixels * sizeof(uchar4);\n if ((pixels == NULL) && ((pixels = (uchar4*) malloc(image_bytes)) == NULL))\n {\n fprintf(stderr, \"malloc of image failed\\n\");\n exit(-1);\n }\n\n for (int i = 0; i < num_pixels; ++i)\n {\n RandomBits(pixels[i].x, entropy_reduction);\n RandomBits(pixels[i].y, entropy_reduction);\n RandomBits(pixels[i].z, entropy_reduction);\n RandomBits(pixels[i].w, entropy_reduction);\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Histogram verification\n//---------------------------------------------------------------------\n\n// Decode float4 pixel into bins\ntemplate \nvoid DecodePixelGold(float4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n{\n float* samples = reinterpret_cast(&pixel);\n\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL] * float(NUM_BINS));\n}\n\n// Decode uchar4 pixel into bins\ntemplate \nvoid DecodePixelGold(uchar4 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n{\n unsigned char* samples = reinterpret_cast(&pixel);\n\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n bins[CHANNEL] = (unsigned int) (samples[CHANNEL]);\n}\n\n// Decode uchar1 pixel into bins\ntemplate \nvoid DecodePixelGold(uchar1 pixel, unsigned int (&bins)[ACTIVE_CHANNELS])\n{\n bins[0] = (unsigned int) pixel.x;\n}\n\n\n// Compute reference histogram. Specialized for uchar4\ntemplate <\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\nvoid HistogramGold(PixelType *image, int width, int height, unsigned int* hist)\n{\n memset(hist, 0, ACTIVE_CHANNELS * NUM_BINS * sizeof(unsigned int));\n\n for (int i = 0; i < width; i++)\n {\n for (int j = 0; j < height; j++)\n {\n PixelType pixel = image[i + j * width];\n\n unsigned int bins[ACTIVE_CHANNELS];\n DecodePixelGold(pixel, bins);\n\n for (int CHANNEL = 0; CHANNEL < ACTIVE_CHANNELS; ++CHANNEL)\n {\n hist[(NUM_BINS * CHANNEL) + bins[CHANNEL]]++;\n }\n }\n }\n}\n\n\n//---------------------------------------------------------------------\n// Test execution\n//---------------------------------------------------------------------\n\n/**\n * Run a specific histogram implementation\n */\ntemplate <\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\nvoid RunTest(\n std::vector >& timings,\n PixelType* d_pixels,\n const int width,\n const int height,\n unsigned int * d_hist,\n unsigned int * h_hist,\n int timing_iterations,\n const char * long_name,\n const char * short_name,\n double (*f)(PixelType*, int, int, unsigned int*, bool))\n{\n if (!g_report) printf(\"%s \", long_name); fflush(stdout);\n\n // Run single test to verify (and code cache)\n (*f)(d_pixels, width, height, d_hist, !g_report);\n\n int compare = CompareDeviceResults(h_hist, d_hist, ACTIVE_CHANNELS * NUM_BINS, true, g_verbose);\n if (!g_report) printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n\n double elapsed_ms = 0;\n for (int i = 0; i < timing_iterations; i++)\n {\n elapsed_ms += (*f)(d_pixels, width, height, d_hist, false);\n }\n double avg_us = (elapsed_ms / timing_iterations) * 1000; // average in us\n timings.push_back(std::pair(short_name, avg_us));\n\n if (!g_report)\n {\n printf(\"Avg time %.3f us (%d iterations)\\n\", avg_us, timing_iterations); fflush(stdout);\n }\n else\n {\n printf(\"%.3f, \", avg_us); fflush(stdout);\n }\n\n AssertEquals(0, compare);\n}\n\n\n/**\n * Evaluate corpus of histogram implementations\n */\ntemplate <\n int NUM_CHANNELS,\n int ACTIVE_CHANNELS,\n int NUM_BINS,\n typename PixelType>\nvoid TestMethods(\n PixelType* h_pixels,\n int height,\n int width,\n int timing_iterations,\n double bandwidth_GBs)\n{\n // Copy data to gpu\n PixelType* d_pixels;\n size_t pixel_bytes = width * height * sizeof(PixelType);\n CubDebugExit(g_allocator.DeviceAllocate((void**) &d_pixels, pixel_bytes));\n CubDebugExit(cudaMemcpy(d_pixels, h_pixels, pixel_bytes, cudaMemcpyHostToDevice));\n\n if (g_report) printf(\"%.3f, \", double(pixel_bytes) / bandwidth_GBs / 1000);\n\n // Allocate results arrays on cpu/gpu\n unsigned int *h_hist;\n unsigned int *d_hist;\n size_t histogram_bytes = NUM_BINS * ACTIVE_CHANNELS * sizeof(unsigned int);\n h_hist = (unsigned int *) malloc(histogram_bytes);\n g_allocator.DeviceAllocate((void **) &d_hist, histogram_bytes);\n\n // Compute reference cpu histogram\n HistogramGold(h_pixels, width, height, h_hist);\n\n // Store timings\n std::vector > timings;\n\n // Run experiments\n RunTest(timings, d_pixels, width, height, d_hist, h_hist, timing_iterations,\n \"CUB\", \"CUB\", run_cub_histogram);\n RunTest(timings, d_pixels, width, height, d_hist, h_hist, timing_iterations,\n \"Shared memory atomics\", \"smem atomics\", run_smem_atomics);\n RunTest(timings, d_pixels, width, height, d_hist, h_hist, timing_iterations,\n \"Global memory atomics\", \"gmem atomics\", run_gmem_atomics);\n\n // Report timings\n if (!g_report)\n {\n std::sort(timings.begin(), timings.end(), less_than_value());\n printf(\"Timings (us):\\n\");\n for (int i = 0; i < timings.size(); i++)\n {\n double bandwidth = height * width * sizeof(PixelType) / timings[i].second / 1000;\n printf(\"\\t %.3f %s (%.3f GB/s, %.3f%% peak)\\n\", timings[i].second, timings[i].first.c_str(), bandwidth, bandwidth / bandwidth_GBs * 100);\n }\n printf(\"\\n\");\n }\n\n // Free data\n CubDebugExit(g_allocator.DeviceFree(d_pixels));\n CubDebugExit(g_allocator.DeviceFree(d_hist));\n free(h_hist);\n}\n\n\n/**\n * Test different problem genres\n */\nvoid TestGenres(\n uchar4* uchar4_pixels,\n int height,\n int width,\n int timing_iterations,\n double bandwidth_GBs)\n{\n int num_pixels = width * height;\n\n {\n if (!g_report) printf(\"1 channel uchar1 tests (256-bin):\\n\\n\"); fflush(stdout);\n\n size_t image_bytes = num_pixels * sizeof(uchar1);\n uchar1* uchar1_pixels = (uchar1*) malloc(image_bytes);\n\n // Convert to 1-channel (averaging first 3 channels)\n for (int i = 0; i < num_pixels; ++i)\n {\n uchar1_pixels[i].x = (unsigned char)\n (((unsigned int) uchar4_pixels[i].x +\n (unsigned int) uchar4_pixels[i].y +\n (unsigned int) uchar4_pixels[i].z) / 3);\n }\n\n TestMethods<1, 1, 256>(uchar1_pixels, width, height, timing_iterations, bandwidth_GBs);\n free(uchar1_pixels);\n if (g_report) printf(\", \");\n }\n\n {\n if (!g_report) printf(\"3/4 channel uchar4 tests (256-bin):\\n\\n\"); fflush(stdout);\n TestMethods<4, 3, 256>(uchar4_pixels, width, height, timing_iterations, bandwidth_GBs);\n if (g_report) printf(\", \");\n }\n\n {\n if (!g_report) printf(\"3/4 channel float4 tests (256-bin):\\n\\n\"); fflush(stdout);\n size_t image_bytes = num_pixels * sizeof(float4);\n float4* float4_pixels = (float4*) malloc(image_bytes);\n\n // Convert to float4 with range [0.0, 1.0)\n for (int i = 0; i < num_pixels; ++i)\n {\n float4_pixels[i].x = float(uchar4_pixels[i].x) / 256;\n float4_pixels[i].y = float(uchar4_pixels[i].y) / 256;\n float4_pixels[i].z = float(uchar4_pixels[i].z) / 256;\n float4_pixels[i].w = float(uchar4_pixels[i].w) / 256;\n }\n TestMethods<4, 3, 256>(float4_pixels, width, height, timing_iterations, bandwidth_GBs);\n free(float4_pixels);\n if (g_report) printf(\"\\n\");\n }\n}\n\n\n/**\n * Main\n */\nint main(int argc, char **argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\n \"%s \"\n \"[--device=] \"\n \"[--v] \"\n \"[--i=] \"\n \"\\n\\t\"\n \"--file=<.tga filename> \"\n \"\\n\\t\"\n \"--entropy=<-1 (0%), 0 (100%), 1 (81%), 2 (54%), 3 (34%), 4 (20%), ...\"\n \"[--height=] \"\n \"[--width=] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n std::string filename;\n int timing_iterations = 100;\n int entropy_reduction = 0;\n int height = 1080;\n int width = 1920;\n\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_report = args.CheckCmdLineFlag(\"report\");\n args.GetCmdLineArgument(\"i\", timing_iterations);\n args.GetCmdLineArgument(\"file\", filename);\n args.GetCmdLineArgument(\"height\", height);\n args.GetCmdLineArgument(\"width\", width);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get GPU device bandwidth (GB/s)\n int device_ordinal, bus_width, mem_clock_khz;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n CubDebugExit(cudaDeviceGetAttribute(&bus_width, cudaDevAttrGlobalMemoryBusWidth, device_ordinal));\n CubDebugExit(cudaDeviceGetAttribute(&mem_clock_khz, cudaDevAttrMemoryClockRate, device_ordinal));\n double bandwidth_GBs = double(bus_width) * mem_clock_khz * 2 / 8 / 1000 / 1000;\n\n // Run test(s)\n uchar4* uchar4_pixels = NULL;\n if (!g_report)\n {\n if (!filename.empty())\n {\n // Parse targa file\n ReadTga(uchar4_pixels, width, height, filename.c_str());\n printf(\"File %s: width(%d) height(%d)\\n\\n\", filename.c_str(), width, height); fflush(stdout);\n }\n else\n {\n // Generate image\n GenerateRandomImage(uchar4_pixels, width, height, entropy_reduction);\n printf(\"Random image: entropy-reduction(%d) width(%d) height(%d)\\n\\n\", entropy_reduction, width, height); fflush(stdout);\n }\n\n TestGenres(uchar4_pixels, height, width, timing_iterations, bandwidth_GBs);\n }\n else\n {\n // Run test suite\n printf(\"Test, MIN, RLE CUB, SMEM, GMEM, , MIN, RLE_CUB, SMEM, GMEM, , MIN, RLE_CUB, SMEM, GMEM\\n\");\n\n // Entropy reduction tests\n for (entropy_reduction = 0; entropy_reduction < 5; ++entropy_reduction)\n {\n printf(\"entropy reduction %d, \", entropy_reduction);\n GenerateRandomImage(uchar4_pixels, width, height, entropy_reduction);\n TestGenres(uchar4_pixels, height, width, timing_iterations, bandwidth_GBs);\n }\n printf(\"entropy reduction -1, \");\n GenerateRandomImage(uchar4_pixels, width, height, -1);\n TestGenres(uchar4_pixels, height, width, timing_iterations, bandwidth_GBs);\n printf(\"\\n\");\n\n // File image tests\n std::vector file_tests;\n file_tests.push_back(\"animals\");\n file_tests.push_back(\"apples\");\n file_tests.push_back(\"sunset\");\n file_tests.push_back(\"cheetah\");\n file_tests.push_back(\"nature\");\n file_tests.push_back(\"operahouse\");\n file_tests.push_back(\"austin\");\n file_tests.push_back(\"cityscape\");\n\n for (int i = 0; i < file_tests.size(); ++i)\n {\n printf(\"%s, \", file_tests[i].c_str());\n std::string filename = std::string(\"histogram/benchmark/\") + file_tests[i] + \".tga\";\n ReadTga(uchar4_pixels, width, height, filename.c_str());\n TestGenres(uchar4_pixels, height, width, timing_iterations, bandwidth_GBs);\n }\n }\n\n free(uchar4_pixels);\n\n CubDebugExit(cudaDeviceSynchronize());\n printf(\"\\n\\n\");\n\n return 0;\n}\n" }, { "path": "external/cub/experimental/sparse_matrix.h", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Matrix data structures and parsing logic\n ******************************************************************************/\n\n#pragma once\n\n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\n#ifdef CUB_MKL\n #include \n #include \n#endif\n\nusing namespace std;\n\n/******************************************************************************\n * COO matrix type\n ******************************************************************************/\n\nstruct GraphStats\n{\n int num_rows;\n int num_cols;\n int num_nonzeros;\n\n double diag_dist_mean; // mean\n double diag_dist_std_dev; // sample std dev\n double pearson_r; // coefficient of variation\n\n double row_length_mean; // mean\n double row_length_std_dev; // sample std_dev\n double row_length_variation; // coefficient of variation\n double row_length_skewness; // skewness\n\n void Display(bool show_labels = true)\n {\n if (show_labels)\n printf(\"\\n\"\n \"\\t num_rows: %d\\n\"\n \"\\t num_cols: %d\\n\"\n \"\\t num_nonzeros: %d\\n\"\n \"\\t diag_dist_mean: %.2f\\n\"\n \"\\t diag_dist_std_dev: %.2f\\n\"\n \"\\t pearson_r: %f\\n\"\n \"\\t row_length_mean: %.5f\\n\"\n \"\\t row_length_std_dev: %.5f\\n\"\n \"\\t row_length_variation: %.5f\\n\"\n \"\\t row_length_skewness: %.5f\\n\",\n num_rows,\n num_cols,\n num_nonzeros,\n diag_dist_mean,\n diag_dist_std_dev,\n pearson_r,\n row_length_mean,\n row_length_std_dev,\n row_length_variation,\n row_length_skewness);\n else\n printf(\n \"%d, \"\n \"%d, \"\n \"%d, \"\n \"%.2f, \"\n \"%.2f, \"\n \"%f, \"\n \"%.5f, \"\n \"%.5f, \"\n \"%.5f, \"\n \"%.5f, \",\n num_rows,\n num_cols,\n num_nonzeros,\n diag_dist_mean,\n diag_dist_std_dev,\n pearson_r,\n row_length_mean,\n row_length_std_dev,\n row_length_variation,\n row_length_skewness);\n }\n};\n\n\n\n/******************************************************************************\n * COO matrix type\n ******************************************************************************/\n\n\n/**\n * COO matrix type. A COO matrix is just a vector of edge tuples. Tuples are sorted\n * first by row, then by column.\n */\ntemplate\nstruct CooMatrix\n{\n //---------------------------------------------------------------------\n // Type definitions and constants\n //---------------------------------------------------------------------\n\n // COO edge tuple\n struct CooTuple\n {\n OffsetT row;\n OffsetT col;\n ValueT val;\n\n CooTuple() {}\n CooTuple(OffsetT row, OffsetT col) : row(row), col(col) {}\n CooTuple(OffsetT row, OffsetT col, ValueT val) : row(row), col(col), val(val) {}\n\n /**\n * Comparator for sorting COO sparse format num_nonzeros\n */\n bool operator<(const CooTuple &other) const\n {\n if ((row < other.row) || ((row == other.row) && (col < other.col)))\n {\n return true;\n }\n\n return false;\n }\n };\n\n\n //---------------------------------------------------------------------\n // Data members\n //---------------------------------------------------------------------\n\n // Fields\n int num_rows;\n int num_cols;\n int num_nonzeros;\n CooTuple* coo_tuples;\n\n //---------------------------------------------------------------------\n // Methods\n //---------------------------------------------------------------------\n\n // Constructor\n CooMatrix() : num_rows(0), num_cols(0), num_nonzeros(0), coo_tuples(NULL) {}\n\n\n /**\n * Clear\n */\n void Clear()\n {\n if (coo_tuples) delete[] coo_tuples;\n coo_tuples = NULL;\n }\n\n\n // Destructor\n ~CooMatrix()\n {\n Clear();\n }\n\n\n // Display matrix to stdout\n void Display()\n {\n cout << \"COO Matrix (\" << num_rows << \" rows, \" << num_cols << \" columns, \" << num_nonzeros << \" non-zeros):\\n\";\n cout << \"Ordinal, Row, Column, Value\\n\";\n for (int i = 0; i < num_nonzeros; i++)\n {\n cout << '\\t' << i << ',' << coo_tuples[i].row << ',' << coo_tuples[i].col << ',' << coo_tuples[i].val << \"\\n\";\n }\n }\n\n\n /**\n * Builds a symmetric COO sparse from an asymmetric CSR matrix.\n */\n template \n void InitCsrSymmetric(CsrMatrixT &csr_matrix)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n num_rows = csr_matrix.num_cols;\n num_cols = csr_matrix.num_rows;\n num_nonzeros = csr_matrix.num_nonzeros * 2;\n coo_tuples = new CooTuple[num_nonzeros];\n\n for (OffsetT row = 0; row < csr_matrix.num_rows; ++row)\n {\n for (OffsetT nonzero = csr_matrix.row_offsets[row]; nonzero < csr_matrix.row_offsets[row + 1]; ++nonzero)\n {\n coo_tuples[nonzero].row = row;\n coo_tuples[nonzero].col = csr_matrix.column_indices[nonzero];\n coo_tuples[nonzero].val = csr_matrix.values[nonzero];\n\n coo_tuples[csr_matrix.num_nonzeros + nonzero].row = coo_tuples[nonzero].col;\n coo_tuples[csr_matrix.num_nonzeros + nonzero].col = coo_tuples[nonzero].row;\n coo_tuples[csr_matrix.num_nonzeros + nonzero].val = csr_matrix.values[nonzero];\n\n }\n }\n\n // Sort by rows, then columns\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n }\n\n /**\n * Builds a COO sparse from a relabeled CSR matrix.\n */\n template \n void InitCsrRelabel(CsrMatrixT &csr_matrix, OffsetT* relabel_indices)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n num_rows = csr_matrix.num_rows;\n num_cols = csr_matrix.num_cols;\n num_nonzeros = csr_matrix.num_nonzeros;\n coo_tuples = new CooTuple[num_nonzeros];\n\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n for (OffsetT nonzero = csr_matrix.row_offsets[row]; nonzero < csr_matrix.row_offsets[row + 1]; ++nonzero)\n {\n coo_tuples[nonzero].row = relabel_indices[row];\n coo_tuples[nonzero].col = relabel_indices[csr_matrix.column_indices[nonzero]];\n coo_tuples[nonzero].val = csr_matrix.values[nonzero];\n }\n }\n\n // Sort by rows, then columns\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n }\n\n\n\n /**\n * Builds a METIS COO sparse from the given file.\n */\n void InitMetis(const string &metis_filename)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n // TODO\n }\n\n\n /**\n * Builds a MARKET COO sparse from the given file.\n */\n void InitMarket(\n const string& market_filename,\n ValueT default_value = 1.0,\n bool verbose = false)\n {\n if (verbose) {\n printf(\"Reading... \"); fflush(stdout);\n }\n\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n std::ifstream ifs;\n ifs.open(market_filename.c_str(), std::ifstream::in);\n if (!ifs.good())\n {\n fprintf(stderr, \"Error opening file\\n\");\n exit(1);\n }\n\n bool array = false;\n bool symmetric = false;\n bool skew = false;\n int current_edge = -1;\n char line[1024];\n\n if (verbose) {\n printf(\"Parsing... \"); fflush(stdout);\n }\n\n while (true)\n {\n ifs.getline(line, 1024);\n if (!ifs.good())\n {\n // Done\n break;\n }\n\n if (line[0] == '%')\n {\n // Comment\n if (line[1] == '%')\n {\n // Banner\n symmetric = (strstr(line, \"symmetric\") != NULL);\n skew = (strstr(line, \"skew\") != NULL);\n array = (strstr(line, \"array\") != NULL);\n\n if (verbose) {\n printf(\"(symmetric: %d, skew: %d, array: %d) \", symmetric, skew, array); fflush(stdout);\n }\n }\n }\n else if (current_edge == -1)\n {\n // Problem description\n int nparsed = sscanf(line, \"%d %d %d\", &num_rows, &num_cols, &num_nonzeros);\n if ((!array) && (nparsed == 3))\n {\n if (symmetric)\n num_nonzeros *= 2;\n\n // Allocate coo matrix\n coo_tuples = new CooTuple[num_nonzeros];\n current_edge = 0;\n\n }\n else if (array && (nparsed == 2))\n {\n // Allocate coo matrix\n num_nonzeros = num_rows * num_cols;\n coo_tuples = new CooTuple[num_nonzeros];\n current_edge = 0;\n }\n else\n {\n fprintf(stderr, \"Error parsing MARKET matrix: invalid problem description: %s\\n\", line);\n exit(1);\n }\n\n }\n else\n {\n // Edge\n if (current_edge >= num_nonzeros)\n {\n fprintf(stderr, \"Error parsing MARKET matrix: encountered more than %d num_nonzeros\\n\", num_nonzeros);\n exit(1);\n }\n\n int row, col;\n double val;\n\n if (array)\n {\n if (sscanf(line, \"%lf\", &val) != 1)\n {\n fprintf(stderr, \"Error parsing MARKET matrix: badly formed current_edge: '%s' at edge %d\\n\", line, current_edge);\n exit(1);\n }\n col = (current_edge / num_rows);\n row = (current_edge - (num_rows * col));\n\n coo_tuples[current_edge] = CooTuple(row, col, val); // Convert indices to zero-based\n }\n else\n {\n // Parse nonzero (note: using strtol and strtod is 2x faster than sscanf or istream parsing)\n char *l = line;\n char *t = NULL;\n\n // parse row\n row = strtol(l, &t, 0);\n if (t == l)\n {\n fprintf(stderr, \"Error parsing MARKET matrix: badly formed row at edge %d\\n\", current_edge);\n exit(1);\n }\n l = t;\n\n // parse col\n col = strtol(l, &t, 0);\n if (t == l)\n {\n fprintf(stderr, \"Error parsing MARKET matrix: badly formed col at edge %d\\n\", current_edge);\n exit(1);\n }\n l = t;\n\n // parse val\n val = strtod(l, &t);\n if (t == l)\n {\n val = default_value;\n }\n/*\n int nparsed = sscanf(line, \"%d %d %lf\", &row, &col, &val);\n if (nparsed == 2)\n {\n // No value specified\n val = default_value;\n \n }\n else if (nparsed != 3)\n {\n fprintf(stderr, \"Error parsing MARKET matrix 1: badly formed current_edge: %d parsed at edge %d\\n\", nparsed, current_edge);\n exit(1);\n }\n*/\n\n coo_tuples[current_edge] = CooTuple(row - 1, col - 1, val); // Convert indices to zero-based\n\n }\n\n current_edge++;\n\n if (symmetric && (row != col))\n {\n coo_tuples[current_edge].row = coo_tuples[current_edge - 1].col;\n coo_tuples[current_edge].col = coo_tuples[current_edge - 1].row;\n coo_tuples[current_edge].val = coo_tuples[current_edge - 1].val * (skew ? -1 : 1);\n current_edge++;\n }\n }\n }\n\n // Adjust nonzero count (nonzeros along the diagonal aren't reversed)\n num_nonzeros = current_edge;\n\n if (verbose) {\n printf(\"done. Ordering...\"); fflush(stdout);\n }\n\n // Sort by rows, then columns\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n\n if (verbose) {\n printf(\"done. \"); fflush(stdout);\n }\n\n ifs.close();\n }\n\n\n /**\n * Builds a dense matrix\n */\n int InitDense(\n OffsetT num_rows,\n OffsetT num_cols,\n ValueT default_value = 1.0,\n bool verbose = false)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n this->num_rows = num_rows;\n this->num_cols = num_cols;\n\n num_nonzeros = num_rows * num_cols;\n coo_tuples = new CooTuple[num_nonzeros];\n\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n for (OffsetT col = 0; col < num_cols; ++col)\n {\n coo_tuples[(row * num_cols) + col] = CooTuple(row, col, default_value);\n }\n }\n\n // Sort by rows, then columns\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n\n return 0;\n }\n\n /**\n * Builds a wheel COO sparse matrix having spokes spokes.\n */\n int InitWheel(\n OffsetT spokes,\n ValueT default_value = 1.0,\n bool verbose = false)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n num_rows = spokes + 1;\n num_cols = num_rows;\n num_nonzeros = spokes * 2;\n coo_tuples = new CooTuple[num_nonzeros];\n\n // Add spoke num_nonzeros\n int current_edge = 0;\n for (OffsetT i = 0; i < spokes; i++)\n {\n coo_tuples[current_edge] = CooTuple(0, i + 1, default_value);\n current_edge++;\n }\n\n // Add rim\n for (OffsetT i = 0; i < spokes; i++)\n {\n OffsetT dest = (i + 1) % spokes;\n coo_tuples[current_edge] = CooTuple(i + 1, dest + 1, default_value);\n current_edge++;\n }\n\n // Sort by rows, then columns\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n\n return 0;\n }\n\n\n /**\n * Builds a square 2D grid CSR matrix. Interior num_vertices have degree 5 when including\n * a self-loop.\n *\n * Returns 0 on success, 1 on failure.\n */\n int InitGrid2d(OffsetT width, bool self_loop, ValueT default_value = 1.0)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n exit(1);\n }\n\n int interior_nodes = (width - 2) * (width - 2);\n int edge_nodes = (width - 2) * 4;\n int corner_nodes = 4;\n num_rows = width * width;\n num_cols = num_rows;\n num_nonzeros = (interior_nodes * 4) + (edge_nodes * 3) + (corner_nodes * 2);\n\n if (self_loop)\n num_nonzeros += num_rows;\n\n coo_tuples = new CooTuple[num_nonzeros];\n int current_edge = 0;\n\n for (OffsetT j = 0; j < width; j++)\n {\n for (OffsetT k = 0; k < width; k++)\n {\n OffsetT me = (j * width) + k;\n\n // West\n OffsetT neighbor = (j * width) + (k - 1);\n if (k - 1 >= 0) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // East\n neighbor = (j * width) + (k + 1);\n if (k + 1 < width) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // North\n neighbor = ((j - 1) * width) + k;\n if (j - 1 >= 0) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // South\n neighbor = ((j + 1) * width) + k;\n if (j + 1 < width) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n if (self_loop)\n {\n neighbor = me;\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n }\n }\n\n // Sort by rows, then columns, update dims\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n\n return 0;\n }\n\n\n /**\n * Builds a square 3D grid COO sparse matrix. Interior num_vertices have degree 7 when including\n * a self-loop. Values are unintialized, coo_tuples are sorted.\n */\n int InitGrid3d(OffsetT width, bool self_loop, ValueT default_value = 1.0)\n {\n if (coo_tuples)\n {\n fprintf(stderr, \"Matrix already constructed\\n\");\n return -1;\n }\n\n OffsetT interior_nodes = (width - 2) * (width - 2) * (width - 2);\n OffsetT face_nodes = (width - 2) * (width - 2) * 6;\n OffsetT edge_nodes = (width - 2) * 12;\n OffsetT corner_nodes = 8;\n num_cols = width * width * width;\n num_rows = num_cols;\n num_nonzeros = (interior_nodes * 6) + (face_nodes * 5) + (edge_nodes * 4) + (corner_nodes * 3);\n\n if (self_loop)\n num_nonzeros += num_rows;\n\n coo_tuples = new CooTuple[num_nonzeros];\n int current_edge = 0;\n\n for (OffsetT i = 0; i < width; i++)\n {\n for (OffsetT j = 0; j < width; j++)\n {\n for (OffsetT k = 0; k < width; k++)\n {\n\n OffsetT me = (i * width * width) + (j * width) + k;\n\n // Up\n OffsetT neighbor = (i * width * width) + (j * width) + (k - 1);\n if (k - 1 >= 0) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // Down\n neighbor = (i * width * width) + (j * width) + (k + 1);\n if (k + 1 < width) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // West\n neighbor = (i * width * width) + ((j - 1) * width) + k;\n if (j - 1 >= 0) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // East\n neighbor = (i * width * width) + ((j + 1) * width) + k;\n if (j + 1 < width) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // North\n neighbor = ((i - 1) * width * width) + (j * width) + k;\n if (i - 1 >= 0) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n // South\n neighbor = ((i + 1) * width * width) + (j * width) + k;\n if (i + 1 < width) {\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n\n if (self_loop)\n {\n neighbor = me;\n coo_tuples[current_edge] = CooTuple(me, neighbor, default_value);\n current_edge++;\n }\n }\n }\n }\n\n // Sort by rows, then columns, update dims\n std::stable_sort(coo_tuples, coo_tuples + num_nonzeros);\n\n return 0;\n }\n};\n\n\n\n/******************************************************************************\n * COO matrix type\n ******************************************************************************/\n\n\n/**\n * CSR sparse format matrix\n */\ntemplate<\n typename ValueT,\n typename OffsetT>\nstruct CsrMatrix\n{\n int num_rows;\n int num_cols;\n int num_nonzeros;\n OffsetT* row_offsets;\n OffsetT* column_indices;\n ValueT* values;\n bool numa_malloc;\n\n /**\n * Constructor\n */\n CsrMatrix() : num_rows(0), num_cols(0), num_nonzeros(0), row_offsets(NULL), column_indices(NULL), values(NULL) \n {\n#ifdef CUB_MKL\n numa_malloc = ((numa_available() >= 0) && (numa_num_task_nodes() > 1));\n#else\n numa_malloc = false;\n#endif\n }\n\n\n /**\n * Clear\n */\n void Clear()\n {\n#ifdef CUB_MKL\n if (numa_malloc) \n {\n numa_free(row_offsets, sizeof(OffsetT) * (num_rows + 1));\n numa_free(values, sizeof(ValueT) * num_nonzeros);\n numa_free(column_indices, sizeof(OffsetT) * num_nonzeros);\n }\n else\n {\n if (row_offsets) mkl_free(row_offsets);\n if (column_indices) mkl_free(column_indices);\n if (values) mkl_free(values);\n }\n\n#else\n if (row_offsets) delete[] row_offsets;\n if (column_indices) delete[] column_indices;\n if (values) delete[] values;\n#endif\n\n row_offsets = NULL;\n column_indices = NULL;\n values = NULL;\n }\n\n /**\n * Destructor\n */\n ~CsrMatrix()\n {\n Clear();\n }\n\n GraphStats Stats()\n {\n GraphStats stats;\n stats.num_rows = num_rows;\n stats.num_cols = num_cols;\n stats.num_nonzeros = num_nonzeros;\n\n //\n // Compute diag-distance statistics\n //\n\n OffsetT samples = 0;\n double mean = 0.0;\n double ss_tot = 0.0;\n\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n OffsetT nz_idx_start = row_offsets[row];\n OffsetT nz_idx_end = row_offsets[row + 1];\n\n for (int nz_idx = nz_idx_start; nz_idx < nz_idx_end; ++nz_idx)\n {\n OffsetT col = column_indices[nz_idx];\n double x = (col > row) ? col - row : row - col;\n\n samples++;\n double delta = x - mean;\n mean = mean + (delta / samples);\n ss_tot += delta * (x - mean);\n }\n }\n stats.diag_dist_mean = mean;\n double variance = ss_tot / samples;\n stats.diag_dist_std_dev = sqrt(variance);\n\n\n //\n // Compute deming statistics\n //\n\n samples = 0;\n double mean_x = 0.0;\n double mean_y = 0.0;\n double ss_x = 0.0;\n double ss_y = 0.0;\n\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n OffsetT nz_idx_start = row_offsets[row];\n OffsetT nz_idx_end = row_offsets[row + 1];\n\n for (int nz_idx = nz_idx_start; nz_idx < nz_idx_end; ++nz_idx)\n {\n OffsetT col = column_indices[nz_idx];\n\n samples++;\n double x = col;\n double y = row;\n double delta;\n\n delta = x - mean_x;\n mean_x = mean_x + (delta / samples);\n ss_x += delta * (x - mean_x);\n\n delta = y - mean_y;\n mean_y = mean_y + (delta / samples);\n ss_y += delta * (y - mean_y);\n }\n }\n\n samples = 0;\n double s_xy = 0.0;\n double s_xxy = 0.0;\n double s_xyy = 0.0;\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n OffsetT nz_idx_start = row_offsets[row];\n OffsetT nz_idx_end = row_offsets[row + 1];\n\n for (int nz_idx = nz_idx_start; nz_idx < nz_idx_end; ++nz_idx)\n {\n OffsetT col = column_indices[nz_idx];\n\n samples++;\n double x = col;\n double y = row;\n\n double xy = (x - mean_x) * (y - mean_y);\n double xxy = (x - mean_x) * (x - mean_x) * (y - mean_y);\n double xyy = (x - mean_x) * (y - mean_y) * (y - mean_y);\n double delta;\n\n delta = xy - s_xy;\n s_xy = s_xy + (delta / samples);\n\n delta = xxy - s_xxy;\n s_xxy = s_xxy + (delta / samples);\n\n delta = xyy - s_xyy;\n s_xyy = s_xyy + (delta / samples);\n }\n }\n\n double s_xx = ss_x / num_nonzeros;\n double s_yy = ss_y / num_nonzeros;\n\n double deming_slope = (s_yy - s_xx + sqrt(((s_yy - s_xx) * (s_yy - s_xx)) + (4 * s_xy * s_xy))) / (2 * s_xy);\n\n stats.pearson_r = (num_nonzeros * s_xy) / (sqrt(ss_x) * sqrt(ss_y));\n\n\n //\n // Compute row-length statistics\n //\n\n // Sample mean\n stats.row_length_mean = double(num_nonzeros) / num_rows;\n variance = 0.0;\n stats.row_length_skewness = 0.0;\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n OffsetT length = row_offsets[row + 1] - row_offsets[row];\n double delta = double(length) - stats.row_length_mean;\n variance += (delta * delta);\n stats.row_length_skewness += (delta * delta * delta);\n }\n variance /= num_rows;\n stats.row_length_std_dev = sqrt(variance);\n stats.row_length_skewness = (stats.row_length_skewness / num_rows) / pow(stats.row_length_std_dev, 3.0);\n stats.row_length_variation = stats.row_length_std_dev / stats.row_length_mean;\n\n return stats;\n }\n\n /**\n * Build CSR matrix from sorted COO matrix\n */\n void FromCoo(const CooMatrix &coo_matrix)\n {\n num_rows = coo_matrix.num_rows;\n num_cols = coo_matrix.num_cols;\n num_nonzeros = coo_matrix.num_nonzeros;\n\n#ifdef CUB_MKL\n\n if (numa_malloc)\n {\n numa_set_strict(1);\n// numa_set_bind_policy(1);\n\n// values = (ValueT*) numa_alloc_interleaved(sizeof(ValueT) * num_nonzeros);\n// row_offsets = (OffsetT*) numa_alloc_interleaved(sizeof(OffsetT) * (num_rows + 1));\n// column_indices = (OffsetT*) numa_alloc_interleaved(sizeof(OffsetT) * num_nonzeros);\n\n row_offsets = (OffsetT*) numa_alloc_onnode(sizeof(OffsetT) * (num_rows + 1), 0);\n column_indices = (OffsetT*) numa_alloc_onnode(sizeof(OffsetT) * num_nonzeros, 0);\n values = (ValueT*) numa_alloc_onnode(sizeof(ValueT) * num_nonzeros, 1);\n }\n else\n {\n values = (ValueT*) mkl_malloc(sizeof(ValueT) * num_nonzeros, 4096);\n row_offsets = (OffsetT*) mkl_malloc(sizeof(OffsetT) * (num_rows + 1), 4096);\n column_indices = (OffsetT*) mkl_malloc(sizeof(OffsetT) * num_nonzeros, 4096);\n\n }\n\n#else\n row_offsets = new OffsetT[num_rows + 1];\n column_indices = new OffsetT[num_nonzeros];\n values = new ValueT[num_nonzeros];\n#endif\n\n OffsetT prev_row = -1;\n for (OffsetT current_edge = 0; current_edge < num_nonzeros; current_edge++)\n {\n OffsetT current_row = coo_matrix.coo_tuples[current_edge].row;\n\n // Fill in rows up to and including the current row\n for (OffsetT row = prev_row + 1; row <= current_row; row++)\n {\n row_offsets[row] = current_edge;\n }\n prev_row = current_row;\n\n column_indices[current_edge] = coo_matrix.coo_tuples[current_edge].col;\n values[current_edge] = coo_matrix.coo_tuples[current_edge].val;\n }\n\n // Fill out any trailing edgeless vertices (and the end-of-list element)\n for (OffsetT row = prev_row + 1; row <= num_rows; row++)\n {\n row_offsets[row] = num_nonzeros;\n }\n }\n\n\n /**\n * Display log-histogram to stdout\n */\n void DisplayHistogram()\n {\n // Initialize\n int log_counts[9];\n for (int i = 0; i < 9; i++)\n {\n log_counts[i] = 0;\n }\n\n // Scan\n int max_log_length = -1;\n for (OffsetT row = 0; row < num_rows; row++)\n {\n OffsetT length = row_offsets[row + 1] - row_offsets[row];\n\n int log_length = -1;\n while (length > 0)\n {\n length /= 10;\n log_length++;\n }\n if (log_length > max_log_length)\n {\n max_log_length = log_length;\n }\n\n log_counts[log_length + 1]++;\n }\n printf(\"CSR matrix (%d rows, %d columns, %d non-zeros):\\n\", (int) num_rows, (int) num_cols, (int) num_nonzeros);\n for (int i = -1; i < max_log_length + 1; i++)\n {\n printf(\"\\tDegree 1e%d: \\t%d (%.2f%%)\\n\", i, log_counts[i + 1], (float) log_counts[i + 1] * 100.0 / num_cols);\n }\n fflush(stdout);\n }\n\n\n /**\n * Display matrix to stdout\n */\n void Display()\n {\n printf(\"Input Matrix:\\n\");\n for (OffsetT row = 0; row < num_rows; row++)\n {\n printf(\"%d [@%d, #%d]: \", row, row_offsets[row], row_offsets[row + 1] - row_offsets[row]);\n for (OffsetT current_edge = row_offsets[row]; current_edge < row_offsets[row + 1]; current_edge++)\n {\n printf(\"%d (%f), \", column_indices[current_edge], values[current_edge]);\n }\n printf(\"\\n\");\n }\n fflush(stdout);\n }\n\n\n};\n\n\n\n/******************************************************************************\n * Matrix transformations\n ******************************************************************************/\n\n// Comparator for ordering rows by degree (lowest first), then by row-id (lowest first)\ntemplate \nstruct OrderByLow\n{\n OffsetT* row_degrees;\n OrderByLow(OffsetT* row_degrees) : row_degrees(row_degrees) {}\n\n bool operator()(const OffsetT &a, const OffsetT &b)\n {\n if (row_degrees[a] < row_degrees[b])\n return true;\n else if (row_degrees[a] > row_degrees[b])\n return false;\n else\n return (a < b);\n }\n};\n\n// Comparator for ordering rows by degree (highest first), then by row-id (lowest first)\ntemplate \nstruct OrderByHigh\n{\n OffsetT* row_degrees;\n OrderByHigh(OffsetT* row_degrees) : row_degrees(row_degrees) {}\n\n bool operator()(const OffsetT &a, const OffsetT &b)\n {\n if (row_degrees[a] > row_degrees[b])\n return true;\n else if (row_degrees[a] < row_degrees[b])\n return false;\n else\n return (a < b);\n }\n};\n\n\n\n/**\n * Reverse Cuthill-McKee\n */\ntemplate \nvoid RcmRelabel(\n CsrMatrix& matrix,\n OffsetT* relabel_indices)\n{\n // Initialize row degrees\n OffsetT* row_degrees_in = new OffsetT[matrix.num_rows];\n OffsetT* row_degrees_out = new OffsetT[matrix.num_rows];\n for (OffsetT row = 0; row < matrix.num_rows; ++row)\n {\n row_degrees_in[row] = 0;\n row_degrees_out[row] = matrix.row_offsets[row + 1] - matrix.row_offsets[row];\n }\n for (OffsetT nonzero = 0; nonzero < matrix.num_nonzeros; ++nonzero)\n {\n row_degrees_in[matrix.column_indices[nonzero]]++;\n }\n\n // Initialize unlabeled set \n typedef std::set > UnlabeledSet;\n typename UnlabeledSet::key_compare unlabeled_comp(row_degrees_in);\n UnlabeledSet unlabeled(unlabeled_comp);\n for (OffsetT row = 0; row < matrix.num_rows; ++row)\n {\n relabel_indices[row] = -1;\n unlabeled.insert(row);\n }\n\n // Initialize queue set\n std::deque q;\n\n // Process unlabeled vertices (traverse connected components)\n OffsetT relabel_idx = 0;\n while (!unlabeled.empty())\n {\n // Seed the unvisited frontier queue with the unlabeled vertex of lowest-degree\n OffsetT vertex = *unlabeled.begin();\n q.push_back(vertex);\n\n while (!q.empty())\n {\n vertex = q.front();\n q.pop_front();\n\n if (relabel_indices[vertex] == -1)\n {\n // Update this vertex\n unlabeled.erase(vertex);\n relabel_indices[vertex] = relabel_idx;\n relabel_idx++;\n\n // Sort neighbors by degree\n OrderByLow neighbor_comp(row_degrees_in);\n std::sort(\n matrix.column_indices + matrix.row_offsets[vertex],\n matrix.column_indices + matrix.row_offsets[vertex + 1],\n neighbor_comp);\n\n // Inspect neighbors, adding to the out frontier if unlabeled\n for (OffsetT neighbor_idx = matrix.row_offsets[vertex];\n neighbor_idx < matrix.row_offsets[vertex + 1];\n ++neighbor_idx)\n {\n OffsetT neighbor = matrix.column_indices[neighbor_idx];\n q.push_back(neighbor);\n }\n }\n }\n }\n\n/*\n // Reverse labels\n for (int row = 0; row < matrix.num_rows; ++row)\n {\n relabel_indices[row] = matrix.num_rows - relabel_indices[row] - 1;\n }\n*/\n\n // Cleanup\n if (row_degrees_in) delete[] row_degrees_in;\n if (row_degrees_out) delete[] row_degrees_out;\n}\n\n\n/**\n * Reverse Cuthill-McKee\n */\ntemplate \nvoid RcmRelabel(\n CsrMatrix& matrix,\n bool verbose = false)\n{\n // Do not process if not square\n if (matrix.num_cols != matrix.num_rows)\n {\n if (verbose) {\n printf(\"RCM transformation ignored (not square)\\n\"); fflush(stdout);\n }\n return;\n }\n\n // Initialize relabel indices\n OffsetT* relabel_indices = new OffsetT[matrix.num_rows];\n\n if (verbose) {\n printf(\"RCM relabeling... \"); fflush(stdout);\n }\n\n RcmRelabel(matrix, relabel_indices);\n\n if (verbose) {\n printf(\"done. Reconstituting... \"); fflush(stdout);\n }\n\n // Create a COO matrix from the relabel indices\n CooMatrix coo_matrix;\n coo_matrix.InitCsrRelabel(matrix, relabel_indices);\n\n // Reconstitute the CSR matrix from the sorted COO tuples\n if (relabel_indices) delete[] relabel_indices;\n matrix.Clear();\n matrix.FromCoo(coo_matrix);\n\n if (verbose) {\n printf(\"done. \"); fflush(stdout);\n }\n}\n\n\n\n\n" }, { "path": "external/cub/experimental/spmv_compare.cu", "content": "/******************************************************************************\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIAeBILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n//---------------------------------------------------------------------\n// SpMV comparison tool\n//---------------------------------------------------------------------\n\n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \n\n#include \"sparse_matrix.h\"\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants, and type declarations\n//---------------------------------------------------------------------\n\nbool g_quiet = false; // Whether to display stats in CSV format\nbool g_verbose = false; // Whether to display output to console\nbool g_verbose2 = false; // Whether to display input to console\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n\n//---------------------------------------------------------------------\n// SpMV verification\n//---------------------------------------------------------------------\n\n// Compute reference SpMV y = Ax\ntemplate <\n typename ValueT,\n typename OffsetT>\nvoid SpmvGold(\n CsrMatrix& a,\n ValueT* vector_x,\n ValueT* vector_y_in,\n ValueT* vector_y_out,\n ValueT alpha,\n ValueT beta)\n{\n for (OffsetT row = 0; row < a.num_rows; ++row)\n {\n ValueT partial = beta * vector_y_in[row];\n for (\n OffsetT offset = a.row_offsets[row];\n offset < a.row_offsets[row + 1];\n ++offset)\n {\n partial += alpha * a.values[offset] * vector_x[a.column_indices[offset]];\n }\n vector_y_out[row] = partial;\n }\n}\n\n\n//---------------------------------------------------------------------\n// GPU I/O proxy\n//---------------------------------------------------------------------\n\n/**\n * Read every matrix nonzero value, read every corresponding vector value\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename ValueT,\n typename OffsetT,\n typename VectorItr>\n__launch_bounds__ (int(BLOCK_THREADS))\n__global__ void NonZeroIoKernel(\n SpmvParams params,\n VectorItr d_vector_x)\n{\n enum\n {\n TILE_ITEMS = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n\n ValueT nonzero = 0.0;\n\n int tile_idx = blockIdx.x;\n\n OffsetT block_offset = tile_idx * TILE_ITEMS;\n\n OffsetT column_indices[ITEMS_PER_THREAD];\n ValueT values[ITEMS_PER_THREAD];\n\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n OffsetT nonzero_idx = block_offset + (ITEM * BLOCK_THREADS) + threadIdx.x;\n\n OffsetT* ci = params.d_column_indices + nonzero_idx;\n ValueT*a = params.d_values + nonzero_idx;\n\n column_indices[ITEM] = (nonzero_idx < params.num_nonzeros) ? *ci : 0;\n values[ITEM] = (nonzero_idx < params.num_nonzeros) ? *a : 0.0;\n }\n\n __syncthreads();\n\n // Read vector\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n ValueT vector_value = ThreadLoad(params.d_vector_x + column_indices[ITEM]);\n nonzero += vector_value * values[ITEM];\n }\n\n __syncthreads();\n\n if (block_offset < params.num_rows)\n {\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n {\n OffsetT row_idx = block_offset + (ITEM * BLOCK_THREADS) + threadIdx.x;\n if (row_idx < params.num_rows)\n {\n OffsetT row_end_offset = ThreadLoad(params.d_row_end_offsets + row_idx);\n\n if ((row_end_offset >= 0) && (nonzero == nonzero))\n params.d_vector_y[row_idx] = nonzero;\n }\n }\n }\n\n}\n\n\n/**\n * Run GPU I/O proxy\n */\ntemplate <\n typename ValueT,\n typename OffsetT>\nfloat TestGpuCsrIoProxy(\n SpmvParams& params,\n int timing_iterations)\n{\n enum {\n BLOCK_THREADS = 128,\n ITEMS_PER_THREAD = 7,\n TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD,\n };\n\n// size_t smem = 1024 * 16;\n size_t smem = 1024 * 0;\n\n unsigned int nonzero_blocks = (params.num_nonzeros + TILE_SIZE - 1) / TILE_SIZE;\n unsigned int row_blocks = (params.num_rows + TILE_SIZE - 1) / TILE_SIZE;\n unsigned int blocks = std::max(nonzero_blocks, row_blocks);\n\n typedef TexRefInputIterator TexItr;\n TexItr x_itr;\n CubDebugExit(x_itr.BindTexture(params.d_vector_x));\n\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get device SM version\n int sm_version;\n CubDebugExit(SmVersion(sm_version, device_ordinal));\n\n void (*kernel)(SpmvParams, TexItr) = NonZeroIoKernel;\n\n\n int spmv_sm_occupancy;\n CubDebugExit(MaxSmOccupancy(spmv_sm_occupancy, kernel, BLOCK_THREADS, smem));\n\n if (!g_quiet)\n printf(\"NonZeroIoKernel<%d,%d><<<%d, %d>>>, sm occupancy %d\\n\", BLOCK_THREADS, ITEMS_PER_THREAD, blocks, BLOCK_THREADS, spmv_sm_occupancy);\n\n // Warmup\n NonZeroIoKernel<<>>(params, x_itr);\n\n // Check for failures\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(SyncStream(0));\n\n // Timing\n GpuTimer timer;\n float elapsed_millis = 0.0;\n timer.Start();\n for (int it = 0; it < timing_iterations; ++it)\n {\n NonZeroIoKernel<<>>(params, x_itr);\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n CubDebugExit(x_itr.UnbindTexture());\n\n return elapsed_millis / timing_iterations;\n}\n\n\n\n//---------------------------------------------------------------------\n// cuSparse HybMV\n//---------------------------------------------------------------------\n\n/**\n * Run cuSparse HYB SpMV (specialized for fp32)\n */\ntemplate <\n typename OffsetT>\nfloat TestCusparseHybmv(\n float* vector_y_in,\n float* reference_vector_y_out,\n SpmvParams& params,\n int timing_iterations,\n cusparseHandle_t cusparse)\n{\n CpuTimer cpu_timer;\n cpu_timer.Start();\n\n // Construct Hyb matrix\n cusparseMatDescr_t mat_desc;\n cusparseHybMat_t hyb_desc;\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&mat_desc));\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateHybMat(&hyb_desc));\n cusparseStatus_t status = cusparseScsr2hyb(\n cusparse,\n params.num_rows, params.num_cols,\n mat_desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n hyb_desc,\n 0,\n CUSPARSE_HYB_PARTITION_AUTO);\n AssertEquals(CUSPARSE_STATUS_SUCCESS, status);\n\n cudaDeviceSynchronize();\n cpu_timer.Stop();\n float elapsed_millis = cpu_timer.ElapsedMillis();\n printf(\"HYB setup ms, %.5f, \", elapsed_millis);\n\n // Reset input/output vector y\n CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));\n\n // Warmup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseShybmv(\n cusparse,\n CUSPARSE_OPERATION_NON_TRANSPOSE,\n ¶ms.alpha, mat_desc,\n hyb_desc,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n\n if (!g_quiet)\n {\n int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);\n printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Timing\n elapsed_millis = 0.0;\n GpuTimer timer;\n\n timer.Start();\n for(int it = 0; it < timing_iterations; ++it)\n {\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseShybmv(\n cusparse,\n CUSPARSE_OPERATION_NON_TRANSPOSE,\n ¶ms.alpha, mat_desc,\n hyb_desc,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n // Cleanup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyHybMat(hyb_desc));\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(mat_desc));\n\n return elapsed_millis / timing_iterations;\n}\n\n\n/**\n * Run cuSparse HYB SpMV (specialized for fp64)\n */\ntemplate <\n typename OffsetT>\nfloat TestCusparseHybmv(\n double* vector_y_in,\n double* reference_vector_y_out,\n SpmvParams& params,\n int timing_iterations,\n cusparseHandle_t cusparse)\n{\n CpuTimer cpu_timer;\n cpu_timer.Start();\n\n // Construct Hyb matrix\n cusparseMatDescr_t mat_desc;\n cusparseHybMat_t hyb_desc;\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&mat_desc));\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateHybMat(&hyb_desc));\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsr2hyb(\n cusparse,\n params.num_rows, params.num_cols,\n mat_desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n hyb_desc,\n 0,\n CUSPARSE_HYB_PARTITION_AUTO));\n\n cudaDeviceSynchronize();\n cpu_timer.Stop();\n float elapsed_millis = cpu_timer.ElapsedMillis();\n printf(\"HYB setup ms, %.5f, \", elapsed_millis);\n\n // Reset input/output vector y\n CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));\n\n // Warmup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDhybmv(\n cusparse,\n CUSPARSE_OPERATION_NON_TRANSPOSE,\n ¶ms.alpha, mat_desc,\n hyb_desc,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n\n if (!g_quiet)\n {\n int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);\n printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Timing\n elapsed_millis = 0.0;\n GpuTimer timer;\n\n timer.Start();\n for(int it = 0; it < timing_iterations; ++it)\n {\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDhybmv(\n cusparse,\n CUSPARSE_OPERATION_NON_TRANSPOSE,\n ¶ms.alpha, mat_desc,\n hyb_desc,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n // Cleanup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyHybMat(hyb_desc));\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(mat_desc));\n\n return elapsed_millis / timing_iterations;\n}\n\n\n\n//---------------------------------------------------------------------\n// cuSparse CsrMV\n//---------------------------------------------------------------------\n\n/**\n * Run cuSparse SpMV (specialized for fp32)\n */\ntemplate <\n typename OffsetT>\nfloat TestCusparseCsrmv(\n float* vector_y_in,\n float* reference_vector_y_out,\n SpmvParams& params,\n int timing_iterations,\n cusparseHandle_t cusparse)\n{\n cusparseMatDescr_t desc;\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&desc));\n\n // Reset input/output vector y\n CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));\n\n // Warmup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseScsrmv(\n cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,\n params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n\n if (!g_quiet)\n {\n int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);\n printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Timing\n float elapsed_millis = 0.0;\n GpuTimer timer;\n\n timer.Start();\n for(int it = 0; it < timing_iterations; ++it)\n {\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseScsrmv(\n cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,\n params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(desc));\n return elapsed_millis / timing_iterations;\n}\n\n\n/**\n * Run cuSparse SpMV (specialized for fp64)\n */\ntemplate <\n typename OffsetT>\nfloat TestCusparseCsrmv(\n double* vector_y_in,\n double* reference_vector_y_out,\n SpmvParams& params,\n int timing_iterations,\n cusparseHandle_t cusparse)\n{\n cusparseMatDescr_t desc;\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreateMatDescr(&desc));\n\n // Reset input/output vector y\n CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(float) * params.num_rows, cudaMemcpyHostToDevice));\n\n // Warmup\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsrmv(\n cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,\n params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n\n if (!g_quiet)\n {\n int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);\n printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Timing\n float elapsed_millis = 0.0;\n GpuTimer timer;\n timer.Start();\n for(int it = 0; it < timing_iterations; ++it)\n {\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDcsrmv(\n cusparse, CUSPARSE_OPERATION_NON_TRANSPOSE,\n params.num_rows, params.num_cols, params.num_nonzeros, ¶ms.alpha, desc,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, ¶ms.beta, params.d_vector_y));\n\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseDestroyMatDescr(desc));\n return elapsed_millis / timing_iterations;\n}\n\n//---------------------------------------------------------------------\n// GPU Merge-based SpMV\n//---------------------------------------------------------------------\n\n/**\n * Run CUB SpMV\n */\ntemplate <\n typename ValueT,\n typename OffsetT>\nfloat TestGpuMergeCsrmv(\n ValueT* vector_y_in,\n ValueT* reference_vector_y_out,\n SpmvParams& params,\n int timing_iterations)\n{\n // Allocate temporary storage\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n\n // Get amount of temporary storage needed\n CubDebugExit(DeviceSpmv::CsrMV(\n d_temp_storage, temp_storage_bytes,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, params.d_vector_y,\n params.num_rows, params.num_cols, params.num_nonzeros,\n// params.alpha, params.beta,\n (cudaStream_t) 0, false));\n\n // Allocate\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Reset input/output vector y\n CubDebugExit(cudaMemcpy(params.d_vector_y, vector_y_in, sizeof(ValueT) * params.num_rows, cudaMemcpyHostToDevice));\n\n // Warmup\n CubDebugExit(DeviceSpmv::CsrMV(\n d_temp_storage, temp_storage_bytes,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, params.d_vector_y,\n params.num_rows, params.num_cols, params.num_nonzeros, \n// params.alpha, params.beta,\n (cudaStream_t) 0, !g_quiet));\n\n if (!g_quiet)\n {\n int compare = CompareDeviceResults(reference_vector_y_out, params.d_vector_y, params.num_rows, true, g_verbose);\n printf(\"\\t%s\\n\", compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Timing\n GpuTimer timer;\n float elapsed_millis = 0.0;\n\n timer.Start();\n for(int it = 0; it < timing_iterations; ++it)\n {\n CubDebugExit(DeviceSpmv::CsrMV(\n d_temp_storage, temp_storage_bytes,\n params.d_values, params.d_row_end_offsets, params.d_column_indices,\n params.d_vector_x, params.d_vector_y,\n params.num_rows, params.num_cols, params.num_nonzeros, \n// params.alpha, params.beta,\n (cudaStream_t) 0, false));\n }\n timer.Stop();\n elapsed_millis += timer.ElapsedMillis();\n\n return elapsed_millis / timing_iterations;\n}\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n/**\n * Display perf\n */\ntemplate \nvoid DisplayPerf(\n float device_giga_bandwidth,\n double avg_millis,\n CsrMatrix& csr_matrix)\n{\n double nz_throughput, effective_bandwidth;\n size_t total_bytes = (csr_matrix.num_nonzeros * (sizeof(ValueT) * 2 + sizeof(OffsetT))) +\n (csr_matrix.num_rows) * (sizeof(OffsetT) + sizeof(ValueT));\n\n nz_throughput = double(csr_matrix.num_nonzeros) / avg_millis / 1.0e6;\n effective_bandwidth = double(total_bytes) / avg_millis / 1.0e6;\n\n if (!g_quiet)\n printf(\"fp%d: %.4f avg ms, %.5f gflops, %.3lf effective GB/s (%.2f%% peak)\\n\",\n sizeof(ValueT) * 8,\n avg_millis,\n 2 * nz_throughput,\n effective_bandwidth,\n effective_bandwidth / device_giga_bandwidth * 100);\n else\n printf(\"%.5f, %.6f, %.3lf, %.2f%%, \",\n avg_millis,\n 2 * nz_throughput,\n effective_bandwidth,\n effective_bandwidth / device_giga_bandwidth * 100);\n\n fflush(stdout);\n}\n\n\n\n/**\n * Run tests\n */\ntemplate <\n typename ValueT,\n typename OffsetT>\nvoid RunTest(\n bool rcm_relabel,\n ValueT alpha,\n ValueT beta,\n CooMatrix& coo_matrix,\n int timing_iterations,\n CommandLineArgs& args)\n{\n // Adaptive timing iterations: run 16 billion nonzeros through\n if (timing_iterations == -1)\n timing_iterations = std::min(50000ull, std::max(100ull, ((16ull << 30) / coo_matrix.num_nonzeros)));\n\n if (!g_quiet)\n printf(\"\\t%d timing iterations\\n\", timing_iterations);\n\n // Convert to CSR\n CsrMatrix csr_matrix;\n csr_matrix.FromCoo(coo_matrix);\n if (!args.CheckCmdLineFlag(\"csrmv\"))\n coo_matrix.Clear();\n\n // Relabel\n if (rcm_relabel)\n {\n if (!g_quiet)\n {\n csr_matrix.Stats().Display();\n printf(\"\\n\");\n csr_matrix.DisplayHistogram();\n printf(\"\\n\");\n if (g_verbose2)\n csr_matrix.Display();\n printf(\"\\n\");\n }\n\n RcmRelabel(csr_matrix, !g_quiet);\n\n if (!g_quiet) printf(\"\\n\");\n }\n\n // Display matrix info\n csr_matrix.Stats().Display(!g_quiet);\n if (!g_quiet)\n {\n printf(\"\\n\");\n csr_matrix.DisplayHistogram();\n printf(\"\\n\");\n if (g_verbose2)\n csr_matrix.Display();\n printf(\"\\n\");\n }\n fflush(stdout);\n\n // Allocate input and output vectors\n ValueT* vector_x = new ValueT[csr_matrix.num_cols];\n ValueT* vector_y_in = new ValueT[csr_matrix.num_rows];\n ValueT* vector_y_out = new ValueT[csr_matrix.num_rows];\n\n for (int col = 0; col < csr_matrix.num_cols; ++col)\n vector_x[col] = 1.0;\n\n for (int row = 0; row < csr_matrix.num_rows; ++row)\n vector_y_in[row] = 1.0;\n\n // Compute reference answer\n SpmvGold(csr_matrix, vector_x, vector_y_in, vector_y_out, alpha, beta);\n\n float avg_millis;\n\n if (g_quiet) {\n printf(\"%s, %s, \", args.deviceProp.name, (sizeof(ValueT) > 4) ? \"fp64\" : \"fp32\"); fflush(stdout);\n }\n\n // Get GPU device bandwidth (GB/s)\n float device_giga_bandwidth = args.device_giga_bandwidth;\n\n // Allocate and initialize GPU problem\n SpmvParams params;\n\n CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_values, sizeof(ValueT) * csr_matrix.num_nonzeros));\n CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_row_end_offsets, sizeof(OffsetT) * (csr_matrix.num_rows + 1)));\n CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_column_indices, sizeof(OffsetT) * csr_matrix.num_nonzeros));\n CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_vector_x, sizeof(ValueT) * csr_matrix.num_cols));\n CubDebugExit(g_allocator.DeviceAllocate((void **) ¶ms.d_vector_y, sizeof(ValueT) * csr_matrix.num_rows));\n params.num_rows = csr_matrix.num_rows;\n params.num_cols = csr_matrix.num_cols;\n params.num_nonzeros = csr_matrix.num_nonzeros;\n params.alpha = alpha;\n params.beta = beta;\n\n CubDebugExit(cudaMemcpy(params.d_values, csr_matrix.values, sizeof(ValueT) * csr_matrix.num_nonzeros, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(params.d_row_end_offsets, csr_matrix.row_offsets, sizeof(OffsetT) * (csr_matrix.num_rows + 1), cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(params.d_column_indices, csr_matrix.column_indices, sizeof(OffsetT) * csr_matrix.num_nonzeros, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(params.d_vector_x, vector_x, sizeof(ValueT) * csr_matrix.num_cols, cudaMemcpyHostToDevice));\n\n if (!g_quiet) printf(\"\\n\\n\");\n printf(\"GPU CSR I/O Prox, \"); fflush(stdout);\n avg_millis = TestGpuCsrIoProxy(params, timing_iterations);\n DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);\n\n if (args.CheckCmdLineFlag(\"csrmv\"))\n {\n if (!g_quiet) printf(\"\\n\\n\");\n printf(\"CUB, \"); fflush(stdout);\n avg_millis = TestGpuMergeCsrmv(vector_y_in, vector_y_out, params, timing_iterations);\n DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);\n }\n\n // Initialize cuSparse\n cusparseHandle_t cusparse;\n AssertEquals(CUSPARSE_STATUS_SUCCESS, cusparseCreate(&cusparse));\n\n if (args.CheckCmdLineFlag(\"csrmv\"))\n {\n if (!g_quiet) printf(\"\\n\\n\");\n printf(\"Cusparse CsrMV, \"); fflush(stdout);\n avg_millis = TestCusparseCsrmv(vector_y_in, vector_y_out, params, timing_iterations, cusparse);\n DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);\n }\n\n if (args.CheckCmdLineFlag(\"hybmv\"))\n {\n if (!g_quiet) printf(\"\\n\\n\");\n printf(\"Cusparse HybMV, \"); fflush(stdout);\n\n avg_millis = TestCusparseHybmv(vector_y_in, vector_y_out, params, timing_iterations, cusparse);\n DisplayPerf(device_giga_bandwidth, avg_millis, csr_matrix);\n }\n\n\n // Cleanup\n if (params.d_values) CubDebugExit(g_allocator.DeviceFree(params.d_values));\n if (params.d_row_end_offsets) CubDebugExit(g_allocator.DeviceFree(params.d_row_end_offsets));\n if (params.d_column_indices) CubDebugExit(g_allocator.DeviceFree(params.d_column_indices));\n if (params.d_vector_x) CubDebugExit(g_allocator.DeviceFree(params.d_vector_x));\n if (params.d_vector_y) CubDebugExit(g_allocator.DeviceFree(params.d_vector_y));\n\n if (vector_x) delete[] vector_x;\n if (vector_y_in) delete[] vector_y_in;\n if (vector_y_out) delete[] vector_y_out;\n}\n\n/**\n * Run tests\n */\ntemplate <\n typename ValueT,\n typename OffsetT>\nvoid RunTests(\n bool rcm_relabel,\n ValueT alpha,\n ValueT beta,\n const std::string& mtx_filename,\n int grid2d,\n int grid3d,\n int wheel,\n int dense,\n int timing_iterations,\n CommandLineArgs& args)\n{\n // Initialize matrix in COO form\n CooMatrix coo_matrix;\n\n if (!mtx_filename.empty())\n {\n // Parse matrix market file\n printf(\"%s, \", mtx_filename.c_str()); fflush(stdout);\n coo_matrix.InitMarket(mtx_filename, 1.0, !g_quiet);\n\n if ((coo_matrix.num_rows == 1) || (coo_matrix.num_cols == 1) || (coo_matrix.num_nonzeros == 1))\n {\n if (!g_quiet) printf(\"Trivial dataset\\n\");\n exit(0);\n }\n }\n else if (grid2d > 0)\n {\n // Generate 2D lattice\n printf(\"grid2d_%d, \", grid2d); fflush(stdout);\n coo_matrix.InitGrid2d(grid2d, false);\n }\n else if (grid3d > 0)\n {\n // Generate 3D lattice\n printf(\"grid3d_%d, \", grid3d); fflush(stdout);\n coo_matrix.InitGrid3d(grid3d, false);\n }\n else if (wheel > 0)\n {\n // Generate wheel graph\n printf(\"wheel_%d, \", grid2d); fflush(stdout);\n coo_matrix.InitWheel(wheel);\n }\n else if (dense > 0)\n {\n // Generate dense graph\n OffsetT size = 1 << 24; // 16M nnz\n args.GetCmdLineArgument(\"size\", size);\n\n OffsetT rows = size / dense;\n printf(\"dense_%d_x_%d, \", rows, dense); fflush(stdout);\n coo_matrix.InitDense(rows, dense);\n }\n else\n {\n fprintf(stderr, \"No graph type specified.\\n\");\n exit(1);\n }\n\n RunTest(\n rcm_relabel,\n alpha,\n beta,\n coo_matrix,\n timing_iterations,\n args);\n}\n\n\n\n/**\n * Main\n */\nint main(int argc, char **argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\n \"%s \"\n \"[--csrmv | --hybmv | --bsrmv ] \"\n \"[--device=] \"\n \"[--quiet] \"\n \"[--v] \"\n \"[--i=] \"\n \"[--fp64] \"\n \"[--rcm] \"\n \"[--alpha=] \"\n \"[--beta=] \"\n \"\\n\\t\"\n \"--mtx= \"\n \"\\n\\t\"\n \"--dense=\"\n \"\\n\\t\"\n \"--grid2d=\"\n \"\\n\\t\"\n \"--grid3d=\"\n \"\\n\\t\"\n \"--wheel=\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n bool fp64;\n bool rcm_relabel;\n std::string mtx_filename;\n int grid2d = -1;\n int grid3d = -1;\n int wheel = -1;\n int dense = -1;\n int timing_iterations = -1;\n float alpha = 1.0;\n float beta = 0.0;\n\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_verbose2 = args.CheckCmdLineFlag(\"v2\");\n g_quiet = args.CheckCmdLineFlag(\"quiet\");\n fp64 = args.CheckCmdLineFlag(\"fp64\");\n rcm_relabel = args.CheckCmdLineFlag(\"rcm\");\n args.GetCmdLineArgument(\"i\", timing_iterations);\n args.GetCmdLineArgument(\"mtx\", mtx_filename);\n args.GetCmdLineArgument(\"grid2d\", grid2d);\n args.GetCmdLineArgument(\"grid3d\", grid3d);\n args.GetCmdLineArgument(\"wheel\", wheel);\n args.GetCmdLineArgument(\"dense\", dense);\n args.GetCmdLineArgument(\"alpha\", alpha);\n args.GetCmdLineArgument(\"beta\", beta);\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Run test(s)\n if (fp64)\n {\n RunTests(rcm_relabel, alpha, beta, mtx_filename, grid2d, grid3d, wheel, dense, timing_iterations, args);\n }\n else\n {\n RunTests(rcm_relabel, alpha, beta, mtx_filename, grid2d, grid3d, wheel, dense, timing_iterations, args);\n }\n\n CubDebugExit(cudaDeviceSynchronize());\n printf(\"\\n\");\n\n return 0;\n}\n" }, { "path": "external/cub/experimental/spmv_script.sh", "content": "#!/bin/bash\n\nfor i in 1 2 4 8 16 32 64 128 256 512 1024 2048 4096 8192 16384 32768 65536 131072 262144 524288 1048576 2097152 4194304 8388608 16777216\ndo\n\techo `date`, `$1 --dense=$i $2 $3 $4 $5 $6 $7`\ndone\n\necho\necho\n\nfor i in `ls /home/dumerrill/graphs/spmv/*.mtx`\ndo\n if [[ ( \"`head -n 50 $i | grep complex`\" = \"\" ) && ( \"`head -n 50 $i | grep array`\" = \"\" ) ]] \n then\n \techo `date`, `$1 --mtx=$i $2 $3 $4 $5 $6 $7 2>/dev/null`\n fi\ndone\n\necho\necho\n\nfor i in `ls /scratch/dumerrill/graphs/mtx/*.mtx`\n#for i in `ls /cygdrive/w/Dev/UFget/mtx/*.mtx`\ndo \n if [[ ( \"`head -n 50 $i | grep complex`\" = \"\" ) && ( \"`head -n 50 $i | grep array`\" = \"\" ) ]] \n then\n \techo `date`, `$1 --mtx=$i $2 $3 $4 $5 $6 $7 2>/dev/null`\n fi\ndone \n\n" }, { "path": "external/cub/test/.gitignore", "content": "/bin\n/link_main.obj\n/dummy/\n" }, { "path": "external/cub/test/Makefile", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n\n\n#-------------------------------------------------------------------------------\n#\n# Makefile usage\n#\n# make [sm=] [cdp=<0|1>] [force32=<0|1>] [abi=<0|1>] [open64=<0|1>] [verbose=<0|1>] [keep=<0|1>] [quicktest=<0|1>] [quickertest=<0|1>]\n#\n#-------------------------------------------------------------------------------\n\ninclude ../common.mk \n \n#-------------------------------------------------------------------------------\n# Commandline Options\n#-------------------------------------------------------------------------------\n\n# Testing mode option (quick/thorough)\nifeq ($(quickertest), 1)\n\tNVCCFLAGS += -DQUICKER_TEST\n\tTEST_SUFFIX = quicker\nelse ifeq ($(quicktest), 1)\n\tNVCCFLAGS += -DQUICK_TEST\n\tTEST_SUFFIX = quick\nelse \n\tTEST_SUFFIX = thorough\n\tNPPI = \nendif\n\n\n# CUDA memcheck (enabled by default) \nifeq ($(memcheck), 0)\n\tMEMCHECK = \nelse \n\tMEMCHECK = cuda-memcheck\nendif\n\n\n#-------------------------------------------------------------------------------\n# Compiler and compilation platform\n#-------------------------------------------------------------------------------\n\n# Includes\nINC += -I$(CUB_DIR) -I$(CUB_DIR)test \n\n# Suffix to append to each binary\nSUFFIX = $(BIN_SUFFIX)_$(TEST_SUFFIX)\n\n# Define test arch\nDEFINES += -DTEST_ARCH=$(TEST_ARCH)\n\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nrwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nDEPS =\t\t\t\t$(CUB_DEPS) \\\n\t\t\t\t\t$(CUB_DIR)test/Makefile \\\n\t\t\t\t\t$(CUB_DIR)test/test_util.h \\\n\t\t\t\t\t$(CUB_DIR)test/mersenne.h \\\n\nBLOCK_REDUCE = \t\ttest_block_reduce_raking \\\n\t \t\t\t\ttest_block_reduce_warp_reductions\t\t\n\n\nBLOCK_SCAN = \t\ttest_block_scan_raking \\\n\t \t\t\t\ttest_block_scan_raking_memoize \\\n\t \t\t\t\ttest_block_scan_warp_scans\t\t\n\n\nBLOCK_RADIX_SORT = \ttest_block_radix_sort_keys \\\n\t \t\t\t\ttest_block_radix_sort_pairs\t\n\n\t\t\nALL = \t\t\t\tlink \\\n\t \t\t\t\ttest_iterator \\\n\t \t\t\t\ttest_allocator \\\n\t \t\t\t\ttest_warp_scan \\\n\t \t\t\t\ttest_warp_reduce \\\n\t \t\t\t\t$(BLOCK_REDUCE) \\\n\t \t\t\t\t$(BLOCK_SCAN) \\\n\t \t\t\t\t$(BLOCK_RADIX_SORT) \\\n\t \t\t\t\ttest_block_load_store \\\n\t \t\t\t\ttest_block_histogram \\\n\t\t\t\t \ttest_device_reduce \\\n\t\t\t \t\ttest_device_histogram \\\n\t\t\t \t\ttest_device_scan \\\n\t\t\t \t\ttest_device_radix_sort \\\n\t\t\t\t\ttest_device_reduce_by_key\\\n\t\t\t\t\ttest_device_run_length_encode\\\n\t\t \t\t\ttest_device_select_unique \\\n\t\t\t\t\ttest_device_select_if \n\t\t\n#\t \ttest_grid_barrier \\\t\tfails on sm110\n#\t \ttest_device_seg_reduce\n\t\t\n\n\n#-------------------------------------------------------------------------------\n# make default\n#-------------------------------------------------------------------------------\n\ndefault:\n\n\n#-------------------------------------------------------------------------------\n# make clean\n#-------------------------------------------------------------------------------\n\nclean :\n\trm -f bin/*$(CPU_ARCH_SUFFIX)* \n\trm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o\n\n\n#-------------------------------------------------------------------------------\n# make all\n#-------------------------------------------------------------------------------\n\nall : $(ALL)\n\n\n#-------------------------------------------------------------------------------\n# make run\n#-------------------------------------------------------------------------------\n\nrun : \n\tfor i in $(ALL); do $(MEMCHECK) ./bin/$${i}_$(SUFFIX) --device=$(device) || exit 1; done\n\nrun_block_reduce : \n\tfor i in $(BLOCK_REDUCE); do $(MEMCHECK) ./bin/$${i}_$(SUFFIX) --device=$(device) || exit 1; done\n\nrun_block_scan : \n\tfor i in $(BLOCK_SCAN); do $(MEMCHECK) ./bin/$${i}_$(SUFFIX) --device=$(device) || exit 1; done\n\nrun_block_radix_sort : \n\tfor i in $(BLOCK_RADIX_SORT); do $(MEMCHECK) ./bin/$${i}_$(SUFFIX) --device=$(device) || exit 1; done\n\n\n\n#-------------------------------------------------------------------------------\n# make link\n#-------------------------------------------------------------------------------\n\nlink : bin/link_$(SUFFIX)\n\nbin/link_$(SUFFIX) : link_a.cu link_b.cu link_main.cpp $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(DEFINES) $(SM_TARGETS) link_a.cu -c -o bin/link_a.obj\n\t$(NVCC) $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(DEFINES) $(SM_TARGETS) link_b.cu -c -o bin/link_b.obj\n\t$(NVCC) $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(DEFINES) $(SM_TARGETS) link_main.cpp bin/link_a.obj bin/link_b.obj -o bin/link_$(SUFFIX)\n\n\n#-------------------------------------------------------------------------------\n# make test_iterator \n#-------------------------------------------------------------------------------\n\ntest_iterator: bin/test_iterator_$(SUFFIX)\n\nbin/test_iterator_$(SUFFIX) : test_iterator.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_iterator_$(SUFFIX) test_iterator.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_allocator \n#-------------------------------------------------------------------------------\n\ntest_allocator: bin/test_allocator_$(SUFFIX)\n\nbin/test_allocator_$(SUFFIX) : test_allocator.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_allocator_$(SUFFIX) test_allocator.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\t\n\t\n#-------------------------------------------------------------------------------\n# make test_grid_barrier \n#-------------------------------------------------------------------------------\n\ntest_grid_barrier: bin/test_grid_barrier_$(SUFFIX)\n\nbin/test_grid_barrier_$(SUFFIX) : test_grid_barrier.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_grid_barrier_$(SUFFIX) test_grid_barrier.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\t\n\n#-------------------------------------------------------------------------------\n# make test_warp_scan \n#-------------------------------------------------------------------------------\n\ntest_warp_scan: bin/test_warp_scan_$(SUFFIX)\n\nbin/test_warp_scan_$(SUFFIX) : test_warp_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_warp_scan_$(SUFFIX) test_warp_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_warp_reduce \n#-------------------------------------------------------------------------------\n\ntest_warp_reduce: bin/test_warp_reduce_$(SUFFIX)\n\nbin/test_warp_reduce_$(SUFFIX) : test_warp_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_warp_reduce_$(SUFFIX) test_warp_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_reduce_raking\n#-------------------------------------------------------------------------------\n\ntest_block_reduce_raking: bin/test_block_reduce_raking_$(SUFFIX)\n\nbin/test_block_reduce_raking_$(SUFFIX) : test_block_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_RAKING $(SM_TARGETS) -o bin/test_block_reduce_raking_$(SUFFIX) test_block_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_reduce_warp_reductions \n#-------------------------------------------------------------------------------\n\ntest_block_reduce_warp_reductions: bin/test_block_reduce_warp_reductions_$(SUFFIX)\n\nbin/test_block_reduce_warp_reductions_$(SUFFIX) : test_block_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_WARP_REDUCTIONS $(SM_TARGETS) -o bin/test_block_reduce_warp_reductions_$(SUFFIX) test_block_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_reduce \n#-------------------------------------------------------------------------------\n\ntest_block_reduce: $(BLOCK_REDUCE)\n\n\n#-------------------------------------------------------------------------------\n# make test_block_scan_raking\n#-------------------------------------------------------------------------------\n\ntest_block_scan_raking: bin/test_block_scan_raking_$(SUFFIX)\n\nbin/test_block_scan_raking_$(SUFFIX) : test_block_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_RAKING $(SM_TARGETS) -o bin/test_block_scan_raking_$(SUFFIX) test_block_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_scan_raking_memoize\n#-------------------------------------------------------------------------------\n\ntest_block_scan_raking_memoize: bin/test_block_scan_raking_memoize_$(SUFFIX)\n\nbin/test_block_scan_raking_memoize_$(SUFFIX) : test_block_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_RAKING_MEMOIZE $(SM_TARGETS) -o bin/test_block_scan_raking_memoize_$(SUFFIX) test_block_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_scan_warp_scans\n#-------------------------------------------------------------------------------\n\ntest_block_scan_warp_scans: bin/test_block_scan_warp_scans_$(SUFFIX)\n\nbin/test_block_scan_warp_scans_$(SUFFIX) : test_block_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_WARP_SCANS $(SM_TARGETS) -o bin/test_block_scan_warp_scans_$(SUFFIX) test_block_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 \n\n\n#-------------------------------------------------------------------------------\n# make test_block_scan \n#-------------------------------------------------------------------------------\n\ntest_block_scan: $(BLOCK_SCAN)\n\n\n#-------------------------------------------------------------------------------\n# make test_block_load_store \n#-------------------------------------------------------------------------------\n\ntest_block_load_store: bin/test_block_load_store_$(SUFFIX)\n\nbin/test_block_load_store_$(SUFFIX) : test_block_load_store.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_block_load_store_$(SUFFIX) test_block_load_store.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\t\n\t\n#-------------------------------------------------------------------------------\n# make test_block_radix_sort_keys \n#-------------------------------------------------------------------------------\n\ntest_block_radix_sort_keys: bin/test_block_radix_sort_keys_$(SUFFIX)\n\nbin/test_block_radix_sort_keys_$(SUFFIX) : test_block_radix_sort.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) -DTEST_KEYS_ONLY $(SM_TARGETS) -o bin/test_block_radix_sort_keys_$(SUFFIX) test_block_radix_sort.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make test_block_radix_sort_pairs \n#-------------------------------------------------------------------------------\n\ntest_block_radix_sort_pairs: bin/test_block_radix_sort_pairs_$(SUFFIX)\n\nbin/test_block_radix_sort_pairs_$(SUFFIX) : test_block_radix_sort.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_block_radix_sort_pairs_$(SUFFIX) test_block_radix_sort.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_block_radix_sort\n#-------------------------------------------------------------------------------\n\ntest_block_radix_sort : $(BLOCK_RADIX_SORT)\n\n\n#-------------------------------------------------------------------------------\n# make test_block_histogram \n#-------------------------------------------------------------------------------\n\ntest_block_histogram: bin/test_block_histogram_$(SUFFIX)\n\nbin/test_block_histogram_$(SUFFIX) : test_block_histogram.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_block_histogram_$(SUFFIX) test_block_histogram.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_reduce\n#-------------------------------------------------------------------------------\n\ntest_device_reduce: bin/test_device_reduce_$(SUFFIX)\n\nbin/test_device_reduce_$(SUFFIX) : test_device_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_reduce_$(SUFFIX) test_device_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_histogram\n#-------------------------------------------------------------------------------\n\ntest_device_histogram: bin/test_device_histogram_$(SUFFIX)\n\nbin/test_device_histogram_$(SUFFIX) : test_device_histogram.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_histogram_$(SUFFIX) test_device_histogram.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) $(NPPI) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_scan\n#-------------------------------------------------------------------------------\n\ntest_device_scan: bin/test_device_scan_$(SUFFIX)\n\nbin/test_device_scan_$(SUFFIX) : test_device_scan.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_scan_$(SUFFIX) test_device_scan.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_radix_sort\n#-------------------------------------------------------------------------------\n\ntest_device_radix_sort: bin/test_device_radix_sort_$(SUFFIX)\n\nbin/test_device_radix_sort_$(SUFFIX) : test_device_radix_sort.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_radix_sort_$(SUFFIX) test_device_radix_sort.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_select_unique\n#-------------------------------------------------------------------------------\n\ntest_device_select_unique: bin/test_device_select_unique_$(SUFFIX)\n\nbin/test_device_select_unique_$(SUFFIX) : test_device_select_unique.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_select_unique_$(SUFFIX) test_device_select_unique.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n#-------------------------------------------------------------------------------\n# make test_device_select_if\n#-------------------------------------------------------------------------------\n\ntest_device_select_if: bin/test_device_select_if_$(SUFFIX)\n\nbin/test_device_select_if_$(SUFFIX) : test_device_select_if.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_select_if_$(SUFFIX) test_device_select_if.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make test_device_reduce_by_key\n#-------------------------------------------------------------------------------\n\ntest_device_reduce_by_key: bin/test_device_reduce_by_key_$(SUFFIX)\n\nbin/test_device_reduce_by_key_$(SUFFIX) : test_device_reduce_by_key.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_reduce_by_key_$(SUFFIX) test_device_reduce_by_key.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n#-------------------------------------------------------------------------------\n# make test_device_run_length_encode\n#-------------------------------------------------------------------------------\n\ntest_device_run_length_encode: bin/test_device_run_length_encode_$(SUFFIX)\n\nbin/test_device_run_length_encode_$(SUFFIX) : test_device_run_length_encode.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_run_length_encode_$(SUFFIX) test_device_run_length_encode.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n\n\n#-------------------------------------------------------------------------------\n# make test_device_seg_reduce\n#-------------------------------------------------------------------------------\n#\n#test_device_seg_reduce: bin/test_device_seg_reduce_$(SUFFIX)\n#\n#bin/test_device_seg_reduce_$(SUFFIX) : test_device_seg_reduce.cu $(DEPS)\n#\tmkdir -p bin\n#\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/test_device_seg_reduce_$(SUFFIX) test_device_seg_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3\n\n\n" }, { "path": "external/cub/test/link_a.cu", "content": "#include \n\nvoid a()\n{\n printf(\"a() called\\n\");\n\n cub::DoubleBuffer d_keys;\n cub::DoubleBuffer d_values;\n size_t temp_storage_bytes = 0;\n cub::DeviceRadixSort::SortPairs(NULL, temp_storage_bytes, d_keys, d_values, 1024);\n}\n" }, { "path": "external/cub/test/link_b.cu", "content": "#include \n\nvoid b()\n{\n printf(\"b() called\\n\");\n\n cub::DoubleBuffer d_keys;\n cub::DoubleBuffer d_values;\n size_t temp_storage_bytes = 0;\n cub::DeviceRadixSort::SortPairs(NULL, temp_storage_bytes, d_keys, d_values, 1024);\n}\n" }, { "path": "external/cub/test/link_main.cpp", "content": "#include \n\nextern void a();\nextern void b();\n\nint main()\n{\n printf(\"hello world\\n\");\n return 0;\n}\n" }, { "path": "external/cub/test/mersenne.h", "content": "/*\n A C-program for MT19937, with initialization improved 2002/1/26.\n Coded by Takuji Nishimura and Makoto Matsumoto.\n\n Before using, initialize the state by using init_genrand(seed)\n or init_by_array(init_key, key_length).\n\n Copyright (C) 1997 - 2002, Makoto Matsumoto and Takuji Nishimura,\n All rights reserved.\n\n Redistribution and use in source and binary forms, with or without\n modification, are permitted provided that the following conditions\n are met:\n\n 1. Redistributions of source code must retain the above copyright\n notice, this list of conditions and the following disclaimer.\n\n 2. Redistributions in binary form must reproduce the above copyright\n notice, this list of conditions and the following disclaimer in the\n documentation and/or other materials provided with the distribution.\n\n 3. The names of its contributors may not be used to endorse or promote\n products derived from this software without specific prior written\n permission.\n\n THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS\n \"AS IS\" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT\n LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR\n A PARTICAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR\n CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,\n EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,\n PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR\n PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF\n LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING\n NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n\n\n Any feedback is very welcome.\n http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/emt.html\n email: m-mat @ math.sci.hiroshima-u.ac.jp (remove space)\n */\n\n#include \n\nnamespace mersenne {\n\n/* Period parameters */\nconst unsigned int N = 624;\nconst unsigned int M = 397;\nconst unsigned int MATRIX_A = 0x9908b0df; /* constant vector a */\nconst unsigned int UPPER_MASK = 0x80000000; /* most significant w-r bits */\nconst unsigned int LOWER_MASK = 0x7fffffff; /* least significant r bits */\n\nstatic unsigned int mt[N]; /* the array for the state vector */\nstatic int mti = N + 1; /* mti==N+1 means mt[N] is not initialized */\n\n/* initializes mt[N] with a seed */\nvoid init_genrand(unsigned int s)\n{\n mt[0] = s & 0xffffffff;\n for (mti = 1; mti < N; mti++)\n {\n mt[mti] = (1812433253 * (mt[mti - 1] ^ (mt[mti - 1] >> 30)) + mti);\n\n /* See Knuth TAOCP Vol2. 3rd Ed. P.106 for mtiplier. */\n /* In the previous versions, MSBs of the seed affect */\n /* only MSBs of the array mt[]. */\n /* 2002/01/09 modified by Makoto Matsumoto */\n\n mt[mti] &= 0xffffffff;\n /* for >32 bit machines */\n }\n}\n\n/* initialize by an array with array-length */\n/* init_key is the array for initializing keys */\n/* key_length is its length */\n/* slight change for C++, 2004/2/26 */\nvoid init_by_array(unsigned int init_key[], int key_length)\n{\n int i, j, k;\n init_genrand(19650218);\n i = 1;\n j = 0;\n k = (N > key_length ? N : key_length);\n for (; k; k--)\n {\n mt[i] = (mt[i] ^ ((mt[i - 1] ^ (mt[i - 1] >> 30)) * 1664525))\n + init_key[j] + j; /* non linear */\n mt[i] &= 0xffffffff; /* for WORDSIZE > 32 machines */\n i++;\n j++;\n if (i >= N)\n {\n mt[0] = mt[N - 1];\n i = 1;\n }\n if (j >= key_length) j = 0;\n }\n for (k = N - 1; k; k--)\n {\n mt[i] = (mt[i] ^ ((mt[i - 1] ^ (mt[i - 1] >> 30)) * 1566083941)) - i; /* non linear */\n mt[i] &= 0xffffffff; /* for WORDSIZE > 32 machines */\n i++;\n if (i >= N)\n {\n mt[0] = mt[N - 1];\n i = 1;\n }\n }\n\n mt[0] = 0x80000000; /* MSB is 1; assuring non-zero initial array */\n}\n\n/* generates a random number on [0,0xffffffff]-interval */\nunsigned int genrand_int32(void)\n{\n unsigned int y;\n static unsigned int mag01[2] = { 0x0, MATRIX_A };\n\n /* mag01[x] = x * MATRIX_A for x=0,1 */\n\n if (mti >= N)\n { /* generate N words at one time */\n int kk;\n\n if (mti == N + 1) /* if init_genrand() has not been called, */\n init_genrand(5489); /* a defat initial seed is used */\n\n for (kk = 0; kk < N - M; kk++)\n {\n y = (mt[kk] & UPPER_MASK) | (mt[kk + 1] & LOWER_MASK);\n mt[kk] = mt[kk + M] ^ (y >> 1) ^ mag01[y & 0x1];\n }\n for (; kk < N - 1; kk++)\n {\n y = (mt[kk] & UPPER_MASK) | (mt[kk + 1] & LOWER_MASK);\n mt[kk] = mt[kk + (M - N)] ^ (y >> 1) ^ mag01[y & 0x1];\n }\n y = (mt[N - 1] & UPPER_MASK) | (mt[0] & LOWER_MASK);\n mt[N - 1] = mt[M - 1] ^ (y >> 1) ^ mag01[y & 0x1];\n\n mti = 0;\n }\n\n y = mt[mti++];\n\n /* Tempering */\n y ^= (y >> 11);\n y ^= (y << 7) & 0x9d2c5680;\n y ^= (y << 15) & 0xefc60000;\n y ^= (y >> 18);\n\n return y;\n}\n\n\n\n} // namespace mersenne\n" }, { "path": "external/cub/test/test_allocator.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test evaluation for caching allocator of device memory\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=]\"\n \"[--bytes=]\"\n \"[--i=]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n#if (CUB_PTX_ARCH == 0)\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get number of GPUs and current GPU\n int num_gpus;\n int initial_gpu;\n int timing_iterations = 10000;\n int timing_bytes = 1024 * 1024;\n\n if (CubDebug(cudaGetDeviceCount(&num_gpus))) exit(1);\n if (CubDebug(cudaGetDevice(&initial_gpu))) exit(1);\n args.GetCmdLineArgument(\"i\", timing_iterations);\n args.GetCmdLineArgument(\"bytes\", timing_bytes);\n\n // Create default allocator (caches up to 6MB in device allocations per GPU)\n CachingDeviceAllocator allocator;\n allocator.debug = true;\n\n printf(\"Running single-gpu tests...\\n\"); fflush(stdout);\n\n //\n // Test0\n //\n\n // Create a new stream\n cudaStream_t other_stream;\n CubDebugExit(cudaStreamCreate(&other_stream));\n\n // Allocate 999 bytes on the current gpu in stream0\n char *d_999B_stream0_a;\n char *d_999B_stream0_b;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_a, 999, 0));\n\n // Run some big kernel in stream 0\n EmptyKernel<<<32000, 512, 1024 * 8, 0>>>();\n\n // Free d_999B_stream0_a\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_a));\n\n // Allocate another 999 bytes in stream 0\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_b, 999, 0));\n\n // Check that that we have 1 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 1);\n\n // Check that that we have no cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 0);\n\n // Run some big kernel in stream 0\n EmptyKernel<<<32000, 512, 1024 * 8, 0>>>();\n\n // Free d_999B_stream0_b\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_b));\n\n // Allocate 999 bytes on the current gpu in other_stream\n char *d_999B_stream_other_a;\n char *d_999B_stream_other_b;\n allocator.DeviceAllocate((void **) &d_999B_stream_other_a, 999, other_stream);\n\n // Check that that we have 1 live blocks on the initial GPU (that we allocated a new one because d_999B_stream0_b is only available for stream 0 until it becomes idle)\n AssertEquals(allocator.live_blocks.size(), 1);\n\n // Check that that we have one cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 1);\n\n // Run some big kernel in other_stream\n EmptyKernel<<<32000, 512, 1024 * 8, other_stream>>>();\n\n // Free d_999B_stream_other\n CubDebugExit(allocator.DeviceFree(d_999B_stream_other_a));\n\n // Check that we can now use both allocations in stream 0 after synchronizing the device\n CubDebugExit(cudaDeviceSynchronize());\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_a, 999, 0));\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_b, 999, 0));\n\n // Check that that we have 2 live blocks on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 2);\n\n // Check that that we have no cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 0);\n\n // Free d_999B_stream0_a and d_999B_stream0_b\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_a));\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_b));\n\n // Check that we can now use both allocations in other_stream\n CubDebugExit(cudaDeviceSynchronize());\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream_other_a, 999, other_stream));\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream_other_b, 999, other_stream));\n\n // Check that that we have 2 live blocks on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 2);\n\n // Check that that we have no cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 0);\n\n // Run some big kernel in other_stream\n EmptyKernel<<<32000, 512, 1024 * 8, other_stream>>>();\n\n // Free d_999B_stream_other_a and d_999B_stream_other_b\n CubDebugExit(allocator.DeviceFree(d_999B_stream_other_a));\n CubDebugExit(allocator.DeviceFree(d_999B_stream_other_b));\n\n // Check that we can now use both allocations in stream 0 after synchronizing the device and destroying the other stream\n CubDebugExit(cudaDeviceSynchronize());\n CubDebugExit(cudaStreamDestroy(other_stream));\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_a, 999, 0));\n CubDebugExit(allocator.DeviceAllocate((void **) &d_999B_stream0_b, 999, 0));\n\n // Check that that we have 2 live blocks on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 2);\n\n // Check that that we have no cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 0);\n\n // Free d_999B_stream0_a and d_999B_stream0_b\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_a));\n CubDebugExit(allocator.DeviceFree(d_999B_stream0_b));\n\n // Free all cached\n CubDebugExit(allocator.FreeAllCached());\n\n //\n // Test1\n //\n\n // Allocate 5 bytes on the current gpu\n char *d_5B;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_5B, 5));\n\n // Check that that we have zero free bytes cached on the initial GPU\n AssertEquals(allocator.cached_bytes[initial_gpu].free, 0);\n\n // Check that that we have 1 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 1);\n\n //\n // Test2\n //\n\n // Allocate 4096 bytes on the current gpu\n char *d_4096B;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_4096B, 4096));\n\n // Check that that we have 2 live blocks on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 2);\n\n //\n // Test3\n //\n\n // DeviceFree d_5B\n CubDebugExit(allocator.DeviceFree(d_5B));\n\n // Check that that we have min_bin_bytes free bytes cached on the initial gpu\n AssertEquals(allocator.cached_bytes[initial_gpu].free, allocator.min_bin_bytes);\n\n // Check that that we have 1 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 1);\n\n // Check that that we have 1 cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 1);\n\n //\n // Test4\n //\n\n // DeviceFree d_4096B\n CubDebugExit(allocator.DeviceFree(d_4096B));\n\n // Check that that we have the 4096 + min_bin free bytes cached on the initial gpu\n AssertEquals(allocator.cached_bytes[initial_gpu].free, allocator.min_bin_bytes + 4096);\n\n // Check that that we have 0 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 0);\n\n // Check that that we have 2 cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 2);\n\n //\n // Test5\n //\n\n // Allocate 768 bytes on the current gpu\n char *d_768B;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_768B, 768));\n\n // Check that that we have the min_bin free bytes cached on the initial gpu (4096 was reused)\n AssertEquals(allocator.cached_bytes[initial_gpu].free, allocator.min_bin_bytes);\n\n // Check that that we have 1 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 1);\n\n // Check that that we have 1 cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 1);\n\n //\n // Test6\n //\n\n // Allocate max_cached_bytes on the current gpu\n char *d_max_cached;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_max_cached, allocator.max_cached_bytes));\n\n // DeviceFree d_max_cached\n CubDebugExit(allocator.DeviceFree(d_max_cached));\n\n // Check that that we have the min_bin free bytes cached on the initial gpu (max cached was not returned because we went over)\n AssertEquals(allocator.cached_bytes[initial_gpu].free, allocator.min_bin_bytes);\n\n // Check that that we have 1 live block on the initial GPU\n AssertEquals(allocator.live_blocks.size(), 1);\n\n // Check that that we still have 1 cached block on the initial GPU\n AssertEquals(allocator.cached_blocks.size(), 1);\n\n //\n // Test7\n //\n\n // Free all cached blocks on all GPUs\n CubDebugExit(allocator.FreeAllCached());\n\n // Check that that we have 0 bytes cached on the initial GPU\n AssertEquals(allocator.cached_bytes[initial_gpu].free, 0);\n\n // Check that that we have 0 cached blocks across all GPUs\n AssertEquals(allocator.cached_blocks.size(), 0);\n\n // Check that that still we have 1 live block across all GPUs\n AssertEquals(allocator.live_blocks.size(), 1);\n\n //\n // Test8\n //\n\n // Allocate max cached bytes + 1 on the current gpu\n char *d_max_cached_plus;\n CubDebugExit(allocator.DeviceAllocate((void **) &d_max_cached_plus, allocator.max_cached_bytes + 1));\n\n // DeviceFree max cached bytes\n CubDebugExit(allocator.DeviceFree(d_max_cached_plus));\n\n // DeviceFree d_768B\n CubDebugExit(allocator.DeviceFree(d_768B));\n\n unsigned int power;\n size_t rounded_bytes;\n allocator.NearestPowerOf(power, rounded_bytes, allocator.bin_growth, 768);\n\n // Check that that we have 4096 free bytes cached on the initial gpu\n AssertEquals(allocator.cached_bytes[initial_gpu].free, rounded_bytes);\n\n // Check that that we have 1 cached blocks across all GPUs\n AssertEquals(allocator.cached_blocks.size(), 1);\n\n // Check that that still we have 0 live block across all GPUs\n AssertEquals(allocator.live_blocks.size(), 0);\n\n#ifndef CUB_CDP\n // BUG: find out why these tests fail when one GPU is CDP compliant and the other is not\n\n if (num_gpus > 1)\n {\n printf(\"\\nRunning multi-gpu tests...\\n\"); fflush(stdout);\n\n //\n // Test9\n //\n\n // Allocate 768 bytes on the next gpu\n int next_gpu = (initial_gpu + 1) % num_gpus;\n char *d_768B_2;\n CubDebugExit(allocator.DeviceAllocate(next_gpu, (void **) &d_768B_2, 768));\n\n // DeviceFree d_768B on the next gpu\n CubDebugExit(allocator.DeviceFree(next_gpu, d_768B_2));\n\n // Re-allocate 768 bytes on the next gpu\n CubDebugExit(allocator.DeviceAllocate(next_gpu, (void **) &d_768B_2, 768));\n\n // Re-free d_768B on the next gpu\n CubDebugExit(allocator.DeviceFree(next_gpu, d_768B_2));\n\n // Check that that we have 4096 free bytes cached on the initial gpu\n AssertEquals(allocator.cached_bytes[initial_gpu].free, rounded_bytes);\n\n // Check that that we have 4096 free bytes cached on the second gpu\n AssertEquals(allocator.cached_bytes[next_gpu].free, rounded_bytes);\n\n // Check that that we have 2 cached blocks across all GPUs\n AssertEquals(allocator.cached_blocks.size(), 2);\n\n // Check that that still we have 0 live block across all GPUs\n AssertEquals(allocator.live_blocks.size(), 0);\n }\n#endif // CUB_CDP\n\n //\n // Performance\n //\n\n printf(\"\\nCPU Performance (%d timing iterations, %d bytes):\\n\", timing_iterations, timing_bytes);\n fflush(stdout); fflush(stderr);\n\n // CPU performance comparisons vs cached. Allocate and free a 1MB block 2000 times\n CpuTimer cpu_timer;\n char *d_1024MB = NULL;\n allocator.debug = false;\n\n // Prime the caching allocator and the kernel\n CubDebugExit(allocator.DeviceAllocate((void **) &d_1024MB, timing_bytes));\n CubDebugExit(allocator.DeviceFree(d_1024MB));\n cub::EmptyKernel<<<1, 32>>>();\n\n // CUDA\n cpu_timer.Start();\n for (int i = 0; i < timing_iterations; ++i)\n {\n CubDebugExit(cudaMalloc((void **) &d_1024MB, timing_bytes));\n CubDebugExit(cudaFree(d_1024MB));\n }\n cpu_timer.Stop();\n float cuda_malloc_elapsed_millis = cpu_timer.ElapsedMillis();\n\n // CUB\n cpu_timer.Start();\n for (int i = 0; i < timing_iterations; ++i)\n {\n CubDebugExit(allocator.DeviceAllocate((void **) &d_1024MB, timing_bytes));\n CubDebugExit(allocator.DeviceFree(d_1024MB));\n }\n cpu_timer.Stop();\n float cub_calloc_elapsed_millis = cpu_timer.ElapsedMillis();\n\n printf(\"\\t CUB CachingDeviceAllocator allocation CPU speedup: %.2f (avg cudaMalloc %.4f ms vs. avg DeviceAllocate %.4f ms)\\n\",\n cuda_malloc_elapsed_millis / cub_calloc_elapsed_millis,\n cuda_malloc_elapsed_millis / timing_iterations,\n cub_calloc_elapsed_millis / timing_iterations);\n\n // GPU performance comparisons. Allocate and free a 1MB block 2000 times\n GpuTimer gpu_timer;\n\n printf(\"\\nGPU Performance (%d timing iterations, %d bytes):\\n\", timing_iterations, timing_bytes);\n fflush(stdout); fflush(stderr);\n\n // Kernel-only\n gpu_timer.Start();\n for (int i = 0; i < timing_iterations; ++i)\n {\n cub::EmptyKernel<<<1, 32>>>();\n }\n gpu_timer.Stop();\n float cuda_empty_elapsed_millis = gpu_timer.ElapsedMillis();\n\n // CUDA\n gpu_timer.Start();\n for (int i = 0; i < timing_iterations; ++i)\n {\n CubDebugExit(cudaMalloc((void **) &d_1024MB, timing_bytes));\n cub::EmptyKernel<<<1, 32>>>();\n CubDebugExit(cudaFree(d_1024MB));\n }\n gpu_timer.Stop();\n cuda_malloc_elapsed_millis = gpu_timer.ElapsedMillis() - cuda_empty_elapsed_millis;\n\n // CUB\n gpu_timer.Start();\n for (int i = 0; i < timing_iterations; ++i)\n {\n CubDebugExit(allocator.DeviceAllocate((void **) &d_1024MB, timing_bytes));\n cub::EmptyKernel<<<1, 32>>>();\n CubDebugExit(allocator.DeviceFree(d_1024MB));\n }\n gpu_timer.Stop();\n cub_calloc_elapsed_millis = gpu_timer.ElapsedMillis() - cuda_empty_elapsed_millis;\n\n printf(\"\\t CUB CachingDeviceAllocator allocation GPU speedup: %.2f (avg cudaMalloc %.4f ms vs. avg DeviceAllocate %.4f ms)\\n\",\n cuda_malloc_elapsed_millis / cub_calloc_elapsed_millis,\n cuda_malloc_elapsed_millis / timing_iterations,\n cub_calloc_elapsed_millis / timing_iterations);\n\n\n#endif\n\n printf(\"Success\\n\");\n\n return 0;\n}\n\n" }, { "path": "external/cub/test/test_block_histogram.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of BlockHistogram utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/**\n * BlockHistogram test kernel.\n */\ntemplate <\n int BINS,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockHistogramAlgorithm ALGORITHM,\n typename T,\n typename HistoCounter>\n__global__ void BlockHistogramKernel(\n T *d_samples,\n HistoCounter *d_histogram)\n{\n // Parameterize BlockHistogram type for our thread block\n typedef BlockHistogram BlockHistogram;\n\n // Allocate temp storage in shared memory\n __shared__ typename BlockHistogram::TempStorage temp_storage;\n\n // Per-thread tile data\n T data[ITEMS_PER_THREAD];\n LoadDirectStriped(threadIdx.x, d_samples, data);\n\n // Test histo (writing directly to histogram buffer in global)\n BlockHistogram(temp_storage).Histogram(data, d_histogram);\n}\n\n\n/**\n * Initialize problem (and solution)\n */\ntemplate <\n int BINS,\n typename SampleT>\nvoid Initialize(\n GenMode gen_mode,\n SampleT *h_samples,\n int *h_histograms_linear,\n int num_samples)\n{\n // Init bins\n for (int bin = 0; bin < BINS; ++bin)\n {\n h_histograms_linear[bin] = 0;\n }\n\n if (g_verbose) printf(\"Samples: \\n\");\n\n // Initialize interleaved channel samples and histogram them correspondingly\n for (int i = 0; i < num_samples; ++i)\n {\n InitValue(gen_mode, h_samples[i], i);\n h_samples[i] %= BINS;\n\n if (g_verbose) std::cout << CoutCast(h_samples[i]) << \", \";\n\n h_histograms_linear[h_samples[i]]++;\n }\n\n if (g_verbose) printf(\"\\n\\n\");\n}\n\n\n/**\n * Test BlockHistogram\n */\ntemplate <\n typename SampleT,\n int BINS,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockHistogramAlgorithm ALGORITHM>\nvoid Test(\n GenMode gen_mode)\n{\n int num_samples = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n printf(\"cub::BlockHistogram %s %d %s samples (%dB), %d bins, %d threads, gen-mode %s\\n\",\n (ALGORITHM == BLOCK_HISTO_SORT) ? \"BLOCK_HISTO_SORT\" : \"BLOCK_HISTO_ATOMIC\",\n num_samples,\n typeid(SampleT).name(),\n (int) sizeof(SampleT),\n BINS,\n BLOCK_THREADS,\n (gen_mode == RANDOM) ? \"RANDOM\" : (gen_mode == INTEGER_SEED) ? \"SEQUENTIAL\" : \"HOMOGENOUS\");\n fflush(stdout);\n\n // Allocate host arrays\n SampleT *h_samples = new SampleT[num_samples];\n int *h_reference = new int[BINS];\n\n // Initialize problem\n Initialize(gen_mode, h_samples, h_reference, num_samples);\n\n // Allocate problem device arrays\n SampleT *d_samples = NULL;\n int *d_histogram = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_samples, sizeof(SampleT) * num_samples));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_histogram, sizeof(int) * BINS));\n\n // Initialize/clear device arrays\n CubDebugExit(cudaMemcpy(d_samples, h_samples, sizeof(SampleT) * num_samples, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_histogram, 0, sizeof(int) * BINS));\n\n // Run kernel\n BlockHistogramKernel<<<1, BLOCK_THREADS>>>(\n d_samples,\n d_histogram);\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults((int*) h_reference, d_histogram, BINS, g_verbose, g_verbose);\n printf(\"\\t%s\\n\\n\", compare ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n fflush(stdout);\n fflush(stderr);\n\n // Cleanup\n if (h_samples) delete[] h_samples;\n if (h_reference) delete[] h_reference;\n if (d_samples) CubDebugExit(g_allocator.DeviceFree(d_samples));\n if (d_histogram) CubDebugExit(g_allocator.DeviceFree(d_histogram));\n\n // Correctness asserts\n AssertEquals(0, compare);\n}\n\n\n/**\n * Test different sample distributions\n */\ntemplate <\n typename SampleT,\n int BINS,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockHistogramAlgorithm ALGORITHM>\nvoid Test()\n{\n Test(UNIFORM);\n Test(INTEGER_SEED);\n Test(RANDOM);\n}\n\n\n/**\n * Test different ALGORITHM\n */\ntemplate <\n typename SampleT,\n int BINS,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid Test()\n{\n Test();\n Test();\n}\n\n\n/**\n * Test different ITEMS_PER_THREAD\n */\ntemplate <\n typename SampleT,\n int BINS,\n int BLOCK_THREADS>\nvoid Test()\n{\n Test();\n Test();\n}\n\n\n/**\n * Test different BLOCK_THREADS\n */\ntemplate <\n typename SampleT,\n int BINS>\nvoid Test()\n{\n Test();\n Test();\n Test();\n}\n\n\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#ifdef QUICK_TEST\n\n // Compile/run quick tests\n Test(RANDOM);\n Test(RANDOM);\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n Test();\n Test();\n Test();\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_block_load_store.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of BlockLoad and BlockStore utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nCachingDeviceAllocator g_allocator(true);\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n\n/**\n * Test load/store kernel.\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM,\n typename InputIteratorT,\n typename OutputIteratorT>\n__launch_bounds__ (BLOCK_THREADS, 1)\n__global__ void Kernel(\n InputIteratorT d_in,\n OutputIteratorT d_out_unguarded,\n OutputIteratorT d_out_guarded,\n int num_items)\n{\n enum\n {\n TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD\n };\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Threadblock load/store abstraction types\n typedef BlockLoad BlockLoad;\n typedef BlockStore BlockStore;\n\n // Shared memory type for this thread block\n union TempStorage\n {\n typename BlockLoad::TempStorage load;\n typename BlockStore::TempStorage store;\n };\n\n // Allocate temp storage in shared memory\n __shared__ TempStorage temp_storage;\n\n // Threadblock work bounds\n int block_offset = blockIdx.x * TILE_SIZE;\n int guarded_elements = num_items - block_offset;\n\n // Tile of items\n OutputT data[ITEMS_PER_THREAD];\n\n // Load data\n BlockLoad(temp_storage.load).Load(d_in + block_offset, data);\n\n __syncthreads();\n\n // Store data\n BlockStore(temp_storage.store).Store(d_out_unguarded + block_offset, data);\n\n __syncthreads();\n\n // reset data\n #pragma unroll\n for (int ITEM = 0; ITEM < ITEMS_PER_THREAD; ++ITEM)\n data[ITEM] = OutputT();\n\n __syncthreads();\n\n // Load data\n BlockLoad(temp_storage.load).Load(d_in + block_offset, data, guarded_elements);\n\n __syncthreads();\n\n // Store data\n BlockStore(temp_storage.store).Store(d_out_guarded + block_offset, data, guarded_elements);\n}\n\n\n//---------------------------------------------------------------------\n// Host testing subroutines\n//---------------------------------------------------------------------\n\n\n/**\n * Test load/store variants\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM,\n typename InputIteratorT,\n typename OutputIteratorT>\nvoid TestKernel(\n T *h_in,\n InputIteratorT d_in,\n OutputIteratorT d_out_unguarded_itr,\n OutputIteratorT d_out_guarded_itr,\n T *d_out_unguarded_ptr,\n T *d_out_guarded_ptr,\n int grid_size,\n int guarded_elements)\n{\n int compare;\n\n int unguarded_elements = grid_size * BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Test with discard output iterator\n typedef typename std::iterator_traits::difference_type OffsetT;\n DiscardOutputIterator discard_itr;\n\n Kernel\n <<>>(\n d_in,\n discard_itr,\n discard_itr,\n guarded_elements);\n\n // Test with regular output iterator\n Kernel\n <<>>(\n d_in,\n d_out_unguarded_itr,\n d_out_guarded_itr,\n guarded_elements);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Check results\n compare = CompareDeviceResults(h_in, d_out_guarded_ptr, guarded_elements, g_verbose, g_verbose);\n printf(\"\\tGuarded: %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Check results\n compare = CompareDeviceResults(h_in, d_out_unguarded_ptr, unguarded_elements, g_verbose, g_verbose);\n printf(\"\\tUnguarded: %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n}\n\n\n/**\n * Test native pointer. Specialized for sufficient resources\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM>\nvoid TestNative(\n int grid_size,\n float fraction_valid,\n Int2Type sufficient_resources)\n{\n int unguarded_elements = grid_size * BLOCK_THREADS * ITEMS_PER_THREAD;\n int guarded_elements = int(fraction_valid * float(unguarded_elements));\n\n // Allocate host arrays\n T *h_in = (T*) malloc(unguarded_elements * sizeof(T));\n\n // Allocate device arrays\n T *d_in = NULL;\n T *d_out_unguarded = NULL;\n T *d_out_guarded = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * unguarded_elements));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out_unguarded, sizeof(T) * unguarded_elements));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out_guarded, sizeof(T) * guarded_elements));\n CubDebugExit(cudaMemset(d_out_unguarded, 0, sizeof(T) * unguarded_elements));\n CubDebugExit(cudaMemset(d_out_guarded, 0, sizeof(T) * guarded_elements));\n\n // Initialize problem on host and device\n for (int i = 0; i < unguarded_elements; ++i)\n {\n InitValue(INTEGER_SEED, h_in[i], i);\n }\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * unguarded_elements, cudaMemcpyHostToDevice));\n\n printf(\"TestNative \"\n \"grid_size(%d) \"\n \"guarded_elements(%d) \"\n \"unguarded_elements(%d) \"\n \"BLOCK_THREADS(%d) \"\n \"ITEMS_PER_THREAD(%d) \"\n \"LOAD_ALGORITHM(%d) \"\n \"STORE_ALGORITHM(%d) \"\n \"sizeof(T)(%d)\\n\",\n grid_size, guarded_elements, unguarded_elements, BLOCK_THREADS, ITEMS_PER_THREAD, LOAD_ALGORITHM, STORE_ALGORITHM, (int) sizeof(T));\n\n TestKernel(\n h_in,\n (T const *) d_in, // Test const\n d_out_unguarded,\n d_out_guarded,\n d_out_unguarded,\n d_out_guarded,\n grid_size,\n guarded_elements);\n\n // Cleanup\n if (h_in) free(h_in);\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out_unguarded) CubDebugExit(g_allocator.DeviceFree(d_out_unguarded));\n if (d_out_guarded) CubDebugExit(g_allocator.DeviceFree(d_out_guarded));\n}\n\n\n/**\n * Test native pointer. Specialized for insufficient resources\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM>\nvoid TestNative(\n int grid_size,\n float fraction_valid,\n Int2Type sufficient_resources)\n{}\n\n\n/**\n * Test iterator. Specialized for sufficient resources.\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM,\n CacheLoadModifier LOAD_MODIFIER,\n CacheStoreModifier STORE_MODIFIER>\nvoid TestIterator(\n int grid_size,\n float fraction_valid,\n Int2Type sufficient_resources)\n{\n int unguarded_elements = grid_size * BLOCK_THREADS * ITEMS_PER_THREAD;\n int guarded_elements = int(fraction_valid * float(unguarded_elements));\n\n // Allocate host arrays\n T *h_in = (T*) malloc(unguarded_elements * sizeof(T));\n\n // Allocate device arrays\n T *d_in = NULL;\n T *d_out_unguarded = NULL;\n T *d_out_guarded = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * unguarded_elements));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out_unguarded, sizeof(T) * unguarded_elements));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out_guarded, sizeof(T) * guarded_elements));\n CubDebugExit(cudaMemset(d_out_unguarded, 0, sizeof(T) * unguarded_elements));\n CubDebugExit(cudaMemset(d_out_guarded, 0, sizeof(T) * guarded_elements));\n\n // Initialize problem on host and device\n for (int i = 0; i < unguarded_elements; ++i)\n {\n InitValue(INTEGER_SEED, h_in[i], i);\n }\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * unguarded_elements, cudaMemcpyHostToDevice));\n\n printf(\"TestIterator \"\n \"grid_size(%d) \"\n \"guarded_elements(%d) \"\n \"unguarded_elements(%d) \"\n \"BLOCK_THREADS(%d) \"\n \"ITEMS_PER_THREAD(%d) \"\n \"LOAD_ALGORITHM(%d) \"\n \"STORE_ALGORITHM(%d) \"\n \"LOAD_MODIFIER(%d) \"\n \"STORE_MODIFIER(%d) \"\n \"sizeof(T)(%d)\\n\",\n grid_size, guarded_elements, unguarded_elements, BLOCK_THREADS, ITEMS_PER_THREAD, LOAD_ALGORITHM, STORE_ALGORITHM, LOAD_MODIFIER, STORE_MODIFIER, (int) sizeof(T));\n\n TestKernel(\n h_in,\n CacheModifiedInputIterator(d_in),\n CacheModifiedOutputIterator(d_out_unguarded),\n CacheModifiedOutputIterator(d_out_guarded),\n d_out_unguarded,\n d_out_guarded,\n grid_size,\n guarded_elements);\n\n // Cleanup\n if (h_in) free(h_in);\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out_unguarded) CubDebugExit(g_allocator.DeviceFree(d_out_unguarded));\n if (d_out_guarded) CubDebugExit(g_allocator.DeviceFree(d_out_guarded));\n}\n\n/**\n * Test iterator. Specialized for insufficient resources.\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM,\n CacheLoadModifier LOAD_MODIFIER,\n CacheStoreModifier STORE_MODIFIER>\nvoid TestIterator(\n int grid_size,\n float fraction_valid,\n Int2Type sufficient_resources)\n{}\n\n\n/**\n * Evaluate different pointer access types\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM>\nvoid TestPointerType(\n int grid_size,\n float fraction_valid)\n{\n // Threadblock load/store abstraction types\n typedef BlockLoad BlockLoad;\n typedef BlockStore BlockStore;\n\n#if defined(SM100) || defined(SM110) || defined(SM130)\n static const bool sufficient_load_smem = sizeof(typename BlockLoad::TempStorage) <= 1024 * 16;\n static const bool sufficient_store_smem = sizeof(typename BlockStore::TempStorage) <= 1024 * 16;\n static const bool sufficient_threads = BLOCK_THREADS <= 512;\n#else\n static const bool sufficient_load_smem = sizeof(typename BlockLoad::TempStorage) <= 1024 * 48;\n static const bool sufficient_store_smem = sizeof(typename BlockStore::TempStorage) <= 1024 * 48;\n static const bool sufficient_threads = BLOCK_THREADS <= 1024;\n#endif\n\n static const bool sufficient_resources = sufficient_load_smem && sufficient_store_smem && sufficient_threads;\n\n TestNative(grid_size, fraction_valid, Int2Type());\n TestIterator(grid_size, fraction_valid, Int2Type());\n}\n\n\n/**\n * Evaluate different time-slicing strategies\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n BlockLoadAlgorithm LOAD_ALGORITHM,\n BlockStoreAlgorithm STORE_ALGORITHM>\nvoid TestSlicedStrategy(\n int grid_size,\n float fraction_valid)\n{\n TestPointerType(grid_size, fraction_valid);\n TestPointerType(grid_size, fraction_valid);\n}\n\n\n\n/**\n * Evaluate different load/store strategies (specialized for block sizes that are not a multiple of 32)\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid TestStrategy(\n int grid_size,\n float fraction_valid,\n Int2Type is_warp_multiple)\n{\n TestPointerType(grid_size, fraction_valid);\n TestPointerType(grid_size, fraction_valid);\n TestPointerType(grid_size, fraction_valid);\n}\n\n\n/**\n * Evaluate different load/store strategies (specialized for block sizes that are a multiple of 32)\n */\ntemplate <\n typename T,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid TestStrategy(\n int grid_size,\n float fraction_valid,\n Int2Type is_warp_multiple)\n{\n TestStrategy(grid_size, fraction_valid, Int2Type());\n TestPointerType(grid_size, fraction_valid);\n TestPointerType(grid_size, fraction_valid);\n}\n\n\n/**\n * Evaluate different register blocking\n */\ntemplate <\n typename T,\n int BLOCK_THREADS>\nvoid TestItemsPerThread(\n int grid_size,\n float fraction_valid)\n{\n Int2Type is_warp_multiple;\n\n TestStrategy(grid_size, fraction_valid, is_warp_multiple);\n TestStrategy(grid_size, fraction_valid, is_warp_multiple);\n TestStrategy(grid_size, fraction_valid, is_warp_multiple);\n TestStrategy(grid_size, fraction_valid, is_warp_multiple);\n}\n\n\n/**\n * Evaluate different thread block sizes\n */\ntemplate \nvoid TestThreads(\n int grid_size,\n float fraction_valid)\n{\n TestItemsPerThread(grid_size, fraction_valid);\n TestItemsPerThread(grid_size, fraction_valid);\n TestItemsPerThread(grid_size, fraction_valid);\n TestItemsPerThread(grid_size, fraction_valid);\n TestItemsPerThread(grid_size, fraction_valid);\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef QUICK_TEST\n\n // Compile/run quick tests\n TestNative< int, 64, 2, BLOCK_LOAD_WARP_TRANSPOSE, BLOCK_STORE_WARP_TRANSPOSE>(1, 0.8f, Int2Type());\n TestIterator< int, 64, 2, BLOCK_LOAD_WARP_TRANSPOSE, BLOCK_STORE_WARP_TRANSPOSE, LOAD_DEFAULT, STORE_DEFAULT>(1, 0.8f, Int2Type());\n\n#else\n\n // Compile/run thorough tests\n TestThreads(2, 0.8f);\n TestThreads(2, 0.8f);\n TestThreads(2, 0.8f);\n TestThreads(2, 0.8f);\n\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n TestThreads(2, 0.8f);\n TestThreads(2, 0.8f);\n TestThreads(2, 0.8f);\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_block_radix_sort.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of BlockRadixSort utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nCachingDeviceAllocator g_allocator(true);\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n\n/// Specialized descending, blocked -> blocked\ntemplate \n__device__ __forceinline__ void TestBlockSort(\n typename BlockRadixSort::TempStorage &temp_storage,\n Key (&keys)[ITEMS_PER_THREAD],\n Value (&values)[ITEMS_PER_THREAD],\n Key *d_keys,\n Value *d_values,\n int begin_bit,\n int end_bit,\n clock_t &stop,\n Int2Type is_descending,\n Int2Type is_blocked_output)\n{\n BlockRadixSort(temp_storage).SortDescending(keys, values, begin_bit, end_bit);\n stop = clock();\n StoreDirectBlocked(threadIdx.x, d_keys, keys);\n StoreDirectBlocked(threadIdx.x, d_values, values);\n}\n\n/// Specialized descending, blocked -> striped\ntemplate \n__device__ __forceinline__ void TestBlockSort(\n typename BlockRadixSort::TempStorage &temp_storage,\n Key (&keys)[ITEMS_PER_THREAD],\n Value (&values)[ITEMS_PER_THREAD],\n Key *d_keys,\n Value *d_values,\n int begin_bit,\n int end_bit,\n clock_t &stop,\n Int2Type is_descending,\n Int2Type is_blocked_output)\n{\n BlockRadixSort(temp_storage).SortDescendingBlockedToStriped(keys, values, begin_bit, end_bit);\n stop = clock();\n StoreDirectStriped(threadIdx.x, d_keys, keys);\n StoreDirectStriped(threadIdx.x, d_values, values);\n}\n\n/// Specialized ascending, blocked -> blocked\ntemplate \n__device__ __forceinline__ void TestBlockSort(\n typename BlockRadixSort::TempStorage &temp_storage,\n Key (&keys)[ITEMS_PER_THREAD],\n Value (&values)[ITEMS_PER_THREAD],\n Key *d_keys,\n Value *d_values,\n int begin_bit,\n int end_bit,\n clock_t &stop,\n Int2Type is_descending,\n Int2Type is_blocked_output)\n{\n BlockRadixSort(temp_storage).Sort(keys, values, begin_bit, end_bit);\n stop = clock();\n StoreDirectBlocked(threadIdx.x, d_keys, keys);\n StoreDirectBlocked(threadIdx.x, d_values, values);\n}\n\n/// Specialized ascending, blocked -> striped\ntemplate \n__device__ __forceinline__ void TestBlockSort(\n typename BlockRadixSort::TempStorage &temp_storage,\n Key (&keys)[ITEMS_PER_THREAD],\n Value (&values)[ITEMS_PER_THREAD],\n Key *d_keys,\n Value *d_values,\n int begin_bit,\n int end_bit,\n clock_t &stop,\n Int2Type is_descending,\n Int2Type is_blocked_output)\n{\n BlockRadixSort(temp_storage).SortBlockedToStriped(keys, values, begin_bit, end_bit);\n stop = clock();\n StoreDirectStriped(threadIdx.x, d_keys, keys);\n StoreDirectStriped(threadIdx.x, d_values, values);\n}\n\n\n\n/**\n * BlockRadixSort kernel\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n cudaSharedMemConfig SMEM_CONFIG,\n int DESCENDING,\n int BLOCKED_OUTPUT,\n typename Key,\n typename Value>\n__launch_bounds__ (BLOCK_THREADS, 1)\n__global__ void Kernel(\n Key *d_keys,\n Value *d_values,\n int begin_bit,\n int end_bit,\n clock_t *d_elapsed)\n{\n // Threadblock load/store abstraction types\n typedef BlockRadixSort<\n Key,\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n Value,\n RADIX_BITS,\n MEMOIZE_OUTER_SCAN,\n INNER_SCAN_ALGORITHM,\n SMEM_CONFIG>\n BlockRadixSortT;\n\n // Allocate temp storage in shared memory\n __shared__ typename BlockRadixSortT::TempStorage temp_storage;\n\n // Items per thread\n Key keys[ITEMS_PER_THREAD];\n Value values[ITEMS_PER_THREAD];\n\n LoadDirectBlocked(threadIdx.x, d_keys, keys);\n LoadDirectBlocked(threadIdx.x, d_values, values);\n\n // Start cycle timer\n clock_t stop;\n clock_t start = clock();\n\n TestBlockSort(\n temp_storage, keys, values, d_keys, d_values, begin_bit, end_bit, stop, Int2Type(), Int2Type());\n\n // Store time\n if (threadIdx.x == 0)\n *d_elapsed = (start > stop) ? start - stop : stop - start;\n}\n\n\n\n//---------------------------------------------------------------------\n// Host testing subroutines\n//---------------------------------------------------------------------\n\n\n/**\n * Simple key-value pairing\n */\ntemplate <\n typename Key,\n typename Value,\n bool IS_FLOAT = (Traits::CATEGORY == FLOATING_POINT)>\nstruct Pair\n{\n Key key;\n Value value;\n\n bool operator<(const Pair &b) const\n {\n return (key < b.key);\n }\n};\n\n/**\n * Simple key-value pairing (specialized for floating point types)\n */\ntemplate \nstruct Pair\n{\n Key key;\n Value value;\n\n bool operator<(const Pair &b) const\n {\n if (key < b.key)\n return true;\n\n if (key > b.key)\n return false;\n\n // Key in unsigned bits\n typedef typename Traits::UnsignedBits UnsignedBits;\n\n // Return true if key is negative zero and b.key is positive zero\n UnsignedBits key_bits = *reinterpret_cast(const_cast(&key));\n UnsignedBits b_key_bits = *reinterpret_cast(const_cast(&b.key));\n UnsignedBits HIGH_BIT = Traits::HIGH_BIT;\n\n return ((key_bits & HIGH_BIT) != 0) && ((b_key_bits & HIGH_BIT) == 0);\n }\n};\n\n\n/**\n * Initialize key-value sorting problem.\n */\ntemplate \nvoid Initialize(\n GenMode gen_mode,\n Key *h_keys,\n Value *h_values,\n Key *h_reference_keys,\n Value *h_reference_values,\n int num_items,\n int entropy_reduction,\n int begin_bit,\n int end_bit)\n{\n Pair *h_pairs = new Pair[num_items];\n\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(gen_mode, h_keys[i], i);\n\n RandomBits(h_values[i]);\n\n // Mask off unwanted portions\n int num_bits = end_bit - begin_bit;\n if ((begin_bit > 0) || (end_bit < sizeof(Key) * 8))\n {\n unsigned long long base = 0;\n memcpy(&base, &h_keys[i], sizeof(Key));\n base &= ((1ull << num_bits) - 1) << begin_bit;\n memcpy(&h_keys[i], &base, sizeof(Key));\n }\n\n h_pairs[i].key = h_keys[i];\n h_pairs[i].value = h_values[i];\n }\n\n if (DESCENDING) std::reverse(h_pairs, h_pairs + num_items);\n std::stable_sort(h_pairs, h_pairs + num_items);\n if (DESCENDING) std::reverse(h_pairs, h_pairs + num_items);\n\n for (int i = 0; i < num_items; ++i)\n {\n h_reference_keys[i] = h_pairs[i].key;\n h_reference_values[i] = h_pairs[i].value;\n }\n\n delete[] h_pairs;\n}\n\n\n\n\n/**\n * Test BlockRadixSort kernel\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n cudaSharedMemConfig SMEM_CONFIG,\n bool DESCENDING,\n bool BLOCKED_OUTPUT,\n typename Key,\n typename Value>\nvoid TestDriver(\n GenMode gen_mode,\n int entropy_reduction,\n int begin_bit,\n int end_bit)\n{\n enum\n {\n TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD,\n KEYS_ONLY = Equals::VALUE,\n };\n\n // Allocate host arrays\n Key *h_keys = new Key[TILE_SIZE];\n Key *h_reference_keys = new Key[TILE_SIZE];\n Value *h_values = new Value[TILE_SIZE];\n Value *h_reference_values = new Value[TILE_SIZE];\n\n // Allocate device arrays\n Key *d_keys = NULL;\n Value *d_values = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys, sizeof(Key) * TILE_SIZE));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values, sizeof(Value) * TILE_SIZE));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(clock_t)));\n\n // Initialize problem and solution on host\n Initialize(gen_mode, h_keys, h_values, h_reference_keys, h_reference_values,\n TILE_SIZE, entropy_reduction, begin_bit, end_bit);\n\n // Copy problem to device\n CubDebugExit(cudaMemcpy(d_keys, h_keys, sizeof(Key) * TILE_SIZE, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_values, h_values, sizeof(Value) * TILE_SIZE, cudaMemcpyHostToDevice));\n\n printf(\"%s \"\n \"BLOCK_THREADS(%d) \"\n \"ITEMS_PER_THREAD(%d) \"\n \"RADIX_BITS(%d) \"\n \"MEMOIZE_OUTER_SCAN(%d) \"\n \"INNER_SCAN_ALGORITHM(%d) \"\n \"SMEM_CONFIG(%d) \"\n \"DESCENDING(%d) \"\n \"BLOCKED_OUTPUT(%d) \"\n \"sizeof(Key)(%d) \"\n \"sizeof(Value)(%d) \"\n \"gen_mode(%d), \"\n \"entropy_reduction(%d) \"\n \"begin_bit(%d) \"\n \"end_bit(%d), \"\n \"samples(%d)\\n\",\n ((KEYS_ONLY) ? \"Keys-only\" : \"Key-value\"),\n BLOCK_THREADS,\n ITEMS_PER_THREAD,\n RADIX_BITS,\n MEMOIZE_OUTER_SCAN,\n INNER_SCAN_ALGORITHM,\n SMEM_CONFIG,\n DESCENDING,\n BLOCKED_OUTPUT,\n (int) sizeof(Key),\n (int) sizeof(Value),\n gen_mode,\n entropy_reduction,\n begin_bit,\n end_bit,\n g_num_rand_samples);\n\n // Set shared memory config\n cudaDeviceSetSharedMemConfig(SMEM_CONFIG);\n\n // Run kernel\n Kernel<<<1, BLOCK_THREADS>>>(\n d_keys, d_values, begin_bit, end_bit, d_elapsed);\n\n // Flush kernel output / errors\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Check keys results\n printf(\"\\tKeys: \");\n int compare = CompareDeviceResults(h_reference_keys, d_keys, TILE_SIZE, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Check value results\n if (!KEYS_ONLY)\n {\n printf(\"\\tValues: \");\n int compare = CompareDeviceResults(h_reference_values, d_values, TILE_SIZE, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n }\n printf(\"\\n\");\n\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n printf(\"\\n\");\n\n // Cleanup\n if (h_keys) delete[] h_keys;\n if (h_reference_keys) delete[] h_reference_keys;\n if (h_values) delete[] h_values;\n if (h_reference_values) delete[] h_reference_values;\n if (d_keys) CubDebugExit(g_allocator.DeviceFree(d_keys));\n if (d_values) CubDebugExit(g_allocator.DeviceFree(d_values));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Test driver (valid tile size <= MAX_SMEM_BYTES)\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n cudaSharedMemConfig SMEM_CONFIG,\n bool DESCENDING,\n bool BLOCKED_OUTPUT,\n typename Key,\n typename Value>\nvoid TestValid(Int2Type fits_smem_capacity)\n{\n // Iterate begin_bit\n for (int begin_bit = 0; begin_bit <= 1; begin_bit++)\n {\n // Iterate end bit\n for (int end_bit = begin_bit + 1; end_bit <= sizeof(Key) * 8; end_bit = end_bit * 2 + begin_bit)\n {\n // Uniform key distribution\n TestDriver(\n UNIFORM, 0, begin_bit, end_bit);\n\n // Sequential key distribution\n TestDriver(\n INTEGER_SEED, 0, begin_bit, end_bit);\n\n // Iterate random with entropy_reduction\n for (int entropy_reduction = 0; entropy_reduction <= 9; entropy_reduction += 3)\n {\n TestDriver(\n RANDOM, entropy_reduction, begin_bit, end_bit);\n }\n }\n }\n}\n\n\n/**\n * Test driver (invalid tile size)\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n cudaSharedMemConfig SMEM_CONFIG,\n bool DESCENDING,\n bool BLOCKED_OUTPUT,\n typename Key,\n typename Value>\nvoid TestValid(Int2Type fits_smem_capacity)\n{}\n\n\n/**\n * Test ascending/descending and to-blocked/to-striped\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n cudaSharedMemConfig SMEM_CONFIG,\n typename Key,\n typename Value>\nvoid Test()\n{\n // Check size of smem storage for the target arch to make sure it will fit\n typedef BlockRadixSort BlockRadixSortT;\n\n#if defined(SM100) || defined(SM110) || defined(SM130)\n Int2Type fits_smem_capacity;\n#else\n Int2Type<(sizeof(typename BlockRadixSortT::TempStorage) <= 48 * 1024)> fits_smem_capacity;\n#endif\n\n // Sort-ascending, to-striped\n TestValid(fits_smem_capacity);\n\n // Sort-descending, to-blocked\n TestValid(fits_smem_capacity);\n\n // Not necessary\n// TestValid(fits_smem_capacity);\n// TestValid(fits_smem_capacity);\n}\n\n\n/**\n * Test value type and smem config\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n typename Key>\nvoid TestKeys()\n{\n // Test keys-only sorting with both smem configs\n Test(); // Keys-only (4-byte smem bank config)\n#if !defined(SM100) && !defined(SM110) && !defined(SM130) && !defined(SM200)\n Test(); // Keys-only (8-byte smem bank config)\n#endif\n}\n\n\n/**\n * Test value type and smem config\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM,\n typename Key>\nvoid TestKeysAndPairs()\n{\n // Test pairs sorting with only 4-byte configs\n Test(); // With small-values\n Test(); // With same-values\n Test(); // With large values\n}\n\n\n/**\n * Test key type\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN,\n BlockScanAlgorithm INNER_SCAN_ALGORITHM>\nvoid Test()\n{\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef TEST_KEYS_ONLY\n\n // Test unsigned types with keys-only\n TestKeys();\n TestKeys();\n TestKeys();\n TestKeys();\n TestKeys();\n\n#else\n\n // Test signed and fp types with paired values\n TestKeysAndPairs();\n TestKeysAndPairs();\n TestKeysAndPairs();\n TestKeysAndPairs();\n TestKeysAndPairs();\n TestKeysAndPairs();\n if (ptx_version > 120)\n {\n // Don't check doubles on PTX120 or below because they're down-converted\n TestKeysAndPairs();\n }\n\n#endif\n}\n\n\n/**\n * Test inner scan algorithm\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS,\n bool MEMOIZE_OUTER_SCAN>\nvoid Test()\n{\n Test();\n Test();\n}\n\n\n/**\n * Test outer scan algorithm\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int RADIX_BITS>\nvoid Test()\n{\n Test();\n Test();\n}\n\n\n/**\n * Test radix bits\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid Test()\n{\n Test();\n Test();\n Test();\n}\n\n\n/**\n * Test items per thread\n */\ntemplate \nvoid Test()\n{\n Test();\n#if defined(SM100) || defined(SM110) || defined(SM130)\n // Open64 compiler can't handle the number of test cases\n#else\n Test();\n#endif\n Test();\n}\n\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#ifdef QUICK_TEST\n\n {\n typedef float T;\n TestDriver<32, 4, 4, true, BLOCK_SCAN_WARP_SCANS, cudaSharedMemBankSizeFourByte, false, false, T, NullType>(INTEGER_SEED, 0, 0, sizeof(T) * 8);\n }\n/*\n // Compile/run quick tests\n typedef unsigned int T;\n TestDriver<64, 17, 4, true, BLOCK_SCAN_WARP_SCANS, cudaSharedMemBankSizeFourByte, false, false, T, NullType>(RANDOM, 0, 0, sizeof(T) * 8);\n TestDriver<96, 8, 4, true, BLOCK_SCAN_WARP_SCANS, cudaSharedMemBankSizeFourByte, false, false, T, NullType>(RANDOM, 0, 0, sizeof(T) * 8);\n TestDriver<128, 2, 4, true, BLOCK_SCAN_WARP_SCANS, cudaSharedMemBankSizeFourByte, false, false, T, NullType>(RANDOM, 0, 0, sizeof(T) * 8);\n*/\n\n#else\n\n // Compile/run thorough tests\n Test<32>();\n Test<64>();\n Test<160>();\n\n\n#endif // QUICK_TEST\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_block_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of BlockReduce utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n\n/// Generic reduction (full, 1)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T (&data)[1], ReductionOp &reduction_op)\n{\n return block_reduce.Reduce(data[0], reduction_op);\n}\n\n/// Generic reduction (full, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T (&data)[ITEMS_PER_THREAD], ReductionOp &reduction_op)\n{\n return block_reduce.Reduce(data, reduction_op);\n}\n\n/// Generic reduction (partial, 1)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T &data, ReductionOp &reduction_op, int valid_threads)\n{\n return block_reduce.Reduce(data, reduction_op, valid_threads);\n}\n\n/// Sum reduction (full, 1)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T (&data)[1], Sum &reduction_op)\n{\n return block_reduce.Sum(data[0]);\n}\n\n/// Sum reduction (full, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T (&data)[ITEMS_PER_THREAD], Sum &reduction_op)\n{\n return block_reduce.Sum(data);\n}\n\n/// Sum reduction (partial, 1)\ntemplate \n__device__ __forceinline__ T DeviceTest(\n BlockReduceT &block_reduce, T &data, Sum &reduction_op, int valid_threads)\n{\n return block_reduce.Sum(data, valid_threads);\n}\n\n\n/**\n * Test full-tile reduction kernel (where num_items is an even\n * multiple of BLOCK_THREADS)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n typename T,\n typename ReductionOp>\n__launch_bounds__ (BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z)\n__global__ void FullTileReduceKernel(\n T *d_in,\n T *d_out,\n ReductionOp reduction_op,\n int tiles,\n clock_t *d_elapsed)\n{\n const int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Cooperative thread block reduction utility type (returns aggregate in thread 0)\n typedef BlockReduce BlockReduceT;\n\n // Allocate temp storage in shared memory\n __shared__ typename BlockReduceT::TempStorage temp_storage;\n\n int linear_tid = RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n\n // Per-thread tile data\n T data[ITEMS_PER_THREAD];\n\n // Load first tile of data\n int block_offset = 0;\n\n if (block_offset < TILE_SIZE * tiles)\n {\n LoadDirectBlocked(linear_tid, d_in + block_offset, data);\n block_offset += TILE_SIZE;\n\n // Start cycle timer\n clock_t start = clock();\n\n // Cooperative reduce first tile\n BlockReduceT block_reduce(temp_storage) ;\n T block_aggregate = DeviceTest(block_reduce, data, reduction_op);\n\n // Stop cycle timer\n #if CUB_PTX_ARCH == 100\n // Bug: recording stop clock causes mis-write of running prefix value\n clock_t stop = 0;\n#else\n clock_t stop = clock();\n#endif // CUB_PTX_ARCH == 100\n clock_t elapsed = (start > stop) ? start - stop : stop - start;\n\n // Loop over input tiles\n while (block_offset < TILE_SIZE * tiles)\n {\n // TestBarrier between thread block reductions\n __syncthreads();\n \n // Load tile of data\n LoadDirectBlocked(linear_tid, d_in + block_offset, data);\n block_offset += TILE_SIZE;\n\n // Start cycle timer\n clock_t start = clock();\n\n // Cooperatively reduce the tile's aggregate\n BlockReduceT block_reduce(temp_storage) ;\n T tile_aggregate = DeviceTest(block_reduce, data, reduction_op);\n\n // Stop cycle timer\n#if CUB_PTX_ARCH == 100\n // Bug: recording stop clock causes mis-write of running prefix value\n clock_t stop = 0;\n#else\n clock_t stop = clock();\n#endif // CUB_PTX_ARCH == 100\n elapsed += (start > stop) ? start - stop : stop - start;\n\n // Reduce thread block aggregate\n block_aggregate = reduction_op(block_aggregate, tile_aggregate);\n }\n\n // Store data\n if (linear_tid == 0)\n {\n d_out[0] = block_aggregate;\n *d_elapsed = elapsed;\n }\n }\n}\n\n\n\n/**\n * Test partial-tile reduction kernel (where num_items < BLOCK_THREADS)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n typename T,\n typename ReductionOp>\n__launch_bounds__ (BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z)\n__global__ void PartialTileReduceKernel(\n T *d_in,\n T *d_out,\n int num_items,\n ReductionOp reduction_op,\n clock_t *d_elapsed)\n{\n // Cooperative thread block reduction utility type (returns aggregate only in thread-0)\n typedef BlockReduce BlockReduceT;\n\n // Allocate temp storage in shared memory\n __shared__ typename BlockReduceT::TempStorage temp_storage;\n\n int linear_tid = RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n\n // Per-thread tile data\n T partial;\n\n // Load partial tile data\n if (linear_tid < num_items)\n {\n partial = d_in[linear_tid];\n }\n\n // Start cycle timer\n clock_t start = clock();\n\n // Cooperatively reduce the tile's aggregate\n BlockReduceT block_reduce(temp_storage) ;\n T tile_aggregate = DeviceTest(block_reduce, partial, reduction_op, num_items);\n\n // Stop cycle timer\n#if CUB_PTX_ARCH == 100\n // Bug: recording stop clock causes mis-write of running prefix value\n clock_t stop = 0;\n#else\n clock_t stop = clock();\n#endif // CUB_PTX_ARCH == 100\n\n clock_t elapsed = (start > stop) ? start - stop : stop - start;\n\n // Store data\n if (linear_tid == 0)\n {\n d_out[0] = tile_aggregate;\n *d_elapsed = elapsed;\n }\n}\n\n\n//---------------------------------------------------------------------\n// Host utility subroutines\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem (and solution)\n */\ntemplate <\n typename T,\n typename ReductionOp>\nvoid Initialize(\n GenMode gen_mode,\n T *h_in,\n T h_reference[1],\n ReductionOp reduction_op,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n if (i == 0)\n h_reference[0] = h_in[0];\n else\n h_reference[0] = reduction_op(h_reference[0], h_in[i]);\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\");\n }\n}\n\n\n//---------------------------------------------------------------------\n// Full tile test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Test full-tile reduction. (Specialized for sufficient resources)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n int tiles,\n ReductionOp reduction_op,\n Int2Type sufficient_resources)\n{\n const int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n int num_items = TILE_SIZE * tiles;\n\n // Allocate host arrays\n T *h_in = new T[num_items];\n T h_reference[1];\n\n // Initialize problem\n Initialize(gen_mode, h_in, h_reference, reduction_op, num_items);\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(unsigned long long)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * 1));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * 1));\n\n // Test multi-tile (unguarded)\n printf(\"TestFullTile %s, %s, gen-mode %d, num_items(%d), BLOCK_THREADS(%d) (%d,%d,%d), ITEMS_PER_THREAD(%d), tiles(%d), %s (%d bytes) elements:\\n\",\n Equals::VALUE ? \"Sum\" : \"Max\",\n (ALGORITHM == BLOCK_REDUCE_RAKING) ? \"BLOCK_REDUCE_RAKING\" : (ALGORITHM == BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY) ? \"BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY\" : \"BLOCK_REDUCE_WARP_REDUCTIONS\",\n gen_mode,\n num_items,\n BLOCK_THREADS, BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z,\n ITEMS_PER_THREAD,\n tiles,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n dim3 block_dims(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n FullTileReduceKernel<<<1, block_dims>>>(\n d_in,\n d_out,\n reduction_op,\n tiles,\n d_elapsed);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tReduction results: \");\n int compare = CompareDeviceResults(h_reference, d_out, 1, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Test full-tile reduction. (Specialized for insufficient resources)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n int tiles,\n ReductionOp reduction_op,\n Int2Type sufficient_resources)\n{}\n\n\n/**\n * Test full-tile reduction.\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n int tiles,\n ReductionOp reduction_op)\n{\n // Check size of smem storage for the target arch to make sure it will fit\n typedef BlockReduce BlockReduceT;\n\n enum \n {\n#if defined(SM100) || defined(SM110) || defined(SM130)\n sufficient_smem = (sizeof(typename BlockReduceT::TempStorage) <= 16 * 1024),\n sufficient_threads = ((BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 512),\n#else\n sufficient_smem = (sizeof(typename BlockReduceT::TempStorage) <= 48 * 1024),\n sufficient_threads = ((BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 1024),\n#endif\n };\n\n TestFullTile(gen_mode, tiles, reduction_op, Int2Type());\n}\n\n\n/**\n * Run battery of tests for different thread block dimensions\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n int tiles,\n ReductionOp reduction_op)\n{\n TestFullTile(gen_mode, tiles, reduction_op);\n TestFullTile(gen_mode, tiles, reduction_op);\n}\n\n/**\n * Run battery of tests for different thread items\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_THREADS,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n int tiles,\n ReductionOp reduction_op)\n{\n TestFullTile(gen_mode, tiles, reduction_op);\n TestFullTile(gen_mode, tiles, reduction_op);\n}\n\n\n/**\n * Run battery of full-tile tests for different numbers of tiles\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_THREADS,\n typename T,\n typename ReductionOp>\nvoid TestFullTile(\n GenMode gen_mode,\n ReductionOp reduction_op)\n{\n for (int tiles = 1; tiles < 3; tiles++)\n {\n TestFullTile(gen_mode, tiles, reduction_op);\n }\n}\n\n\n//---------------------------------------------------------------------\n// Partial-tile test generation\n//---------------------------------------------------------------------\n\n/**\n * Test partial-tile reduction. (Specialized for sufficient resources)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n typename T,\n typename ReductionOp>\nvoid TestPartialTile(\n GenMode gen_mode,\n int num_items,\n ReductionOp reduction_op,\n Int2Type sufficient_resources)\n{\n const int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;\n const int TILE_SIZE = BLOCK_THREADS;\n\n // Allocate host arrays\n T *h_in = new T[num_items];\n T h_reference[1];\n\n // Initialize problem\n Initialize(gen_mode, h_in, h_reference, reduction_op, num_items);\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(unsigned long long)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * TILE_SIZE));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * 1));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * 1));\n\n printf(\"TestPartialTile %s, gen-mode %d, num_items(%d), BLOCK_THREADS(%d) (%d,%d,%d), %s (%d bytes) elements:\\n\",\n (ALGORITHM == BLOCK_REDUCE_RAKING) ? \"BLOCK_REDUCE_RAKING\" : (ALGORITHM == BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY) ? \"BLOCK_REDUCE_RAKING_COMMUTATIVE_ONLY\" : \"BLOCK_REDUCE_WARP_REDUCTIONS\",\n gen_mode,\n num_items,\n BLOCK_THREADS, BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n dim3 block_dims(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n PartialTileReduceKernel<<<1, block_dims>>>(\n d_in,\n d_out,\n num_items,\n reduction_op,\n d_elapsed);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tReduction results: \");\n int compare = CompareDeviceResults(h_reference, d_out, 1, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n\n/**\n * Test partial-tile reduction (specialized for insufficient resources)\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n typename T,\n typename ReductionOp>\nvoid TestPartialTile(\n GenMode gen_mode,\n int num_items,\n ReductionOp reduction_op,\n Int2Type sufficient_resources)\n{}\n\n\n/**\n * Run battery of partial-tile tests for different numbers of effective threads and thread dimensions\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n typename T,\n typename ReductionOp>\nvoid TestPartialTile(\n GenMode gen_mode,\n int num_items,\n ReductionOp reduction_op)\n{\n // Check size of smem storage for the target arch to make sure it will fit\n typedef BlockReduce BlockReduceT;\n\n enum \n {\n#if defined(SM100) || defined(SM110) || defined(SM130)\n sufficient_smem = sizeof(typename BlockReduceT::TempStorage) <= 16 * 1024,\n sufficient_threads = (BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 512,\n#else\n sufficient_smem = sizeof(typename BlockReduceT::TempStorage) <= 48 * 1024,\n sufficient_threads = (BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 1024,\n#endif\n };\n\n TestPartialTile(gen_mode, num_items, reduction_op, Int2Type());\n}\n\n\n\n/**\n * Run battery of partial-tile tests for different numbers of effective threads and thread dimensions\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_THREADS,\n typename T,\n typename ReductionOp>\nvoid TestPartialTile(\n GenMode gen_mode,\n ReductionOp reduction_op)\n{\n for (\n int num_items = 1;\n num_items < BLOCK_THREADS;\n num_items += CUB_MAX(1, BLOCK_THREADS / 5))\n {\n TestPartialTile(gen_mode, num_items, reduction_op);\n TestPartialTile(gen_mode, num_items, reduction_op);\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Run battery of full-tile tests for different gen modes\n */\ntemplate <\n BlockReduceAlgorithm ALGORITHM,\n int BLOCK_THREADS,\n typename T,\n typename ReductionOp>\nvoid Test(\n ReductionOp reduction_op)\n{\n TestFullTile(UNIFORM, reduction_op);\n TestPartialTile(UNIFORM, reduction_op);\n\n TestFullTile(INTEGER_SEED, reduction_op);\n TestPartialTile(INTEGER_SEED, reduction_op);\n\n if (Traits::CATEGORY != FLOATING_POINT)\n {\n // Don't test randomly-generated floats b/c of stability\n TestFullTile(RANDOM, reduction_op);\n TestPartialTile(RANDOM, reduction_op);\n }\n}\n\n\n/**\n * Run battery of tests for different block-reduction algorithmic variants\n */\ntemplate <\n int BLOCK_THREADS,\n typename T,\n typename ReductionOp>\nvoid Test(\n ReductionOp reduction_op)\n{\n#ifdef TEST_RAKING\n Test(reduction_op);\n Test(reduction_op);\n#endif\n#ifdef TEST_WARP_REDUCTIONS\n Test(reduction_op);\n#endif\n}\n\n\n/**\n * Run battery of tests for different block sizes\n */\ntemplate <\n typename T,\n typename ReductionOp>\nvoid Test(\n ReductionOp reduction_op)\n{\n Test<7, T>(reduction_op);\n Test<32, T>(reduction_op);\n Test<63, T>(reduction_op);\n Test<97, T>(reduction_op);\n Test<128, T>(reduction_op);\n Test<238, T>(reduction_op);\n}\n\n\n/**\n * Run battery of tests for different block sizes\n */\ntemplate \nvoid Test()\n{\n Test(Sum());\n Test(Max());\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef QUICK_TEST\n\n // Compile/run quick tests\n\n\n printf(\"\\n full tile ------------------------\\n\\n\");\n\n TestFullTile(RANDOM, 1, Sum());\n TestFullTile(RANDOM, 1, Sum());\n TestFullTile(RANDOM, 1, Sum());\n\n TestFullTile(RANDOM, 1, Sum());\n TestFullTile(RANDOM, 1, Sum());\n TestFullTile(RANDOM, 1, Sum());\n\n printf(\"\\n partial tile ------------------------\\n\\n\");\n\n TestPartialTile(RANDOM, 7, Sum());\n TestPartialTile(RANDOM, 7, Sum());\n TestPartialTile(RANDOM, 7, Sum());\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // primitives\n Test();\n Test();\n Test();\n Test();\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test();\n\n Test();\n\n // vector types\n Test();\n Test();\n Test();\n Test();\n\n Test();\n Test();\n Test();\n Test();\n\n // Complex types\n Test();\n Test();\n }\n\n#endif\n\n return 0;\n}\n\n\n" }, { "path": "external/cub/test/test_block_scan.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of BlockScan utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n/**\n * Primitive variant to test\n */\nenum TestMode\n{\n BASIC,\n AGGREGATE,\n PREFIX,\n};\n\n\n/**\n * Scan mode to test\n */\nenum ScanMode\n{\n EXCLUSIVE,\n INCLUSIVE\n};\n\n\n/**\n * \\brief WrapperFunctor (for precluding test-specialized dispatch to *Sum variants)\n */\ntemplate\nstruct WrapperFunctor\n{\n OpT op;\n\n WrapperFunctor(OpT op) : op(op) {}\n\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return op(a, b);\n }\n};\n\n\n/**\n * Stateful prefix functor\n */\ntemplate <\n typename T,\n typename ScanOpT>\nstruct BlockPrefixCallbackOp\n{\n int linear_tid;\n T prefix;\n ScanOpT scan_op;\n\n __device__ __forceinline__\n BlockPrefixCallbackOp(int linear_tid, T prefix, ScanOpT scan_op) :\n linear_tid(linear_tid),\n prefix(prefix),\n scan_op(scan_op)\n {}\n\n __device__ __forceinline__\n T operator()(T block_aggregate)\n {\n // For testing purposes\n T retval = (linear_tid == 0) ? prefix : T();\n prefix = scan_op(prefix, block_aggregate);\n return retval;\n }\n};\n\n\n//---------------------------------------------------------------------\n// Exclusive scan\n//---------------------------------------------------------------------\n\n/// Exclusive scan (BASIC, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data[0], data[0], initial_value, scan_op);\n}\n\n/// Exclusive scan (BASIC, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data, data, initial_value, scan_op);\n}\n\n/// Exclusive scan (AGGREGATE, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data[0], data[0], initial_value, scan_op, block_aggregate);\n}\n\n/// Exclusive scan (AGGREGATE, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data, data, initial_value, scan_op, block_aggregate);\n}\n\n/// Exclusive scan (PREFIX, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data[0], data[0], scan_op, prefix_op);\n}\n\n/// Exclusive scan (PREFIX, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.ExclusiveScan(data, data, scan_op, prefix_op);\n}\n\n\n//---------------------------------------------------------------------\n// Exclusive sum\n//---------------------------------------------------------------------\n\n/// Exclusive sum (BASIC, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data[0], data[0]);\n}\n\n/// Exclusive sum (BASIC, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data, data);\n}\n\n/// Exclusive sum (AGGREGATE, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data[0], data[0], block_aggregate);\n}\n\n/// Exclusive sum (AGGREGATE, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data, data, block_aggregate);\n}\n\n/// Exclusive sum (PREFIX, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data[0], data[0], prefix_op);\n}\n\n/// Exclusive sum (PREFIX, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.ExclusiveSum(data, data, prefix_op);\n}\n\n\n//---------------------------------------------------------------------\n// Inclusive scan\n//---------------------------------------------------------------------\n\n/// Inclusive scan (BASIC, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data[0], data[0], scan_op);\n}\n\n/// Inclusive scan (BASIC, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data, data, scan_op);\n}\n\n/// Inclusive scan (AGGREGATE, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data[0], data[0], scan_op, block_aggregate);\n}\n\n/// Inclusive scan (AGGREGATE, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data, data, scan_op, block_aggregate);\n}\n\n/// Inclusive scan (PREFIX, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data[0], data[0], scan_op, prefix_op);\n}\n\n/// Inclusive scan (PREFIX, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, ScanOpT &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, IsPrimitiveT is_primitive)\n{\n block_scan.InclusiveScan(data, data, scan_op, prefix_op);\n}\n\n\n//---------------------------------------------------------------------\n// Inclusive sum\n//---------------------------------------------------------------------\n\n/// Inclusive sum (BASIC, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data[0], data[0]);\n}\n\n/// Inclusive sum (BASIC, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data, data);\n}\n\n/// Inclusive sum (AGGREGATE, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data[0], data[0], block_aggregate);\n}\n\n/// Inclusive sum (AGGREGATE, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data, data, block_aggregate);\n}\n\n/// Inclusive sum (PREFIX, 1)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[1], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data[0], data[0], prefix_op);\n}\n\n/// Inclusive sum (PREFIX, ITEMS_PER_THREAD)\ntemplate \n__device__ __forceinline__ void DeviceTest(\n BlockScanT &block_scan, T (&data)[ITEMS_PER_THREAD], T &initial_value, Sum &scan_op, T &block_aggregate, PrefixCallbackOp &prefix_op,\n Int2Type scan_mode, Int2Type test_mode, Int2Type is_primitive)\n{\n block_scan.InclusiveSum(data, data, prefix_op);\n}\n\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/**\n * BlockScan test kernel.\n */\ntemplate <\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n BlockScanAlgorithm ALGORITHM,\n typename T,\n typename ScanOpT>\n__launch_bounds__ (BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z)\n__global__ void BlockScanKernel(\n T *d_in,\n T *d_out,\n T *d_aggregate,\n ScanOpT scan_op,\n T initial_value,\n clock_t *d_elapsed)\n{\n const int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Parameterize BlockScan type for our thread block\n typedef BlockScan BlockScanT;\n\n // Allocate temp storage in shared memory\n __shared__ typename BlockScanT::TempStorage temp_storage;\n\n int linear_tid = RowMajorTid(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n\n // Per-thread tile data\n T data[ITEMS_PER_THREAD];\n LoadDirectBlocked(linear_tid, d_in, data);\n\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Test scan\n T block_aggregate;\n BlockScanT block_scan(temp_storage);\n BlockPrefixCallbackOp prefix_op(linear_tid, initial_value, scan_op);\n\n DeviceTest(block_scan, data, initial_value, scan_op, block_aggregate, prefix_op,\n Int2Type(), Int2Type(), Int2Type::PRIMITIVE>());\n\n // Stop cycle timer\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Store output\n StoreDirectBlocked(linear_tid, d_out, data);\n\n // Store block_aggregate\n if (TEST_MODE != BASIC)\n d_aggregate[linear_tid] = block_aggregate;\n\n // Store prefix\n if (TEST_MODE == PREFIX)\n {\n if (linear_tid == 0)\n d_out[TILE_SIZE] = prefix_op.prefix;\n }\n\n // Store time\n if (linear_tid == 0)\n *d_elapsed = (start > stop) ? start - stop : stop - start;\n}\n\n\n\n//---------------------------------------------------------------------\n// Host utility subroutines\n//---------------------------------------------------------------------\n\n/**\n * Initialize exclusive-scan problem (and solution)\n */\ntemplate \nT Initialize(\n GenMode gen_mode,\n T *h_in,\n T *h_reference,\n int num_items,\n ScanOpT scan_op,\n T initial_value,\n Int2Type)\n{\n InitValue(gen_mode, h_in[0], 0);\n\n T block_aggregate = h_in[0];\n h_reference[0] = initial_value;\n T inclusive = scan_op(initial_value, h_in[0]);\n\n for (int i = 1; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n h_reference[i] = inclusive;\n inclusive = scan_op(inclusive, h_in[i]);\n block_aggregate = scan_op(block_aggregate, h_in[i]);\n }\n\n return block_aggregate;\n}\n\n\n/**\n * Initialize inclusive-scan problem (and solution)\n */\ntemplate \nT Initialize(\n GenMode gen_mode,\n T *h_in,\n T *h_reference,\n int num_items,\n ScanOpT scan_op,\n T initial_value,\n Int2Type)\n{\n InitValue(gen_mode, h_in[0], 0);\n\n T block_aggregate = h_in[0];\n T inclusive = scan_op(initial_value, h_in[0]);\n h_reference[0] = inclusive;\n\n for (int i = 1; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n inclusive = scan_op(inclusive, h_in[i]);\n block_aggregate = scan_op(block_aggregate, h_in[i]);\n h_reference[i] = inclusive;\n }\n\n return block_aggregate;\n}\n\n\n/**\n * Test thread block scan. (Specialized for sufficient resources)\n */\ntemplate <\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n BlockScanAlgorithm ALGORITHM,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value,\n Int2Type sufficient_resources)\n{\n const int BLOCK_THREADS = BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z;\n const int TILE_SIZE = BLOCK_THREADS * ITEMS_PER_THREAD;\n\n // Allocate host arrays\n T *h_in = new T[TILE_SIZE];\n T *h_reference = new T[TILE_SIZE];\n T *h_aggregate = new T[BLOCK_THREADS];\n\n // Initialize problem\n T block_aggregate = Initialize(\n gen_mode,\n h_in,\n h_reference,\n TILE_SIZE,\n scan_op,\n initial_value,\n Int2Type());\n\n // Test reference block_aggregate is returned in all threads\n for (int i = 0; i < BLOCK_THREADS; ++i)\n {\n h_aggregate[i] = block_aggregate;\n }\n\n // Run kernel\n printf(\"Test-mode %d, gen-mode %d, policy %d, %s %s BlockScan, %d (%d,%d,%d) thread block threads, %d items per thread, %d tile size, %s (%d bytes) elements:\\n\",\n TEST_MODE, gen_mode, ALGORITHM,\n (SCAN_MODE == INCLUSIVE) ? \"Inclusive\" : \"Exclusive\", typeid(ScanOpT).name(),\n BLOCK_THREADS, BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z,\n ITEMS_PER_THREAD, TILE_SIZE,\n typeid(T).name(), (int) sizeof(T));\n fflush(stdout);\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n T *d_aggregate = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(unsigned long long)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * TILE_SIZE));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * (TILE_SIZE + 2)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_aggregate, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * TILE_SIZE, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * (TILE_SIZE + 1)));\n CubDebugExit(cudaMemset(d_aggregate, 0, sizeof(T) * BLOCK_THREADS));\n\n // Display input problem data\n if (g_verbose)\n {\n printf(\"Input data: \");\n for (int i = 0; i < TILE_SIZE; i++)\n {\n std::cout << CoutCast(h_in[i]) << \", \";\n }\n printf(\"\\n\\n\");\n }\n\n // Run block_aggregate/prefix kernel\n dim3 block_dims(BLOCK_DIM_X, BLOCK_DIM_Y, BLOCK_DIM_Z);\n BlockScanKernel<<<1, block_dims>>>(\n d_in,\n d_out,\n d_aggregate,\n scan_op,\n initial_value,\n d_elapsed);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tScan results: \");\n int compare = CompareDeviceResults(h_reference, d_out, TILE_SIZE, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n if (TEST_MODE == AGGREGATE)\n {\n // Copy out and display block_aggregate\n printf(\"\\tScan block aggregate: \");\n compare = CompareDeviceResults(h_aggregate, d_aggregate, BLOCK_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n }\n\n if (TEST_MODE == PREFIX)\n {\n // Copy out and display updated prefix\n printf(\"\\tScan running total: \");\n T running_total = scan_op(initial_value, block_aggregate);\n compare = CompareDeviceResults(&running_total, d_out + TILE_SIZE, 1, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n }\n\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (h_aggregate) delete[] h_aggregate;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_aggregate) CubDebugExit(g_allocator.DeviceFree(d_aggregate));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Test thread block scan. (Specialized for insufficient resources)\n */\ntemplate <\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n BlockScanAlgorithm ALGORITHM,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value,\n Int2Type sufficient_resources)\n{}\n\n\n/**\n * Test thread block scan.\n */\ntemplate <\n int BLOCK_DIM_X,\n int BLOCK_DIM_Y,\n int BLOCK_DIM_Z,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n BlockScanAlgorithm ALGORITHM,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value)\n{\n // Check size of smem storage for the target arch to make sure it will fit\n typedef BlockScan BlockScanT;\n\n enum\n {\n#if defined(SM100) || defined(SM110) || defined(SM130)\n sufficient_smem = (sizeof(typename BlockScanT::TempStorage) <= 16 * 1024),\n sufficient_threads = ((BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 512),\n#else\n sufficient_smem = (sizeof(typename BlockScanT::TempStorage) <= 16 * 1024),\n sufficient_threads = ((BLOCK_DIM_X * BLOCK_DIM_Y * BLOCK_DIM_Z) <= 1024),\n#endif\n\n#if defined(_WIN32) || defined(_WIN64)\n // Accommodate ptxas crash bug (access violation) on Windows\n special_skip = ((TEST_ARCH <= 130) && (Equals::VALUE) && (BLOCK_DIM_Z > 1)),\n#else\n special_skip = false,\n#endif\n sufficient_resources = (sufficient_smem && sufficient_threads && !special_skip),\n };\n\n Test(\n gen_mode, scan_op, initial_value, Int2Type());\n}\n\n\n\n/**\n * Run test for different thread block dimensions\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n BlockScanAlgorithm ALGORITHM,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value)\n{\n Test(gen_mode, scan_op, initial_value);\n Test(gen_mode, scan_op, initial_value);\n}\n\n\n/**\n * Run test for different policy types\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n ScanMode SCAN_MODE,\n TestMode TEST_MODE,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value)\n{\n#ifdef TEST_RAKING\n Test(gen_mode, scan_op, initial_value);\n#endif\n#ifdef TEST_RAKING_MEMOIZE\n Test(gen_mode, scan_op, initial_value);\n#endif\n#ifdef TEST_WARP_SCANS\n Test(gen_mode, scan_op, initial_value);\n#endif\n}\n\n\n/**\n * Run tests for different primitive variants\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T identity,\n T initial_value)\n{\n // Exclusive (use identity as initial value because it will dispatch to *Sum variants that don't take initial values)\n Test(gen_mode, scan_op, identity);\n Test(gen_mode, scan_op, identity);\n Test(gen_mode, scan_op, identity);\n\n // Exclusive (non-specialized, so we can use initial-value)\n Test(gen_mode, WrapperFunctor(scan_op), initial_value);\n Test(gen_mode, WrapperFunctor(scan_op), initial_value);\n Test(gen_mode, WrapperFunctor(scan_op), initial_value);\n\n // Inclusive\n Test(gen_mode, scan_op, identity); // This scan doesn't take an initial value\n Test(gen_mode, scan_op, identity); // This scan doesn't take an initial value\n Test(gen_mode, scan_op, initial_value);\n}\n\n\n/**\n * Run tests for different problem-generation options\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n typename ScanOpT,\n typename T>\nvoid Test(\n ScanOpT scan_op,\n T identity,\n T initial_value)\n{\n Test(UNIFORM, scan_op, identity, initial_value);\n Test(INTEGER_SEED, scan_op, identity, initial_value);\n\n // Don't test randomly-generated floats b/c of stability\n if (Traits::CATEGORY != FLOATING_POINT)\n Test(RANDOM, scan_op, identity, initial_value);\n}\n\n\n/**\n * Run tests for different data types and scan ops\n */\ntemplate <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\nvoid Test()\n{\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n // primitive\n Test(Sum(), (unsigned char) 0, (unsigned char) 99);\n Test(Sum(), (unsigned short) 0, (unsigned short) 99);\n Test(Sum(), (unsigned int) 0, (unsigned int) 99);\n Test(Sum(), (unsigned long long) 0, (unsigned long long) 99);\n Test(Sum(), (float) 0, (float) 99);\n\n // primitive (alternative scan op)\n Test(Max(), std::numeric_limits::min(), (char) 99);\n Test(Max(), std::numeric_limits::min(), (short) 99);\n Test(Max(), std::numeric_limits::min(), (int) 99);\n Test(Max(), std::numeric_limits::min(), (long long) 99);\n\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test(Max(), std::numeric_limits::max() * -1, (double) 99);\n\n // vec-1\n Test(Sum(), make_uchar1(0), make_uchar1(17));\n\n // vec-2\n Test(Sum(), make_uchar2(0, 0), make_uchar2(17, 21));\n Test(Sum(), make_ushort2(0, 0), make_ushort2(17, 21));\n Test(Sum(), make_uint2(0, 0), make_uint2(17, 21));\n Test(Sum(), make_ulonglong2(0, 0), make_ulonglong2(17, 21));\n\n // vec-4\n Test(Sum(), make_char4(0, 0, 0, 0), make_char4(17, 21, 32, 85));\n Test(Sum(), make_short4(0, 0, 0, 0), make_short4(17, 21, 32, 85));\n Test(Sum(), make_int4(0, 0, 0, 0), make_int4(17, 21, 32, 85));\n Test(Sum(), make_longlong4(0, 0, 0, 0), make_longlong4(17, 21, 32, 85));\n\n // complex\n Test(Sum(), TestFoo::MakeTestFoo(0, 0, 0, 0), TestFoo::MakeTestFoo(17, 21, 32, 85));\n Test(Sum(), TestBar(0, 0), TestBar(17, 21));\n\n}\n\n\n/**\n * Run tests for different items per thread\n */\ntemplate \nvoid Test()\n{\n Test();\n Test();\n Test();\n}\n\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#ifdef QUICK_TEST\n\n Test<128, 1, 1, 1, EXCLUSIVE, AGGREGATE, BLOCK_SCAN_WARP_SCANS>(UNIFORM, Sum(), int(0));\n\n // Compile/run quick tests\n Test<128, 1, 1, 4, EXCLUSIVE, AGGREGATE, BLOCK_SCAN_WARP_SCANS>(UNIFORM, Sum(), int(0));\n Test<128, 1, 1, 4, EXCLUSIVE, AGGREGATE, BLOCK_SCAN_RAKING>(UNIFORM, Sum(), int(0));\n Test<128, 1, 1, 4, EXCLUSIVE, AGGREGATE, BLOCK_SCAN_RAKING_MEMOIZE>(UNIFORM, Sum(), int(0));\n\n Test<128, 1, 1, 2, INCLUSIVE, PREFIX, BLOCK_SCAN_RAKING>(INTEGER_SEED, Sum(), TestFoo::MakeTestFoo(17, 21, 32, 85));\n Test<128, 1, 1, 1, EXCLUSIVE, AGGREGATE, BLOCK_SCAN_WARP_SCANS>(UNIFORM, Sum(), make_longlong4(17, 21, 32, 85));\n\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Run tests for different thread block sizes\n Test<17>();\n Test<32>();\n Test<62>();\n Test<65>();\n// Test<96>(); // TODO: file bug for UNREACHABLE error for Test<96, 9, BASIC, BLOCK_SCAN_RAKING>(UNIFORM, Sum(), NullType(), make_ulonglong2(17, 21));\n Test<128>();\n }\n\n#endif\n\n return 0;\n}\n\n\n\n\n" }, { "path": "external/cub/test/test_device_histogram.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceHistogram utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n\n#if defined(QUICK_TEST) || defined(QUICKER_TEST)\n #include \n#endif\n\n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n NPP, // NPP method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\nbool g_verbose_input = false;\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n\n\n//---------------------------------------------------------------------\n// Dispatch to NPP histogram\n//---------------------------------------------------------------------\n\n#if defined(QUICK_TEST) || defined(QUICKER_TEST)\n\n/**\n * Dispatch to single-channel 8b NPP histo-even\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchEven(\n Int2Type<1> num_channels,\n Int2Type<1> num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n unsigned char *d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[1], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[1], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[1], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[1], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef unsigned char SampleT;\n\n cudaError_t error = cudaSuccess;\n NppiSize oSizeROI = {\n num_row_pixels,\n num_rows\n };\n\n if (d_temp_storage_bytes == NULL)\n {\n int nDeviceBufferSize;\n nppiHistogramEvenGetBufferSize_8u_C1R(oSizeROI, num_levels[0] ,&nDeviceBufferSize);\n temp_storage_bytes = nDeviceBufferSize;\n }\n else\n {\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n // compute the histogram\n nppiHistogramEven_8u_C1R(\n d_samples,\n row_stride_bytes,\n oSizeROI,\n d_histogram[0],\n num_levels[0],\n lower_level[0],\n upper_level[0],\n (Npp8u*) d_temp_storage);\n }\n }\n\n return error;\n}\n\n\n/**\n * Dispatch to 3/4 8b NPP histo-even\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchEven(\n Int2Type<4> num_channels,\n Int2Type<3> num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n unsigned char *d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[3], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[3], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[3], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[3], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef unsigned char SampleT;\n\n cudaError_t error = cudaSuccess;\n NppiSize oSizeROI = {\n num_row_pixels,\n num_rows\n };\n\n if (d_temp_storage_bytes == NULL)\n {\n int nDeviceBufferSize;\n nppiHistogramEvenGetBufferSize_8u_AC4R(oSizeROI, num_levels ,&nDeviceBufferSize);\n temp_storage_bytes = nDeviceBufferSize;\n }\n else\n {\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n // compute the histogram\n nppiHistogramEven_8u_AC4R(\n d_samples,\n row_stride_bytes,\n oSizeROI,\n d_histogram,\n num_levels,\n lower_level,\n upper_level,\n (Npp8u*) d_temp_storage);\n }\n }\n\n return error;\n}\n\n\n#endif // #if defined(QUICK_TEST) || defined(QUICKER_TEST)\n\n\n//---------------------------------------------------------------------\n// Dispatch to different DeviceHistogram entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to CUB single histogram-even entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchEven(\n Int2Type<1> num_channels,\n Int2Type<1> num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[1], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[1], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[1], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[1], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef typename std::iterator_traits::value_type SampleT;\n\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceHistogram::HistogramEven(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram[0],\n num_levels[0],\n lower_level[0],\n upper_level[0],\n num_row_pixels,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to CUB multi histogram-even entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchEven(\n Int2Type num_channels,\n Int2Type num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef typename std::iterator_traits::value_type SampleT;\n\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceHistogram::MultiHistogramEven(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram,\n num_levels,\n lower_level,\n upper_level,\n num_row_pixels,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to CUB single histogram-range entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchRange(\n Int2Type<1> num_channels,\n Int2Type<1> num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[1], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[1], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT *d_levels[1], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef typename std::iterator_traits::value_type SampleT;\n\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceHistogram::HistogramRange(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram[0],\n num_levels[0],\n d_levels[0],\n num_row_pixels,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to CUB multi histogram-range entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t DispatchRange(\n Int2Type num_channels,\n Int2Type num_active_channels,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n SampleIteratorT d_samples, ///< [in] The pointer to the multi-channel input sequence of data samples. The samples from different channels are assumed to be interleaved (e.g., an array of 32-bit pixels where each pixel consists of four RGBA 8-bit samples).\n CounterT *d_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT *d_levels[NUM_ACTIVE_CHANNELS], ///< [in] The pointers to the arrays of boundaries (levels), one for each active channel. Bin ranges are defined by consecutive boundary pairings: lower sample value boundaries are inclusive and upper sample value boundaries are exclusive.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef typename std::iterator_traits::value_type SampleT;\n\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceHistogram::MultiHistogramRange(\n d_temp_storage,\n temp_storage_bytes,\n d_samples,\n d_histogram,\n num_levels,\n d_levels,\n num_row_pixels,\n num_rows,\n row_stride_bytes,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n\n//---------------------------------------------------------------------\n// CUDA nested-parallelism test kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceHistogram\n * /\ntemplate \n__global__ void CnpDispatchKernel(\n Int2Type algorithm,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n SampleT *d_samples,\n SampleIteratorT d_sample_itr,\n ArrayWrapper d_out_histograms,\n int num_samples,\n bool debug_synchronous)\n{\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(algorithm, Int2Type(), timing_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_samples, d_sample_itr, d_out_histograms.array, num_samples, 0, debug_synchronous);\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/ **\n * Dispatch to CDP kernel\n * /\ntemplate \ncudaError_t Dispatch(\n Int2Type algorithm,\n Int2Type use_cdp,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n SampleT *d_samples,\n SampleIteratorT d_sample_itr,\n CounterT *d_histograms[NUM_ACTIVE_CHANNELS],\n int num_samples,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Setup array wrapper for histogram channel output (because we can't pass static arrays as kernel parameters)\n ArrayWrapper d_histo_wrapper;\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n d_histo_wrapper.array[CHANNEL] = d_histograms[CHANNEL];\n\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(algorithm, timing_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_samples, d_sample_itr, d_histo_wrapper, num_samples, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n*/\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n// Searches for bin given a list of bin-boundary levels\ntemplate \nstruct SearchTransform\n{\n LevelT *levels; // Pointer to levels array\n int num_levels; // Number of levels in array\n\n // Functor for converting samples to bin-ids (num_levels is returned if sample is out of range)\n template \n int operator()(SampleT sample)\n {\n int bin = int(std::upper_bound(levels, levels + num_levels, (LevelT) sample) - levels - 1);\n if (bin < 0)\n {\n // Sample out of range\n return num_levels;\n }\n return bin;\n }\n};\n\n\n// Scales samples to evenly-spaced bins\ntemplate \nstruct ScaleTransform\n{\n int num_levels; // Number of levels in array\n LevelT max; // Max sample level (exclusive)\n LevelT min; // Min sample level (inclusive)\n LevelT scale; // Bin scaling factor\n\n void Init(\n int num_levels, // Number of levels in array\n LevelT max, // Max sample level (exclusive)\n LevelT min, // Min sample level (inclusive)\n LevelT scale) // Bin scaling factor\n {\n this->num_levels = num_levels;\n this->max = max;\n this->min = min;\n this->scale = scale;\n }\n\n // Functor for converting samples to bin-ids (num_levels is returned if sample is out of range)\n template \n int operator()(SampleT sample)\n {\n if ((sample < min) || (sample >= max))\n {\n // Sample out of range\n return num_levels;\n }\n\n return (int) ((((LevelT) sample) - min) / scale);\n }\n};\n\n// Scales samples to evenly-spaced bins\ntemplate <>\nstruct ScaleTransform\n{\n int num_levels; // Number of levels in array\n float max; // Max sample level (exclusive)\n float min; // Min sample level (inclusive)\n float scale; // Bin scaling factor\n\n void Init(\n int num_levels, // Number of levels in array\n float max, // Max sample level (exclusive)\n float min, // Min sample level (inclusive)\n float scale) // Bin scaling factor\n {\n this->num_levels = num_levels;\n this->max = max;\n this->min = min;\n this->scale = 1.0f / scale;\n }\n\n // Functor for converting samples to bin-ids (num_levels is returned if sample is out of range)\n template \n int operator()(SampleT sample)\n {\n if ((sample < min) || (sample >= max))\n {\n // Sample out of range\n return num_levels;\n }\n\n return (int) ((((float) sample) - min) * scale);\n }\n};\n\n\n/**\n * Generate sample\n */\ntemplate \nvoid Sample(T &datum, LevelT max_level, int entropy_reduction)\n{\n unsigned int max = (unsigned int) -1;\n unsigned int bits;\n RandomBits(bits, entropy_reduction);\n float fraction = (float(bits) / max);\n\n datum = (T) (fraction * max_level);\n}\n\n\n/**\n * Initialize histogram samples\n */\ntemplate <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename LevelT,\n typename SampleT,\n typename OffsetT>\nvoid InitializeSamples(\n LevelT max_level,\n int entropy_reduction,\n SampleT *h_samples,\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes) ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n{\n // Initialize samples\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n for (OffsetT pixel = 0; pixel < num_row_pixels; ++pixel)\n {\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n // Sample offset\n OffsetT offset = (row * (row_stride_bytes / sizeof(SampleT))) + (pixel * NUM_CHANNELS) + channel;\n\n // Init sample value\n Sample(h_samples[offset], max_level, entropy_reduction);\n if (g_verbose_input)\n {\n if (channel > 0) printf(\", \");\n std::cout << CoutCast(h_samples[offset]);\n }\n }\n }\n }\n}\n\n\n/**\n * Initialize histogram solutions\n */\ntemplate <\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename SampleIteratorT,\n typename TransformOp,\n typename OffsetT>\nvoid InitializeBins(\n SampleIteratorT h_samples,\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n TransformOp transform_op[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n CounterT *h_histogram[NUM_ACTIVE_CHANNELS], ///< [out] The pointers to the histogram counter output arrays, one for each active channel. For channeli, the allocation length of d_histograms[i] should be num_levels[i] - 1.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes) ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n{\n typedef typename std::iterator_traits::value_type SampleT;\n\n // Init bins\n for (int CHANNEL = 0; CHANNEL < NUM_ACTIVE_CHANNELS; ++CHANNEL)\n {\n for (int bin = 0; bin < num_levels[CHANNEL] - 1; ++bin)\n {\n h_histogram[CHANNEL][bin] = 0;\n }\n }\n\n // Initialize samples\n if (g_verbose_input) printf(\"Samples: \\n\");\n for (OffsetT row = 0; row < num_rows; ++row)\n {\n for (OffsetT pixel = 0; pixel < num_row_pixels; ++pixel)\n {\n if (g_verbose_input) printf(\"[\");\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n // Sample offset\n OffsetT offset = (row * (row_stride_bytes / sizeof(SampleT))) + (pixel * NUM_CHANNELS) + channel;\n\n // Update sample bin\n int bin = transform_op[channel](h_samples[offset]);\n if (g_verbose_input) printf(\" (%d)\", bin); fflush(stdout);\n if ((bin >= 0) && (bin < num_levels[channel] - 1))\n {\n // valid bin\n h_histogram[channel][bin]++;\n }\n }\n if (g_verbose_input) printf(\"]\");\n }\n if (g_verbose_input) printf(\"\\n\\n\");\n }\n}\n\n\n\n/**\n * Test histogram-even\n */\ntemplate <\n Backend BACKEND,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT,\n typename SampleIteratorT>\nvoid TestEven(\n LevelT max_level,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes, ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n SampleIteratorT h_samples,\n SampleIteratorT d_samples)\n{\n OffsetT total_samples = num_rows * (row_stride_bytes / sizeof(SampleT));\n\n printf(\"\\n----------------------------\\n\");\n printf(\"%s cub::DeviceHistogramEven (%s) %d pixels (%d height, %d width, %d-byte row stride), %d %d-byte %s samples (entropy reduction %d), %s counters, %d/%d channels, max sample \",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == NPP) ? \"NPP\" : \"CUB\",\n (IsPointer::VALUE) ? \"pointer\" : \"iterator\",\n (int) (num_row_pixels * num_rows),\n (int) num_rows,\n (int) num_row_pixels,\n (int) row_stride_bytes,\n (int) total_samples,\n (int) sizeof(SampleT),\n typeid(SampleT).name(),\n entropy_reduction,\n typeid(CounterT).name(),\n NUM_ACTIVE_CHANNELS,\n NUM_CHANNELS);\n std::cout << CoutCast(max_level) << \"\\n\";\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n std::cout << \"\\n\\tChannel \" << channel << \": \" << num_levels[channel] - 1 << \" bins [\" << lower_level[channel] << \", \" << upper_level[channel] << \")\\n\";\n fflush(stdout);\n\n // Allocate and initialize host and device data\n\n typedef SampleT Foo; // rename type to quelch gcc warnings (bug?)\n CounterT* h_histogram[NUM_ACTIVE_CHANNELS];\n ScaleTransform transform_op[NUM_ACTIVE_CHANNELS];\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n int bins = num_levels[channel] - 1;\n h_histogram[channel] = new CounterT[bins];\n\n transform_op[channel].Init(\n num_levels[channel],\n upper_level[channel],\n lower_level[channel],\n ((upper_level[channel] - lower_level[channel]) / bins));\n }\n\n InitializeBins(\n h_samples, num_levels, transform_op, h_histogram, num_row_pixels, num_rows, row_stride_bytes);\n\n // Allocate and initialize device data\n\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS];\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_histogram[channel], sizeof(CounterT) * (num_levels[channel] - 1)));\n CubDebugExit(cudaMemset(d_histogram[channel], 0, sizeof(CounterT) * (num_levels[channel] - 1)));\n }\n\n // Allocate CDP device arrays\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n\n DispatchEven(\n Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes,\n d_samples, d_histogram, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, true);\n\n // Allocate temporary storage with \"canary\" zones\n int canary_bytes = 256;\n char canary_token = 8;\n char* canary_zone = new char[canary_bytes];\n\n memset(canary_zone, canary_token, canary_bytes);\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes + (canary_bytes * 2)));\n CubDebugExit(cudaMemset(d_temp_storage, canary_token, temp_storage_bytes + (canary_bytes * 2)));\n\n // Run warmup/correctness iteration\n DispatchEven(\n Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n ((char *) d_temp_storage) + canary_bytes, temp_storage_bytes,\n d_samples, d_histogram, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, true);\n\n // Check canary zones\n int error = CompareDeviceResults(canary_zone, (char *) d_temp_storage, canary_bytes, true, g_verbose);\n AssertEquals(0, error);\n error = CompareDeviceResults(canary_zone, ((char *) d_temp_storage) + canary_bytes + temp_storage_bytes, canary_bytes, true, g_verbose);\n AssertEquals(0, error);\n\n // Flush any stdout/stderr\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n fflush(stdout);\n fflush(stderr);\n\n // Check for correctness (and display results, if specified)\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n int channel_error = CompareDeviceResults(h_histogram[channel], d_histogram[channel], num_levels[channel] - 1, true, g_verbose);\n printf(\"\\tChannel %d %s\", channel, channel_error ? \"FAIL\" : \"PASS\\n\");\n error |= channel_error;\n }\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n DispatchEven(\n Int2Type(), Int2Type(), Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes,\n d_samples, d_histogram, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, false);\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(total_samples) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = giga_rate * sizeof(SampleT);\n printf(\"\\t%.3f avg ms, %.3f billion samples/s, %.3f billion bins/s, %.3f billion pixels/s, %.3f logical GB/s\",\n avg_millis,\n giga_rate,\n giga_rate * NUM_ACTIVE_CHANNELS / NUM_CHANNELS,\n giga_rate / NUM_CHANNELS,\n giga_bandwidth);\n }\n\n printf(\"\\n\\n\");\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n if (h_histogram[channel])\n delete[] h_histogram[channel];\n\n if (d_histogram[channel])\n CubDebugExit(g_allocator.DeviceFree(d_histogram[channel]));\n }\n\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, error);\n}\n\n\n/**\n * Test histogram-even (native pointer input)\n */\ntemplate <\n Backend BACKEND,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestEvenNative(\n LevelT max_level,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes) ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n{\n OffsetT total_samples = num_rows * (row_stride_bytes / sizeof(SampleT));\n\n // Allocate and initialize host sample data\n typedef SampleT Foo; // rename type to quelch gcc warnings (bug?)\n SampleT* h_samples = new Foo[total_samples];\n\n InitializeSamples(\n max_level, entropy_reduction, h_samples, num_row_pixels, num_rows, row_stride_bytes);\n\n // Allocate and initialize device data\n SampleT* d_samples = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_samples, sizeof(SampleT) * total_samples));\n CubDebugExit(cudaMemcpy(d_samples, h_samples, sizeof(SampleT) * total_samples, cudaMemcpyHostToDevice));\n\n TestEven(\n max_level, entropy_reduction, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, row_stride_bytes,\n h_samples, d_samples);\n\n // Cleanup\n if (h_samples) delete[] h_samples;\n if (d_samples) CubDebugExit(g_allocator.DeviceFree(d_samples));\n}\n\n\n/**\n * Test histogram-even (native pointer input)\n */\ntemplate <\n Backend BACKEND,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestEvenIterator(\n LevelT max_level,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT lower_level[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n LevelT upper_level[NUM_ACTIVE_CHANNELS], ///< [in] The upper sample value bound (exclusive) for the highest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes) ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n{\n SampleT sample = (SampleT) lower_level[0];\n ConstantInputIterator sample_itr(sample);\n\n TestEven(\n max_level, entropy_reduction, num_levels, lower_level, upper_level,\n num_row_pixels, num_rows, row_stride_bytes,\n sample_itr, sample_itr);\n\n}\n\n\n/**\n * Test histogram-range\n */\ntemplate <\n Backend BACKEND,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestRange(\n LevelT max_level,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS], ///< [in] The number of boundaries (levels) for delineating histogram samples in each active channel. Implies that the number of bins for channeli is num_levels[i] - 1.\n LevelT* levels[NUM_ACTIVE_CHANNELS], ///< [in] The lower sample value bound (inclusive) for the lowest histogram bin in each active channel.\n OffsetT num_row_pixels, ///< [in] The number of multi-channel pixels per row in the region of interest\n OffsetT num_rows, ///< [in] The number of rows in the region of interest\n OffsetT row_stride_bytes) ///< [in] The number of bytes between starts of consecutive rows in the region of interest\n{\n OffsetT total_samples = num_rows * (row_stride_bytes / sizeof(SampleT));\n\n printf(\"\\n----------------------------\\n\");\n printf(\"%s cub::DeviceHistogramRange %d pixels (%d height, %d width, %d-byte row stride), %d %d-byte %s samples (entropy reduction %d), %s counters, %d/%d channels, max sample \",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == NPP) ? \"NPP\" : \"CUB\",\n (int) (num_row_pixels * num_rows),\n (int) num_rows,\n (int) num_row_pixels,\n (int) row_stride_bytes,\n (int) total_samples,\n (int) sizeof(SampleT),\n typeid(SampleT).name(),\n entropy_reduction,\n typeid(CounterT).name(),\n NUM_ACTIVE_CHANNELS,\n NUM_CHANNELS);\n std::cout << CoutCast(max_level) << \"\\n\";\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n printf(\"Channel %d: %d bins [\", channel, num_levels[channel] - 1);\n std::cout << levels[channel][0];\n for (int level = 1; level < num_levels[channel]; ++level)\n std::cout << \", \" << levels[channel][level];\n printf(\"]\\n\");\n }\n fflush(stdout);\n\n // Allocate and initialize host and device data\n typedef SampleT Foo; // rename type to quelch gcc warnings (bug?)\n SampleT* h_samples = new Foo[total_samples];\n CounterT* h_histogram[NUM_ACTIVE_CHANNELS];\n SearchTransform transform_op[NUM_ACTIVE_CHANNELS];\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n transform_op[channel].levels = levels[channel];\n transform_op[channel].num_levels = num_levels[channel];\n\n int bins = num_levels[channel] - 1;\n h_histogram[channel] = new CounterT[bins];\n }\n\n InitializeSamples(\n max_level, entropy_reduction, h_samples, num_row_pixels, num_rows, row_stride_bytes);\n\n InitializeBins(\n h_samples, num_levels, transform_op, h_histogram, num_row_pixels, num_rows, row_stride_bytes);\n\n // Allocate and initialize device data\n SampleT* d_samples = NULL;\n LevelT* d_levels[NUM_ACTIVE_CHANNELS];\n CounterT* d_histogram[NUM_ACTIVE_CHANNELS];\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_samples, sizeof(SampleT) * total_samples));\n CubDebugExit(cudaMemcpy(d_samples, h_samples, sizeof(SampleT) * total_samples, cudaMemcpyHostToDevice));\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_levels[channel], sizeof(LevelT) * num_levels[channel]));\n CubDebugExit(cudaMemcpy(d_levels[channel], levels[channel], sizeof(LevelT) * num_levels[channel], cudaMemcpyHostToDevice));\n\n int bins = num_levels[channel] - 1;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_histogram[channel], sizeof(CounterT) * bins));\n CubDebugExit(cudaMemset(d_histogram[channel], 0, sizeof(CounterT) * bins));\n }\n\n // Allocate CDP device arrays\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n\n DispatchRange(\n Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes,\n d_samples, d_histogram, num_levels, d_levels,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, true);\n\n // Allocate temporary storage with \"canary\" zones\n int canary_bytes = 256;\n char canary_token = 9;\n char* canary_zone = new char[canary_bytes];\n\n memset(canary_zone, canary_token, canary_bytes);\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes + (canary_bytes * 2)));\n CubDebugExit(cudaMemset(d_temp_storage, canary_token, temp_storage_bytes + (canary_bytes * 2)));\n\n // Run warmup/correctness iteration\n DispatchRange(\n Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n ((char *) d_temp_storage) + canary_bytes, temp_storage_bytes,\n d_samples, d_histogram, num_levels, d_levels,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, true);\n\n // Check canary zones\n int error = CompareDeviceResults(canary_zone, (char *) d_temp_storage, canary_bytes, true, g_verbose);\n AssertEquals(0, error);\n error = CompareDeviceResults(canary_zone, ((char *) d_temp_storage) + canary_bytes + temp_storage_bytes, canary_bytes, true, g_verbose);\n AssertEquals(0, error);\n\n // Flush any stdout/stderr\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n fflush(stdout);\n fflush(stderr);\n\n // Check for correctness (and display results, if specified)\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n int channel_error = CompareDeviceResults(h_histogram[channel], d_histogram[channel], num_levels[channel] - 1, true, g_verbose);\n printf(\"\\tChannel %d %s\", channel, channel_error ? \"FAIL\" : \"PASS\\n\");\n error |= channel_error;\n }\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n DispatchRange(\n Int2Type(), Int2Type(), Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes,\n d_samples, d_histogram, num_levels, d_levels,\n num_row_pixels, num_rows, row_stride_bytes,\n 0, false);\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(total_samples) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = giga_rate * sizeof(SampleT);\n printf(\"\\t%.3f avg ms, %.3f billion samples/s, %.3f billion bins/s, %.3f billion pixels/s, %.3f logical GB/s\",\n avg_millis,\n giga_rate,\n giga_rate * NUM_ACTIVE_CHANNELS / NUM_CHANNELS,\n giga_rate / NUM_CHANNELS,\n giga_bandwidth);\n }\n\n printf(\"\\n\\n\");\n\n // Cleanup\n if (h_samples) delete[] h_samples;\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n if (h_histogram[channel])\n delete[] h_histogram[channel];\n\n if (d_histogram[channel])\n CubDebugExit(g_allocator.DeviceFree(d_histogram[channel]));\n\n if (d_levels[channel])\n CubDebugExit(g_allocator.DeviceFree(d_levels[channel]));\n }\n\n if (d_samples) CubDebugExit(g_allocator.DeviceFree(d_samples));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, error);\n}\n\n\n/**\n * Test histogram-even\n */\ntemplate <\n Backend BACKEND,\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestEven(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n OffsetT row_stride_bytes,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS],\n LevelT max_level,\n int max_num_levels)\n{\n LevelT lower_level[NUM_ACTIVE_CHANNELS];\n LevelT upper_level[NUM_ACTIVE_CHANNELS];\n\n // Find smallest level increment\n int max_bins = max_num_levels - 1;\n LevelT min_level_increment = max_level / max_bins;\n\n // Set upper and lower levels for each channel\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n int num_bins = num_levels[channel] - 1;\n lower_level[channel] = (max_level - (num_bins * min_level_increment)) / 2;\n upper_level[channel] = (max_level + (num_bins * min_level_increment)) / 2;\n }\n\n // Test pointer-based samples\n TestEvenNative(\n max_level, entropy_reduction, num_levels, lower_level, upper_level, num_row_pixels, num_rows, row_stride_bytes);\n\n // Test iterator-based samples (CUB-only)\n TestEvenIterator(\n max_level, entropy_reduction, num_levels, lower_level, upper_level, num_row_pixels, num_rows, row_stride_bytes);\n}\n\n\n\n/**\n * Test histogram-range\n */\ntemplate <\n Backend BACKEND,\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestRange(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n OffsetT row_stride_bytes,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS],\n LevelT max_level,\n int max_num_levels)\n{\n // Find smallest level increment\n int max_bins = max_num_levels - 1;\n LevelT min_level_increment = max_level / max_bins;\n\n LevelT* levels[NUM_ACTIVE_CHANNELS];\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n levels[channel] = new LevelT[num_levels[channel]];\n\n int num_bins = num_levels[channel] - 1;\n LevelT lower_level = (max_level - (num_bins * min_level_increment)) / 2;\n\n for (int level = 0; level < num_levels[channel]; ++level)\n levels[channel][level] = lower_level + (level * min_level_increment);\n }\n\n TestRange(\n max_level, entropy_reduction, num_levels, levels, num_row_pixels, num_rows, row_stride_bytes);\n\n for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n delete[] levels[channel];\n\n}\n\n\n\n/**\n * Test different entrypoints\n */\ntemplate <\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid Test(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n OffsetT row_stride_bytes,\n int entropy_reduction,\n int num_levels[NUM_ACTIVE_CHANNELS],\n LevelT max_level,\n int max_num_levels)\n{\n TestEven(\n num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, max_num_levels);\n\n TestRange(\n num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, max_num_levels);\n}\n\n\n/**\n * Test different number of levels\n */\ntemplate <\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid Test(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n OffsetT row_stride_bytes,\n int entropy_reduction,\n LevelT max_level,\n int max_num_levels)\n{\n int num_levels[NUM_ACTIVE_CHANNELS];\n\n// Unnecessary testing\n// // All the same level\n// for (int channel = 0; channel < NUM_ACTIVE_CHANNELS; ++channel)\n// {\n// num_levels[channel] = max_num_levels;\n// }\n// Test(\n// num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, max_num_levels);\n\n // All different levels\n num_levels[0] = max_num_levels;\n for (int channel = 1; channel < NUM_ACTIVE_CHANNELS; ++channel)\n {\n num_levels[channel] = (num_levels[channel - 1] / 2) + 1;\n }\n Test(\n num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, max_num_levels);\n}\n\n\n\n/**\n * Test different entropy-levels\n */\ntemplate <\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid Test(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n OffsetT row_stride_bytes,\n LevelT max_level,\n int max_num_levels)\n{\n Test(\n num_row_pixels, num_rows, row_stride_bytes, 0, max_level, max_num_levels);\n\n Test(\n num_row_pixels, num_rows, row_stride_bytes, -1, max_level, max_num_levels);\n\n Test(\n num_row_pixels, num_rows, row_stride_bytes, 5, max_level, max_num_levels);\n}\n\n\n/**\n * Test different row strides\n */\ntemplate <\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid Test(\n OffsetT num_row_pixels,\n OffsetT num_rows,\n LevelT max_level,\n int max_num_levels)\n{\n OffsetT row_stride_bytes = num_row_pixels * NUM_CHANNELS * sizeof(SampleT);\n\n // No padding\n Test(\n num_row_pixels, num_rows, row_stride_bytes, max_level, max_num_levels);\n\n // 13 samples padding\n Test(\n num_row_pixels, num_rows, row_stride_bytes + (13 * sizeof(SampleT)), max_level, max_num_levels);\n}\n\n\n/**\n * Test different problem sizes\n */\ntemplate <\n typename SampleT,\n int NUM_CHANNELS,\n int NUM_ACTIVE_CHANNELS,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid Test(\n LevelT max_level,\n int max_num_levels)\n{\n // 0 row/col images\n Test(\n OffsetT(1920), OffsetT(0), max_level, max_num_levels);\n Test(\n OffsetT(0), OffsetT(0), max_level, max_num_levels);\n\n // 1080 image\n Test(\n OffsetT(1920), OffsetT(1080), max_level, max_num_levels);\n\n // Sample different aspect ratios sizes\n for (OffsetT rows = 1; rows < 1000000; rows *= 1000)\n {\n for (OffsetT cols = 1; cols < (1000000 / rows); cols *= 1000)\n {\n Test(\n cols, rows, max_level, max_num_levels);\n }\n }\n\n // Randomly select linear problem size between 1:10,000,000\n unsigned int max_int = (unsigned int) -1;\n for (int i = 0; i < 4; ++i)\n {\n unsigned int num_items;\n RandomBits(num_items);\n num_items = (unsigned int) ((double(num_items) * double(10000000)) / double(max_int));\n num_items = CUB_MAX(1, num_items);\n\n Test(\n OffsetT(num_items), 1, max_level, max_num_levels);\n }\n}\n\n\n\n/**\n * Test different channel interleavings (valid specialiation)\n */\ntemplate <\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestChannels(\n LevelT max_level,\n int max_num_levels,\n Int2Type is_valid_tag)\n{\n Test(max_level, max_num_levels);\n Test(max_level, max_num_levels);\n Test(max_level, max_num_levels);\n Test(max_level, max_num_levels);\n}\n\n\n/**\n * Test different channel interleavings (invalid specialiation)\n */\ntemplate <\n typename SampleT,\n typename CounterT,\n typename LevelT,\n typename OffsetT>\nvoid TestChannels(\n LevelT max_level,\n int max_num_levels,\n Int2Type is_valid_tag)\n{}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_row_pixels = -1;\n int entropy_reduction = 0;\n int num_rows = 1;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_verbose_input = args.CheckCmdLineFlag(\"v2\");\n args.GetCmdLineArgument(\"n\", num_row_pixels);\n\n int row_stride_pixels = num_row_pixels;\n\n args.GetCmdLineArgument(\"rows\", num_rows);\n args.GetCmdLineArgument(\"stride\", row_stride_pixels);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n#if defined(QUICK_TEST) || defined(QUICKER_TEST)\n bool compare_npp = args.CheckCmdLineFlag(\"npp\");\n#endif\n\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--rows= \"\n \"[--stride= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--entropy=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"[--npp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n if (num_row_pixels < 0)\n {\n num_row_pixels = 1920 * 1080;\n row_stride_pixels = num_row_pixels;\n }\n\n#if defined(QUICKER_TEST)\n\n // Compile/run quick tests\n {\n // HistogramEven: unsigned char 256 bins\n typedef unsigned char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[1] = {257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n if (compare_npp)\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n\n#elif defined(QUICK_TEST)\n\n // Compile/run quick tests\n {\n // HistogramEven: unsigned char 256 bins\n typedef unsigned char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[1] = {257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n if (compare_npp)\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: 4/4 multichannel Unsigned char 256 bins\n typedef unsigned char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[4] = {257, 257, 257, 257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 4;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: 3/4 multichannel Unsigned char 256 bins\n typedef unsigned char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[3] = {257, 257, 257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 4;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n if (compare_npp)\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: short [0,1024] 256 bins\n typedef unsigned short SampleT;\n typedef unsigned short LevelT;\n\n LevelT max_level = 1024;\n int num_levels[1] = {257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: float [0,1.0] 256 bins\n typedef float SampleT;\n typedef float LevelT;\n\n LevelT max_level = 1.0;\n int num_levels[1] = {257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: 3/4 multichannel float [0,1.0] 256 bins\n typedef float SampleT;\n typedef float LevelT;\n\n LevelT max_level = 1.0;\n int num_levels[3] = {257, 257, 257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 4;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramRange: signed char 256 bins\n typedef signed char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[1] = {257};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestRange(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramRange: 3/4 channel, unsigned char, varied bins (256, 128, 64)\n typedef unsigned char SampleT;\n typedef int LevelT;\n\n LevelT max_level = 256;\n int num_levels[3] = {257, 129, 65};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 4;\n\n TestRange(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n {\n // HistogramEven: double [0,1.0] 64 bins\n typedef double SampleT;\n typedef double LevelT;\n\n LevelT max_level = 1.0;\n int num_levels[1] = {65};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n {\n // HistogramEven: short [0,1024] 512 bins\n typedef unsigned short SampleT;\n typedef unsigned short LevelT;\n\n LevelT max_level = 1024;\n int num_levels[1] = {513};\n int row_stride_bytes = sizeof(SampleT) * row_stride_pixels * 1;\n\n TestEven(num_row_pixels, num_rows, row_stride_bytes, entropy_reduction, num_levels, max_level, num_levels[0]);\n }\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n TestChannels (256, 256 + 1, Int2Type());\n TestChannels (256, 256 + 1, Int2Type());\n TestChannels (128, 128 + 1, Int2Type());\n TestChannels (8192, 8192 + 1, Int2Type());\n TestChannels (1.0, 256 + 1, Int2Type());\n\n\t\t// Test down-conversion of size_t offsets to int\n TestChannels (256, 256 + 1, Int2Type<(sizeof(size_t) != sizeof(int))>());\n }\n\n#endif\n\n return 0;\n}\n\n" }, { "path": "external/cub/test/test_device_radix_sort.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceRadixSort utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n\n#include \"test_util.h\"\n\n#include \n#include \n#include \n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method (allows overwriting of input)\n CUB_NO_OVERWRITE, // CUB method (disallows overwriting of input)\n\n CUB_SEGMENTED, // CUB method (allows overwriting of input)\n CUB_SEGMENTED_NO_OVERWRITE, // CUB method (disallows overwriting of input)\n\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different DeviceRadixSort entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to CUB sorting entrypoint (specialized for ascending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n return DeviceRadixSort::SortPairs(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, begin_bit, end_bit, stream, debug_synchronous);\n}\n\n/**\n * Dispatch to CUB_NO_OVERWRITE sorting entrypoint (specialized for ascending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n KeyT const *const_keys_itr = d_keys.Current();\n ValueT const *const_values_itr = d_values.Current();\n\n cudaError_t retval = DeviceRadixSort::SortPairs(\n d_temp_storage, temp_storage_bytes,\n const_keys_itr, d_keys.Alternate(), const_values_itr, d_values.Alternate(),\n num_items, begin_bit, end_bit, stream, debug_synchronous);\n\n d_keys.selector ^= 1;\n d_values.selector ^= 1;\n return retval;\n}\n\n/**\n * Dispatch to CUB sorting entrypoint (specialized for descending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n return DeviceRadixSort::SortPairsDescending(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, begin_bit, end_bit, stream, debug_synchronous);\n}\n\n\n/**\n * Dispatch to CUB_NO_OVERWRITE sorting entrypoint (specialized for descending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n KeyT const *const_keys_itr = d_keys.Current();\n ValueT const *const_values_itr = d_values.Current();\n\n cudaError_t retval = DeviceRadixSort::SortPairsDescending(\n d_temp_storage, temp_storage_bytes,\n const_keys_itr, d_keys.Alternate(), const_values_itr, d_values.Alternate(),\n num_items, begin_bit, end_bit, stream, debug_synchronous);\n\n d_keys.selector ^= 1;\n d_values.selector ^= 1;\n return retval;\n}\n\n//---------------------------------------------------------------------\n// Dispatch to different DeviceRadixSort entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to CUB_SEGMENTED sorting entrypoint (specialized for ascending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n return DeviceSegmentedRadixSort::SortPairs(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, num_segments, d_segment_offsets, d_segment_offsets + 1,\n begin_bit, end_bit, stream, debug_synchronous);\n}\n\n/**\n * Dispatch to CUB_SEGMENTED_NO_OVERWRITE sorting entrypoint (specialized for ascending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n KeyT const *const_keys_itr = d_keys.Current();\n ValueT const *const_values_itr = d_values.Current();\n\n cudaError_t retval = DeviceSegmentedRadixSort::SortPairs(\n d_temp_storage, temp_storage_bytes,\n const_keys_itr, d_keys.Alternate(), const_values_itr, d_values.Alternate(),\n num_items, num_segments, d_segment_offsets, d_segment_offsets + 1,\n begin_bit, end_bit, stream, debug_synchronous);\n\n d_keys.selector ^= 1;\n d_values.selector ^= 1;\n return retval;\n}\n\n\n/**\n * Dispatch to CUB_SEGMENTED sorting entrypoint (specialized for descending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n return DeviceSegmentedRadixSort::SortPairsDescending(\n d_temp_storage, temp_storage_bytes,\n d_keys, d_values,\n num_items, num_segments, d_segment_offsets, d_segment_offsets + 1,\n begin_bit, end_bit, stream, debug_synchronous);\n}\n\n/**\n * Dispatch to CUB_SEGMENTED_NO_OVERWRITE sorting entrypoint (specialized for descending)\n */\ntemplate \nCUB_RUNTIME_FUNCTION\n__forceinline__\ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n KeyT const *const_keys_itr = d_keys.Current();\n ValueT const *const_values_itr = d_values.Current();\n\n cudaError_t retval = DeviceSegmentedRadixSort::SortPairsDescending(\n d_temp_storage, temp_storage_bytes,\n const_keys_itr, d_keys.Alternate(), const_values_itr, d_values.Alternate(),\n num_items, num_segments, d_segment_offsets, d_segment_offsets + 1,\n begin_bit, end_bit, stream, debug_synchronous);\n\n d_keys.selector ^= 1;\n d_values.selector ^= 1;\n return retval;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch keys-only to Thrust sorting entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_keys_wrapper(d_keys.Current());\n\n if (IS_DESCENDING) thrust::reverse(d_keys_wrapper, d_keys_wrapper + num_items);\n thrust::sort(d_keys_wrapper, d_keys_wrapper + num_items);\n if (IS_DESCENDING) thrust::reverse(d_keys_wrapper, d_keys_wrapper + num_items);\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch key-value pairs to Thrust sorting entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_keys_wrapper(d_keys.Current());\n thrust::device_ptr d_values_wrapper(d_values.Current());\n\n if (IS_DESCENDING) {\n thrust::reverse(d_keys_wrapper, d_keys_wrapper + num_items);\n thrust::reverse(d_values_wrapper, d_values_wrapper + num_items);\n }\n\n thrust::sort_by_key(d_keys_wrapper, d_keys_wrapper + num_items, d_values_wrapper);\n\n if (IS_DESCENDING) {\n thrust::reverse(d_keys_wrapper, d_keys_wrapper + num_items);\n thrust::reverse(d_values_wrapper, d_values_wrapper + num_items);\n }\n }\n\n return cudaSuccess;\n}\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceRadixSort\n */\ntemplate \n__global__ void CnpDispatchKernel(\n Int2Type is_descending,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t temp_storage_bytes,\n DoubleBuffer d_keys,\n DoubleBuffer d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n bool debug_synchronous)\n{\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(\n is_descending, Int2Type(), d_selector, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_keys, d_values,\n num_items, num_segments, d_segment_offsets,\n begin_bit, end_bit, 0, debug_synchronous);\n *d_temp_storage_bytes = temp_storage_bytes;\n *d_selector = d_keys.selector;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type is_descending,\n Int2Type dispatch_to,\n int *d_selector,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n DoubleBuffer &d_keys,\n DoubleBuffer &d_values,\n int num_items,\n int num_segments,\n const int *d_segment_offsets,\n int begin_bit,\n int end_bit,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(\n is_descending, d_selector, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_keys, d_values,\n num_items, num_segments, d_segment_offsets,\n begin_bit, end_bit, debug_synchronous);\n\n // Copy out selector\n CubDebugExit(cudaMemcpy(&d_keys.selector, d_selector, sizeof(int) * 1, cudaMemcpyDeviceToHost));\n d_values.selector = d_keys.selector;\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Problem generation\n//---------------------------------------------------------------------\n\n\n/**\n * Simple key-value pairing\n */\ntemplate <\n typename KeyT,\n typename ValueT,\n bool IS_FLOAT = (Traits::CATEGORY == FLOATING_POINT)>\nstruct Pair\n{\n KeyT key;\n ValueT value;\n\n bool operator<(const Pair &b) const\n {\n return (key < b.key);\n }\n};\n\n\n/**\n * Simple key-value pairing (specialized for bool types)\n */\ntemplate \nstruct Pair\n{\n bool key;\n ValueT value;\n\n bool operator<(const Pair &b) const\n {\n return (!key && b.key);\n }\n};\n\n\n/**\n * Simple key-value pairing (specialized for floating point types)\n */\ntemplate \nstruct Pair\n{\n KeyT key;\n ValueT value;\n\n bool operator<(const Pair &b) const\n {\n if (key < b.key)\n return true;\n\n if (key > b.key)\n return false;\n\n // KeyT in unsigned bits\n typedef typename Traits::UnsignedBits UnsignedBits;\n\n // Return true if key is negative zero and b.key is positive zero\n UnsignedBits key_bits = *reinterpret_cast(const_cast(&key));\n UnsignedBits b_key_bits = *reinterpret_cast(const_cast(&b.key));\n UnsignedBits HIGH_BIT = Traits::HIGH_BIT;\n\n return ((key_bits & HIGH_BIT) != 0) && ((b_key_bits & HIGH_BIT) == 0);\n }\n};\n\n\n/**\n * Initialize key data\n */\ntemplate \nvoid InitializeKeyBits(\n GenMode gen_mode,\n KeyT *h_keys,\n int num_items,\n int entropy_reduction)\n{\n for (int i = 0; i < num_items; ++i)\n InitValue(gen_mode, h_keys[i], i);\n}\n\n\n/**\n * Initialize solution\n */\ntemplate \nvoid InitializeSolution(\n KeyT *h_keys,\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n int begin_bit,\n int end_bit,\n int *&h_reference_ranks,\n KeyT *&h_reference_keys)\n{\n typedef Pair PairT;\n\n PairT *h_pairs = new PairT[num_items];\n\n int num_bits = end_bit - begin_bit;\n for (int i = 0; i < num_items; ++i)\n {\n\n // Mask off unwanted portions\n if (num_bits < sizeof(KeyT) * 8)\n {\n unsigned long long base = 0;\n memcpy(&base, &h_keys[i], sizeof(KeyT));\n base &= ((1ull << num_bits) - 1) << begin_bit;\n memcpy(&h_pairs[i].key, &base, sizeof(KeyT));\n }\n else\n {\n h_pairs[i].key = h_keys[i];\n }\n\n h_pairs[i].value = i;\n }\n\n printf(\"\\nSorting reference solution on CPU (%d segments)...\", num_segments); fflush(stdout);\n\n for (int i = 0; i < num_segments; ++i)\n {\n if (IS_DESCENDING) std::reverse(h_pairs + h_segment_offsets[i], h_pairs + h_segment_offsets[i + 1]);\n std::stable_sort( h_pairs + h_segment_offsets[i], h_pairs + h_segment_offsets[i + 1]);\n if (IS_DESCENDING) std::reverse(h_pairs + h_segment_offsets[i], h_pairs + h_segment_offsets[i + 1]);\n }\n\n printf(\" Done.\\n\"); fflush(stdout);\n\n h_reference_ranks = new int[num_items];\n h_reference_keys = new KeyT[num_items];\n\n for (int i = 0; i < num_items; ++i)\n {\n h_reference_ranks[i] = h_pairs[i].value;\n h_reference_keys[i] = h_keys[h_pairs[i].value];\n }\n\n if (h_pairs) delete[] h_pairs;\n}\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Test DeviceRadixSort\n */\ntemplate <\n Backend BACKEND,\n bool IS_DESCENDING,\n typename KeyT,\n typename ValueT>\nvoid Test(\n KeyT *h_keys,\n ValueT *h_values,\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n int begin_bit,\n int end_bit,\n KeyT *h_reference_keys,\n ValueT *h_reference_values)\n{\n const bool KEYS_ONLY = Equals::VALUE;\n\n printf(\"%s %s cub::DeviceRadixSort %d items, %d segments, %d-byte keys (%s) %d-byte values (%s), descending %d, begin_bit %d, end_bit %d\\n\",\n (BACKEND == CUB_NO_OVERWRITE) ? \"CUB_NO_OVERWRITE\" : (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n (KEYS_ONLY) ? \"keys-only\" : \"key-value\",\n num_items, num_segments,\n (int) sizeof(KeyT), typeid(KeyT).name(), (KEYS_ONLY) ? 0 : (int) sizeof(ValueT), typeid(ValueT).name(),\n IS_DESCENDING, begin_bit, end_bit);\n fflush(stdout);\n\n if (g_verbose)\n {\n printf(\"Input keys:\\n\");\n DisplayResults(h_keys, num_items);\n printf(\"\\n\\n\");\n }\n\n // Allocate device arrays\n DoubleBuffer d_keys;\n DoubleBuffer d_values;\n int *d_selector;\n int *d_segment_offsets;\n size_t *d_temp_storage_bytes;\n cudaError_t *d_cdp_error;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[0], sizeof(KeyT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(KeyT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_selector, sizeof(int) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_segment_offsets, sizeof(int) * (num_segments + 1)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n if (!KEYS_ONLY)\n {\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[0], sizeof(ValueT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values.d_buffers[1], sizeof(ValueT) * num_items));\n }\n\n // Allocate temporary storage (and make it un-aligned)\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n CubDebugExit(Dispatch(\n Int2Type(), Int2Type(), d_selector, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_keys, d_values,\n num_items, num_segments, d_segment_offsets,\n begin_bit, end_bit, 0, true));\n\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes + 1));\n void* mis_aligned_temp = static_cast(d_temp_storage) + 1;\n\n // Initialize/clear device arrays\n d_keys.selector = 0;\n CubDebugExit(cudaMemcpy(d_keys.d_buffers[0], h_keys, sizeof(KeyT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_keys.d_buffers[1], 0, sizeof(KeyT) * num_items));\n if (!KEYS_ONLY)\n {\n d_values.selector = 0;\n CubDebugExit(cudaMemcpy(d_values.d_buffers[0], h_values, sizeof(ValueT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_values.d_buffers[1], 0, sizeof(ValueT) * num_items));\n }\n CubDebugExit(cudaMemcpy(d_segment_offsets, h_segment_offsets, sizeof(int) * (num_segments + 1), cudaMemcpyHostToDevice));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(\n Int2Type(), Int2Type(), d_selector, d_temp_storage_bytes, d_cdp_error,\n mis_aligned_temp, temp_storage_bytes, d_keys, d_values,\n num_items, num_segments, d_segment_offsets,\n begin_bit, end_bit, 0, true));\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Check for correctness (and display results, if specified)\n printf(\"Warmup done. Checking results:\\n\"); fflush(stdout);\n int compare = CompareDeviceResults(h_reference_keys, d_keys.Current(), num_items, true, g_verbose);\n printf(\"\\t Compare keys (selector %d): %s \", d_keys.selector, compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n if (!KEYS_ONLY)\n {\n int values_compare = CompareDeviceResults(h_reference_values, d_values.Current(), num_items, true, g_verbose);\n compare |= values_compare;\n printf(\"\\t Compare values (selector %d): %s \", d_values.selector, values_compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n if (BACKEND == CUB_NO_OVERWRITE)\n {\n // Check that input isn't overwritten\n int input_compare = CompareDeviceResults(h_keys, d_keys.d_buffers[0], num_items, true, g_verbose);\n compare |= input_compare;\n printf(\"\\t Compare input keys: %s \", input_compare ? \"FAIL\" : \"PASS\"); fflush(stdout);\n }\n\n // Performance\n if (g_timing_iterations)\n printf(\"\\nPerforming timing iterations:\\n\"); fflush(stdout);\n\n GpuTimer gpu_timer;\n float elapsed_millis = 0.0f;\n for (int i = 0; i < g_timing_iterations; ++i)\n {\n // Initialize/clear device arrays\n CubDebugExit(cudaMemcpy(d_keys.d_buffers[d_keys.selector], h_keys, sizeof(KeyT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_keys.d_buffers[d_keys.selector ^ 1], 0, sizeof(KeyT) * num_items));\n if (!KEYS_ONLY)\n {\n CubDebugExit(cudaMemcpy(d_values.d_buffers[d_values.selector], h_values, sizeof(ValueT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_values.d_buffers[d_values.selector ^ 1], 0, sizeof(ValueT) * num_items));\n }\n\n gpu_timer.Start();\n CubDebugExit(Dispatch(\n Int2Type(), Int2Type(), d_selector, d_temp_storage_bytes, d_cdp_error,\n mis_aligned_temp, temp_storage_bytes, d_keys, d_values,\n num_items, num_segments, d_segment_offsets,\n begin_bit, end_bit, 0, false));\n gpu_timer.Stop();\n elapsed_millis += gpu_timer.ElapsedMillis();\n }\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = (KEYS_ONLY) ?\n giga_rate * sizeof(KeyT) * 2 :\n giga_rate * (sizeof(KeyT) + sizeof(ValueT)) * 2;\n printf(\"\\n%.3f elapsed ms, %.3f avg ms, %.3f billion items/s, %.3f logical GB/s\", elapsed_millis, avg_millis, giga_rate, giga_bandwidth);\n }\n\n printf(\"\\n\\n\");\n\n // Cleanup\n if (d_keys.d_buffers[0]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[0]));\n if (d_keys.d_buffers[1]) CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[1]));\n if (d_values.d_buffers[0]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[0]));\n if (d_values.d_buffers[1]) CubDebugExit(g_allocator.DeviceFree(d_values.d_buffers[1]));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_selector) CubDebugExit(g_allocator.DeviceFree(d_selector));\n if (d_segment_offsets) CubDebugExit(g_allocator.DeviceFree(d_segment_offsets));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n\n // Correctness asserts\n AssertEquals(0, compare);\n}\n\n\n/**\n * Test backend\n */\ntemplate \nvoid TestBackend(\n KeyT *h_keys,\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n int begin_bit,\n int end_bit,\n KeyT *h_reference_keys,\n int *h_reference_ranks)\n{\n const bool KEYS_ONLY = Equals::VALUE;\n\n ValueT *h_values = NULL;\n ValueT *h_reference_values = NULL;\n\n if (!KEYS_ONLY)\n {\n h_values = new ValueT[num_items];\n h_reference_values = new ValueT[num_items];\n\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(INTEGER_SEED, h_values[i], i);\n InitValue(INTEGER_SEED, h_reference_values[i], h_reference_ranks[i]);\n }\n }\n\n if (num_segments == 1)\n {\n // Test single-segment implementations\n Test( h_keys, h_values, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_values);\n Test( h_keys, h_values, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_values);\n#ifdef CUB_CDP\n Test( h_keys, h_values, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_values);\n#endif\n }\n\n // Test multi-segment implementations\n Test( h_keys, h_values, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_values);\n Test( h_keys, h_values, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_values);\n\n if (h_values) delete[] h_values;\n if (h_reference_values) delete[] h_reference_values;\n}\n\n\n\n\n/**\n * Test value type\n */\ntemplate \nvoid TestValueTypes(\n KeyT *h_keys,\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n int begin_bit,\n int end_bit)\n{\n // Initialize the solution\n\n int *h_reference_ranks = NULL;\n KeyT *h_reference_keys = NULL;\n InitializeSolution(h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_ranks, h_reference_keys);\n\n // Test keys-only\n TestBackend (h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_ranks);\n\n // Test with 8b value\n TestBackend (h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_ranks);\n\n // Test with 32b value\n TestBackend (h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_ranks);\n\n // Test with 64b value\n TestBackend(h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_ranks);\n\n // Test with non-trivially-constructable value\n TestBackend (h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit, h_reference_keys, h_reference_ranks);\n\n // Cleanup\n if (h_reference_ranks) delete[] h_reference_ranks;\n if (h_reference_keys) delete[] h_reference_keys;\n}\n\n\n\n/**\n * Test ascending/descending\n */\ntemplate \nvoid TestDirection(\n KeyT *h_keys,\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n int begin_bit,\n int end_bit)\n{\n TestValueTypes(h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit);\n TestValueTypes(h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit);\n}\n\n\n/**\n * Test different bit ranges\n */\ntemplate \nvoid TestBits(\n KeyT *h_keys,\n int num_items,\n int num_segments,\n int *h_segment_offsets)\n{\n // Don't test partial-word sorting for boolean, fp, or signed types (the bit-flipping techniques get in the way)\n if ((Traits::CATEGORY == UNSIGNED_INTEGER) && (!Equals::VALUE))\n {\n // Partial bits\n int begin_bit = 1;\n int end_bit = (sizeof(KeyT) * 8) - 1;\n printf(\"Testing key bits [%d,%d)\\n\", begin_bit, end_bit); fflush(stdout);\n TestDirection(h_keys, num_items, num_segments, h_segment_offsets, begin_bit, end_bit);\n\n // Across subword boundaries\n int mid_bit = sizeof(KeyT) * 4;\n printf(\"Testing key bits [%d,%d)\\n\", mid_bit - 1, mid_bit + 1); fflush(stdout);\n TestDirection(h_keys, num_items, num_segments, h_segment_offsets, mid_bit - 1, mid_bit + 1);\n }\n\n printf(\"Testing key bits [%d,%d)\\n\", 0, int(sizeof(KeyT)) * 8); fflush(stdout);\n TestDirection(h_keys, num_items, num_segments, h_segment_offsets, 0, sizeof(KeyT) * 8);\n}\n\n\n/**\n * Test different segment compositions\n */\ntemplate \nvoid TestSegments(\n KeyT *h_keys,\n int num_items,\n int max_segments)\n{\n int *h_segment_offsets = new int[max_segments + 1];\n\n for (int num_segments = max_segments; num_segments > 1; num_segments = (num_segments + 32 - 1) / 32)\n {\n if (num_items / num_segments < 128 * 1000) {\n // Right now we assign a single thread block to each segment, so lets keep it to under 128K items per segment\n InitializeSegments(num_items, num_segments, h_segment_offsets);\n TestBits(h_keys, num_items, num_segments, h_segment_offsets);\n }\n }\n\n // Test single segment\n if (num_items < 128 * 1000) {\n // Right now we assign a single thread block to each segment, so lets keep it to under 128K items per segment\n InitializeSegments(num_items, 1, h_segment_offsets);\n TestBits(h_keys, num_items, 1, h_segment_offsets);\n }\n\n if (h_segment_offsets) delete[] h_segment_offsets;\n}\n\n\n/**\n * Test different (sub)lengths and number of segments\n */\ntemplate \nvoid TestSizes(\n KeyT *h_keys,\n int max_items,\n int max_segments)\n{\n for (int num_items = max_items; num_items > 1; num_items = (num_items + 32 - 1) / 32)\n {\n TestSegments(h_keys, num_items, max_segments);\n }\n TestSegments(h_keys, 1, max_segments);\n TestSegments(h_keys, 0, max_segments);\n}\n\n\n/**\n * Test key sampling distributions\n */\ntemplate \nvoid TestGen(\n int max_items,\n int max_segments)\n{\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n if (max_items < 0)\n max_items = (ptx_version > 100) ? 9000003 : max_items = 5000003;\n\n if (max_segments < 0)\n max_segments = 5003;\n\n KeyT *h_keys = new KeyT[max_items];\n\n for (int entropy_reduction = 0; entropy_reduction <= 6; entropy_reduction += 3)\n {\n printf(\"\\nTesting random %s keys with entropy reduction factor %d\\n\", typeid(KeyT).name(), entropy_reduction); fflush(stdout);\n InitializeKeyBits(RANDOM, h_keys, max_items, entropy_reduction);\n TestSizes(h_keys, max_items, max_segments);\n }\n\n printf(\"\\nTesting uniform %s keys\\n\", typeid(KeyT).name()); fflush(stdout);\n InitializeKeyBits(UNIFORM, h_keys, max_items, 0);\n TestSizes(h_keys, max_items, max_segments);\n\n printf(\"\\nTesting natural number %s keys\\n\", typeid(KeyT).name()); fflush(stdout);\n InitializeKeyBits(INTEGER_SEED, h_keys, max_items, 0);\n TestSizes(h_keys, max_items, max_segments);\n\n if (h_keys) delete[] h_keys;\n}\n\n\n//---------------------------------------------------------------------\n// Simple test\n//---------------------------------------------------------------------\n\ntemplate <\n Backend BACKEND,\n typename KeyT,\n typename ValueT,\n bool IS_DESCENDING>\nvoid Test(\n int num_items,\n int num_segments,\n GenMode gen_mode,\n int entropy_reduction,\n int begin_bit,\n int end_bit)\n{\n const bool KEYS_ONLY = Equals::VALUE;\n\n KeyT *h_keys = new KeyT[num_items];\n int *h_reference_ranks = NULL;\n KeyT *h_reference_keys = NULL;\n ValueT *h_values = NULL;\n ValueT *h_reference_values = NULL;\n int *h_segment_offsets = new int[num_segments + 1];\n\n if (end_bit < 0)\n end_bit = sizeof(KeyT) * 8;\n\n InitializeKeyBits(gen_mode, h_keys, num_items, entropy_reduction);\n InitializeSegments(num_items, num_segments, h_segment_offsets);\n InitializeSolution(\n h_keys, num_items, num_segments, h_segment_offsets,\n begin_bit, end_bit, h_reference_ranks, h_reference_keys);\n\n if (!KEYS_ONLY)\n {\n h_values = new ValueT[num_items];\n h_reference_values = new ValueT[num_items];\n\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(INTEGER_SEED, h_values[i], i);\n InitValue(INTEGER_SEED, h_reference_values[i], h_reference_ranks[i]);\n }\n }\n if (h_reference_ranks) delete[] h_reference_ranks;\n\n printf(\"\\nTesting bits [%d,%d) of %s keys with gen-mode %d\\n\", begin_bit, end_bit, typeid(KeyT).name(), gen_mode); fflush(stdout);\n Test(\n h_keys, h_values,\n num_items, num_segments, h_segment_offsets,\n begin_bit, end_bit, h_reference_keys, h_reference_values);\n\n if (h_keys) delete[] h_keys;\n if (h_reference_keys) delete[] h_reference_keys;\n if (h_values) delete[] h_values;\n if (h_reference_values) delete[] h_reference_values;\n if (h_segment_offsets) delete[] h_segment_offsets;\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int bits = -1;\n int num_items = -1;\n int num_segments = -1;\n int entropy_reduction = 0;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"s\", num_segments);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"bits\", bits);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--bits=]\"\n \"[--n= \"\n \"[--s= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--entropy=]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef QUICKER_TEST\n\n enum {\n IS_DESCENDING = false\n };\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 48000000;\n if (num_segments < 0) num_segments = 5000;\n\n\n Test( num_items, num_segments, RANDOM, entropy_reduction, 0, bits);\n\n Test( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n Test( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n#elif defined(QUICK_TEST)\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 48000000;\n if (num_segments < 0) num_segments = 5000;\n\n // Compare CUB and thrust on 32b keys-only\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n // Compare CUB and thrust on 64b keys-only\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n\n // Compare CUB and thrust on 32b key-value pairs\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n // Compare CUB and thrust on 64b key-value pairs\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n Test ( num_items, 1, RANDOM, entropy_reduction, 0, bits);\n\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n TestGen (num_items, num_segments);\n\n TestGen (num_items, num_segments);\n\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n TestGen (num_items, num_segments);\n\n }\n\n#endif\n\n return 0;\n}\n\n" }, { "path": "external/cub/test/test_device_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceReduce utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n\n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nint g_ptx_version;\nint g_sm_count;\nbool g_verbose = false;\nbool g_verbose_input = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n CUB_SEGMENTED, // CUB segmented method\n CUB_CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n THRUST, // Thrust method\n};\n\n\n// Custom max functor\nstruct CustomMax\n{\n /// Boolean max operator, returns (a > b) ? a : b\n template \n __host__ __device__ __forceinline__ OutputT operator()(const OutputT &a, const OutputT &b)\n {\n return CUB_MAX(a, b);\n }\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB DeviceReduce entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to reduce entrypoint (custom-max)\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Max-identity\n OutputT identity = Traits::Lowest(); // replace with std::numeric_limits::lowest() when C++ support is more prevalent\n\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::Reduce(d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, reduction_op, identity,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to sum entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Sum reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::Sum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to min entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Min reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::Min(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to max entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Max reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::Max(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to argmin entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::ArgMin reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::ArgMin(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to argmax entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::ArgMax reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::ArgMax(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB DeviceSegmentedReduce entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to reduce entrypoint (custom-max)\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n // Max-identity\n OutputT identity = Traits::Lowest(); // replace with std::numeric_limits::lowest() when C++ support is more prevalent\n\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::Reduce(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1, reduction_op, identity,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to sum entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Sum reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::Sum(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to min entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Min reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::Min(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to max entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::Max reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::Max(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to argmin entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::ArgMin reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::ArgMin(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1,\n stream, debug_synchronous);\n }\n return error;\n}\n\n/**\n * Dispatch to argmax entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n cub::ArgMax reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to device reduction directly\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSegmentedReduce::ArgMax(d_temp_storage, temp_storage_bytes,\n d_in, d_out, max_segments, d_segment_offsets, d_segment_offsets + 1,\n stream, debug_synchronous);\n }\n return error;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to reduction entrypoint (min or max specialization)\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n OutputT init;\n CubDebugExit(cudaMemcpy(&init, d_in + 0, sizeof(OutputT), cudaMemcpyDeviceToHost));\n\n thrust::device_ptr d_in_wrapper(d_in);\n OutputT retval;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n retval = thrust::reduce(d_in_wrapper, d_in_wrapper + num_items, init, reduction_op);\n }\n\n if (!Equals >::VALUE)\n CubDebugExit(cudaMemcpy(d_out, &retval, sizeof(OutputT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n/**\n * Dispatch to reduction entrypoint (sum specialization)\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n Sum reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n OutputT retval;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n retval = thrust::reduce(d_in_wrapper, d_in_wrapper + num_items);\n }\n\n if (!Equals >::VALUE)\n CubDebugExit(cudaMemcpy(d_out, &retval, sizeof(OutputT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n//---------------------------------------------------------------------\n// CUDA nested-parallelism test kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceReduce\n */\ntemplate <\n typename InputIteratorT,\n typename OutputIteratorT,\n typename OffsetIteratorT,\n typename ReductionOpT>\n__global__ void CnpDispatchKernel(\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n bool debug_synchronous)\n{\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(Int2Type(), timing_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, max_segments, d_segment_offsets, reduction_op, 0, debug_synchronous);\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CUB_CDP kernel\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n int num_items,\n int max_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(timing_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, max_segments, d_segment_offsets, reduction_op, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Problem generation\n//---------------------------------------------------------------------\n\n/// Initialize problem\ntemplate \nvoid Initialize(\n GenMode gen_mode,\n InputT *h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n }\n\n if (g_verbose_input)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/// Solve problem (max/custom-max functor)\ntemplate \nstruct Solution\n{\n typedef _OutputT OutputT;\n\n template \n static void Solve(HostInputIteratorT h_in, OutputT *h_reference, OffsetT num_segments, OffsetIteratorT h_segment_offsets,\n ReductionOpT reduction_op)\n {\n for (int i = 0; i < num_segments; ++i)\n {\n OutputT aggregate = Traits::Lowest(); // replace with std::numeric_limits::lowest() when C++ support is more prevalent\n for (int j = h_segment_offsets[i]; j < h_segment_offsets[i + 1]; ++j)\n aggregate = reduction_op(aggregate, OutputT(h_in[j]));\n h_reference[i] = aggregate;\n }\n }\n};\n\n/// Solve problem (min functor)\ntemplate \nstruct Solution\n{\n typedef _OutputT OutputT;\n\n template \n static void Solve(HostInputIteratorT h_in, OutputT *h_reference, OffsetT num_segments, OffsetIteratorT h_segment_offsets,\n cub::Min reduction_op)\n {\n for (int i = 0; i < num_segments; ++i)\n {\n OutputT aggregate = Traits::Max(); // replace with std::numeric_limits::max() when C++ support is more prevalent\n for (int j = h_segment_offsets[i]; j < h_segment_offsets[i + 1]; ++j)\n aggregate = reduction_op(aggregate, OutputT(h_in[j]));\n h_reference[i] = aggregate;\n }\n }\n};\n\n\n/// Solve problem (sum functor)\ntemplate \nstruct Solution\n{\n typedef _OutputT OutputT;\n\n template \n static void Solve(HostInputIteratorT h_in, OutputT *h_reference, OffsetT num_segments, OffsetIteratorT h_segment_offsets,\n cub::Sum reduction_op)\n {\n for (int i = 0; i < num_segments; ++i)\n {\n OutputT aggregate;\n InitValue(INTEGER_SEED, aggregate, 0);\n for (int j = h_segment_offsets[i]; j < h_segment_offsets[i + 1]; ++j)\n aggregate = reduction_op(aggregate, OutputT(h_in[j]));\n h_reference[i] = aggregate;\n }\n }\n};\n\n/// Solve problem (argmin functor)\ntemplate \nstruct Solution\n{\n typedef KeyValuePair OutputT;\n\n template \n static void Solve(HostInputIteratorT h_in, OutputT *h_reference, OffsetT num_segments, OffsetIteratorT h_segment_offsets,\n cub::ArgMin reduction_op)\n {\n for (int i = 0; i < num_segments; ++i)\n {\n OutputT aggregate(1, Traits::Max()); // replace with std::numeric_limits::max() when C++ support is more prevalent\n for (int j = h_segment_offsets[i]; j < h_segment_offsets[i + 1]; ++j)\n {\n OutputT item(j - h_segment_offsets[i], OutputValueT(h_in[j]));\n aggregate = reduction_op(aggregate, item);\n }\n h_reference[i] = aggregate;\n }\n }\n};\n\n\n/// Solve problem (argmax functor)\ntemplate \nstruct Solution\n{\n typedef KeyValuePair OutputT;\n\n template \n static void Solve(HostInputIteratorT h_in, OutputT *h_reference, OffsetT num_segments, OffsetIteratorT h_segment_offsets,\n cub::ArgMax reduction_op)\n {\n for (int i = 0; i < num_segments; ++i)\n {\n OutputT aggregate(1, Traits::Lowest()); // replace with std::numeric_limits::lowest() when C++ support is more prevalent\n for (int j = h_segment_offsets[i]; j < h_segment_offsets[i + 1]; ++j)\n {\n OutputT item(j - h_segment_offsets[i], OutputValueT(h_in[j]));\n aggregate = reduction_op(aggregate, item);\n }\n h_reference[i] = aggregate;\n }\n }\n};\n\n\n//---------------------------------------------------------------------\n// Problem generation\n//---------------------------------------------------------------------\n\n/// Test DeviceReduce for a given problem input\ntemplate <\n typename BackendT,\n typename DeviceInputIteratorT,\n typename HostReferenceIteratorT,\n typename OffsetT,\n typename OffsetIteratorT,\n typename ReductionOpT>\nvoid Test(\n BackendT backend,\n DeviceInputIteratorT d_in,\n OffsetT num_items,\n OffsetT num_segments,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op,\n HostReferenceIteratorT h_reference)\n{\n // Input and output data types\n typedef typename std::iterator_traits::value_type InputT;\n typedef typename std::iterator_traits::value_type OutputT;\n\n // Allocate CUB_CDP device arrays for temp storage size and error\n OutputT *d_out = NULL;\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(OutputT) * num_segments));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Inquire temp device storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(backend, 1,\n d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, num_segments, d_segment_offsets,\n reduction_op, 0, true));\n\n // Allocate temp device storage\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output\n CubDebugExit(cudaMemset(d_out, 0, sizeof(OutputT) * num_segments));\n\n // Run once with discard iterator\n DiscardOutputIterator discard_itr;\n CubDebugExit(Dispatch(backend, 1,\n d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, discard_itr, num_items, num_segments, d_segment_offsets,\n reduction_op, 0, true));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(backend, 1,\n d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, num_segments, d_segment_offsets,\n reduction_op, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_segments, g_verbose, g_verbose);\n printf(\"\\t%s\", compare ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n if (g_timing_iterations > 0)\n {\n GpuTimer gpu_timer;\n gpu_timer.Start();\n\n CubDebugExit(Dispatch(backend, g_timing_iterations,\n d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes,\n d_in, d_out, num_items, num_segments, d_segment_offsets,\n reduction_op, 0, false));\n\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = giga_rate * sizeof(InputT);\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s\", avg_millis, giga_rate, giga_bandwidth);\n }\n\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare);\n}\n\n\n/// Test DeviceReduce\ntemplate <\n Backend BACKEND,\n typename OutputValueT,\n typename HostInputIteratorT,\n typename DeviceInputIteratorT,\n typename OffsetT,\n typename OffsetIteratorT,\n typename ReductionOpT>\nvoid SolveAndTest(\n HostInputIteratorT h_in,\n DeviceInputIteratorT d_in,\n OffsetT num_items,\n OffsetT num_segments,\n OffsetIteratorT h_segment_offsets,\n OffsetIteratorT d_segment_offsets,\n ReductionOpT reduction_op)\n{\n typedef typename std::iterator_traits::value_type InputValueT;\n typedef Solution SolutionT;\n typedef typename SolutionT::OutputT OutputT;\n\n printf(\"\\n\\n%s cub::DeviceReduce<%s> %d items (%s), %d segments\\n\",\n (BACKEND == CUB_CDP) ? \"CUB_CDP\" : (BACKEND == THRUST) ? \"Thrust\" : (BACKEND == CUB_SEGMENTED) ? \"CUB_SEGMENTED\" : \"CUB\",\n typeid(ReductionOpT).name(), num_items, typeid(HostInputIteratorT).name(), num_segments);\n fflush(stdout);\n\n // Allocate and solve solution\n OutputT *h_reference = new OutputT[num_segments];\n SolutionT::Solve(h_in, h_reference, num_segments, h_segment_offsets, reduction_op);\n\n // Run test\n Test(Int2Type(), d_in, num_items, num_segments, d_segment_offsets, reduction_op, h_reference);\n\n // Cleanup\n if (h_reference) delete[] h_reference;\n}\n\n\n/// Test specific problem type\ntemplate <\n Backend BACKEND,\n typename InputT,\n typename OutputT,\n typename OffsetT,\n typename ReductionOpT>\nvoid TestProblem(\n OffsetT num_items,\n OffsetT num_segments,\n GenMode gen_mode,\n ReductionOpT reduction_op)\n{\n printf(\"\\n\\nInitializing %d %s->%s (gen mode %d)... \", num_items, typeid(InputT).name(), typeid(OutputT).name(), gen_mode); fflush(stdout);\n fflush(stdout);\n\n // Initialize value data\n InputT* h_in = new InputT[num_items];\n Initialize(gen_mode, h_in, num_items);\n\n // Initialize segment data\n OffsetT *h_segment_offsets = new OffsetT[num_segments + 1];\n InitializeSegments(num_items, num_segments, h_segment_offsets, g_verbose_input);\n\n // Initialize device data\n OffsetT *d_segment_offsets = NULL;\n InputT *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(InputT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_segment_offsets, sizeof(OffsetT) * (num_segments + 1)));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(InputT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_segment_offsets, h_segment_offsets, sizeof(OffsetT) * (num_segments + 1), cudaMemcpyHostToDevice));\n\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, reduction_op);\n\n if (h_segment_offsets) delete[] h_segment_offsets;\n if (d_segment_offsets) CubDebugExit(g_allocator.DeviceFree(d_segment_offsets));\n if (h_in) delete[] h_in;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n}\n\n\n/// Test different operators\ntemplate <\n Backend BACKEND,\n typename OutputT,\n typename HostInputIteratorT,\n typename DeviceInputIteratorT,\n typename OffsetT,\n typename OffsetIteratorT>\nvoid TestByOp(\n HostInputIteratorT h_in,\n DeviceInputIteratorT d_in,\n OffsetT num_items,\n OffsetT num_segments,\n OffsetIteratorT h_segment_offsets,\n OffsetIteratorT d_segment_offsets)\n{\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, CustomMax());\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, Sum());\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, Min());\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, ArgMin());\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, Max());\n SolveAndTest(h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets, ArgMax());\n}\n\n\n/// Test different backends\ntemplate <\n typename InputT,\n typename OutputT,\n typename OffsetT>\nvoid TestByBackend(\n OffsetT num_items,\n OffsetT max_segments,\n GenMode gen_mode)\n{\n // Initialize host data\n printf(\"\\n\\nInitializing %d %s -> %s (gen mode %d)... \",\n num_items, typeid(InputT).name(), typeid(OutputT).name(), gen_mode); fflush(stdout);\n\n InputT *h_in = new InputT[num_items];\n OffsetT *h_segment_offsets = new OffsetT[max_segments + 1];\n Initialize(gen_mode, h_in, num_items);\n\n // Initialize device data\n InputT *d_in = NULL;\n OffsetT *d_segment_offsets = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(InputT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_segment_offsets, sizeof(OffsetT) * (max_segments + 1)));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(InputT) * num_items, cudaMemcpyHostToDevice));\n\n //\n // Test single-segment implementations\n //\n\n InitializeSegments(num_items, 1, h_segment_offsets, g_verbose_input);\n\n // Page-aligned-input tests\n TestByOp(h_in, d_in, num_items, 1, h_segment_offsets, (OffsetT*) NULL); // Host-dispatch\n#ifdef CUB_CDP\n TestByOp(h_in, d_in, num_items, 1, h_segment_offsets, (OffsetT*) NULL); // Device-dispatch\n#endif\n\n // Non-page-aligned-input tests\n if (num_items > 1)\n {\n InitializeSegments(num_items - 1, 1, h_segment_offsets, g_verbose_input);\n TestByOp(h_in + 1, d_in + 1, num_items - 1, 1, h_segment_offsets, (OffsetT*) NULL);\n }\n\n //\n // Test segmented implementation\n //\n\n // Right now we assign a single thread block to each segment, so lets keep it to under 128K items per segment\n int max_items_per_segment = 128000;\n\n for (int num_segments = (num_items + max_items_per_segment - 1) / max_items_per_segment;\n num_segments < max_segments;\n num_segments = (num_segments * 32) + 1)\n {\n // Test with segment pointer\n InitializeSegments(num_items, num_segments, h_segment_offsets, g_verbose_input);\n CubDebugExit(cudaMemcpy(d_segment_offsets, h_segment_offsets, sizeof(OffsetT) * (num_segments + 1), cudaMemcpyHostToDevice));\n TestByOp(\n h_in, d_in, num_items, num_segments, h_segment_offsets, d_segment_offsets);\n\n // Test with segment iterator\n typedef CastOp IdentityOpT;\n IdentityOpT identity_op;\n TransformInputIterator h_segment_offsets_itr(\n h_segment_offsets,\n identity_op);\n TransformInputIterator d_segment_offsets_itr(\n d_segment_offsets,\n identity_op);\n\n TestByOp(\n h_in, d_in, num_items, num_segments, h_segment_offsets_itr, d_segment_offsets_itr);\n }\n\n if (h_in) delete[] h_in;\n if (h_segment_offsets) delete[] h_segment_offsets;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_segment_offsets) CubDebugExit(g_allocator.DeviceFree(d_segment_offsets));\n}\n\n\n/// Test different input-generation modes\ntemplate <\n typename InputT,\n typename OutputT,\n typename OffsetT>\nvoid TestByGenMode(\n OffsetT num_items,\n OffsetT max_segments)\n{\n //\n // Test pointer support using different input-generation modes\n //\n\n TestByBackend(num_items, max_segments, UNIFORM);\n TestByBackend(num_items, max_segments, INTEGER_SEED);\n TestByBackend(num_items, max_segments, RANDOM);\n\n //\n // Test iterator support using a constant-iterator and SUM\n //\n\n InputT val;\n InitValue(UNIFORM, val, 0);\n ConstantInputIterator h_in(val);\n\n OffsetT *h_segment_offsets = new OffsetT[1 + 1];\n InitializeSegments(num_items, 1, h_segment_offsets, g_verbose_input);\n\n SolveAndTest(h_in, h_in, num_items, 1, h_segment_offsets, (OffsetT*) NULL, Sum());\n#ifdef CUB_CDP\n SolveAndTest(h_in, h_in, num_items, 1, h_segment_offsets, (OffsetT*) NULL, Sum());\n#endif\n\n if (h_segment_offsets) delete[] h_segment_offsets;\n}\n\n\n/// Test different problem sizes\ntemplate <\n typename InputT,\n typename OutputT,\n typename OffsetT>\nstruct TestBySize\n{\n OffsetT max_items;\n OffsetT max_segments;\n\n TestBySize(OffsetT max_items, OffsetT max_segments) :\n max_items(max_items),\n max_segments(max_segments)\n {}\n\n template \n cudaError_t Invoke()\n {\n //\n // Black-box testing on all backends\n //\n\n // Test 0, 1, many\n TestByGenMode(0, max_segments);\n TestByGenMode(1, max_segments);\n TestByGenMode(max_items, max_segments);\n\n // Test random problem sizes from a log-distribution [8, max_items-ish)\n int num_iterations = 8;\n double max_exp = log(double(max_items)) / log(double(2.0));\n for (int i = 0; i < num_iterations; ++i)\n {\n OffsetT num_items = (OffsetT) pow(2.0, RandomValue(max_exp - 3.0) + 3.0);\n TestByGenMode(num_items, max_segments);\n }\n\n //\n // White-box testing of single-segment problems around specific sizes\n //\n\n // Tile-boundaries: multiple blocks, one tile per block\n OffsetT tile_size = ActivePolicyT::ReducePolicy::BLOCK_THREADS * ActivePolicyT::ReducePolicy::ITEMS_PER_THREAD;\n TestProblem(tile_size * 4, 1, RANDOM, Sum());\n TestProblem(tile_size * 4 + 1, 1, RANDOM, Sum());\n TestProblem(tile_size * 4 - 1, 1, RANDOM, Sum());\n\n // Tile-boundaries: multiple blocks, multiple tiles per block\n OffsetT sm_occupancy = 32;\n OffsetT occupancy = tile_size * sm_occupancy * g_sm_count;\n TestProblem(occupancy, 1, RANDOM, Sum());\n TestProblem(occupancy + 1, 1, RANDOM, Sum());\n TestProblem(occupancy - 1, 1, RANDOM, Sum());\n\n return cudaSuccess;\n }\n};\n\n\n/// Test problem type\ntemplate <\n typename InputT,\n typename OutputT,\n typename OffsetT>\nvoid TestType(\n OffsetT max_items,\n OffsetT max_segments)\n{\n typedef typename DeviceReducePolicy::MaxPolicy MaxPolicyT;\n\n TestBySize dispatch(max_items, max_segments);\n\n MaxPolicyT::Invoke(g_ptx_version, dispatch);\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n typedef int OffsetT;\n\n OffsetT max_items = 27000000;\n OffsetT max_segments = 34000;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_verbose_input = args.CheckCmdLineFlag(\"v2\");\n args.GetCmdLineArgument(\"n\", max_items);\n args.GetCmdLineArgument(\"s\", max_segments);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--s= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n CubDebugExit(PtxVersion(g_ptx_version));\n\n // Get SM count\n g_sm_count = args.deviceProp.multiProcessorCount;\n\n std::numeric_limits::max();\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic test\n\n\n\n TestProblem( max_items, 1, RANDOM, Sum());\n\n TestProblem( max_items, 1, RANDOM, Sum());\n\n TestProblem( max_items, 1, RANDOM, ArgMax());\n\n TestProblem( max_items, 1, RANDOM, Sum());\n\n TestProblem(max_items, max_segments, RANDOM, Sum());\n\n\n#elif defined(QUICK_TEST)\n\n // Compile/run quick comparison tests\n\n TestProblem( max_items * 4, 1, UNIFORM, Sum());\n TestProblem( max_items * 4, 1, UNIFORM, Sum());\n\n printf(\"\\n----------------------------\\n\");\n TestProblem( max_items * 2, 1, UNIFORM, Sum());\n TestProblem( max_items * 2, 1, UNIFORM, Sum());\n\n printf(\"\\n----------------------------\\n\");\n TestProblem( max_items, 1, UNIFORM, Sum());\n TestProblem( max_items, 1, UNIFORM, Sum());\n\n printf(\"\\n----------------------------\\n\");\n TestProblem( max_items / 2, 1, UNIFORM, Sum());\n TestProblem( max_items / 2, 1, UNIFORM, Sum());\n\n printf(\"\\n----------------------------\\n\");\n TestProblem( max_items / 4, 1, UNIFORM, Max());\n TestProblem( max_items / 4, 1, UNIFORM, Max());\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test different input types\n TestType(max_items, max_segments);\n\n TestType(max_items, max_segments);\n\n TestType(max_items, max_segments);\n\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n\n TestType(max_items, max_segments);\n TestType(max_items, max_segments);\n\n }\n\n#endif\n\n\n printf(\"\\n\");\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_device_reduce_by_key.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceReduce::ReduceByKey utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to reduce-by-key entrypoint\n */\ntemplate <\n typename KeyInputIteratorT,\n typename KeyOutputIteratorT,\n typename ValueInputIteratorT,\n typename ValueOutputIteratorT,\n typename NumRunsIteratorT,\n typename EqualityOpT,\n typename ReductionOpT,\n typename OffsetT>\nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n KeyInputIteratorT d_keys_in,\n KeyOutputIteratorT d_keys_out,\n ValueInputIteratorT d_values_in,\n ValueOutputIteratorT d_values_out,\n NumRunsIteratorT d_num_runs,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceReduce::ReduceByKey(\n d_temp_storage,\n temp_storage_bytes,\n d_keys_in,\n d_keys_out,\n d_values_in,\n d_values_out,\n d_num_runs,\n reduction_op,\n num_items,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to reduce-by-key entrypoint\n */\ntemplate <\n typename KeyInputIteratorT,\n typename KeyOutputIteratorT,\n typename ValueInputIteratorT,\n typename ValueOutputIteratorT,\n typename NumRunsIteratorT,\n typename EqualityOpT,\n typename ReductionOpT,\n typename OffsetT>\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n KeyInputIteratorT d_keys_in,\n KeyOutputIteratorT d_keys_out,\n ValueInputIteratorT d_values_in,\n ValueOutputIteratorT d_values_out,\n NumRunsIteratorT d_num_runs,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input keys type\n typedef typename std::iterator_traits::value_type KeyInputT;\n\n // The output keys type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type KeyOutputT; // ... else the output iterator's value type\n\n // The input values type\n typedef typename std::iterator_traits::value_type ValueInputT;\n\n // The output values type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type ValueOuputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_keys_in_wrapper(d_keys_in);\n thrust::device_ptr d_keys_out_wrapper(d_keys_out);\n\n thrust::device_ptr d_values_in_wrapper(d_values_in);\n thrust::device_ptr d_values_out_wrapper(d_values_out);\n\n thrust::pair, thrust::device_ptr > d_out_ends;\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_ends = thrust::reduce_by_key(\n d_keys_in_wrapper,\n d_keys_in_wrapper + num_items,\n d_values_in_wrapper,\n d_keys_out_wrapper,\n d_values_out_wrapper);\n }\n\n OffsetT num_segments = OffsetT(d_out_ends.first - d_keys_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_runs, &num_segments, sizeof(OffsetT), cudaMemcpyHostToDevice));\n\n }\n\n return cudaSuccess;\n}\n\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceSelect\n */\ntemplate <\n typename KeyInputIteratorT,\n typename KeyOutputIteratorT,\n typename ValueInputIteratorT,\n typename ValueOutputIteratorT,\n typename NumRunsIteratorT,\n typename EqualityOpT,\n typename ReductionOpT,\n typename OffsetT>\n__global__ void CnpDispatchKernel(\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t temp_storage_bytes,\n KeyInputIteratorT d_keys_in,\n KeyOutputIteratorT d_keys_out,\n ValueInputIteratorT d_values_in,\n ValueOutputIteratorT d_values_out,\n NumRunsIteratorT d_num_runs,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(Int2Type(), timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_runs, equality_op, reduction_op, num_items, 0, debug_synchronous);\n\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate <\n typename KeyInputIteratorT,\n typename KeyOutputIteratorT,\n typename ValueInputIteratorT,\n typename ValueOutputIteratorT,\n typename NumRunsIteratorT,\n typename EqualityOpT,\n typename ReductionOpT,\n typename OffsetT>\nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n KeyInputIteratorT d_keys_in,\n KeyOutputIteratorT d_keys_out,\n ValueInputIteratorT d_values_in,\n ValueOutputIteratorT d_values_out,\n NumRunsIteratorT d_num_runs,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_runs, equality_op, reduction_op, num_items, 0, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n int entropy_reduction,\n T *h_in,\n int num_items,\n int max_segment)\n{\n unsigned int max_int = (unsigned int) -1;\n\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Select number of repeating occurrences\n\n int repeat;\n\n if (max_segment < 0)\n {\n repeat = num_items;\n }\n else if (max_segment < 2)\n {\n repeat = 1;\n }\n else\n {\n RandomBits(repeat, entropy_reduction);\n repeat = (int) ((double(repeat) * double(max_segment)) / double(max_int));\n repeat = CUB_MAX(1, repeat);\n }\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n InitValue(INTEGER_SEED, h_in[j], key);\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve problem. Returns total number of segments identified\n */\ntemplate <\n typename KeyInputIteratorT,\n typename ValueInputIteratorT,\n typename KeyT,\n typename ValueT,\n typename EqualityOpT,\n typename ReductionOpT>\nint Solve(\n KeyInputIteratorT h_keys_in,\n KeyT *h_keys_reference,\n ValueInputIteratorT h_values_in,\n ValueT *h_values_reference,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n int num_items)\n{\n // First item\n KeyT previous = h_keys_in[0];\n ValueT aggregate = h_values_in[0];\n int num_segments = 0;\n\n // Subsequent items\n for (int i = 1; i < num_items; ++i)\n {\n if (!equality_op(previous, h_keys_in[i]))\n {\n h_keys_reference[num_segments] = previous;\n h_values_reference[num_segments] = aggregate;\n num_segments++;\n aggregate = h_values_in[i];\n }\n else\n {\n aggregate = reduction_op(aggregate, h_values_in[i]);\n }\n previous = h_keys_in[i];\n }\n\n h_keys_reference[num_segments] = previous;\n h_values_reference[num_segments] = aggregate;\n num_segments++;\n\n return num_segments;\n}\n\n\n\n/**\n * Test DeviceSelect for a given problem input\n */\ntemplate <\n Backend BACKEND,\n typename DeviceKeyInputIteratorT,\n typename DeviceValueInputIteratorT,\n typename KeyT,\n typename ValueT,\n typename EqualityOpT,\n typename ReductionOpT>\nvoid Test(\n DeviceKeyInputIteratorT d_keys_in,\n DeviceValueInputIteratorT d_values_in,\n KeyT* h_keys_reference,\n ValueT* h_values_reference,\n EqualityOpT equality_op,\n ReductionOpT reduction_op,\n int num_segments,\n int num_items)\n{\n // Allocate device output arrays and number of segments\n KeyT* d_keys_out = NULL;\n ValueT* d_values_out = NULL;\n int* d_num_runs = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys_out, sizeof(KeyT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values_out, sizeof(ValueT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_runs, sizeof(int)));\n\n // Allocate CDP device arrays\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_runs, equality_op, reduction_op, num_items, 0, true));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output arrays\n CubDebugExit(cudaMemset(d_keys_out, 0, sizeof(KeyT) * num_items));\n CubDebugExit(cudaMemset(d_values_out, 0, sizeof(ValueT) * num_items));\n CubDebugExit(cudaMemset(d_num_runs, 0, sizeof(int)));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_runs, equality_op, reduction_op, num_items, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare1 = CompareDeviceResults(h_keys_reference, d_keys_out, num_segments, true, g_verbose);\n printf(\"\\t Keys %s \", compare1 ? \"FAIL\" : \"PASS\");\n\n int compare2 = CompareDeviceResults(h_values_reference, d_values_out, num_segments, true, g_verbose);\n printf(\"\\t Values %s \", compare2 ? \"FAIL\" : \"PASS\");\n\n int compare3 = CompareDeviceResults(&num_segments, d_num_runs, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare3 ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n CubDebugExit(Dispatch(Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_keys_in, d_keys_out, d_values_in, d_values_out, d_num_runs, equality_op, reduction_op, num_items, 0, false));\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n int bytes_moved = ((num_items + num_segments) * sizeof(KeyT)) + ((num_items + num_segments) * sizeof(ValueT));\n float giga_bandwidth = float(bytes_moved) / avg_millis / 1000.0f / 1000.0f;\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s\", avg_millis, giga_rate, giga_bandwidth);\n }\n printf(\"\\n\\n\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Cleanup\n if (d_keys_out) CubDebugExit(g_allocator.DeviceFree(d_keys_out));\n if (d_values_out) CubDebugExit(g_allocator.DeviceFree(d_values_out));\n if (d_num_runs) CubDebugExit(g_allocator.DeviceFree(d_num_runs));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare1 | compare2 | compare3);\n}\n\n\n/**\n * Test DeviceSelect on pointer type\n */\ntemplate <\n Backend BACKEND,\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid TestPointer(\n int num_items,\n int entropy_reduction,\n int max_segment,\n ReductionOpT reduction_op)\n{\n // Allocate host arrays\n KeyT* h_keys_in = new KeyT[num_items];\n KeyT* h_keys_reference = new KeyT[num_items];\n\n ValueT* h_values_in = new ValueT[num_items];\n ValueT* h_values_reference = new ValueT[num_items];\n\n for (int i = 0; i < num_items; ++i)\n InitValue(INTEGER_SEED, h_values_in[i], 1);\n\n // Initialize problem and solution\n Equality equality_op;\n Initialize(entropy_reduction, h_keys_in, num_items, max_segment);\n int num_segments = Solve(h_keys_in, h_keys_reference, h_values_in, h_values_reference, equality_op, reduction_op, num_items);\n\n printf(\"\\nPointer %s cub::DeviceReduce::ReduceByKey %s reduction of %d items, %d segments (avg run length %.3f), {%s,%s} key value pairs, max_segment %d, entropy_reduction %d\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n (Equals::VALUE) ? \"Sum\" : \"Max\",\n num_items, num_segments, float(num_items) / num_segments,\n typeid(KeyT).name(), typeid(ValueT).name(),\n max_segment, entropy_reduction);\n fflush(stdout);\n\n // Allocate problem device arrays\n KeyT *d_keys_in = NULL;\n ValueT *d_values_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys_in, sizeof(KeyT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_values_in, sizeof(ValueT) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_keys_in, h_keys_in, sizeof(KeyT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_values_in, h_values_in, sizeof(ValueT) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_keys_in, d_values_in, h_keys_reference, h_values_reference, equality_op, reduction_op, num_segments, num_items);\n\n // Cleanup\n if (h_keys_in) delete[] h_keys_in;\n if (h_values_in) delete[] h_values_in;\n if (h_keys_reference) delete[] h_keys_reference;\n if (h_values_reference) delete[] h_values_reference;\n if (d_keys_in) CubDebugExit(g_allocator.DeviceFree(d_keys_in));\n if (d_values_in) CubDebugExit(g_allocator.DeviceFree(d_values_in));\n}\n\n\n/**\n * Test on iterator type\n */\ntemplate <\n Backend BACKEND,\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid TestIterator(\n int num_items,\n int entropy_reduction,\n int max_segment,\n ReductionOpT reduction_op)\n{\n // Allocate host arrays\n KeyT* h_keys_in = new KeyT[num_items];\n KeyT* h_keys_reference = new KeyT[num_items];\n\n ValueT one_val;\n InitValue(INTEGER_SEED, one_val, 1);\n ConstantInputIterator h_values_in(one_val);\n ValueT* h_values_reference = new ValueT[num_items];\n\n // Initialize problem and solution\n Equality equality_op;\n Initialize(entropy_reduction, h_keys_in, num_items, max_segment);\n int num_segments = Solve(h_keys_in, h_keys_reference, h_values_in, h_values_reference, equality_op, reduction_op, num_items);\n\n printf(\"\\nIterator %s cub::DeviceReduce::ReduceByKey %s reduction of %d items, %d segments (avg run length %.3f), {%s,%s} key value pairs, max_segment %d, entropy_reduction %d\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n (Equals::VALUE) ? \"Sum\" : \"Max\",\n num_items, num_segments, float(num_items) / num_segments,\n typeid(KeyT).name(), typeid(ValueT).name(),\n max_segment, entropy_reduction);\n fflush(stdout);\n\n // Allocate problem device arrays\n KeyT *d_keys_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys_in, sizeof(KeyT) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_keys_in, h_keys_in, sizeof(KeyT) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_keys_in, h_values_in, h_keys_reference, h_values_reference, equality_op, reduction_op, num_segments, num_items);\n\n // Cleanup\n if (h_keys_in) delete[] h_keys_in;\n if (h_keys_reference) delete[] h_keys_reference;\n if (h_values_reference) delete[] h_values_reference;\n if (d_keys_in) CubDebugExit(g_allocator.DeviceFree(d_keys_in));\n}\n\n\n/**\n * Test different gen modes\n */\ntemplate <\n Backend BACKEND,\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid Test(\n int num_items,\n ReductionOpT reduction_op,\n int max_segment)\n{\n // 0 key-bit entropy reduction rounds\n TestPointer(num_items, 0, max_segment, reduction_op);\n\n if (max_segment > 1)\n {\n // 2 key-bit entropy reduction rounds\n TestPointer(num_items, 2, max_segment, reduction_op);\n\n // 7 key-bit entropy reduction rounds\n TestPointer(num_items, 7, max_segment, reduction_op);\n }\n}\n\n\n/**\n * Test different avg segment lengths modes\n */\ntemplate <\n Backend BACKEND,\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid Test(\n int num_items,\n ReductionOpT reduction_op)\n{\n Test(num_items, reduction_op, -1);\n Test(num_items, reduction_op, 1);\n\n // Evaluate different max-segment lengths\n for (int max_segment = 3; max_segment < CUB_MIN(num_items, (unsigned short) -1); max_segment *= 11)\n {\n Test(num_items, reduction_op, max_segment);\n }\n}\n\n\n\n/**\n * Test different dispatch\n */\ntemplate <\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid TestDispatch(\n int num_items,\n ReductionOpT reduction_op)\n{\n Test(num_items, reduction_op);\n#ifdef CUB_CDP\n Test(num_items, reduction_op);\n#endif\n}\n\n\n/**\n * Test different input sizes\n */\ntemplate <\n typename KeyT,\n typename ValueT,\n typename ReductionOpT>\nvoid TestSize(\n int num_items,\n ReductionOpT reduction_op)\n{\n if (num_items < 0)\n {\n TestDispatch(1, reduction_op);\n TestDispatch(100, reduction_op);\n TestDispatch(10000, reduction_op);\n TestDispatch(1000000, reduction_op);\n }\n else\n {\n TestDispatch(num_items, reduction_op);\n }\n\n}\n\n\ntemplate <\n typename KeyT,\n typename ValueT>\nvoid TestOp(\n int num_items)\n{\n TestSize(num_items, cub::Sum());\n TestSize(num_items, cub::Max());\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = -1;\n int entropy_reduction = 0;\n int maxseg = 1000;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"maxseg\", maxseg);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--entropy=]\"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n printf(\"\\n\");\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 32000000;\n\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n TestIterator(num_items, entropy_reduction, maxseg, cub::Sum());\n\n#elif defined(QUICK_TEST)\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 32000000;\n\n printf(\"---- RLE int ---- \\n\");\n TestIterator(num_items, entropy_reduction, maxseg, cub::Sum());\n\n printf(\"---- RLE long long ---- \\n\");\n TestIterator(num_items, entropy_reduction, maxseg, cub::Sum());\n\n printf(\"---- int ---- \\n\");\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n\n printf(\"---- float ---- \\n\");\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n {\n printf(\"---- double ---- \\n\");\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n TestPointer(num_items, entropy_reduction, maxseg, cub::Sum());\n }\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n\n // Test different input types\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n TestOp(num_items);\n\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n TestOp(num_items);\n\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_device_run_length_encode.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceReduce::RunLengthEncode utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n// Operation types\nenum RleMethod\n{\n RLE, // Run length encode\n NON_TRIVIAL,\n CSR,\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB entrypoints\n//---------------------------------------------------------------------\n\n\n/**\n * Dispatch to run-length encode entrypoint\n */\ntemplate <\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsIterator,\n typename OffsetT>\nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type method,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n UniqueOutputIteratorT d_unique_out,\n OffsetsOutputIteratorT d_offsets_out,\n LengthsOutputIteratorT d_lengths_out,\n NumRunsIterator d_num_runs,\n cub::Equality equality_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceRunLengthEncode::Encode(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_unique_out,\n d_lengths_out,\n d_num_runs,\n num_items,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to non-trivial runs entrypoint\n */\ntemplate <\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsIterator,\n typename OffsetT>\nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type method,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n UniqueOutputIteratorT d_unique_out,\n OffsetsOutputIteratorT d_offsets_out,\n LengthsOutputIteratorT d_lengths_out,\n NumRunsIterator d_num_runs,\n cub::Equality equality_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceRunLengthEncode::NonTrivialRuns(\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_offsets_out,\n d_lengths_out,\n d_num_runs,\n num_items,\n stream,\n debug_synchronous);\n }\n return error;\n}\n\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to run-length encode entrypoint\n */\ntemplate <\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsIterator,\n typename OffsetT>\ncudaError_t Dispatch(\n Int2Type method,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n UniqueOutputIteratorT d_unique_out,\n OffsetsOutputIteratorT d_offsets_out,\n LengthsOutputIteratorT d_lengths_out,\n NumRunsIterator d_num_runs,\n cub::Equality equality_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type UniqueT; // ... else the output iterator's value type\n\n // The lengths output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // LengthT = (if output iterator's value type is void) ?\n OffsetT, // ... then the OffsetT type,\n typename std::iterator_traits::value_type>::Type LengthT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_unique_out_wrapper(d_unique_out);\n thrust::device_ptr d_lengths_out_wrapper(d_lengths_out);\n\n thrust::pair, thrust::device_ptr > d_out_ends;\n\n LengthT one_val;\n InitValue(INTEGER_SEED, one_val, 1);\n thrust::constant_iterator constant_one(one_val);\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_ends = thrust::reduce_by_key(\n d_in_wrapper,\n d_in_wrapper + num_items,\n constant_one,\n d_unique_out_wrapper,\n d_lengths_out_wrapper);\n }\n\n OffsetT num_runs = OffsetT(d_out_ends.first - d_unique_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_runs, &num_runs, sizeof(OffsetT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceRunLengthEncode\n */\ntemplate <\n int RLE_METHOD,\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsIterator,\n typename EqualityOp,\n typename OffsetT>\n__global__ void CnpDispatchKernel(\n Int2Type method,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n InputIteratorT d_in,\n UniqueOutputIteratorT d_unique_out,\n OffsetsOutputIteratorT d_offsets_out,\n LengthsOutputIteratorT d_lengths_out,\n NumRunsIterator d_num_runs,\n cub::Equality equality_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(method, Int2Type(), timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_offsets_out, d_lengths_out, d_num_runs, equality_op, num_items, 0, debug_synchronous);\n\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate <\n int RLE_METHOD,\n typename InputIteratorT,\n typename UniqueOutputIteratorT,\n typename OffsetsOutputIteratorT,\n typename LengthsOutputIteratorT,\n typename NumRunsIterator,\n typename EqualityOp,\n typename OffsetT>\nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type method,\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n UniqueOutputIteratorT d_unique_out,\n OffsetsOutputIteratorT d_offsets_out,\n LengthsOutputIteratorT d_lengths_out,\n NumRunsIterator d_num_runs,\n EqualityOp equality_op,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(method, timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_offsets_out, d_lengths_out, d_num_runs, equality_op, num_items, 0, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n int entropy_reduction,\n T *h_in,\n int num_items,\n int max_segment)\n{\n unsigned int max_int = (unsigned int) -1;\n\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Select number of repeating occurrences for the current run\n int repeat;\n if (max_segment < 0)\n {\n repeat = num_items;\n }\n else if (max_segment < 2)\n {\n repeat = 1;\n }\n else\n {\n RandomBits(repeat, entropy_reduction);\n repeat = (int) ((double(repeat) * double(max_segment)) / double(max_int));\n repeat = CUB_MAX(1, repeat);\n }\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n InitValue(INTEGER_SEED, h_in[j], key);\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve problem. Returns total number of segments identified\n */\ntemplate <\n RleMethod RLE_METHOD,\n typename InputIteratorT,\n typename T,\n typename OffsetT,\n typename LengthT,\n typename EqualityOp>\nint Solve(\n InputIteratorT h_in,\n T *h_unique_reference,\n OffsetT *h_offsets_reference,\n LengthT *h_lengths_reference,\n EqualityOp equality_op,\n int num_items)\n{\n if (num_items == 0) \n return 0;\n\n // First item\n T previous = h_in[0];\n LengthT length = 1;\n int num_runs = 0;\n int run_begin = 0;\n\n // Subsequent items\n for (int i = 1; i < num_items; ++i)\n {\n if (!equality_op(previous, h_in[i]))\n {\n if ((RLE_METHOD != NON_TRIVIAL) || (length > 1))\n {\n h_unique_reference[num_runs] = previous;\n h_offsets_reference[num_runs] = run_begin;\n h_lengths_reference[num_runs] = length;\n num_runs++;\n }\n length = 1;\n run_begin = i;\n }\n else\n {\n length++;\n }\n previous = h_in[i];\n }\n\n if ((RLE_METHOD != NON_TRIVIAL) || (length > 1))\n {\n h_unique_reference[num_runs] = previous;\n h_offsets_reference[num_runs] = run_begin;\n h_lengths_reference[num_runs] = length;\n num_runs++;\n }\n\n return num_runs;\n}\n\n\n\n/**\n * Test DeviceRunLengthEncode for a given problem input\n */\ntemplate <\n RleMethod RLE_METHOD,\n Backend BACKEND,\n typename DeviceInputIteratorT,\n typename T,\n typename OffsetT,\n typename LengthT,\n typename EqualityOp>\nvoid Test(\n DeviceInputIteratorT d_in,\n T *h_unique_reference,\n OffsetT *h_offsets_reference,\n LengthT *h_lengths_reference,\n EqualityOp equality_op,\n int num_runs,\n int num_items)\n{\n // Allocate device output arrays and number of segments\n T* d_unique_out = NULL;\n LengthT* d_offsets_out = NULL;\n OffsetT* d_lengths_out = NULL;\n int* d_num_runs = NULL;\n\n if (RLE_METHOD == RLE)\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_unique_out, sizeof(T) * num_items));\n if (RLE_METHOD == NON_TRIVIAL)\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_offsets_out, sizeof(OffsetT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_lengths_out, sizeof(LengthT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_runs, sizeof(int)));\n\n // Allocate CDP device arrays\n size_t* d_temp_storage_bytes = NULL;\n cudaError_t* d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void* d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_offsets_out, d_lengths_out, d_num_runs, equality_op, num_items, 0, true));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output arrays\n if (RLE_METHOD == RLE)\n CubDebugExit(cudaMemset(d_unique_out, 0, sizeof(T) * num_items));\n if (RLE_METHOD == NON_TRIVIAL)\n CubDebugExit(cudaMemset(d_offsets_out, 0, sizeof(OffsetT) * num_items));\n CubDebugExit(cudaMemset(d_lengths_out, 0, sizeof(LengthT) * num_items));\n CubDebugExit(cudaMemset(d_num_runs, 0, sizeof(int)));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_offsets_out, d_lengths_out, d_num_runs, equality_op, num_items, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare0 = 0;\n int compare1 = 0;\n int compare2 = 0;\n int compare3 = 0;\n\n if (RLE_METHOD == RLE)\n {\n compare0 = CompareDeviceResults(h_unique_reference, d_unique_out, num_runs, true, g_verbose);\n printf(\"\\t Keys %s\\n\", compare0 ? \"FAIL\" : \"PASS\");\n }\n\n if (RLE_METHOD != RLE)\n {\n compare1 = CompareDeviceResults(h_offsets_reference, d_offsets_out, num_runs, true, g_verbose);\n printf(\"\\t Offsets %s\\n\", compare1 ? \"FAIL\" : \"PASS\");\n }\n\n if (RLE_METHOD != CSR)\n {\n compare2 = CompareDeviceResults(h_lengths_reference, d_lengths_out, num_runs, true, g_verbose);\n printf(\"\\t Lengths %s\\n\", compare2 ? \"FAIL\" : \"PASS\");\n }\n\n compare3 = CompareDeviceResults(&num_runs, d_num_runs, 1, true, g_verbose);\n printf(\"\\t Count %s\\n\", compare3 ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_unique_out, d_offsets_out, d_lengths_out, d_num_runs, equality_op, num_items, 0, false));\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n int bytes_moved = (num_items * sizeof(T)) + (num_runs * (sizeof(OffsetT) + sizeof(LengthT)));\n float giga_bandwidth = float(bytes_moved) / avg_millis / 1000.0f / 1000.0f;\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s\", avg_millis, giga_rate, giga_bandwidth);\n }\n printf(\"\\n\\n\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Cleanup\n if (d_unique_out) CubDebugExit(g_allocator.DeviceFree(d_unique_out));\n if (d_offsets_out) CubDebugExit(g_allocator.DeviceFree(d_offsets_out));\n if (d_lengths_out) CubDebugExit(g_allocator.DeviceFree(d_lengths_out));\n if (d_num_runs) CubDebugExit(g_allocator.DeviceFree(d_num_runs));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare0 | compare1 | compare2 | compare3);\n}\n\n\n/**\n * Test DeviceRunLengthEncode on pointer type\n */\ntemplate <\n RleMethod RLE_METHOD,\n Backend BACKEND,\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid TestPointer(\n int num_items,\n int entropy_reduction,\n int max_segment)\n{\n // Allocate host arrays\n T* h_in = new T[num_items];\n T* h_unique_reference = new T[num_items];\n OffsetT* h_offsets_reference = new OffsetT[num_items];\n LengthT* h_lengths_reference = new LengthT[num_items];\n\n for (int i = 0; i < num_items; ++i)\n InitValue(INTEGER_SEED, h_offsets_reference[i], 1);\n\n // Initialize problem and solution\n Equality equality_op;\n Initialize(entropy_reduction, h_in, num_items, max_segment);\n\n int num_runs = Solve(h_in, h_unique_reference, h_offsets_reference, h_lengths_reference, equality_op, num_items);\n\n printf(\"\\nPointer %s cub::%s on %d items, %d segments (avg run length %.3f), {%s key, %s offset, %s length}, max_segment %d, entropy_reduction %d\\n\",\n (RLE_METHOD == RLE) ? \"DeviceReduce::RunLengthEncode\" : (RLE_METHOD == NON_TRIVIAL) ? \"DeviceRunLengthEncode::NonTrivialRuns\" : \"Other\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_runs, float(num_items) / num_runs,\n typeid(T).name(), typeid(OffsetT).name(), typeid(LengthT).name(),\n max_segment, entropy_reduction);\n fflush(stdout);\n\n // Allocate problem device arrays\n T* d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_in, h_unique_reference, h_offsets_reference, h_lengths_reference, equality_op, num_runs, num_items);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_unique_reference) delete[] h_unique_reference;\n if (h_offsets_reference) delete[] h_offsets_reference;\n if (h_lengths_reference) delete[] h_lengths_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n}\n\n\n/**\n * Test on iterator type\n */\ntemplate <\n RleMethod RLE_METHOD,\n Backend BACKEND,\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid TestIterator(\n int num_items,\n Int2Type is_primitive)\n{\n // Allocate host arrays\n T* h_unique_reference = new T[num_items];\n OffsetT* h_offsets_reference = new OffsetT[num_items];\n LengthT* h_lengths_reference = new LengthT[num_items];\n\n T one_val;\n InitValue(INTEGER_SEED, one_val, 1);\n ConstantInputIterator h_in(one_val);\n\n // Initialize problem and solution\n Equality equality_op;\n int num_runs = Solve(h_in, h_unique_reference, h_offsets_reference, h_lengths_reference, equality_op, num_items);\n\n printf(\"\\nIterator %s cub::%s on %d items, %d segments (avg run length %.3f), {%s key, %s offset, %s length}\\n\",\n (RLE_METHOD == RLE) ? \"DeviceReduce::RunLengthEncode\" : (RLE_METHOD == NON_TRIVIAL) ? \"DeviceRunLengthEncode::NonTrivialRuns\" : \"Other\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_runs, float(num_items) / num_runs,\n typeid(T).name(), typeid(OffsetT).name(), typeid(LengthT).name());\n fflush(stdout);\n\n // Run Test\n Test(h_in, h_unique_reference, h_offsets_reference, h_lengths_reference, equality_op, num_runs, num_items);\n\n // Cleanup\n if (h_unique_reference) delete[] h_unique_reference;\n if (h_offsets_reference) delete[] h_offsets_reference;\n if (h_lengths_reference) delete[] h_lengths_reference;\n}\n\n\ntemplate <\n RleMethod RLE_METHOD,\n Backend BACKEND,\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid TestIterator(\n int num_items,\n Int2Type is_primitive)\n{}\n\n\n/**\n * Test different gen modes\n */\ntemplate <\n RleMethod RLE_METHOD,\n Backend BACKEND,\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid Test(\n int num_items)\n{\n // Test iterator (one run)\n TestIterator(num_items, Int2Type::PRIMITIVE>());\n\n // num_items runs\n TestPointer(num_items, 0, 1);\n\n // Evaluate different run lengths\n for (int max_segment = 3; max_segment < CUB_MIN(num_items, (unsigned short) -1); max_segment *= 3)\n {\n // Uniform selection run length\n TestPointer(num_items, 0, max_segment);\n\n // Reduced-entropy run length\n TestPointer(num_items, 4, max_segment);\n }\n}\n\n\n/**\n * Test different dispatch\n */\ntemplate <\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid TestDispatch(\n int num_items)\n{\n Test(num_items);\n Test(num_items);\n\n#ifdef CUB_CDP\n Test(num_items);\n Test(num_items);\n#endif\n}\n\n\n/**\n * Test different input sizes\n */\ntemplate <\n typename T,\n typename OffsetT,\n typename LengthT>\nvoid TestSize(\n int num_items)\n{\n if (num_items < 0)\n {\n TestDispatch(0);\n TestDispatch(1);\n TestDispatch(100);\n TestDispatch(10000);\n TestDispatch(1000000);\n\n // Randomly select problem size between 1:10,000,000\n unsigned int max_int = (unsigned int) -1;\n for (int i = 0; i < 10; ++i)\n {\n unsigned int num_items;\n RandomBits(num_items);\n num_items = (unsigned int) ((double(num_items) * double(10000000)) / double(max_int));\n num_items = CUB_MAX(1, num_items);\n TestDispatch(num_items);\n }\n }\n else\n {\n TestDispatch(num_items);\n }\n\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = -1;\n int entropy_reduction = 0;\n int max_segment = 1000;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"maxseg\", max_segment);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--entropy=]\"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n printf(\"\\n\");\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 32000000;\n\n TestPointer( num_items, entropy_reduction, max_segment);\n TestPointer( num_items, entropy_reduction, max_segment);\n TestIterator( num_items, Int2Type::PRIMITIVE>());\n\n\n#elif defined(QUICK_TEST)\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 32000000;\n\n TestPointer( num_items, entropy_reduction, max_segment);\n TestPointer( num_items, entropy_reduction, max_segment);\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test different input types\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n TestSize(num_items);\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_device_scan.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceScan utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n\n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\ndouble g_device_giga_bandwidth;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\n/**\n * \\brief WrapperFunctor (for precluding test-specialized dispatch to *Sum variants)\n */\ntemplate\nstruct WrapperFunctor\n{\n OpT op;\n\n WrapperFunctor(OpT op) : op(op) {}\n\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return op(a, b);\n }\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB DeviceScan entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to exclusive scan entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceScan::ExclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, scan_op, initial_value, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to exclusive sum entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n Sum scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceScan::ExclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to inclusive scan entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n NullType initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceScan::InclusiveScan(d_temp_storage, temp_storage_bytes, d_in, d_out, scan_op, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to inclusive sum entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n Sum scan_op,\n NullType initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to exclusive scan entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n thrust::exclusive_scan(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper, initial_value, scan_op);\n }\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to exclusive sum entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n Sum scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n thrust::exclusive_scan(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper);\n }\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to inclusive scan entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n NullType initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n thrust::inclusive_scan(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper, scan_op);\n }\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to inclusive sum entrypoint\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n Sum scan_op,\n NullType initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n thrust::inclusive_scan(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper);\n }\n }\n\n return cudaSuccess;\n}\n\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceScan\n */\ntemplate \n__global__ void CnpDispatchKernel(\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n bool debug_synchronous)\n{\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(\n Int2Type(),\n is_primitive,\n timing_timing_iterations,\n d_temp_storage_bytes,\n d_cdp_error,\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n initial_value,\n num_items,\n 0,\n debug_synchronous);\n\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsPrimitiveT is_primitive,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n ScanOpT scan_op,\n InitialValueT initial_value,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(\n is_primitive,\n timing_timing_iterations,\n d_temp_storage_bytes,\n d_cdp_error,\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n initial_value,\n num_items,\n debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n GenMode gen_mode,\n T *h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n/**\n * Solve exclusive-scan problem\n */\ntemplate <\n typename InputIteratorT,\n typename OutputT,\n typename ScanOpT>\nvoid Solve(\n InputIteratorT h_in,\n OutputT *h_reference,\n int num_items,\n ScanOpT scan_op,\n OutputT initial_value)\n{\n if (num_items > 0)\n {\n OutputT val = h_in[0];\n h_reference[0] = initial_value;\n OutputT inclusive = scan_op(initial_value, val);\n\n for (int i = 1; i < num_items; ++i)\n {\n val = h_in[i];\n h_reference[i] = inclusive;\n inclusive = scan_op(inclusive, val);\n }\n }\n}\n\n\n/**\n * Solve inclusive-scan problem\n */\ntemplate <\n typename InputIteratorT,\n typename OutputT,\n typename ScanOpT>\nvoid Solve(\n InputIteratorT h_in,\n OutputT *h_reference,\n int num_items,\n ScanOpT scan_op,\n NullType)\n{\n if (num_items > 0)\n {\n OutputT inclusive = h_in[0];\n h_reference[0] = inclusive;\n\n for (int i = 1; i < num_items; ++i)\n {\n OutputT val = h_in[i];\n inclusive = scan_op(inclusive, val);\n h_reference[i] = inclusive;\n }\n }\n}\n\n\n/**\n * Test DeviceScan for a given problem input\n */\ntemplate <\n Backend BACKEND,\n typename DeviceInputIteratorT,\n typename OutputT,\n typename ScanOpT,\n typename InitialValueT>\nvoid Test(\n DeviceInputIteratorT d_in,\n OutputT *h_reference,\n int num_items,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n typedef typename std::iterator_traits::value_type InputT;\n\n // Allocate device output array\n OutputT *d_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(OutputT) * num_items));\n\n // Allocate CDP device arrays\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(\n Int2Type(),\n Int2Type::PRIMITIVE>(),\n 1,\n d_temp_storage_bytes,\n d_cdp_error,\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n initial_value,\n num_items,\n 0,\n true));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output array\n CubDebugExit(cudaMemset(d_out, 0, sizeof(OutputT) * num_items));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(\n Int2Type(),\n Int2Type::PRIMITIVE>(),\n 1,\n d_temp_storage_bytes,\n d_cdp_error,\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n initial_value,\n num_items,\n 0,\n true));\n\n // Check for correctness (and display results, if specified)\n int compare = CompareDeviceResults(h_reference, d_out, num_items, true, g_verbose);\n printf(\"\\t%s\", compare ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n CubDebugExit(Dispatch(Int2Type(),\n Int2Type::PRIMITIVE>(),\n g_timing_iterations,\n d_temp_storage_bytes,\n d_cdp_error,\n d_temp_storage,\n temp_storage_bytes,\n d_in,\n d_out,\n scan_op,\n initial_value,\n num_items,\n 0,\n false));\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = giga_rate * (sizeof(InputT) + sizeof(OutputT));\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s, %.1f%% peak\", avg_millis, giga_rate, giga_bandwidth, giga_bandwidth / g_device_giga_bandwidth * 100.0);\n }\n\n printf(\"\\n\\n\");\n\n // Cleanup\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare);\n}\n\n\n/**\n * Test DeviceScan on pointer type\n */\ntemplate <\n Backend BACKEND,\n typename InputT,\n typename OutputT,\n typename ScanOpT,\n typename InitialValueT>\nvoid TestPointer(\n int num_items,\n GenMode gen_mode,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n printf(\"\\nPointer %s %s cub::DeviceScan::%s %d items, %s->%s (%d->%d bytes) , gen-mode %s\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n (Equals::VALUE) ? \"Inclusive\" : \"Exclusive\",\n (Equals::VALUE) ? \"Sum\" : \"Scan\",\n num_items,\n typeid(InputT).name(), typeid(OutputT).name(), (int) sizeof(InputT), (int) sizeof(OutputT),\n (gen_mode == RANDOM) ? \"RANDOM\" : (gen_mode == INTEGER_SEED) ? \"SEQUENTIAL\" : \"HOMOGENOUS\");\n fflush(stdout);\n\n // Allocate host arrays\n InputT* h_in = new InputT[num_items];\n OutputT* h_reference = new OutputT[num_items];\n\n // Initialize problem and solution\n Initialize(gen_mode, h_in, num_items);\n Solve(h_in, h_reference, num_items, scan_op, initial_value);\n\n // Allocate problem device arrays\n InputT *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(InputT) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(InputT) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_in, h_reference, num_items, scan_op, initial_value);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n}\n\n\n/**\n * Test DeviceScan on iterator type\n */\ntemplate <\n Backend BACKEND,\n typename InputT,\n typename OutputT,\n typename ScanOpT,\n typename InitialValueT>\nvoid TestIterator(\n int num_items,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n printf(\"\\nIterator %s %s cub::DeviceScan::%s %d items, %s->%s (%d->%d bytes)\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n (Equals::VALUE) ? \"Inclusive\" : \"Exclusive\",\n (Equals::VALUE) ? \"Sum\" : \"Scan\",\n num_items,\n typeid(InputT).name(), typeid(OutputT).name(), (int) sizeof(InputT), (int) sizeof(OutputT));\n fflush(stdout);\n\n // Use a constant iterator as the input\n InputT val = InputT();\n ConstantInputIterator h_in(val);\n\n // Allocate host arrays\n OutputT* h_reference = new OutputT[num_items];\n\n // Initialize problem and solution\n Solve(h_in, h_reference, num_items, scan_op, initial_value);\n\n // Run Test\n Test(h_in, h_reference, num_items, scan_op, initial_value);\n\n // Cleanup\n if (h_reference) delete[] h_reference;\n}\n\n\n/**\n * Test different gen modes\n */\ntemplate <\n Backend BACKEND,\n typename InputT,\n typename OutputT,\n typename ScanOpT,\n typename InitialValueT>\nvoid Test(\n int num_items,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n TestPointer( num_items, UNIFORM, scan_op, initial_value);\n TestPointer( num_items, RANDOM, scan_op, initial_value);\n TestIterator( num_items, scan_op, initial_value);\n}\n\n\n/**\n * Test different dispatch\n */\ntemplate <\n typename InputT,\n typename OutputT,\n typename ScanOpT,\n typename InitialValueT>\nvoid Test(\n int num_items,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n Test(num_items, scan_op, initial_value);\n#ifdef CUB_CDP\n Test(num_items, scan_op, initial_value);\n#endif\n}\n\n\n/**\n * Test different operators\n */\ntemplate \nvoid TestOp(\n int num_items,\n OutputT identity,\n OutputT initial_value)\n{\n // Exclusive (use identity as initial value because it will dispatch to *Sum variants that don't take initial values)\n Test(num_items, cub::Sum(), identity);\n Test(num_items, cub::Max(), identity);\n\n // Exclusive (non-specialized, so we can test initial-value)\n Test(num_items, WrapperFunctor(cub::Sum()), initial_value);\n Test(num_items, WrapperFunctor(cub::Max()), initial_value);\n\n // Inclusive (no initial value)\n Test(num_items, cub::Sum(), NullType());\n Test(num_items, cub::Max(), NullType());\n}\n\n\n/**\n * Test different input sizes\n */\ntemplate <\n typename InputT,\n typename OutputT>\nvoid TestSize(\n int num_items,\n OutputT identity,\n OutputT initial_value)\n{\n if (num_items < 0)\n {\n TestOp(0, identity, initial_value);\n TestOp(1, identity, initial_value);\n TestOp(100, identity, initial_value);\n TestOp(10000, identity, initial_value);\n TestOp(1000000, identity, initial_value);\n\n // Randomly select problem size between 1:10,000,000\n unsigned int max_int = (unsigned int) -1;\n for (int i = 0; i < 10; ++i)\n {\n unsigned int num_items;\n RandomBits(num_items);\n num_items = (unsigned int) ((double(num_items) * double(10000000)) / double(max_int));\n num_items = CUB_MAX(1, num_items);\n TestOp(num_items, identity, initial_value);\n }\n }\n else\n {\n TestOp(num_items, identity, initial_value);\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = -1;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n g_device_giga_bandwidth = args.device_giga_bandwidth;\n printf(\"\\n\");\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 32000000;\n\n TestPointer( num_items , UNIFORM, Sum(), (int) (0));\n TestPointer( num_items , UNIFORM, Sum(), (int) (0));\n\n#elif defined(QUICK_TEST)\n\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get device SM version\n int sm_version;\n CubDebugExit(SmVersion(sm_version, device_ordinal));\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 32000000;\n\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 4), UNIFORM, Sum(), char(0));\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 4), UNIFORM, Sum(), char(0));\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 2), UNIFORM, Sum(), short(0));\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 2), UNIFORM, Sum(), short(0));\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items , UNIFORM, Sum(), (int) (0));\n TestPointer( num_items , UNIFORM, Sum(), (int) (0));\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items / 2, UNIFORM, Sum(), (long long) (0));\n TestPointer(num_items / 2, UNIFORM, Sum(), (long long) (0));\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items / 4, UNIFORM, Sum(), TestBar());\n TestPointer( num_items / 4, UNIFORM, Sum(), TestBar());\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test different input+output data types\n TestSize(num_items, (int) 0, (int) 99);\n\n // Test same intput+output data types\n TestSize(num_items, (unsigned char) 0, (unsigned char) 99);\n TestSize(num_items, (char) 0, (char) 99);\n TestSize(num_items, (unsigned short) 0, (unsigned short)99);\n TestSize(num_items, (unsigned int) 0, (unsigned int) 99);\n TestSize(num_items, (unsigned long long) 0, (unsigned long long) 99);\n\n TestSize(num_items, make_uchar2(0, 0), make_uchar2(17, 21));\n TestSize(num_items, make_char2(0, 0), make_char2(17, 21));\n TestSize(num_items, make_ushort2(0, 0), make_ushort2(17, 21));\n TestSize(num_items, make_uint2(0, 0), make_uint2(17, 21));\n TestSize(num_items, make_ulonglong2(0, 0), make_ulonglong2(17, 21));\n TestSize(num_items, make_uchar4(0, 0, 0, 0), make_uchar4(17, 21, 32, 85));\n TestSize(num_items, make_char4(0, 0, 0, 0), make_char4(17, 21, 32, 85));\n\n TestSize(num_items, make_ushort4(0, 0, 0, 0), make_ushort4(17, 21, 32, 85));\n TestSize(num_items, make_uint4(0, 0, 0, 0), make_uint4(17, 21, 32, 85));\n TestSize(num_items, make_ulonglong4(0, 0, 0, 0), make_ulonglong4(17, 21, 32, 85));\n\n TestSize(num_items,\n TestFoo::MakeTestFoo(0, 0, 0, 0),\n TestFoo::MakeTestFoo(1ll << 63, 1 << 31, short(1 << 15), char(1 << 7)));\n\n TestSize(num_items,\n TestBar(0, 0),\n TestBar(1ll << 63, 1 << 31));\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_device_select_if.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceSelect::If and DevicePartition::If utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nfloat g_device_giga_bandwidth;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\n// Selection functor type\ntemplate \nstruct LessThan\n{\n T compare;\n\n __host__ __device__ __forceinline__\n LessThan(T compare) : compare(compare) {}\n\n __host__ __device__ __forceinline__\n bool operator()(const T &a) const {\n return (a < compare);\n }\n};\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB DeviceSelect entrypoints\n//---------------------------------------------------------------------\n\n\n/**\n * Dispatch to select if entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSelect::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to partition if entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DevicePartition::If(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, select_op, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to select flagged entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSelect::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n/**\n * Dispatch to partition flagged entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DevicePartition::Flagged(d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n/**\n * Dispatch to select if entrypoint\n */\ntemplate \n__host__ __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_out_wrapper_end;\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_wrapper_end = thrust::copy_if(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper, select_op);\n }\n\n OffsetT num_selected = OffsetT(d_out_wrapper_end - d_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_selected_out, &num_selected, sizeof(OffsetT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to partition if entrypoint\n */\ntemplate \n__host__ __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n typedef thrust::reverse_iterator > ReverseOutputIteratorT;\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::pair, ReverseOutputIteratorT> d_out_wrapper_end;\n\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n\n ReverseOutputIteratorT d_out_unselected(d_out_wrapper + num_items);\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_wrapper_end = thrust::partition_copy(\n d_in_wrapper,\n d_in_wrapper + num_items,\n d_out_wrapper,\n d_out_unselected,\n select_op);\n }\n\n OffsetT num_selected = OffsetT(d_out_wrapper_end.first - d_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_selected_out, &num_selected, sizeof(OffsetT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to select flagged entrypoint\n */\ntemplate \n__host__ __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The flag type\n typedef typename std::iterator_traits::value_type FlagT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_out_wrapper_end;\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n thrust::device_ptr d_flags_wrapper(d_flags);\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_wrapper_end = thrust::copy_if(d_in_wrapper, d_in_wrapper + num_items, d_flags_wrapper, d_out_wrapper, CastOp());\n }\n\n OffsetT num_selected = OffsetT(d_out_wrapper_end - d_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_selected_out, &num_selected, sizeof(OffsetT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n/**\n * Dispatch to partition flagged entrypoint\n */\ntemplate \n__host__ __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n Int2Type is_flagged,\n Int2Type is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The flag type\n typedef typename std::iterator_traits::value_type FlagT;\n\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n typedef thrust::reverse_iterator > ReverseOutputIteratorT;\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::pair, ReverseOutputIteratorT> d_out_wrapper_end;\n\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n thrust::device_ptr d_flags_wrapper(d_flags);\n ReverseOutputIteratorT d_out_unselected(d_out_wrapper + num_items);\n\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_wrapper_end = thrust::partition_copy(\n d_in_wrapper,\n d_in_wrapper + num_items,\n d_flags_wrapper,\n d_out_wrapper,\n d_out_unselected,\n CastOp());\n }\n\n OffsetT num_selected = OffsetT(d_out_wrapper_end.first - d_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_selected_out, &num_selected, sizeof(OffsetT), cudaMemcpyHostToDevice));\n }\n\n return cudaSuccess;\n}\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceSelect\n */\ntemplate \n__global__ void CnpDispatchKernel(\n IsFlaggedTag is_flagged,\n IsPartitionTag is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n bool debug_synchronous)\n{\n\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(Int2Type(), is_flagged, is_partition, timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, select_op, 0, debug_synchronous);\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n IsFlaggedTag is_flagged,\n IsPartitionTag is_partition,\n int timing_timing_iterations,\n size_t* d_temp_storage_bytes,\n cudaError_t* d_cdp_error,\n\n void* d_temp_storage,\n size_t& temp_storage_bytes,\n InputIteratorT d_in,\n FlagIteratorT d_flags,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n SelectOpT select_op,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(is_flagged, is_partition, timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, select_op, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n T* h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n // Initialize each item to a randomly selected value from [0..126]\n unsigned int value;\n RandomBits(value, 0, 0, 7);\n if (value == 127)\n value = 126;\n InitValue(INTEGER_SEED, h_in[i], value);\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve selection problem (and set corresponding flags)\n */\ntemplate <\n typename InputIteratorT,\n typename FlagIteratorT,\n typename SelectOpT,\n typename T>\nint Solve(\n InputIteratorT h_in,\n SelectOpT select_op,\n T* h_reference,\n FlagIteratorT h_flags,\n int num_items)\n{\n int num_selected = 0;\n for (int i = 0; i < num_items; ++i)\n {\n if ((h_flags[i] = select_op(h_in[i])))\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n else\n {\n h_reference[num_items - (i - num_selected) - 1] = h_in[i];\n }\n }\n\n return num_selected;\n}\n\n\n\n/**\n * Test DeviceSelect for a given problem input\n */\ntemplate <\n Backend BACKEND,\n bool IS_FLAGGED,\n bool IS_PARTITION,\n typename DeviceInputIteratorT,\n typename FlagT,\n typename SelectOpT,\n typename T>\nvoid Test(\n DeviceInputIteratorT d_in,\n FlagT* h_flags,\n SelectOpT select_op,\n T* h_reference,\n int num_selected,\n int num_items)\n{\n // Allocate device flags, output, and num-selected\n FlagT* d_flags = NULL;\n T* d_out = NULL;\n int* d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_flags, sizeof(FlagT) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate CDP device arrays\n size_t* d_temp_storage_bytes = NULL;\n cudaError_t* d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, select_op, 0, true));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Copy flags and clear device output array\n CubDebugExit(cudaMemcpy(d_flags, h_flags, sizeof(FlagT) * num_items, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * num_items));\n CubDebugExit(cudaMemset(d_num_selected_out, 0, sizeof(int)));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), Int2Type(), 1, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, select_op, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare1 = (IS_PARTITION) ?\n CompareDeviceResults(h_reference, d_out, num_items, true, g_verbose) :\n CompareDeviceResults(h_reference, d_out, num_selected, true, g_verbose);\n printf(\"\\t Data %s\\n\", compare1 ? \"FAIL\" : \"PASS\");\n\n int compare2 = CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s\\n\", compare2 ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n CubDebugExit(Dispatch(Int2Type(), Int2Type(), Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_flags, d_out, d_num_selected_out, num_items, select_op, 0, false));\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n int num_output_items = (IS_PARTITION) ? num_items : num_selected;\n int num_flag_items = (IS_FLAGGED) ? num_items : 0;\n size_t num_bytes = sizeof(T) * (num_items + num_output_items) + sizeof(FlagT) * num_flag_items;\n float giga_bandwidth = float(num_bytes) / avg_millis / 1000.0f / 1000.0f;\n\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s, %.1f%% peak\", avg_millis, giga_rate, giga_bandwidth, giga_bandwidth / g_device_giga_bandwidth * 100.0);\n }\n printf(\"\\n\\n\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Cleanup\n if (d_flags) CubDebugExit(g_allocator.DeviceFree(d_flags));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare1 | compare2);\n}\n\n\n/**\n * Test on pointer type\n */\ntemplate <\n Backend BACKEND,\n bool IS_FLAGGED,\n bool IS_PARTITION,\n typename T>\nvoid TestPointer(\n int num_items,\n float select_ratio)\n{\n typedef char FlagT;\n\n // Allocate host arrays\n T* h_in = new T[num_items];\n FlagT* h_flags = new FlagT[num_items];\n T* h_reference = new T[num_items];\n\n // Initialize input\n Initialize(h_in, num_items);\n\n // Select a comparison value that is select_ratio through the space of [0,127]\n T compare;\n if (select_ratio <= 0.0)\n InitValue(INTEGER_SEED, compare, 0); // select none\n else if (select_ratio >= 1.0)\n InitValue(INTEGER_SEED, compare, 127); // select all\n else\n InitValue(INTEGER_SEED, compare, int(double(double(127) * select_ratio)));\n\n LessThan select_op(compare);\n int num_selected = Solve(h_in, select_op, h_reference, h_flags, num_items);\n\n if (g_verbose) std::cout << \"\\nComparison item: \" << compare << \"\\n\";\n printf(\"\\nPointer %s cub::%s::%s %d items, %d selected (select ratio %.3f), %s %d-byte elements\\n\",\n (IS_PARTITION) ? \"DevicePartition\" : \"DeviceSelect\",\n (IS_FLAGGED) ? \"Flagged\" : \"If\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_selected, float(num_selected) / num_items, typeid(T).name(), (int) sizeof(T));\n fflush(stdout);\n\n // Allocate problem device arrays\n T *d_in = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_in, h_flags, select_op, h_reference, num_selected, num_items);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (h_flags) delete[] h_flags;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n}\n\n\n/**\n * Test on iterator type\n */\ntemplate <\n Backend BACKEND,\n bool IS_FLAGGED,\n bool IS_PARTITION,\n typename T>\nvoid TestIterator(\n int num_items,\n float select_ratio)\n{\n typedef char FlagT;\n\n // Allocate host arrays\n T* h_reference = new T[num_items];\n FlagT* h_flags = new FlagT[num_items];\n\n // Use counting iterator as the input\n CountingInputIterator h_in(0);\n\n // Select a comparison value that is select_ratio through the space of [0,127]\n T compare;\n if (select_ratio <= 0.0)\n InitValue(INTEGER_SEED, compare, 0); // select none\n else if (select_ratio >= 1.0)\n InitValue(INTEGER_SEED, compare, 127); // select all\n else\n InitValue(INTEGER_SEED, compare, int(double(double(127) * select_ratio)));\n\n LessThan select_op(compare);\n int num_selected = Solve(h_in, select_op, h_reference, h_flags, num_items);\n\n if (g_verbose) std::cout << \"\\nComparison item: \" << compare << \"\\n\";\n printf(\"\\nIterator %s cub::%s::%s %d items, %d selected (select ratio %.3f), %s %d-byte elements\\n\",\n (IS_PARTITION) ? \"DevicePartition\" : \"DeviceSelect\",\n (IS_FLAGGED) ? \"Flagged\" : \"If\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_selected, float(num_selected) / num_items, typeid(T).name(), (int) sizeof(T));\n fflush(stdout);\n\n // Run Test\n Test(h_in, h_flags, select_op, h_reference, num_selected, num_items);\n\n // Cleanup\n if (h_reference) delete[] h_reference;\n if (h_flags) delete[] h_flags;\n}\n\n\n/**\n * Test different selection ratios\n */\ntemplate <\n Backend BACKEND,\n bool IS_FLAGGED,\n bool IS_PARTITION,\n typename T>\nvoid Test(\n int num_items)\n{\n for (float select_ratio = 0.0f; select_ratio <= 1.0f; select_ratio += 0.2f)\n {\n TestPointer(num_items, select_ratio);\n }\n}\n\n\n/**\n * Test (select vs. partition) and (flagged vs. functor)\n */\ntemplate <\n Backend BACKEND,\n typename T>\nvoid TestMethod(\n int num_items)\n{\n // Functor\n Test(num_items);\n Test(num_items);\n\n // Flagged\n Test(num_items);\n Test(num_items);\n}\n\n\n/**\n * Test different dispatch\n */\ntemplate <\n typename T>\nvoid TestOp(\n int num_items)\n{\n TestMethod(num_items);\n#ifdef CUB_CDP\n TestMethod(num_items);\n#endif\n}\n\n\n/**\n * Test different input sizes\n */\ntemplate \nvoid Test(\n int num_items)\n{\n if (num_items < 0)\n {\n TestOp(0);\n TestOp(1);\n TestOp(100);\n TestOp(10000);\n TestOp(1000000);\n }\n else\n {\n TestOp(num_items);\n }\n}\n\n/**\n * Test select/partition on pointer types\n */\ntemplate \nvoid ComparePointer(\n int num_items,\n float select_ratio)\n{\n printf(\"-- Select-if ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Partition-if ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Select-flagged ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Partition-flagged ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n TestPointer(num_items, select_ratio);\n\n}\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = -1;\n float select_ratio = 0.5;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"ratio\", select_ratio);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--ratio=] \"\n \"[--repeat=] \"\n \"[--v] \"\n \"[--cdp] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n g_device_giga_bandwidth = args.device_giga_bandwidth;\n printf(\"\\n\");\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 32000000;\n\n printf(\"-- Select-if ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Partition-if ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Select-flagged ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n\n printf(\"-- Partition-flagged ----------------------------\\n\");\n TestPointer(num_items, select_ratio);\n\n\n#elif defined(QUICK_TEST)\n\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get device SM version\n int sm_version;\n CubDebugExit(SmVersion(sm_version, device_ordinal));\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 32000000;\n\n printf(\"-- Iterator ----------------------------\\n\");\n TestIterator(num_items, select_ratio);\n\n ComparePointer( num_items * ((sm_version <= 130) ? 1 : 4), select_ratio);\n ComparePointer( num_items * ((sm_version <= 130) ? 1 : 2), select_ratio);\n ComparePointer( num_items, select_ratio);\n ComparePointer( num_items / 2, select_ratio);\n ComparePointer( num_items / 4, select_ratio);\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test different input types\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_device_select_unique.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of DeviceSelect::Unique utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n\n#include \n#include \n#include \n\n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_timing_iterations = 0;\nint g_repeat = 0;\nfloat g_device_giga_bandwidth;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\n//---------------------------------------------------------------------\n// Dispatch to different CUB DeviceSelect entrypoints\n//---------------------------------------------------------------------\n\n\n/**\n * Dispatch to unique entrypoint\n */\ntemplate \nCUB_RUNTIME_FUNCTION __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n cudaError_t error = cudaSuccess;\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n error = DeviceSelect::Unique(d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, stream, debug_synchronous);\n }\n return error;\n}\n\n\n//---------------------------------------------------------------------\n// Dispatch to different Thrust entrypoints\n//---------------------------------------------------------------------\n\n\n/**\n * Dispatch to unique entrypoint\n */\ntemplate \n__host__ __forceinline__\ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void *d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // The input value type\n typedef typename std::iterator_traits::value_type InputT;\n\n // The output value type\n typedef typename If<(Equals::value_type, void>::VALUE), // OutputT = (if output iterator's value type is void) ?\n typename std::iterator_traits::value_type, // ... then the input iterator's value type,\n typename std::iterator_traits::value_type>::Type OutputT; // ... else the output iterator's value type\n\n if (d_temp_storage == 0)\n {\n temp_storage_bytes = 1;\n }\n else\n {\n thrust::device_ptr d_out_wrapper_end;\n thrust::device_ptr d_in_wrapper(d_in);\n thrust::device_ptr d_out_wrapper(d_out);\n for (int i = 0; i < timing_timing_iterations; ++i)\n {\n d_out_wrapper_end = thrust::unique_copy(d_in_wrapper, d_in_wrapper + num_items, d_out_wrapper);\n }\n\n OffsetT num_selected = OffsetT(d_out_wrapper_end - d_out_wrapper);\n CubDebugExit(cudaMemcpy(d_num_selected_out, &num_selected, sizeof(OffsetT), cudaMemcpyHostToDevice));\n\n }\n\n return cudaSuccess;\n}\n\n\n\n//---------------------------------------------------------------------\n// CUDA Nested Parallelism Test Kernel\n//---------------------------------------------------------------------\n\n/**\n * Simple wrapper kernel to invoke DeviceSelect\n */\ntemplate \n__global__ void CnpDispatchKernel(\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n bool debug_synchronous)\n{\n\n#ifndef CUB_CDP\n *d_cdp_error = cudaErrorNotSupported;\n#else\n *d_cdp_error = Dispatch(Int2Type(), timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, 0, debug_synchronous);\n *d_temp_storage_bytes = temp_storage_bytes;\n#endif\n}\n\n\n/**\n * Dispatch to CDP kernel\n */\ntemplate \ncudaError_t Dispatch(\n Int2Type dispatch_to,\n int timing_timing_iterations,\n size_t *d_temp_storage_bytes,\n cudaError_t *d_cdp_error,\n\n void* d_temp_storage,\n size_t &temp_storage_bytes,\n InputIteratorT d_in,\n OutputIteratorT d_out,\n NumSelectedIteratorT d_num_selected_out,\n OffsetT num_items,\n cudaStream_t stream,\n bool debug_synchronous)\n{\n // Invoke kernel to invoke device-side dispatch\n CnpDispatchKernel<<<1,1>>>(timing_timing_iterations, d_temp_storage_bytes, d_cdp_error,\n d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, debug_synchronous);\n\n // Copy out temp_storage_bytes\n CubDebugExit(cudaMemcpy(&temp_storage_bytes, d_temp_storage_bytes, sizeof(size_t) * 1, cudaMemcpyDeviceToHost));\n\n // Copy out error\n cudaError_t retval;\n CubDebugExit(cudaMemcpy(&retval, d_cdp_error, sizeof(cudaError_t) * 1, cudaMemcpyDeviceToHost));\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Test generation\n//---------------------------------------------------------------------\n\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n int entropy_reduction,\n T *h_in,\n int num_items,\n int max_segment)\n{\n unsigned int max_int = (unsigned int) -1;\n\n int key = 0;\n int i = 0;\n while (i < num_items)\n {\n // Select number of repeating occurrences for the current run\n int repeat;\n if (max_segment < 0)\n {\n repeat = num_items;\n }\n else if (max_segment < 2)\n {\n repeat = 1;\n }\n else\n {\n RandomBits(repeat, entropy_reduction);\n repeat = (int) ((double(repeat) * double(max_segment)) / double(max_int));\n repeat = CUB_MAX(1, repeat);\n }\n\n int j = i;\n while (j < CUB_MIN(i + repeat, num_items))\n {\n InitValue(INTEGER_SEED, h_in[j], key);\n j++;\n }\n\n i = j;\n key++;\n }\n\n if (g_verbose)\n {\n printf(\"Input:\\n\");\n DisplayResults(h_in, num_items);\n printf(\"\\n\\n\");\n }\n}\n\n\n/**\n * Solve unique problem\n */\ntemplate <\n typename InputIteratorT,\n typename T>\nint Solve(\n InputIteratorT h_in,\n T *h_reference,\n int num_items)\n{\n int num_selected = 0;\n if (num_items > 0)\n {\n h_reference[num_selected] = h_in[0];\n num_selected++;\n }\n\n for (int i = 1; i < num_items; ++i)\n {\n if (h_in[i] != h_in[i - 1])\n {\n h_reference[num_selected] = h_in[i];\n num_selected++;\n }\n }\n\n return num_selected;\n}\n\n\n\n/**\n * Test DeviceSelect for a given problem input\n */\ntemplate <\n Backend BACKEND,\n typename DeviceInputIteratorT,\n typename T>\nvoid Test(\n DeviceInputIteratorT d_in,\n T *h_reference,\n int num_selected,\n int num_items)\n{\n // Allocate device output array and num selected\n T *d_out = NULL;\n int *d_num_selected_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * num_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_num_selected_out, sizeof(int)));\n\n // Allocate CDP device arrays\n size_t *d_temp_storage_bytes = NULL;\n cudaError_t *d_cdp_error = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_temp_storage_bytes, sizeof(size_t) * 1));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_cdp_error, sizeof(cudaError_t) * 1));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(Dispatch(Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, 0, true));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Clear device output array\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * num_items));\n CubDebugExit(cudaMemset(d_num_selected_out, 0, sizeof(int)));\n\n // Run warmup/correctness iteration\n CubDebugExit(Dispatch(Int2Type(), 1, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, 0, true));\n\n // Check for correctness (and display results, if specified)\n int compare1 = CompareDeviceResults(h_reference, d_out, num_selected, true, g_verbose);\n printf(\"\\t Data %s \", compare1 ? \"FAIL\" : \"PASS\");\n\n int compare2 = CompareDeviceResults(&num_selected, d_num_selected_out, 1, true, g_verbose);\n printf(\"\\t Count %s \", compare2 ? \"FAIL\" : \"PASS\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Performance\n GpuTimer gpu_timer;\n gpu_timer.Start();\n CubDebugExit(Dispatch(Int2Type(), g_timing_iterations, d_temp_storage_bytes, d_cdp_error, d_temp_storage, temp_storage_bytes, d_in, d_out, d_num_selected_out, num_items, 0, false));\n gpu_timer.Stop();\n float elapsed_millis = gpu_timer.ElapsedMillis();\n\n // Display performance\n if (g_timing_iterations > 0)\n {\n float avg_millis = elapsed_millis / g_timing_iterations;\n float giga_rate = float(num_items) / avg_millis / 1000.0f / 1000.0f;\n float giga_bandwidth = float((num_items + num_selected) * sizeof(T)) / avg_millis / 1000.0f / 1000.0f;\n printf(\", %.3f avg ms, %.3f billion items/s, %.3f logical GB/s, %.1f%% peak\", avg_millis, giga_rate, giga_bandwidth, giga_bandwidth / g_device_giga_bandwidth * 100.0);\n }\n printf(\"\\n\\n\");\n\n // Flush any stdout/stderr\n fflush(stdout);\n fflush(stderr);\n\n // Cleanup\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_num_selected_out) CubDebugExit(g_allocator.DeviceFree(d_num_selected_out));\n if (d_temp_storage_bytes) CubDebugExit(g_allocator.DeviceFree(d_temp_storage_bytes));\n if (d_cdp_error) CubDebugExit(g_allocator.DeviceFree(d_cdp_error));\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n // Correctness asserts\n AssertEquals(0, compare1 | compare2);\n}\n\n\n/**\n * Test DeviceSelect on pointer type\n */\ntemplate <\n Backend BACKEND,\n typename T>\nvoid TestPointer(\n int num_items,\n int entropy_reduction,\n int max_segment)\n{\n // Allocate host arrays\n T* h_in = new T[num_items];\n T* h_reference = new T[num_items];\n\n // Initialize problem and solution\n Initialize(entropy_reduction, h_in, num_items, max_segment);\n int num_selected = Solve(h_in, h_reference, num_items);\n\n printf(\"\\nPointer %s cub::DeviceSelect::Unique %d items, %d selected (avg run length %.3f), %s %d-byte elements, entropy_reduction %d\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_selected, float(num_items) / num_selected,\n typeid(T).name(),\n (int) sizeof(T),\n entropy_reduction);\n fflush(stdout);\n\n // Allocate problem device arrays\n T *d_in = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * num_items));\n\n // Initialize device input\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * num_items, cudaMemcpyHostToDevice));\n\n // Run Test\n Test(d_in, h_reference, num_selected, num_items);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n}\n\n\n/**\n * Test DeviceSelect on iterator type\n */\ntemplate <\n Backend BACKEND,\n typename T>\nvoid TestIterator(\n int num_items)\n{\n // Use a counting iterator as the input\n CountingInputIterator h_in(0);\n\n // Allocate host arrays\n T* h_reference = new T[num_items];\n\n // Initialize problem and solution\n int num_selected = Solve(h_in, h_reference, num_items);\n\n printf(\"\\nIterator %s cub::DeviceSelect::Unique %d items, %d selected (avg run length %.3f), %s %d-byte elements\\n\",\n (BACKEND == CDP) ? \"CDP CUB\" : (BACKEND == THRUST) ? \"Thrust\" : \"CUB\",\n num_items, num_selected, float(num_items) / num_selected,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n // Run Test\n Test(h_in, h_reference, num_selected, num_items);\n\n // Cleanup\n if (h_reference) delete[] h_reference;\n}\n\n\n/**\n * Test different gen modes\n */\ntemplate <\n Backend BACKEND,\n typename T>\nvoid Test(\n int num_items)\n{\n for (int max_segment = 1; ((max_segment > 0) && (max_segment < num_items)); max_segment *= 11)\n {\n TestPointer(num_items, 0, max_segment);\n TestPointer(num_items, 2, max_segment);\n TestPointer(num_items, 7, max_segment);\n }\n}\n\n\n/**\n * Test different dispatch\n */\ntemplate <\n typename T>\nvoid TestOp(\n int num_items)\n{\n Test(num_items);\n#ifdef CUB_CDP\n Test(num_items);\n#endif\n}\n\n\n/**\n * Test different input sizes\n */\ntemplate \nvoid Test(\n int num_items)\n{\n if (num_items < 0)\n {\n TestOp(0);\n TestOp(1);\n TestOp(100);\n TestOp(10000);\n TestOp(1000000);\n }\n else\n {\n TestOp(num_items);\n }\n}\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n int num_items = -1;\n int entropy_reduction = 0;\n int maxseg = 1000;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"n\", num_items);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n args.GetCmdLineArgument(\"maxseg\", maxseg);\n args.GetCmdLineArgument(\"entropy\", entropy_reduction);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--n= \"\n \"[--i= \"\n \"[--device=] \"\n \"[--maxseg=]\"\n \"[--entropy=]\"\n \"[--repeat=]\"\n \"[--v] \"\n \"[--cdp]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n g_device_giga_bandwidth = args.device_giga_bandwidth;\n printf(\"\\n\");\n\n#ifdef QUICKER_TEST\n\n // Compile/run basic CUB test\n if (num_items < 0) num_items = 32000000;\n TestPointer( num_items, entropy_reduction, maxseg);\n\n#elif defined(QUICK_TEST)\n\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get device SM version\n int sm_version;\n CubDebugExit(SmVersion(sm_version, device_ordinal));\n\n // Compile/run quick tests\n if (num_items < 0) num_items = 32000000;\n\n printf(\"-- Iterator ----------------------------\\n\");\n TestIterator( num_items);\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 4), entropy_reduction, maxseg);\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 4), entropy_reduction, maxseg);\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 2), entropy_reduction, maxseg);\n TestPointer( num_items * ((sm_version <= 130) ? 1 : 2), entropy_reduction, maxseg);\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items, entropy_reduction, maxseg);\n TestPointer( num_items, entropy_reduction, maxseg);\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items / 2, entropy_reduction, maxseg);\n TestPointer(num_items / 2, entropy_reduction, maxseg);\n\n printf(\"----------------------------\\n\");\n TestPointer( num_items / 4, entropy_reduction, maxseg);\n TestPointer( num_items / 4, entropy_reduction, maxseg);\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test different input types\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n Test(num_items);\n Test(num_items);\n\n Test(num_items);\n Test(num_items);\n }\n\n#endif\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_grid_barrier.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test evaluation for software global barrier throughput\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n\n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/**\n * Kernel that iterates through the specified number of software global barriers\n */\n__global__ void Kernel(\n GridBarrier global_barrier,\n int iterations)\n{\n for (int i = 0; i < iterations; i++)\n {\n global_barrier.Sync();\n }\n}\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n cudaError_t retval = cudaSuccess;\n\n // Defaults\n int iterations = 10000;\n int block_size = 128;\n int grid_size = -1;\n\n // Initialize command line\n CommandLineArgs args(argc, argv);\n\n // Get args\n args.GetCmdLineArgument(\"i\", iterations);\n args.GetCmdLineArgument(\"grid-size\", grid_size);\n args.GetCmdLineArgument(\"block-size\", block_size);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=]\"\n \"[--i=]\"\n \"[--grid-size]\"\n \"[--block-size]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get device SM version\n int sm_version;\n CubDebugExit(SmVersion(sm_version, device_ordinal));\n\n // Get SM properties\n int sm_count, max_block_threads, max_sm_occupancy;\n CubDebugExit(cudaDeviceGetAttribute(&sm_count, cudaDevAttrMultiProcessorCount, device_ordinal));\n CubDebugExit(cudaDeviceGetAttribute(&max_block_threads, cudaDevAttrMaxThreadsPerBlock, device_ordinal));\n CubDebugExit(MaxSmOccupancy(max_sm_occupancy, EmptyKernel, 32));\n\n // Compute grid size and occupancy\n int occupancy = CUB_MIN((max_block_threads / block_size), max_sm_occupancy);\n\n if (grid_size == -1)\n {\n grid_size = occupancy * sm_count;\n }\n else\n {\n occupancy = grid_size / sm_count;\n }\n\n printf(\"Initializing software global barrier for Kernel<<<%d,%d>>> with %d occupancy\\n\",\n grid_size, block_size, occupancy);\n fflush(stdout);\n\n // Init global barrier\n GridBarrierLifetime global_barrier;\n global_barrier.Setup(grid_size);\n\n // Time kernel\n GpuTimer gpu_timer;\n gpu_timer.Start();\n Kernel<<>>(global_barrier, iterations);\n gpu_timer.Stop();\n\n retval = CubDebug(cudaThreadSynchronize());\n\n // Output timing results\n float avg_elapsed = gpu_timer.ElapsedMillis() / float(iterations);\n printf(\"%d iterations, %f total elapsed millis, %f avg elapsed millis\\n\",\n iterations,\n gpu_timer.ElapsedMillis(),\n avg_elapsed);\n\n return retval;\n}\n" }, { "path": "external/cub/test/test_iterator.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of iterator utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \n#include \n\n#include \"test_util.h\"\n\n#include \n#include \n\nusing namespace cub;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nCachingDeviceAllocator g_allocator(true);\n\n// Dispatch types\nenum Backend\n{\n CUB, // CUB method\n THRUST, // Thrust method\n CDP, // GPU-based (dynamic parallelism) dispatch to CUB method\n};\n\n\ntemplate \nstruct TransformOp\n{\n // Increment transform\n __host__ __device__ __forceinline__ T operator()(T input) const\n {\n T addend;\n InitValue(INTEGER_SEED, addend, 1);\n return input + addend;\n }\n};\n\nstruct SelectOp\n{\n template \n __host__ __device__ __forceinline__ bool operator()(T input)\n {\n return true;\n }\n};\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/**\n * Test random access input iterator\n */\ntemplate <\n typename InputIteratorT,\n typename T>\n__global__ void Kernel(\n InputIteratorT d_in,\n T *d_out,\n InputIteratorT *d_itrs)\n{\n d_out[0] = *d_in; // Value at offset 0\n d_out[1] = d_in[100]; // Value at offset 100\n d_out[2] = *(d_in + 1000); // Value at offset 1000\n d_out[3] = *(d_in + 10000); // Value at offset 10000\n\n d_in++;\n d_out[4] = d_in[0]; // Value at offset 1\n\n d_in += 20;\n d_out[5] = d_in[0]; // Value at offset 21\n d_itrs[0] = d_in; // Iterator at offset 21\n\n d_in -= 10;\n d_out[6] = d_in[0]; // Value at offset 11;\n\n d_in -= 11;\n d_out[7] = d_in[0]; // Value at offset 0\n d_itrs[1] = d_in; // Iterator at offset 0\n}\n\n\n\n//---------------------------------------------------------------------\n// Host testing subroutines\n//---------------------------------------------------------------------\n\n\n/**\n * Run iterator test on device\n */\ntemplate <\n typename InputIteratorT,\n typename T,\n int TEST_VALUES>\nvoid Test(\n InputIteratorT d_in,\n T (&h_reference)[TEST_VALUES])\n{\n // Allocate device arrays\n T *d_out = NULL;\n InputIteratorT *d_itrs = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * TEST_VALUES));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_itrs, sizeof(InputIteratorT) * 2));\n\n int compare;\n\n // Run unguarded kernel\n Kernel<<<1, 1>>>(d_in, d_out, d_itrs);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Check results\n compare = CompareDeviceResults(h_reference, d_out, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tValues: %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Check iterator at offset 21\n InputIteratorT h_itr = d_in + 21;\n compare = CompareDeviceResults(&h_itr, d_itrs, 1, g_verbose, g_verbose);\n printf(\"\\tIterators: %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Check iterator at offset 0\n compare = CompareDeviceResults(&d_in, d_itrs + 1, 1, g_verbose, g_verbose);\n printf(\"\\tIterators: %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_itrs) CubDebugExit(g_allocator.DeviceFree(d_itrs));\n}\n\n\n/**\n * Test constant iterator\n */\ntemplate \nvoid TestConstant(T base)\n{\n printf(\"\\nTesting constant iterator on type %s (base: %lld)\\n\", typeid(T).name(), (unsigned long long) (base)); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n T h_reference[8] = {base, base, base, base, base, base, base, base};\n ConstantInputIterator d_itr(base);\n Test(d_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n int copy_items = 100;\n T *h_copy = new T[copy_items];\n T *d_copy = NULL;\n\n for (int i = 0; i < copy_items; ++i)\n h_copy[i] = d_itr[i];\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * copy_items));\n thrust::device_ptr d_copy_wrapper(d_copy);\n\n thrust::copy_if(d_itr, d_itr + copy_items, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_copy, d_copy, copy_items, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n if (h_copy) delete[] h_copy;\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n}\n\n\n/**\n * Test counting iterator\n */\ntemplate \nvoid TestCounting(T base)\n{\n printf(\"\\nTesting counting iterator on type %s (base: %d) \\n\", typeid(T).name(), int(base)); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = base + 0; // Value at offset 0\n h_reference[1] = base + 100; // Value at offset 100\n h_reference[2] = base + 1000; // Value at offset 1000\n h_reference[3] = base + 10000; // Value at offset 10000\n h_reference[4] = base + 1; // Value at offset 1\n h_reference[5] = base + 21; // Value at offset 21\n h_reference[6] = base + 11; // Value at offset 11\n h_reference[7] = base + 0; // Value at offset 0;\n\n CountingInputIterator d_itr(base);\n Test(d_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n unsigned long long max_items = ((1ull << ((sizeof(T) * 8) - 1)) - 1);\n size_t copy_items = (size_t) CUB_MIN(max_items - base, 100); // potential issue with differencing overflows when T is a smaller type than can handle the offset\n T *h_copy = new T[copy_items];\n T *d_copy = NULL;\n\n for (unsigned long long i = 0; i < copy_items; ++i)\n h_copy[i] = d_itr[i];\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * copy_items));\n thrust::device_ptr d_copy_wrapper(d_copy);\n thrust::copy_if(d_itr, d_itr + copy_items, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_copy, d_copy, copy_items, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n if (h_copy) delete[] h_copy;\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n}\n\n\n/**\n * Test modified iterator\n */\ntemplate \nvoid TestModified()\n{\n printf(\"\\nTesting cache-modified iterator on type %s\\n\", typeid(T).name()); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n const unsigned int TEST_VALUES = 11000;\n\n T *h_data = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n {\n RandomBits(h_data[i]);\n }\n\n // Allocate device arrays\n T *d_data = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_data, sizeof(T) * TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_data, h_data, sizeof(T) * TEST_VALUES, cudaMemcpyHostToDevice));\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = h_data[0]; // Value at offset 0\n h_reference[1] = h_data[100]; // Value at offset 100\n h_reference[2] = h_data[1000]; // Value at offset 1000\n h_reference[3] = h_data[10000]; // Value at offset 10000\n h_reference[4] = h_data[1]; // Value at offset 1\n h_reference[5] = h_data[21]; // Value at offset 21\n h_reference[6] = h_data[11]; // Value at offset 11\n h_reference[7] = h_data[0]; // Value at offset 0;\n\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n Test(CacheModifiedInputIterator((CastT*) d_data), h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n T *d_copy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * TEST_VALUES));\n\n CacheModifiedInputIterator d_in_itr((CastT*) d_data);\n CacheModifiedOutputIterator d_out_itr((CastT*) d_copy);\n\n thrust::copy_if(d_in_itr, d_in_itr + TEST_VALUES, d_out_itr, SelectOp());\n\n int compare = CompareDeviceResults(h_data, d_copy, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n\n if (h_data) delete[] h_data;\n if (d_data) CubDebugExit(g_allocator.DeviceFree(d_data));\n}\n\n\n/**\n * Test transform iterator\n */\ntemplate \nvoid TestTransform()\n{\n printf(\"\\nTesting transform iterator on type %s\\n\", typeid(T).name()); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n const unsigned int TEST_VALUES = 11000;\n\n T *h_data = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n {\n InitValue(INTEGER_SEED, h_data[i], i);\n }\n\n // Allocate device arrays\n T *d_data = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_data, sizeof(T) * TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_data, h_data, sizeof(T) * TEST_VALUES, cudaMemcpyHostToDevice));\n\n TransformOp op;\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = op(h_data[0]); // Value at offset 0\n h_reference[1] = op(h_data[100]); // Value at offset 100\n h_reference[2] = op(h_data[1000]); // Value at offset 1000\n h_reference[3] = op(h_data[10000]); // Value at offset 10000\n h_reference[4] = op(h_data[1]); // Value at offset 1\n h_reference[5] = op(h_data[21]); // Value at offset 21\n h_reference[6] = op(h_data[11]); // Value at offset 11\n h_reference[7] = op(h_data[0]); // Value at offset 0;\n\n TransformInputIterator, CastT*> d_itr((CastT*) d_data, op);\n Test(d_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n T *h_copy = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n h_copy[i] = op(h_data[i]);\n\n T *d_copy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * TEST_VALUES));\n thrust::device_ptr d_copy_wrapper(d_copy);\n\n thrust::copy_if(d_itr, d_itr + TEST_VALUES, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_copy, d_copy, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_copy) delete[] h_copy;\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n\n if (h_data) delete[] h_data;\n if (d_data) CubDebugExit(g_allocator.DeviceFree(d_data));\n}\n\n\n/**\n * Test tex-obj texture iterator\n */\ntemplate \nvoid TestTexObj()\n{\n printf(\"\\nTesting tex-obj iterator on type %s\\n\", typeid(T).name()); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n const unsigned int TEST_VALUES = 11000;\n const unsigned int DUMMY_OFFSET = 500;\n const unsigned int DUMMY_TEST_VALUES = TEST_VALUES - DUMMY_OFFSET;\n\n T *h_data = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n {\n RandomBits(h_data[i]);\n }\n\n // Allocate device arrays\n T *d_data = NULL;\n T *d_dummy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_data, sizeof(T) * TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_data, h_data, sizeof(T) * TEST_VALUES, cudaMemcpyHostToDevice));\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_dummy, sizeof(T) * DUMMY_TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_dummy, h_data + DUMMY_OFFSET, sizeof(T) * DUMMY_TEST_VALUES, cudaMemcpyHostToDevice));\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = h_data[0]; // Value at offset 0\n h_reference[1] = h_data[100]; // Value at offset 100\n h_reference[2] = h_data[1000]; // Value at offset 1000\n h_reference[3] = h_data[10000]; // Value at offset 10000\n h_reference[4] = h_data[1]; // Value at offset 1\n h_reference[5] = h_data[21]; // Value at offset 21\n h_reference[6] = h_data[11]; // Value at offset 11\n h_reference[7] = h_data[0]; // Value at offset 0;\n\n // Create and bind obj-based test iterator\n TexObjInputIterator d_obj_itr;\n CubDebugExit(d_obj_itr.BindTexture((CastT*) d_data, sizeof(T) * TEST_VALUES));\n\n Test(d_obj_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n T *d_copy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * TEST_VALUES));\n thrust::device_ptr d_copy_wrapper(d_copy);\n\n CubDebugExit(cudaMemset(d_copy, 0, sizeof(T) * TEST_VALUES));\n thrust::copy_if(d_obj_itr, d_obj_itr + TEST_VALUES, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_data, d_copy, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n CubDebugExit(d_obj_itr.UnbindTexture());\n\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n\n if (h_data) delete[] h_data;\n if (d_data) CubDebugExit(g_allocator.DeviceFree(d_data));\n if (d_dummy) CubDebugExit(g_allocator.DeviceFree(d_dummy));\n}\n\n\n#if CUDA_VERSION >= 5050\n\n/**\n * Test tex-ref texture iterator\n */\ntemplate \nvoid TestTexRef()\n{\n printf(\"\\nTesting tex-ref iterator on type %s\\n\", typeid(T).name()); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n const unsigned int TEST_VALUES = 11000;\n const unsigned int DUMMY_OFFSET = 500;\n const unsigned int DUMMY_TEST_VALUES = TEST_VALUES - DUMMY_OFFSET;\n\n T *h_data = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n {\n RandomBits(h_data[i]);\n }\n\n // Allocate device arrays\n T *d_data = NULL;\n T *d_dummy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_data, sizeof(T) * TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_data, h_data, sizeof(T) * TEST_VALUES, cudaMemcpyHostToDevice));\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_dummy, sizeof(T) * DUMMY_TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_dummy, h_data + DUMMY_OFFSET, sizeof(T) * DUMMY_TEST_VALUES, cudaMemcpyHostToDevice));\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = h_data[0]; // Value at offset 0\n h_reference[1] = h_data[100]; // Value at offset 100\n h_reference[2] = h_data[1000]; // Value at offset 1000\n h_reference[3] = h_data[10000]; // Value at offset 10000\n h_reference[4] = h_data[1]; // Value at offset 1\n h_reference[5] = h_data[21]; // Value at offset 21\n h_reference[6] = h_data[11]; // Value at offset 11\n h_reference[7] = h_data[0]; // Value at offset 0;\n\n // Create and bind ref-based test iterator\n TexRefInputIterator d_ref_itr;\n CubDebugExit(d_ref_itr.BindTexture((CastT*) d_data, sizeof(T) * TEST_VALUES));\n\n // Create and bind dummy iterator of same type to check with interferance\n TexRefInputIterator d_ref_itr2;\n CubDebugExit(d_ref_itr2.BindTexture((CastT*) d_dummy, sizeof(T) * DUMMY_TEST_VALUES));\n\n Test(d_ref_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n T *d_copy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * TEST_VALUES));\n thrust::device_ptr d_copy_wrapper(d_copy);\n\n CubDebugExit(cudaMemset(d_copy, 0, sizeof(T) * TEST_VALUES));\n thrust::copy_if(d_ref_itr, d_ref_itr + TEST_VALUES, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_data, d_copy, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n\n CubDebugExit(d_ref_itr.UnbindTexture());\n CubDebugExit(d_ref_itr2.UnbindTexture());\n\n if (h_data) delete[] h_data;\n if (d_data) CubDebugExit(g_allocator.DeviceFree(d_data));\n if (d_dummy) CubDebugExit(g_allocator.DeviceFree(d_dummy));\n}\n\n\n/**\n * Test texture transform iterator\n */\ntemplate \nvoid TestTexTransform()\n{\n printf(\"\\nTesting tex-transform iterator on type %s\\n\", typeid(T).name()); fflush(stdout);\n\n //\n // Test iterator manipulation in kernel\n //\n\n const unsigned int TEST_VALUES = 11000;\n\n T *h_data = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n {\n InitValue(INTEGER_SEED, h_data[i], i);\n }\n\n // Allocate device arrays\n T *d_data = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_data, sizeof(T) * TEST_VALUES));\n CubDebugExit(cudaMemcpy(d_data, h_data, sizeof(T) * TEST_VALUES, cudaMemcpyHostToDevice));\n\n TransformOp op;\n\n // Initialize reference data\n T h_reference[8];\n h_reference[0] = op(h_data[0]); // Value at offset 0\n h_reference[1] = op(h_data[100]); // Value at offset 100\n h_reference[2] = op(h_data[1000]); // Value at offset 1000\n h_reference[3] = op(h_data[10000]); // Value at offset 10000\n h_reference[4] = op(h_data[1]); // Value at offset 1\n h_reference[5] = op(h_data[21]); // Value at offset 21\n h_reference[6] = op(h_data[11]); // Value at offset 11\n h_reference[7] = op(h_data[0]); // Value at offset 0;\n\n // Create and bind texture iterator\n typedef TexRefInputIterator TextureIterator;\n\n TextureIterator d_tex_itr;\n CubDebugExit(d_tex_itr.BindTexture((CastT*) d_data, sizeof(T) * TEST_VALUES));\n\n // Create transform iterator\n TransformInputIterator, TextureIterator> xform_itr(d_tex_itr, op);\n\n Test(xform_itr, h_reference);\n\n#if (THRUST_VERSION >= 100700) // Thrust 1.7 or newer\n\n //\n // Test with thrust::copy_if()\n //\n\n T *h_copy = new T[TEST_VALUES];\n for (int i = 0; i < TEST_VALUES; ++i)\n h_copy[i] = op(h_data[i]);\n\n T *d_copy = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_copy, sizeof(T) * TEST_VALUES));\n thrust::device_ptr d_copy_wrapper(d_copy);\n\n thrust::copy_if(xform_itr, xform_itr + TEST_VALUES, d_copy_wrapper, SelectOp());\n\n int compare = CompareDeviceResults(h_copy, d_copy, TEST_VALUES, g_verbose, g_verbose);\n printf(\"\\tthrust::copy_if(): %s\\n\", (compare) ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Cleanup\n if (h_copy) delete[] h_copy;\n if (d_copy) CubDebugExit(g_allocator.DeviceFree(d_copy));\n\n#endif // THRUST_VERSION\n\n CubDebugExit(d_tex_itr.UnbindTexture());\n if (h_data) delete[] h_data;\n if (d_data) CubDebugExit(g_allocator.DeviceFree(d_data));\n}\n\n#endif // CUDA_VERSION\n\n\n\n\n/**\n * Run non-integer tests\n */\ntemplate \nvoid Test(Int2Type is_integer)\n{\n TestModified();\n TestTransform();\n\n#if CUB_CDP\n // Test tex-obj iterators if CUDA dynamic parallelism enabled\n TestTexObj(type_string);\n#endif // CUB_CDP\n\n#if CUDA_VERSION >= 5050\n // Test tex-ref iterators for CUDA 5.5\n TestTexRef();\n TestTexTransform();\n#endif // CUDA_VERSION\n}\n\n/**\n * Run integer tests\n */\ntemplate \nvoid Test(Int2Type is_integer)\n{\n TestConstant(0);\n TestConstant(99);\n\n TestCounting(0);\n TestCounting(99);\n\n // Run non-integer tests\n Test(Int2Type());\n}\n\n/**\n * Run tests\n */\ntemplate \nvoid Test()\n{\n enum {\n IS_INTEGER = (Traits::CATEGORY == SIGNED_INTEGER) || (Traits::CATEGORY == UNSIGNED_INTEGER)\n };\n\n // Test non-const type\n Test(Int2Type());\n\n // Test non-const type\n Test(Int2Type());\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n // Evaluate different data types\n Test();\n Test();\n Test();\n Test();\n Test();\n Test();\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test();\n\n Test();\n Test();\n Test();\n Test();\n Test();\n Test();\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test();\n\n Test();\n Test();\n Test();\n Test();\n Test();\n Test();\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test();\n\n Test();\n Test();\n Test();\n Test();\n Test();\n Test();\n if (ptx_version > 120) // Don't check doubles on PTX120 or below because they're down-converted\n Test();\n\n Test();\n Test();\n\n printf(\"\\nTest complete\\n\"); fflush(stdout);\n\n return 0;\n}\n\n\n\n" }, { "path": "external/cub/test/test_util.h", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n * \n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n * \n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n\n#pragma once\n\n#if defined(_WIN32) || defined(_WIN64)\n #include \n #undef small // Windows is terrible for polluting macro namespace\n#else\n #include \n#endif\n\n#include \n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n\n#include \"mersenne.h\"\n\n#include \"cub/util_debug.cuh\"\n#include \"cub/util_device.cuh\"\n#include \"cub/util_type.cuh\"\n#include \"cub/util_macro.cuh\"\n\n/******************************************************************************\n * Assertion macros\n ******************************************************************************/\n\n/**\n * Assert equals\n */\n#define AssertEquals(a, b) if ((a) != (b)) { std::cerr << \"\\n(\" << __FILE__ << \": \" << __LINE__ << \")\\n\"; exit(1);}\n\n\n/******************************************************************************\n * Command-line parsing functionality\n ******************************************************************************/\n\n/**\n * Utility for parsing command line arguments\n */\nstruct CommandLineArgs\n{\n\n std::vector keys;\n std::vector values;\n std::vector args;\n cudaDeviceProp deviceProp;\n float device_giga_bandwidth;\n size_t device_free_physmem;\n size_t device_total_physmem;\n\n /**\n * Constructor\n */\n CommandLineArgs(int argc, char **argv) :\n keys(10),\n values(10)\n {\n using namespace std;\n\n // Initialize mersenne generator\n unsigned int mersenne_init[4]= {0x123, 0x234, 0x345, 0x456};\n mersenne::init_by_array(mersenne_init, 4);\n\n for (int i = 1; i < argc; i++)\n {\n string arg = argv[i];\n\n if ((arg[0] != '-') || (arg[1] != '-'))\n {\n args.push_back(arg);\n continue;\n }\n\n string::size_type pos;\n string key, val;\n if ((pos = arg.find('=')) == string::npos) {\n key = string(arg, 2, arg.length() - 2);\n val = \"\";\n } else {\n key = string(arg, 2, pos - 2);\n val = string(arg, pos + 1, arg.length() - 1);\n }\n\n keys.push_back(key);\n values.push_back(val);\n }\n }\n\n\n /**\n * Checks whether a flag \"--\" is present in the commandline\n */\n bool CheckCmdLineFlag(const char* arg_name)\n {\n using namespace std;\n\n for (int i = 0; i < int(keys.size()); ++i)\n {\n if (keys[i] == string(arg_name))\n return true;\n }\n return false;\n }\n\n\n /**\n * Returns number of naked (non-flag and non-key-value) commandline parameters\n */\n template \n int NumNakedArgs()\n {\n return args.size();\n }\n\n\n /**\n * Returns the commandline parameter for a given index (not including flags)\n */\n template \n void GetCmdLineArgument(int index, T &val)\n {\n using namespace std;\n if (index < args.size()) {\n istringstream str_stream(args[index]);\n str_stream >> val;\n }\n }\n\n /**\n * Returns the value specified for a given commandline parameter --=\n */\n template \n void GetCmdLineArgument(const char *arg_name, T &val)\n {\n using namespace std;\n\n for (int i = 0; i < int(keys.size()); ++i)\n {\n if (keys[i] == string(arg_name))\n {\n istringstream str_stream(values[i]);\n str_stream >> val;\n }\n }\n }\n\n\n /**\n * Returns the values specified for a given commandline parameter --=,*\n */\n template \n void GetCmdLineArguments(const char *arg_name, std::vector &vals)\n {\n using namespace std;\n\n if (CheckCmdLineFlag(arg_name))\n {\n // Clear any default values\n vals.clear();\n\n // Recover from multi-value string\n for (int i = 0; i < keys.size(); ++i)\n {\n if (keys[i] == string(arg_name))\n {\n string val_string(values[i]);\n istringstream str_stream(val_string);\n string::size_type old_pos = 0;\n string::size_type new_pos = 0;\n\n // Iterate comma-separated values\n T val;\n while ((new_pos = val_string.find(',', old_pos)) != string::npos)\n {\n if (new_pos != old_pos)\n {\n str_stream.width(new_pos - old_pos);\n str_stream >> val;\n vals.push_back(val);\n }\n\n // skip over comma\n str_stream.ignore(1);\n old_pos = new_pos + 1;\n }\n\n // Read last value\n str_stream >> val;\n vals.push_back(val);\n }\n }\n }\n }\n\n\n /**\n * The number of pairs parsed\n */\n int ParsedArgc()\n {\n return (int) keys.size();\n }\n\n /**\n * Initialize device\n */\n cudaError_t DeviceInit(int dev = -1)\n {\n cudaError_t error = cudaSuccess;\n\n do\n {\n int deviceCount;\n error = CubDebug(cudaGetDeviceCount(&deviceCount));\n if (error) break;\n\n if (deviceCount == 0) {\n fprintf(stderr, \"No devices supporting CUDA.\\n\");\n exit(1);\n }\n if (dev < 0)\n {\n GetCmdLineArgument(\"device\", dev);\n }\n if ((dev > deviceCount - 1) || (dev < 0))\n {\n dev = 0;\n }\n\n error = CubDebug(cudaSetDevice(dev));\n if (error) break;\n\n CubDebugExit(cudaMemGetInfo(&device_free_physmem, &device_total_physmem));\n\n int ptx_version;\n error = CubDebug(cub::PtxVersion(ptx_version));\n if (error) break;\n\n error = CubDebug(cudaGetDeviceProperties(&deviceProp, dev));\n if (error) break;\n\n if (deviceProp.major < 1) {\n fprintf(stderr, \"Device does not support CUDA.\\n\");\n exit(1);\n }\n\n device_giga_bandwidth = float(deviceProp.memoryBusWidth) * deviceProp.memoryClockRate * 2 / 8 / 1000 / 1000;\n\n if (!CheckCmdLineFlag(\"quiet\"))\n {\n printf(\n \"Using device %d: %s (PTX version %d, SM%d, %d SMs, \"\n \"%lld free / %lld total MB physmem, \"\n \"%.3f GB/s @ %d kHz mem clock, ECC %s)\\n\",\n dev,\n deviceProp.name,\n ptx_version,\n deviceProp.major * 100 + deviceProp.minor * 10,\n deviceProp.multiProcessorCount,\n (unsigned long long) device_free_physmem / 1024 / 1024,\n (unsigned long long) device_total_physmem / 1024 / 1024,\n device_giga_bandwidth,\n deviceProp.memoryClockRate,\n (deviceProp.ECCEnabled) ? \"on\" : \"off\");\n fflush(stdout);\n }\n\n } while (0);\n\n return error;\n }\n};\n\n/******************************************************************************\n * Random bits generator\n ******************************************************************************/\n\nint g_num_rand_samples = 0;\n\n\ntemplate \nbool IsNaN(T val) { return false; }\n\ntemplate<>\n__noinline__ bool IsNaN(float val)\n{\n volatile unsigned int bits = reinterpret_cast(val);\n\n return (((bits >= 0x7F800001) && (bits <= 0x7FFFFFFF)) || \n ((bits >= 0xFF800001) && (bits <= 0xFFFFFFFF)));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(float1 val)\n{\n return (IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(float2 val)\n{\n return (IsNaN(val.y) || IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(float3 val)\n{\n return (IsNaN(val.z) || IsNaN(val.y) || IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(float4 val)\n{\n return (IsNaN(val.y) || IsNaN(val.x) || IsNaN(val.w) || IsNaN(val.z));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(double val)\n{\n volatile unsigned long long bits = *reinterpret_cast(&val);\n\n return (((bits >= 0x7FF0000000000001) && (bits <= 0x7FFFFFFFFFFFFFFF)) || \n ((bits >= 0xFFF0000000000001) && (bits <= 0xFFFFFFFFFFFFFFFF)));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(double1 val)\n{\n return (IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(double2 val)\n{\n return (IsNaN(val.y) || IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(double3 val)\n{\n return (IsNaN(val.z) || IsNaN(val.y) || IsNaN(val.x));\n}\n\ntemplate<>\n__noinline__ bool IsNaN(double4 val)\n{\n return (IsNaN(val.y) || IsNaN(val.x) || IsNaN(val.w) || IsNaN(val.z));\n}\n\n\n/**\n * Generates random keys.\n *\n * We always take the second-order byte from rand() because the higher-order\n * bits returned by rand() are commonly considered more uniformly distributed\n * than the lower-order bits.\n *\n * We can decrease the entropy level of keys by adopting the technique\n * of Thearling and Smith in which keys are computed from the bitwise AND of\n * multiple random samples:\n *\n * entropy_reduction | Effectively-unique bits per key\n * -----------------------------------------------------\n * -1 | 0\n * 0 | 32\n * 1 | 25.95 (81%)\n * 2 | 17.41 (54%)\n * 3 | 10.78 (34%)\n * 4 | 6.42 (20%)\n * ... | ...\n *\n */\ntemplate \nvoid RandomBits(\n K &key,\n int entropy_reduction = 0,\n int begin_bit = 0,\n int end_bit = sizeof(K) * 8)\n{\n const int NUM_BYTES = sizeof(K);\n const int WORD_BYTES = sizeof(unsigned int);\n const int NUM_WORDS = (NUM_BYTES + WORD_BYTES - 1) / WORD_BYTES;\n\n unsigned int word_buff[NUM_WORDS];\n\n if (entropy_reduction == -1)\n {\n memset((void *) &key, 0, sizeof(key));\n return;\n }\n\n if (end_bit < 0)\n end_bit = sizeof(K) * 8;\n\n while (true) \n {\n // Generate random word_buff\n for (int j = 0; j < NUM_WORDS; j++)\n {\n int current_bit = j * WORD_BYTES * 8;\n\n unsigned int word = 0xffffffff;\n word &= 0xffffffff << CUB_MAX(0, begin_bit - current_bit);\n word &= 0xffffffff >> CUB_MAX(0, (current_bit + (WORD_BYTES * 8)) - end_bit);\n\n for (int i = 0; i <= entropy_reduction; i++)\n {\n // Grab some of the higher bits from rand (better entropy, supposedly)\n word &= mersenne::genrand_int32();\n g_num_rand_samples++; \n }\n\n word_buff[j] = word;\n }\n\n memcpy(&key, word_buff, sizeof(K));\n\n K copy = key;\n if (!IsNaN(copy))\n break; // avoids NaNs when generating random floating point numbers\n }\n}\n\n/// Randomly select number between [0:max)\ntemplate \nT RandomValue(T max)\n{\n unsigned int bits;\n unsigned int max_int = (unsigned int) -1;\n do {\n RandomBits(bits);\n } while (bits == max_int);\n\n return (T) ((double(bits) / double(max_int)) * double(max));\n}\n\n\n/******************************************************************************\n * Console printing utilities\n ******************************************************************************/\n\n/**\n * Helper for casting character types to integers for cout printing\n */\ntemplate \nT CoutCast(T val) { return val; }\n\nint CoutCast(char val) { return val; }\n\nint CoutCast(unsigned char val) { return val; }\n\nint CoutCast(signed char val) { return val; }\n\n\n\n/******************************************************************************\n * Test value initialization utilities\n ******************************************************************************/\n\n/**\n * Test problem generation options\n */\nenum GenMode\n{\n UNIFORM, // Assign to '2', regardless of integer seed\n INTEGER_SEED, // Assign to integer seed\n RANDOM, // Assign to random, regardless of integer seed\n};\n\n/**\n * Initialize value\n */\ntemplate \n__host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, T &value, int index = 0)\n{\n switch (gen_mode)\n {\n#if (CUB_PTX_ARCH == 0)\n case RANDOM:\n RandomBits(value);\n break;\n#endif\n case UNIFORM:\n value = 2;\n break;\n case INTEGER_SEED:\n default:\n value = (T) index;\n break;\n }\n}\n\n\n/**\n * Initialize value (bool)\n */\n__host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, bool &value, int index = 0)\n{\n switch (gen_mode)\n {\n#if (CUB_PTX_ARCH == 0)\n case RANDOM:\n char c;\n RandomBits(c, 0, 0, 1);\n value = (c > 0);\n break;\n#endif\n case UNIFORM:\n value = true;\n break;\n case INTEGER_SEED:\n default:\n value = (index > 0);\n break;\n }\n}\n\n\n/**\n * cub::NullType test initialization\n */\n__host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, cub::NullType &value, int index = 0)\n{}\n\n\n/**\n * cub::KeyValuePairtest initialization\n */\ntemplate \n__host__ __device__ __forceinline__ void InitValue(\n GenMode gen_mode,\n cub::KeyValuePair& value,\n int index = 0)\n{\n InitValue(gen_mode, value.value, index);\n\n // Assign corresponding flag with a likelihood of the last bit being set with entropy-reduction level 3\n RandomBits(value.key, 3);\n value.key = (value.key & 0x1);\n}\n\n\n\n/******************************************************************************\n * Comparison and ostream operators\n ******************************************************************************/\n\n/**\n * KeyValuePair ostream operator\n */\ntemplate \nstd::ostream& operator<<(std::ostream& os, const cub::KeyValuePair &val)\n{\n os << '(' << CoutCast(val.key) << ',' << CoutCast(val.value) << ')';\n return os;\n}\n\n\n/******************************************************************************\n * Comparison and ostream operators for CUDA vector types\n ******************************************************************************/\n\n/**\n * Vector1 overloads\n */\n#define CUB_VEC_OVERLOAD_1(T, BaseT) \\\n /* Ostream output */ \\\n std::ostream& operator<<( \\\n std::ostream& os, \\\n const T& val) \\\n { \\\n os << '(' << CoutCast(val.x) << ')'; \\\n return os; \\\n } \\\n /* Inequality */ \\\n __host__ __device__ __forceinline__ bool operator!=( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x != b.x); \\\n } \\\n /* Equality */ \\\n __host__ __device__ __forceinline__ bool operator==( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x == b.x); \\\n } \\\n /* Test initialization */ \\\n __host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, T &value, int index = 0) \\\n { \\\n InitValue(gen_mode, value.x, index); \\\n } \\\n /* Max */ \\\n __host__ __device__ __forceinline__ bool operator>( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x > b.x); \\\n } \\\n /* Min */ \\\n __host__ __device__ __forceinline__ bool operator<( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x < b.x); \\\n } \\\n /* Summation (non-reference addends for VS2003 -O3 warpscan workaround */ \\\n __host__ __device__ __forceinline__ T operator+( \\\n T a, \\\n T b) \\\n { \\\n T retval = make_##T(a.x + b.x); \\\n return retval; \\\n } \\\n namespace cub { \\\n template<> \\\n struct NumericTraits \\\n { \\\n static const Category CATEGORY = NOT_A_NUMBER; \\\n enum { \\\n PRIMITIVE = false, \\\n NULL_TYPE = false, \\\n }; \\\n static T Max() \\\n { \\\n T retval = { \\\n NumericTraits::Max()}; \\\n return retval; \\\n } \\\n static T Lowest() \\\n { \\\n T retval = { \\\n NumericTraits::Lowest()}; \\\n return retval; \\\n } \\\n }; \\\n } /* namespace std */\n\n\n\n/**\n * Vector2 overloads\n */\n#define CUB_VEC_OVERLOAD_2(T, BaseT) \\\n /* Ostream output */ \\\n std::ostream& operator<<( \\\n std::ostream& os, \\\n const T& val) \\\n { \\\n os << '(' \\\n << CoutCast(val.x) << ',' \\\n << CoutCast(val.y) << ')'; \\\n return os; \\\n } \\\n /* Inequality */ \\\n __host__ __device__ __forceinline__ bool operator!=( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x != b.x) || \\\n (a.y != b.y); \\\n } \\\n /* Equality */ \\\n __host__ __device__ __forceinline__ bool operator==( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x == b.x) && \\\n (a.y == b.y); \\\n } \\\n /* Test initialization */ \\\n __host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, T &value, int index = 0) \\\n { \\\n InitValue(gen_mode, value.x, index); \\\n InitValue(gen_mode, value.y, index); \\\n } \\\n /* Max */ \\\n __host__ __device__ __forceinline__ bool operator>( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x > b.x) return true; else if (b.x > a.x) return false; \\\n return a.y > b.y; \\\n } \\\n /* Min */ \\\n __host__ __device__ __forceinline__ bool operator<( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x < b.x) return true; else if (b.x < a.x) return false; \\\n return a.y < b.y; \\\n } \\\n /* Summation (non-reference addends for VS2003 -O3 warpscan workaround */ \\\n __host__ __device__ __forceinline__ T operator+( \\\n T a, \\\n T b) \\\n { \\\n T retval = make_##T( \\\n a.x + b.x, \\\n a.y + b.y); \\\n return retval; \\\n } \\\n namespace cub { \\\n template<> \\\n struct NumericTraits \\\n { \\\n static const Category CATEGORY = NOT_A_NUMBER; \\\n enum { \\\n PRIMITIVE = false, \\\n NULL_TYPE = false, \\\n }; \\\n static T Max() \\\n { \\\n T retval = { \\\n NumericTraits::Max(), \\\n NumericTraits::Max()}; \\\n return retval; \\\n } \\\n static T Lowest() \\\n { \\\n T retval = { \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest()}; \\\n return retval; \\\n } \\\n }; \\\n } /* namespace cub */\n\n\n\n/**\n * Vector3 overloads\n */\n#define CUB_VEC_OVERLOAD_3(T, BaseT) \\\n /* Ostream output */ \\\n std::ostream& operator<<( \\\n std::ostream& os, \\\n const T& val) \\\n { \\\n os << '(' \\\n << CoutCast(val.x) << ',' \\\n << CoutCast(val.y) << ',' \\\n << CoutCast(val.z) << ')'; \\\n return os; \\\n } \\\n /* Inequality */ \\\n __host__ __device__ __forceinline__ bool operator!=( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x != b.x) || \\\n (a.y != b.y) || \\\n (a.z != b.z); \\\n } \\\n /* Equality */ \\\n __host__ __device__ __forceinline__ bool operator==( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x == b.x) && \\\n (a.y == b.y) && \\\n (a.z == b.z); \\\n } \\\n /* Test initialization */ \\\n __host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, T &value, int index = 0) \\\n { \\\n InitValue(gen_mode, value.x, index); \\\n InitValue(gen_mode, value.y, index); \\\n InitValue(gen_mode, value.z, index); \\\n } \\\n /* Max */ \\\n __host__ __device__ __forceinline__ bool operator>( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x > b.x) return true; else if (b.x > a.x) return false; \\\n if (a.y > b.y) return true; else if (b.y > a.y) return false; \\\n return a.z > b.z; \\\n } \\\n /* Min */ \\\n __host__ __device__ __forceinline__ bool operator<( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x < b.x) return true; else if (b.x < a.x) return false; \\\n if (a.y < b.y) return true; else if (b.y < a.y) return false; \\\n return a.z < b.z; \\\n } \\\n /* Summation (non-reference addends for VS2003 -O3 warpscan workaround */ \\\n __host__ __device__ __forceinline__ T operator+( \\\n T a, \\\n T b) \\\n { \\\n T retval = make_##T( \\\n a.x + b.x, \\\n a.y + b.y, \\\n a.z + b.z); \\\n return retval; \\\n } \\\n namespace cub { \\\n template<> \\\n struct NumericTraits \\\n { \\\n static const Category CATEGORY = NOT_A_NUMBER; \\\n enum { \\\n PRIMITIVE = false, \\\n NULL_TYPE = false, \\\n }; \\\n static T Max() \\\n { \\\n T retval = { \\\n NumericTraits::Max(), \\\n NumericTraits::Max(), \\\n NumericTraits::Max()}; \\\n return retval; \\\n } \\\n static T Lowest() \\\n { \\\n T retval = { \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest()}; \\\n return retval; \\\n } \\\n }; \\\n } /* namespace cub */\n\n\n/**\n * Vector4 overloads\n */\n#define CUB_VEC_OVERLOAD_4(T, BaseT) \\\n /* Ostream output */ \\\n std::ostream& operator<<( \\\n std::ostream& os, \\\n const T& val) \\\n { \\\n os << '(' \\\n << CoutCast(val.x) << ',' \\\n << CoutCast(val.y) << ',' \\\n << CoutCast(val.z) << ',' \\\n << CoutCast(val.w) << ')'; \\\n return os; \\\n } \\\n /* Inequality */ \\\n __host__ __device__ __forceinline__ bool operator!=( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x != b.x) || \\\n (a.y != b.y) || \\\n (a.z != b.z) || \\\n (a.w != b.w); \\\n } \\\n /* Equality */ \\\n __host__ __device__ __forceinline__ bool operator==( \\\n const T &a, \\\n const T &b) \\\n { \\\n return (a.x == b.x) && \\\n (a.y == b.y) && \\\n (a.z == b.z) && \\\n (a.w == b.w); \\\n } \\\n /* Test initialization */ \\\n __host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, T &value, int index = 0) \\\n { \\\n InitValue(gen_mode, value.x, index); \\\n InitValue(gen_mode, value.y, index); \\\n InitValue(gen_mode, value.z, index); \\\n InitValue(gen_mode, value.w, index); \\\n } \\\n /* Max */ \\\n __host__ __device__ __forceinline__ bool operator>( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x > b.x) return true; else if (b.x > a.x) return false; \\\n if (a.y > b.y) return true; else if (b.y > a.y) return false; \\\n if (a.z > b.z) return true; else if (b.z > a.z) return false; \\\n return a.w > b.w; \\\n } \\\n /* Min */ \\\n __host__ __device__ __forceinline__ bool operator<( \\\n const T &a, \\\n const T &b) \\\n { \\\n if (a.x < b.x) return true; else if (b.x < a.x) return false; \\\n if (a.y < b.y) return true; else if (b.y < a.y) return false; \\\n if (a.z < b.z) return true; else if (b.z < a.z) return false; \\\n return a.w < b.w; \\\n } \\\n /* Summation (non-reference addends for VS2003 -O3 warpscan workaround */ \\\n __host__ __device__ __forceinline__ T operator+( \\\n T a, \\\n T b) \\\n { \\\n T retval = make_##T( \\\n a.x + b.x, \\\n a.y + b.y, \\\n a.z + b.z, \\\n a.w + b.w); \\\n return retval; \\\n } \\\n namespace cub { \\\n template<> \\\n struct NumericTraits \\\n { \\\n static const Category CATEGORY = NOT_A_NUMBER; \\\n enum { \\\n PRIMITIVE = false, \\\n NULL_TYPE = false, \\\n }; \\\n static T Max() \\\n { \\\n T retval = { \\\n NumericTraits::Max(), \\\n NumericTraits::Max(), \\\n NumericTraits::Max(), \\\n NumericTraits::Max()}; \\\n return retval; \\\n } \\\n static T Lowest() \\\n { \\\n T retval = { \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest(), \\\n NumericTraits::Lowest()}; \\\n return retval; \\\n } \\\n }; \\\n } /* namespace cub */\n\n/**\n * All vector overloads\n */\n#define CUB_VEC_OVERLOAD(COMPONENT_T, BaseT) \\\n CUB_VEC_OVERLOAD_1(COMPONENT_T##1, BaseT) \\\n CUB_VEC_OVERLOAD_2(COMPONENT_T##2, BaseT) \\\n CUB_VEC_OVERLOAD_3(COMPONENT_T##3, BaseT) \\\n CUB_VEC_OVERLOAD_4(COMPONENT_T##4, BaseT)\n\n/**\n * Define for types\n */\nCUB_VEC_OVERLOAD(char, char)\nCUB_VEC_OVERLOAD(short, short)\nCUB_VEC_OVERLOAD(int, int)\nCUB_VEC_OVERLOAD(long, long)\nCUB_VEC_OVERLOAD(longlong, long long)\nCUB_VEC_OVERLOAD(uchar, unsigned char)\nCUB_VEC_OVERLOAD(ushort, unsigned short)\nCUB_VEC_OVERLOAD(uint, unsigned int)\nCUB_VEC_OVERLOAD(ulong, unsigned long)\nCUB_VEC_OVERLOAD(ulonglong, unsigned long long)\nCUB_VEC_OVERLOAD(float, float)\nCUB_VEC_OVERLOAD(double, double)\n\n\n//---------------------------------------------------------------------\n// Complex data type TestFoo\n//---------------------------------------------------------------------\n\n/**\n * TestFoo complex data type\n */\nstruct TestFoo\n{\n long long x;\n int y;\n short z;\n char w;\n\n // Factory\n static __host__ __device__ __forceinline__ TestFoo MakeTestFoo(long long x, int y, short z, char w)\n {\n TestFoo retval = {x, y, z, w};\n return retval;\n }\n\n // Assignment from int operator\n __host__ __device__ __forceinline__ TestFoo& operator =(int b)\n {\n x = b;\n y = b;\n z = b;\n w = b;\n return *this;\n }\n\n // Summation operator\n __host__ __device__ __forceinline__ TestFoo operator+(const TestFoo &b) const\n {\n return MakeTestFoo(x + b.x, y + b.y, z + b.z, w + b.w);\n }\n\n // Inequality operator\n __host__ __device__ __forceinline__ bool operator !=(const TestFoo &b) const\n {\n return (x != b.x) || (y != b.y) || (z != b.z) || (w != b.w);\n }\n\n // Equality operator\n __host__ __device__ __forceinline__ bool operator ==(const TestFoo &b) const\n {\n return (x == b.x) && (y == b.y) && (z == b.z) && (w == b.w);\n }\n\n // Less than operator\n __host__ __device__ __forceinline__ bool operator <(const TestFoo &b) const\n {\n if (x < b.x) return true; else if (b.x < x) return false;\n if (y < b.y) return true; else if (b.y < y) return false;\n if (z < b.z) return true; else if (b.z < z) return false;\n return w < b.w;\n }\n\n // Greater than operator\n __host__ __device__ __forceinline__ bool operator >(const TestFoo &b) const\n {\n if (x > b.x) return true; else if (b.x > x) return false;\n if (y > b.y) return true; else if (b.y > y) return false;\n if (z > b.z) return true; else if (b.z > z) return false;\n return w > b.w;\n }\n\n};\n\n/**\n * TestFoo ostream operator\n */\nstd::ostream& operator<<(std::ostream& os, const TestFoo& val)\n{\n os << '(' << val.x << ',' << val.y << ',' << val.z << ',' << CoutCast(val.w) << ')';\n return os;\n}\n\n/**\n * TestFoo test initialization\n */\n__host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, TestFoo &value, int index = 0)\n{\n InitValue(gen_mode, value.x, index);\n InitValue(gen_mode, value.y, index);\n InitValue(gen_mode, value.z, index);\n InitValue(gen_mode, value.w, index);\n}\n\n\n/// numeric_limits specialization\nnamespace cub {\ntemplate<>\nstruct NumericTraits\n{\n static const Category CATEGORY = NOT_A_NUMBER;\n enum {\n PRIMITIVE = false,\n NULL_TYPE = false,\n };\n static TestFoo Max()\n {\n return TestFoo::MakeTestFoo(\n NumericTraits::Max(),\n NumericTraits::Max(),\n NumericTraits::Max(),\n NumericTraits::Max());\n }\n\n static TestFoo Lowest()\n {\n return TestFoo::MakeTestFoo(\n NumericTraits::Lowest(),\n NumericTraits::Lowest(),\n NumericTraits::Lowest(),\n NumericTraits::Lowest());\n }\n};\n} // namespace cub\n\n\n//---------------------------------------------------------------------\n// Complex data type TestBar (with optimizations for fence-free warp-synchrony)\n//---------------------------------------------------------------------\n\n/**\n * TestBar complex data type\n */\nstruct TestBar\n{\n long long x;\n int y;\n\n // Constructor\n __host__ __device__ __forceinline__ TestBar() : x(0), y(0)\n {}\n\n // Constructor\n __host__ __device__ __forceinline__ TestBar(int b) : x(b), y(b)\n {}\n\n // Constructor\n __host__ __device__ __forceinline__ TestBar(long long x, int y) : x(x), y(y)\n {}\n\n // Assignment from int operator\n __host__ __device__ __forceinline__ TestBar& operator =(int b)\n {\n x = b;\n y = b;\n return *this;\n }\n\n // Summation operator\n __host__ __device__ __forceinline__ TestBar operator+(const TestBar &b) const\n {\n return TestBar(x + b.x, y + b.y);\n }\n\n // Inequality operator\n __host__ __device__ __forceinline__ bool operator !=(const TestBar &b) const\n {\n return (x != b.x) || (y != b.y);\n }\n\n // Equality operator\n __host__ __device__ __forceinline__ bool operator ==(const TestBar &b) const\n {\n return (x == b.x) && (y == b.y);\n }\n\n // Less than operator\n __host__ __device__ __forceinline__ bool operator <(const TestBar &b) const\n {\n if (x < b.x) return true; else if (b.x < x) return false;\n return y < b.y;\n }\n\n // Greater than operator\n __host__ __device__ __forceinline__ bool operator >(const TestBar &b) const\n {\n if (x > b.x) return true; else if (b.x > x) return false;\n return y > b.y;\n }\n\n};\n\n\n/**\n * TestBar ostream operator\n */\nstd::ostream& operator<<(std::ostream& os, const TestBar& val)\n{\n os << '(' << val.x << ',' << val.y << ')';\n return os;\n}\n\n/**\n * TestBar test initialization\n */\n__host__ __device__ __forceinline__ void InitValue(GenMode gen_mode, TestBar &value, int index = 0)\n{\n InitValue(gen_mode, value.x, index);\n InitValue(gen_mode, value.y, index);\n}\n\n/// numeric_limits specialization\nnamespace cub {\ntemplate<>\nstruct NumericTraits\n{\n static const Category CATEGORY = NOT_A_NUMBER;\n enum {\n PRIMITIVE = false,\n NULL_TYPE = false,\n };\n static TestBar Max()\n {\n return TestBar(\n NumericTraits::Max(),\n NumericTraits::Max());\n }\n\n static TestBar Lowest()\n {\n return TestBar(\n NumericTraits::Lowest(),\n NumericTraits::Lowest());\n }\n};\n} // namespace cub\n\n\n/******************************************************************************\n * Helper routines for list comparison and display\n ******************************************************************************/\n\n\n/**\n * Compares the equivalence of two arrays\n */\ntemplate \nint CompareResults(T* computed, S* reference, OffsetT len, bool verbose = true)\n{\n for (OffsetT i = 0; i < len; i++)\n {\n if (computed[i] != reference[i])\n {\n if (verbose) std::cout << \"INCORRECT: [\" << i << \"]: \"\n << CoutCast(computed[i]) << \" != \"\n << CoutCast(reference[i]);\n return 1;\n }\n }\n return 0;\n}\n\n\n/**\n * Compares the equivalence of two arrays\n */\ntemplate \nint CompareResults(float* computed, float* reference, OffsetT len, bool verbose = true)\n{\n for (OffsetT i = 0; i < len; i++)\n {\n if (computed[i] != reference[i])\n {\n float difference = std::abs(computed[i]-reference[i]);\n float fraction = difference / std::abs(reference[i]);\n\n if (fraction > 0.0001)\n {\n if (verbose) std::cout << \"INCORRECT: [\" << i << \"]: \"\n << \"(computed) \" << CoutCast(computed[i]) << \" != \"\n << CoutCast(reference[i]) << \" (difference:\" << difference << \", fraction: \" << fraction << \")\";\n return 1;\n }\n }\n }\n return 0;\n}\n\n\n/**\n * Compares the equivalence of two arrays\n */\ntemplate \nint CompareResults(cub::NullType* computed, cub::NullType* reference, OffsetT len, bool verbose = true)\n{\n return 0;\n}\n\n/**\n * Compares the equivalence of two arrays\n */\ntemplate \nint CompareResults(double* computed, double* reference, OffsetT len, bool verbose = true)\n{\n for (OffsetT i = 0; i < len; i++)\n {\n if (computed[i] != reference[i])\n {\n double difference = std::abs(computed[i]-reference[i]);\n double fraction = difference / std::abs(reference[i]);\n\n if (fraction > 0.0001)\n {\n if (verbose) std::cout << \"INCORRECT: [\" << i << \"]: \"\n << CoutCast(computed[i]) << \" != \"\n << CoutCast(reference[i]) << \" (difference:\" << difference << \", fraction: \" << fraction << \")\";\n return 1;\n }\n }\n }\n return 0;\n}\n\n\n/**\n * Verify the contents of a device array match those\n * of a host array\n */\nint CompareDeviceResults(\n cub::NullType *h_reference,\n cub::NullType *d_data,\n size_t num_items,\n bool verbose = true,\n bool display_data = false)\n{\n return 0;\n}\n\n\n/**\n * Verify the contents of a device array match those\n * of a host array\n */\ntemplate \nint CompareDeviceResults(\n S *h_reference,\n T *d_data,\n size_t num_items,\n bool verbose = true,\n bool display_data = false)\n{\n // Allocate array on host\n T *h_data = (T*) malloc(num_items * sizeof(T));\n\n // Copy data back\n cudaMemcpy(h_data, d_data, sizeof(T) * num_items, cudaMemcpyDeviceToHost);\n\n // Display data\n if (display_data)\n {\n printf(\"Reference:\\n\");\n for (int i = 0; i < int(num_items); i++)\n {\n std::cout << CoutCast(h_reference[i]) << \", \";\n }\n printf(\"\\n\\nComputed:\\n\");\n for (int i = 0; i < int(num_items); i++)\n {\n std::cout << CoutCast(h_data[i]) << \", \";\n }\n printf(\"\\n\\n\");\n }\n\n // Check\n int retval = CompareResults(h_data, h_reference, num_items, verbose);\n\n // Cleanup\n if (h_data) free(h_data);\n\n return retval;\n}\n\n\n/**\n * Verify the contents of a device array match those\n * of a device array\n */\ntemplate \nint CompareDeviceDeviceResults(\n T *d_reference,\n T *d_data,\n size_t num_items,\n bool verbose = true,\n bool display_data = false)\n{\n // Allocate array on host\n T *h_reference = (T*) malloc(num_items * sizeof(T));\n T *h_data = (T*) malloc(num_items * sizeof(T));\n\n // Copy data back\n cudaMemcpy(h_reference, d_reference, sizeof(T) * num_items, cudaMemcpyDeviceToHost);\n cudaMemcpy(h_data, d_data, sizeof(T) * num_items, cudaMemcpyDeviceToHost);\n\n // Display data\n if (display_data) {\n printf(\"Reference:\\n\");\n for (int i = 0; i < num_items; i++)\n {\n std::cout << CoutCast(h_reference[i]) << \", \";\n }\n printf(\"\\n\\nComputed:\\n\");\n for (int i = 0; i < num_items; i++)\n {\n std::cout << CoutCast(h_data[i]) << \", \";\n }\n printf(\"\\n\\n\");\n }\n\n // Check\n int retval = CompareResults(h_data, h_reference, num_items, verbose);\n\n // Cleanup\n if (h_reference) free(h_reference);\n if (h_data) free(h_data);\n\n return retval;\n}\n\n\n/**\n * Print the contents of a host array\n */\nvoid DisplayResults(\n cub::NullType *h_data,\n size_t num_items)\n{}\n\n\n/**\n * Print the contents of a host array\n */\ntemplate \nvoid DisplayResults(\n InputIteratorT h_data,\n size_t num_items)\n{\n // Display data\n for (int i = 0; i < int(num_items); i++)\n {\n std::cout << CoutCast(h_data[i]) << \", \";\n }\n printf(\"\\n\");\n}\n\n\n/**\n * Print the contents of a device array\n */\ntemplate \nvoid DisplayDeviceResults(\n T *d_data,\n size_t num_items)\n{\n // Allocate array on host\n T *h_data = (T*) malloc(num_items * sizeof(T));\n\n // Copy data back\n cudaMemcpy(h_data, d_data, sizeof(T) * num_items, cudaMemcpyDeviceToHost);\n\n DisplayResults(h_data, num_items);\n\n // Cleanup\n if (h_data) free(h_data);\n}\n\n\n/******************************************************************************\n * Segment descriptor generation\n ******************************************************************************/\n\n/**\n * Initialize segments\n */\nvoid InitializeSegments(\n int num_items,\n int num_segments,\n int *h_segment_offsets,\n bool verbose = false)\n{\n if (num_segments <= 0)\n return;\n\n unsigned int expected_segment_length = (num_items + num_segments - 1) / num_segments;\n int offset = 0;\n for (int i = 0; i < num_segments; ++i)\n {\n h_segment_offsets[i] = offset;\n\n unsigned int segment_length = RandomValue((expected_segment_length * 2) + 1);\n offset += segment_length;\n offset = CUB_MIN(offset, num_items);\n }\n h_segment_offsets[num_segments] = num_items;\n\n if (verbose)\n {\n printf(\"Segment offsets: \");\n DisplayResults(h_segment_offsets, num_segments + 1);\n }\n}\n\n\n/******************************************************************************\n * Timing\n ******************************************************************************/\n\n\nstruct CpuTimer\n{\n#if defined(_WIN32) || defined(_WIN64)\n\n LARGE_INTEGER ll_freq;\n LARGE_INTEGER ll_start;\n LARGE_INTEGER ll_stop;\n\n CpuTimer()\n {\n QueryPerformanceFrequency(&ll_freq);\n }\n\n void Start()\n {\n QueryPerformanceCounter(&ll_start);\n }\n\n void Stop()\n {\n QueryPerformanceCounter(&ll_stop);\n }\n\n float ElapsedMillis()\n {\n double start = double(ll_start.QuadPart) / double(ll_freq.QuadPart);\n double stop = double(ll_stop.QuadPart) / double(ll_freq.QuadPart);\n\n return float((stop - start) * 1000);\n }\n\n#else\n\n rusage start;\n rusage stop;\n\n void Start()\n {\n getrusage(RUSAGE_SELF, &start);\n }\n\n void Stop()\n {\n getrusage(RUSAGE_SELF, &stop);\n }\n\n float ElapsedMillis()\n {\n float sec = stop.ru_utime.tv_sec - start.ru_utime.tv_sec;\n float usec = stop.ru_utime.tv_usec - start.ru_utime.tv_usec;\n\n return (sec * 1000) + (usec / 1000);\n }\n\n#endif\n};\n\nstruct GpuTimer\n{\n cudaEvent_t start;\n cudaEvent_t stop;\n\n GpuTimer()\n {\n cudaEventCreate(&start);\n cudaEventCreate(&stop);\n }\n\n ~GpuTimer()\n {\n cudaEventDestroy(start);\n cudaEventDestroy(stop);\n }\n\n void Start()\n {\n cudaEventRecord(start, 0);\n }\n\n void Stop()\n {\n cudaEventRecord(stop, 0);\n }\n\n float ElapsedMillis()\n {\n float elapsed;\n cudaEventSynchronize(stop);\n cudaEventElapsedTime(&elapsed, start, stop);\n return elapsed;\n }\n};\n" }, { "path": "external/cub/test/test_warp_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of WarpReduce utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n/**\n * \\brief WrapperFunctor (for precluding test-specialized dispatch to *Sum variants)\n */\ntemplate<\n typename OpT,\n int LOGICAL_WARP_THREADS>\nstruct WrapperFunctor\n{\n OpT op;\n int num_valid;\n\n inline __host__ __device__ WrapperFunctor(OpT op, int num_valid) : op(op), num_valid(num_valid) {}\n\n template \n inline __host__ __device__ T operator()(const T &a, const T &b) const\n {\n#if CUB_PTX_ARCH != 0\n if ((cub::LaneId() % LOGICAL_WARP_THREADS) >= num_valid)\n cub::ThreadTrap();\n#endif\n\n return op(a, b);\n }\n\n};\n\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/**\n * Generic reduction\n */\ntemplate <\n typename T,\n typename ReductionOp,\n typename WarpReduce,\n bool PRIMITIVE = Traits::PRIMITIVE>\nstruct DeviceTest\n{\n static __device__ __forceinline__ T Reduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n ReductionOp &reduction_op)\n {\n return WarpReduce(temp_storage).Reduce(data, reduction_op);\n }\n\n static __device__ __forceinline__ T Reduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n ReductionOp &reduction_op,\n const int &valid_warp_threads)\n {\n return WarpReduce(temp_storage).Reduce(data, reduction_op, valid_warp_threads);\n }\n\n template \n static __device__ __forceinline__ T HeadSegmentedReduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n FlagT &flag,\n ReductionOp &reduction_op)\n {\n return WarpReduce(temp_storage).HeadSegmentedReduce(data, flag, reduction_op);\n }\n\n template \n static __device__ __forceinline__ T TailSegmentedReduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n FlagT &flag,\n ReductionOp &reduction_op)\n {\n return WarpReduce(temp_storage).TailSegmentedReduce(data, flag, reduction_op);\n }\n\n};\n\n\n/**\n * Summation\n */\ntemplate <\n typename T,\n typename WarpReduce>\nstruct DeviceTest\n{\n static __device__ __forceinline__ T Reduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n Sum &reduction_op)\n {\n return WarpReduce(temp_storage).Sum(data);\n }\n\n static __device__ __forceinline__ T Reduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n Sum &reduction_op,\n const int &valid_warp_threads)\n {\n return WarpReduce(temp_storage).Sum(data, valid_warp_threads);\n }\n\n template \n static __device__ __forceinline__ T HeadSegmentedReduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n FlagT &flag,\n Sum &reduction_op)\n {\n return WarpReduce(temp_storage).HeadSegmentedSum(data, flag);\n }\n\n template \n static __device__ __forceinline__ T TailSegmentedReduce(\n typename WarpReduce::TempStorage &temp_storage,\n T &data,\n FlagT &flag,\n Sum &reduction_op)\n {\n return WarpReduce(temp_storage).TailSegmentedSum(data, flag);\n }\n\n};\n\n\n/**\n * Full-tile warp reduction kernel\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename ReductionOp>\n__global__ void FullWarpReduceKernel(\n T *d_in,\n T *d_out,\n ReductionOp reduction_op,\n clock_t *d_elapsed)\n{\n // Cooperative warp-reduce utility type (1 warp)\n typedef WarpReduce WarpReduce;\n\n // Allocate temp storage in shared memory\n __shared__ typename WarpReduce::TempStorage temp_storage[WARPS];\n\n // Per-thread tile data\n T input = d_in[threadIdx.x];\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Test warp reduce\n int warp_id = threadIdx.x / LOGICAL_WARP_THREADS;\n\n T output = DeviceTest::Reduce(\n temp_storage[warp_id], input, reduction_op);\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n *d_elapsed = stop - start;\n\n // Store aggregate\n d_out[threadIdx.x] = (threadIdx.x % LOGICAL_WARP_THREADS == 0) ?\n output :\n input;\n}\n\n/**\n * Partially-full warp reduction kernel\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename ReductionOp>\n__global__ void PartialWarpReduceKernel(\n T *d_in,\n T *d_out,\n ReductionOp reduction_op,\n clock_t *d_elapsed,\n int valid_warp_threads)\n{\n // Cooperative warp-reduce utility type\n typedef WarpReduce WarpReduce;\n\n // Allocate temp storage in shared memory\n __shared__ typename WarpReduce::TempStorage temp_storage[WARPS];\n\n // Per-thread tile data\n T input = d_in[threadIdx.x];\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Test partial-warp reduce\n int warp_id = threadIdx.x / LOGICAL_WARP_THREADS;\n T output = DeviceTest::Reduce(\n temp_storage[warp_id], input, reduction_op, valid_warp_threads);\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n *d_elapsed = stop - start;\n\n // Store aggregate\n d_out[threadIdx.x] = (threadIdx.x % LOGICAL_WARP_THREADS == 0) ?\n output :\n input;\n}\n\n\n/**\n * Head-based segmented warp reduction test kernel\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename FlagT,\n typename ReductionOp>\n__global__ void WarpHeadSegmentedReduceKernel(\n T *d_in,\n FlagT *d_head_flags,\n T *d_out,\n ReductionOp reduction_op,\n clock_t *d_elapsed)\n{\n // Cooperative warp-reduce utility type\n typedef WarpReduce WarpReduce;\n\n // Allocate temp storage in shared memory\n __shared__ typename WarpReduce::TempStorage temp_storage[WARPS];\n\n // Per-thread tile data\n T input = d_in[threadIdx.x];\n FlagT head_flag = d_head_flags[threadIdx.x];\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Test segmented warp reduce\n int warp_id = threadIdx.x / LOGICAL_WARP_THREADS;\n T output = DeviceTest::HeadSegmentedReduce(\n temp_storage[warp_id], input, head_flag, reduction_op);\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n *d_elapsed = stop - start;\n\n // Store aggregate\n d_out[threadIdx.x] = ((threadIdx.x % LOGICAL_WARP_THREADS == 0) || head_flag) ?\n output :\n input;\n}\n\n\n/**\n * Tail-based segmented warp reduction test kernel\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename FlagT,\n typename ReductionOp>\n__global__ void WarpTailSegmentedReduceKernel(\n T *d_in,\n FlagT *d_tail_flags,\n T *d_out,\n ReductionOp reduction_op,\n clock_t *d_elapsed)\n{\n // Cooperative warp-reduce utility type\n typedef WarpReduce WarpReduce;\n\n // Allocate temp storage in shared memory\n __shared__ typename WarpReduce::TempStorage temp_storage[WARPS];\n\n // Per-thread tile data\n T input = d_in[threadIdx.x];\n FlagT tail_flag = d_tail_flags[threadIdx.x];\n FlagT head_flag = (threadIdx.x == 0) ?\n 0 :\n d_tail_flags[threadIdx.x - 1];\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Test segmented warp reduce\n int warp_id = threadIdx.x / LOGICAL_WARP_THREADS;\n T output = DeviceTest::TailSegmentedReduce(\n temp_storage[warp_id], input, tail_flag, reduction_op);\n\n // Record elapsed clocks\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n *d_elapsed = stop - start;\n\n // Store aggregate\n d_out[threadIdx.x] = ((threadIdx.x % LOGICAL_WARP_THREADS == 0) || head_flag) ?\n output :\n input;\n}\n\n\n//---------------------------------------------------------------------\n// Host utility subroutines\n//---------------------------------------------------------------------\n\n/**\n * Initialize reduction problem (and solution)\n */\ntemplate <\n typename T,\n typename ReductionOp>\nvoid Initialize(\n GenMode gen_mode,\n int flag_entropy,\n T *h_in,\n int *h_flags,\n int warps,\n int warp_threads,\n int valid_warp_threads,\n ReductionOp reduction_op,\n T *h_head_out,\n T *h_tail_out)\n{\n for (int i = 0; i < warps * warp_threads; ++i)\n {\n // Sample a value for this item\n InitValue(gen_mode, h_in[i], i);\n h_head_out[i] = h_in[i];\n h_tail_out[i] = h_in[i];\n\n // Sample whether or not this item will be a segment head\n char bits;\n RandomBits(bits, flag_entropy);\n h_flags[i] = bits & 0x1;\n }\n\n // Accumulate segments (lane 0 of each warp is implicitly a segment head)\n for (int warp = 0; warp < warps; ++warp)\n {\n int warp_offset = warp * warp_threads;\n int item_offset = warp_offset + valid_warp_threads - 1;\n\n // Last item in warp\n T head_aggregate = h_in[item_offset];\n T tail_aggregate = h_in[item_offset];\n\n if (h_flags[item_offset])\n h_head_out[item_offset] = head_aggregate;\n item_offset--;\n\n // Work backwards\n while (item_offset >= warp_offset)\n {\n if (h_flags[item_offset + 1])\n {\n head_aggregate = h_in[item_offset];\n }\n else\n {\n head_aggregate = reduction_op(head_aggregate, h_in[item_offset]);\n }\n\n if (h_flags[item_offset])\n {\n h_head_out[item_offset] = head_aggregate;\n h_tail_out[item_offset + 1] = tail_aggregate;\n tail_aggregate = h_in[item_offset];\n }\n else\n {\n tail_aggregate = reduction_op(tail_aggregate, h_in[item_offset]);\n }\n\n item_offset--;\n }\n\n // Record last segment head_aggregate to head offset\n h_head_out[warp_offset] = head_aggregate;\n h_tail_out[warp_offset] = tail_aggregate;\n }\n}\n\n\n/**\n * Test warp reduction\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename ReductionOp>\nvoid TestReduce(\n GenMode gen_mode,\n ReductionOp reduction_op,\n int valid_warp_threads = LOGICAL_WARP_THREADS)\n{\n const int BLOCK_THREADS = LOGICAL_WARP_THREADS * WARPS;\n\n // Allocate host arrays\n T *h_in = new T[BLOCK_THREADS];\n int *h_flags = new int[BLOCK_THREADS];\n T *h_out = new T[BLOCK_THREADS];\n T *h_tail_out = new T[BLOCK_THREADS];\n\n // Initialize problem\n Initialize(gen_mode, -1, h_in, h_flags, WARPS, LOGICAL_WARP_THREADS, valid_warp_threads, reduction_op, h_out, h_tail_out);\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n clock_t *d_elapsed = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(clock_t)));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * BLOCK_THREADS, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * BLOCK_THREADS));\n\n if (g_verbose)\n {\n printf(\"Data:\\n\");\n for (int i = 0; i < WARPS; ++i)\n DisplayResults(h_in + (i * LOGICAL_WARP_THREADS), valid_warp_threads);\n }\n\n // Run kernel\n printf(\"\\nGen-mode %d, %d warps, %d warp threads, %d valid lanes, %s (%d bytes) elements:\\n\",\n gen_mode,\n WARPS,\n LOGICAL_WARP_THREADS,\n valid_warp_threads,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n if (valid_warp_threads == LOGICAL_WARP_THREADS)\n {\n // Run full-warp kernel\n FullWarpReduceKernel<<<1, BLOCK_THREADS>>>(\n d_in,\n d_out,\n reduction_op,\n d_elapsed);\n }\n else\n {\n // Run partial-warp kernel\n PartialWarpReduceKernel<<<1, BLOCK_THREADS>>>(\n d_in,\n d_out,\n reduction_op,\n d_elapsed,\n valid_warp_threads);\n }\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tReduction results: \");\n int compare = CompareDeviceResults(h_out, d_out, BLOCK_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_flags) delete[] h_flags;\n if (h_out) delete[] h_out;\n if (h_tail_out) delete[] h_tail_out;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Test warp segmented reduction\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename ReductionOp>\nvoid TestSegmentedReduce(\n GenMode gen_mode,\n int flag_entropy,\n ReductionOp reduction_op)\n{\n const int BLOCK_THREADS = LOGICAL_WARP_THREADS * WARPS;\n\n // Allocate host arrays\n int compare;\n T *h_in = new T[BLOCK_THREADS];\n int *h_flags = new int[BLOCK_THREADS];\n T *h_head_out = new T[BLOCK_THREADS];\n T *h_tail_out = new T[BLOCK_THREADS];\n\n // Initialize problem\n Initialize(gen_mode, flag_entropy, h_in, h_flags, WARPS, LOGICAL_WARP_THREADS, LOGICAL_WARP_THREADS, reduction_op, h_head_out, h_tail_out);\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n int *d_flags = NULL;\n T *d_head_out = NULL;\n T *d_tail_out = NULL;\n clock_t *d_elapsed = NULL;\n\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_flags, sizeof(int) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_head_out, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_tail_out, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(clock_t)));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * BLOCK_THREADS, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemcpy(d_flags, h_flags, sizeof(int) * BLOCK_THREADS, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_head_out, 0, sizeof(T) * BLOCK_THREADS));\n CubDebugExit(cudaMemset(d_tail_out, 0, sizeof(T) * BLOCK_THREADS));\n\n if (g_verbose)\n {\n printf(\"Data:\\n\");\n for (int i = 0; i < WARPS; ++i)\n DisplayResults(h_in + (i * LOGICAL_WARP_THREADS), LOGICAL_WARP_THREADS);\n\n printf(\"\\nFlags:\\n\");\n for (int i = 0; i < WARPS; ++i)\n DisplayResults(h_flags + (i * LOGICAL_WARP_THREADS), LOGICAL_WARP_THREADS);\n }\n\n printf(\"\\nGen-mode %d, head flag entropy reduction %d, %d warps, %d warp threads, %s (%d bytes) elements:\\n\",\n gen_mode,\n flag_entropy,\n WARPS,\n LOGICAL_WARP_THREADS,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n // Run head-based kernel\n WarpHeadSegmentedReduceKernel<<<1, BLOCK_THREADS>>>(\n d_in,\n d_flags,\n d_head_out,\n reduction_op,\n d_elapsed);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tHead-based segmented reduction results: \");\n compare = CompareDeviceResults(h_head_out, d_head_out, BLOCK_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Run tail-based kernel\n WarpTailSegmentedReduceKernel<<<1, BLOCK_THREADS>>>(\n d_in,\n d_flags,\n d_tail_out,\n reduction_op,\n d_elapsed);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tTail-based segmented reduction results: \");\n compare = CompareDeviceResults(h_tail_out, d_tail_out, BLOCK_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_flags) delete[] h_flags;\n if (h_head_out) delete[] h_head_out;\n if (h_tail_out) delete[] h_tail_out;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_flags) CubDebugExit(g_allocator.DeviceFree(d_flags));\n if (d_head_out) CubDebugExit(g_allocator.DeviceFree(d_head_out));\n if (d_tail_out) CubDebugExit(g_allocator.DeviceFree(d_tail_out));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Run battery of tests for different full and partial tile sizes\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS,\n typename T,\n typename ReductionOp>\nvoid Test(\n GenMode gen_mode,\n ReductionOp reduction_op)\n{\n // Partial tiles\n for (\n int valid_warp_threads = 1;\n valid_warp_threads < LOGICAL_WARP_THREADS;\n valid_warp_threads += CUB_MAX(1, LOGICAL_WARP_THREADS / 5))\n {\n // Without wrapper (to test non-excepting PTX POD-op specializations)\n TestReduce(gen_mode, reduction_op, valid_warp_threads);\n\n // With wrapper to ensure no ops called on OOB lanes\n WrapperFunctor wrapped_op(reduction_op, valid_warp_threads);\n TestReduce(gen_mode, wrapped_op, valid_warp_threads);\n }\n\n // Full tile\n TestReduce(gen_mode, reduction_op, LOGICAL_WARP_THREADS);\n\n // Segmented reduction with different head flags\n for (int flag_entropy = 0; flag_entropy < 10; ++flag_entropy)\n {\n TestSegmentedReduce(gen_mode, flag_entropy, reduction_op);\n }\n}\n\n\n/**\n * Run battery of tests for different data types and reduce ops\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS>\nvoid Test(GenMode gen_mode)\n{\n // primitive\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n\n if (gen_mode != RANDOM)\n {\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n }\n\n // primitive (alternative reduce op)\n Test( gen_mode, Max());\n Test( gen_mode, Max());\n Test( gen_mode, Max());\n Test( gen_mode, Max());\n\n // vec-1\n Test( gen_mode, Sum());\n\n // vec-2\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n\n // vec-4\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n\n // complex\n Test( gen_mode, Sum());\n Test( gen_mode, Sum());\n}\n\n\n/**\n * Run battery of tests for different problem generation options\n */\ntemplate <\n int WARPS,\n int LOGICAL_WARP_THREADS>\nvoid Test()\n{\n Test(UNIFORM);\n Test(INTEGER_SEED);\n Test(RANDOM);\n}\n\n\n/**\n * Run battery of tests for different number of active warps\n */\ntemplate \nvoid Test()\n{\n Test<1, LOGICAL_WARP_THREADS>();\n\n // Only power-of-two subwarps can be tiled\n if ((LOGICAL_WARP_THREADS == 32) || PowerOfTwo::VALUE)\n Test<2, LOGICAL_WARP_THREADS>();\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#ifdef QUICK_TEST\n\n // Compile/run quick tests\n TestReduce<1, 32, int>(UNIFORM, Sum());\n\n TestReduce<1, 32, double>(UNIFORM, Sum());\n TestReduce<2, 16, TestBar>(UNIFORM, Sum());\n TestSegmentedReduce<1, 32, int>(UNIFORM, 1, Sum());\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test logical warp sizes\n Test<32>();\n Test<16>();\n Test<9>();\n Test<7>();\n }\n\n#endif\n\n return 0;\n}\n\n\n\n\n" }, { "path": "external/cub/test/test_warp_scan.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Test of WarpScan utilities\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n\n#include \n#include \n\n#include \"test_util.h\"\n\nusing namespace cub;\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\nbool g_verbose = false;\nint g_repeat = 0;\nCachingDeviceAllocator g_allocator(true);\n\n\n/**\n * Primitive variant to test\n */\nenum TestMode\n{\n BASIC,\n AGGREGATE,\n};\n\n\n\n/**\n * \\brief WrapperFunctor (for precluding test-specialized dispatch to *Sum variants)\n */\ntemplate\nstruct WrapperFunctor\n{\n OpT op;\n\n WrapperFunctor(OpT op) : op(op) {}\n\n template \n __host__ __device__ __forceinline__ T operator()(const T &a, const T &b) const\n {\n return op(a, b);\n }\n};\n\n//---------------------------------------------------------------------\n// Test kernels\n//---------------------------------------------------------------------\n\n/// Exclusive scan basic\ntemplate \n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n T &initial_value,\n ScanOpT &scan_op,\n T &aggregate,\n Int2Type test_mode,\n IsPrimitiveT is_primitive)\n{\n // Test basic warp scan\n warp_scan.ExclusiveScan(data, data, initial_value, scan_op);\n}\n\n/// Exclusive scan aggregate\ntemplate <\n typename WarpScanT,\n typename T,\n typename ScanOpT,\n typename IsPrimitiveT>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n T &initial_value,\n ScanOpT &scan_op,\n T &aggregate,\n Int2Type test_mode,\n IsPrimitiveT is_primitive)\n{\n // Test with cumulative aggregate\n warp_scan.ExclusiveScan(data, data, initial_value, scan_op, aggregate);\n}\n\n\n/// Exclusive sum basic\ntemplate <\n typename WarpScanT,\n typename T>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n T &initial_value,\n Sum &scan_op,\n T &aggregate,\n Int2Type test_mode,\n Int2Type is_primitive)\n{\n // Test basic warp scan\n warp_scan.ExclusiveSum(data, data);\n}\n\n\n/// Exclusive sum aggregate\ntemplate <\n typename WarpScanT,\n typename T>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n T &initial_value,\n Sum &scan_op,\n T &aggregate,\n Int2Type test_mode,\n Int2Type is_primitive)\n{\n // Test with cumulative aggregate\n warp_scan.ExclusiveSum(data, data, aggregate);\n}\n\n\n/// Inclusive scan basic\ntemplate <\n typename WarpScanT,\n typename T,\n typename ScanOpT,\n typename IsPrimitiveT>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n NullType &initial_value,\n ScanOpT &scan_op,\n T &aggregate,\n Int2Type test_mode,\n IsPrimitiveT is_primitive)\n{\n // Test basic warp scan\n warp_scan.InclusiveScan(data, data, scan_op);\n}\n\n/// Inclusive scan aggregate\ntemplate <\n typename WarpScanT,\n typename T,\n typename ScanOpT,\n typename IsPrimitiveT>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n NullType &initial_value,\n ScanOpT &scan_op,\n T &aggregate,\n Int2Type test_mode,\n IsPrimitiveT is_primitive)\n{\n // Test with cumulative aggregate\n warp_scan.InclusiveScan(data, data, scan_op, aggregate);\n}\n\n/// Inclusive sum basic\ntemplate <\n typename WarpScanT,\n typename T,\n typename InitialValueT>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n NullType &initial_value,\n Sum &scan_op,\n T &aggregate,\n Int2Type test_mode,\n Int2Type is_primitive)\n{\n // Test basic warp scan\n warp_scan.InclusiveSum(data, data);\n}\n\n/// Inclusive sum aggregate\ntemplate <\n typename WarpScanT,\n typename T,\n typename InitialValueT>\n__device__ __forceinline__ void DeviceTest(\n WarpScanT &warp_scan,\n T &data,\n NullType &initial_value,\n Sum &scan_op,\n T &aggregate,\n Int2Type test_mode,\n Int2Type is_primitive)\n{\n // Test with cumulative aggregate\n warp_scan.InclusiveSum(data, data, aggregate);\n}\n\n\n/**\n * WarpScan test kernel\n */\ntemplate <\n int LOGICAL_WARP_THREADS,\n TestMode TEST_MODE,\n typename T,\n typename ScanOpT,\n typename InitialValueT>\n__global__ void WarpScanKernel(\n T *d_in,\n T *d_out,\n T *d_aggregate,\n ScanOpT scan_op,\n InitialValueT initial_value,\n clock_t *d_elapsed)\n{\n // Cooperative warp-scan utility type (1 warp)\n typedef WarpScan WarpScanT;\n\n // Allocate temp storage in shared memory\n __shared__ typename WarpScanT::TempStorage temp_storage;\n\n // Per-thread tile data\n T data = d_in[threadIdx.x];\n\n // Start cycle timer\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t start = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n T aggregate;\n\n // Test scan\n WarpScanT warp_scan(temp_storage);\n DeviceTest(\n warp_scan,\n data,\n initial_value,\n scan_op,\n aggregate,\n Int2Type(),\n Int2Type::PRIMITIVE>());\n\n // Stop cycle timer\n __threadfence_block(); // workaround to prevent clock hoisting\n clock_t stop = clock();\n __threadfence_block(); // workaround to prevent clock hoisting\n\n // Store data\n d_out[threadIdx.x] = data;\n\n if (TEST_MODE != BASIC)\n {\n // Store aggregate\n d_aggregate[threadIdx.x] = aggregate;\n }\n\n // Store time\n if (threadIdx.x == 0)\n {\n *d_elapsed = (start > stop) ? start - stop : stop - start;\n }\n}\n\n\n//---------------------------------------------------------------------\n// Host utility subroutines\n//---------------------------------------------------------------------\n\n/**\n * Initialize exclusive-scan problem (and solution)\n */\ntemplate <\n typename T,\n typename ScanOpT>\nT Initialize(\n GenMode gen_mode,\n T *h_in,\n T *h_reference,\n int num_items,\n ScanOpT scan_op,\n T initial_value)\n{\n InitValue(gen_mode, h_in[0], 0);\n\n T block_aggregate = h_in[0];\n h_reference[0] = initial_value;\n T inclusive = scan_op(initial_value, h_in[0]);\n\n for (int i = 1; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n h_reference[i] = inclusive;\n inclusive = scan_op(inclusive, h_in[i]);\n block_aggregate = scan_op(block_aggregate, h_in[i]);\n }\n\n return block_aggregate;\n}\n\n\n/**\n * Initialize inclusive-scan problem (and solution)\n */\ntemplate <\n typename T,\n typename ScanOpT>\nT Initialize(\n GenMode gen_mode,\n T *h_in,\n T *h_reference,\n int num_items,\n ScanOpT scan_op,\n NullType)\n{\n InitValue(gen_mode, h_in[0], 0);\n\n T block_aggregate = h_in[0];\n T inclusive = h_in[0];\n h_reference[0] = inclusive;\n\n for (int i = 1; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n inclusive = scan_op(inclusive, h_in[i]);\n block_aggregate = scan_op(block_aggregate, h_in[i]);\n h_reference[i] = inclusive;\n }\n\n return block_aggregate;\n}\n\n\n/**\n * Test warp scan\n */\ntemplate <\n int LOGICAL_WARP_THREADS,\n TestMode TEST_MODE,\n typename T,\n typename ScanOpT,\n typename InitialValueT> // NullType implies inclusive-scan, otherwise inclusive scan\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n InitialValueT initial_value)\n{\n // Allocate host arrays\n T *h_in = new T[LOGICAL_WARP_THREADS];\n T *h_reference = new T[LOGICAL_WARP_THREADS];\n T *h_aggregate = new T[LOGICAL_WARP_THREADS];\n\n // Initialize problem\n T aggregate = Initialize(\n gen_mode,\n h_in,\n h_reference,\n LOGICAL_WARP_THREADS,\n scan_op,\n initial_value);\n\n if (g_verbose)\n {\n printf(\"Input: \\n\");\n DisplayResults(h_in, LOGICAL_WARP_THREADS);\n printf(\"\\n\");\n }\n\n for (int i = 0; i < LOGICAL_WARP_THREADS; ++i)\n {\n h_aggregate[i] = aggregate;\n }\n\n // Initialize/clear device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n T *d_aggregate = NULL;\n clock_t *d_elapsed = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * LOGICAL_WARP_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * (LOGICAL_WARP_THREADS + 1)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_aggregate, sizeof(T) * LOGICAL_WARP_THREADS));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_elapsed, sizeof(clock_t)));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * LOGICAL_WARP_THREADS, cudaMemcpyHostToDevice));\n CubDebugExit(cudaMemset(d_out, 0, sizeof(T) * (LOGICAL_WARP_THREADS + 1)));\n CubDebugExit(cudaMemset(d_aggregate, 0, sizeof(T) * LOGICAL_WARP_THREADS));\n\n // Run kernel\n printf(\"Test-mode %d (%s), gen-mode %d (%s), %s warpscan, %d warp threads, %s (%d bytes) elements:\\n\",\n TEST_MODE, typeid(TEST_MODE).name(),\n gen_mode, typeid(gen_mode).name(),\n (Equals::VALUE) ? \"Inclusive\" : \"Exclusive\",\n LOGICAL_WARP_THREADS,\n typeid(T).name(),\n (int) sizeof(T));\n fflush(stdout);\n\n // Run aggregate/prefix kernel\n WarpScanKernel<<<1, LOGICAL_WARP_THREADS>>>(\n d_in,\n d_out,\n d_aggregate,\n scan_op,\n initial_value,\n d_elapsed);\n\n printf(\"\\tElapsed clocks: \");\n DisplayDeviceResults(d_elapsed, 1);\n\n CubDebugExit(cudaPeekAtLastError());\n CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n printf(\"\\tScan results: \");\n int compare = CompareDeviceResults(h_reference, d_out, LOGICAL_WARP_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n\n // Copy out and display aggregate\n if (TEST_MODE == AGGREGATE)\n {\n printf(\"\\tScan aggregate: \");\n compare = CompareDeviceResults(h_aggregate, d_aggregate, LOGICAL_WARP_THREADS, g_verbose, g_verbose);\n printf(\"%s\\n\", compare ? \"FAIL\" : \"PASS\");\n AssertEquals(0, compare);\n }\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (h_reference) delete[] h_reference;\n if (h_aggregate) delete[] h_aggregate;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n if (d_aggregate) CubDebugExit(g_allocator.DeviceFree(d_aggregate));\n if (d_elapsed) CubDebugExit(g_allocator.DeviceFree(d_elapsed));\n}\n\n\n/**\n * Run battery of tests for different primitive variants\n */\ntemplate <\n int LOGICAL_WARP_THREADS,\n typename ScanOpT,\n typename T>\nvoid Test(\n GenMode gen_mode,\n ScanOpT scan_op,\n T initial_value)\n{\n // Exclusive\n Test(gen_mode, scan_op, T());\n Test(gen_mode, scan_op, T());\n\n // Exclusive (non-specialized, so we can use initial-value)\n Test(gen_mode, WrapperFunctor(scan_op), initial_value);\n Test(gen_mode, WrapperFunctor(scan_op), initial_value);\n\n // Inclusive\n Test(gen_mode, scan_op, NullType());\n Test(gen_mode, scan_op, NullType());\n}\n\n\n/**\n * Run battery of tests for different data types and scan ops\n */\ntemplate \nvoid Test(GenMode gen_mode)\n{\n // Get device ordinal\n int device_ordinal;\n CubDebugExit(cudaGetDevice(&device_ordinal));\n\n // Get ptx version\n int ptx_version;\n CubDebugExit(PtxVersion(ptx_version));\n\n // primitive\n Test(gen_mode, Sum(), (char) 99);\n Test(gen_mode, Sum(), (short) 99);\n Test(gen_mode, Sum(), (int) 99);\n Test(gen_mode, Sum(), (long) 99);\n Test(gen_mode, Sum(), (long long) 99);\n if (gen_mode != RANDOM) {\n // Only test numerically stable inputs\n Test(gen_mode, Sum(), (float) 99);\n if (ptx_version > 100)\n Test(gen_mode, Sum(), (double) 99);\n }\n\n // primitive (alternative scan op)\n Test(gen_mode, Max(), (unsigned char) 99);\n Test(gen_mode, Max(), (unsigned short) 99);\n Test(gen_mode, Max(), (unsigned int) 99);\n Test(gen_mode, Max(), (unsigned long long) 99);\n\n // vec-2\n Test(gen_mode, Sum(), make_uchar2(17, 21));\n Test(gen_mode, Sum(), make_ushort2(17, 21));\n Test(gen_mode, Sum(), make_uint2(17, 21));\n Test(gen_mode, Sum(), make_ulong2(17, 21));\n Test(gen_mode, Sum(), make_ulonglong2(17, 21));\n if (gen_mode != RANDOM) {\n // Only test numerically stable inputs\n Test(gen_mode, Sum(), make_float2(17, 21));\n if (ptx_version > 100)\n Test(gen_mode, Sum(), make_double2(17, 21));\n }\n\n // vec-4\n Test(gen_mode, Sum(), make_char4(17, 21, 32, 85));\n Test(gen_mode, Sum(), make_short4(17, 21, 32, 85));\n Test(gen_mode, Sum(), make_int4(17, 21, 32, 85));\n Test(gen_mode, Sum(), make_long4(17, 21, 32, 85));\n Test(gen_mode, Sum(), make_longlong4(17, 21, 32, 85));\n if (gen_mode != RANDOM) {\n // Only test numerically stable inputs\n Test(gen_mode, Sum(), make_float4(17, 21, 32, 85));\n if (ptx_version > 100)\n Test(gen_mode, Sum(), make_double4(17, 21, 32, 85));\n }\n\n // complex\n Test(gen_mode, Sum(), TestFoo::MakeTestFoo(17, 21, 32, 85));\n Test(gen_mode, Sum(), TestBar(17, 21));\n\n}\n\n\n/**\n * Run battery of tests for different problem generation options\n */\ntemplate \nvoid Test()\n{\n Test(UNIFORM);\n Test(INTEGER_SEED);\n Test(RANDOM);\n}\n\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n args.GetCmdLineArgument(\"repeat\", g_repeat);\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--repeat=]\"\n \"[--v] \"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#ifdef QUICK_TEST\n\n // Compile/run quick tests\n Test<32, AGGREGATE, int>(UNIFORM, Sum(), (int) 0);\n Test<32, AGGREGATE, float>(UNIFORM, Sum(), (float) 0);\n Test<32, AGGREGATE, long long>(UNIFORM, Sum(), (long long) 0);\n Test<32, AGGREGATE, double>(UNIFORM, Sum(), (double) 0);\n\n typedef KeyValuePair T;\n cub::Sum sum_op;\n Test<32, AGGREGATE, T>(UNIFORM, ReduceBySegmentOp(sum_op), T());\n\n#else\n\n // Compile/run thorough tests\n for (int i = 0; i <= g_repeat; ++i)\n {\n // Test logical warp sizes\n Test<32>();\n Test<16>();\n Test<9>();\n Test<7>();\n }\n\n#endif\n\n return 0;\n}\n\n\n\n\n" }, { "path": "external/cub/tune/.gitignore", "content": "/bin\n" }, { "path": "external/cub/tune/Makefile", "content": "#/******************************************************************************\n# * Copyright (c) 2011, Duane Merrill. All rights reserved.\n# * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n# * \n# * Redistribution and use in source and binary forms, with or without\n# * modification, are permitted provided that the following conditions are met:\n# *\t * Redistributions of source code must retain the above copyright\n# *\t notice, this list of conditions and the following disclaimer.\n# *\t * Redistributions in binary form must reproduce the above copyright\n# *\t notice, this list of conditions and the following disclaimer in the\n# *\t documentation and/or other materials provided with the distribution.\n# *\t * Neither the name of the NVIDIA CORPORATION nor the\n# *\t names of its contributors may be used to endorse or promote products\n# *\t derived from this software without specific prior written permission.\n# * \n# * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n# * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n# * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n# * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n# * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n# * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n# * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n# * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n# * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n# * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n# *\n#******************************************************************************/\n \n#-------------------------------------------------------------------------------\n# Build script for project\n#-------------------------------------------------------------------------------\n\nNVCC = \"$(shell which nvcc)\"\nNVCC_VERSION = $(strip $(shell nvcc --version | grep release | sed 's/.*release //' | sed 's/,.*//'))\n\n# detect OS\nOSUPPER = $(shell uname -s 2>/dev/null | tr [:lower:] [:upper:])\n\n#-------------------------------------------------------------------------------\n# Libs\n#-------------------------------------------------------------------------------\n\n\n#-------------------------------------------------------------------------------\n# Includes\n#-------------------------------------------------------------------------------\n\nINC = -I. -I.. -I../test\n\n#-------------------------------------------------------------------------------\n# Libs\n#-------------------------------------------------------------------------------\n\nLIBS += -lcudart \n\n#-------------------------------------------------------------------------------\n# Defines\n#-------------------------------------------------------------------------------\n\nDEFINES = \n\n#-------------------------------------------------------------------------------\n# SM Arch\n#-------------------------------------------------------------------------------\n\nifdef sm\n\tSM_ARCH = $(sm)\nelse \n SM_ARCH = 200\nendif\n\n# Only one arch per tuning binary\nifeq (350, $(findstring 350, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_35\n SM_ARCH = 350\nendif\nifeq (300, $(findstring 300, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_30\n SM_ARCH = 300\nendif\nifeq (200, $(findstring 200, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_20\n SM_ARCH = 200\nendif\nifeq (130, $(findstring 130, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_13\n SM_ARCH = 130\nendif\nifeq (110, $(findstring 110, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_11 \n SM_ARCH = 110\nendif\nifeq (100, $(findstring 100, $(SM_ARCH)))\n SM_TARGETS = -arch=sm_10 \n SM_ARCH = 100\nendif\n\n\n#-------------------------------------------------------------------------------\n# Compiler Flags\n#-------------------------------------------------------------------------------\n\nNVCCFLAGS = -Xptxas -v -Xcudafe -\\#\n\n# Help the compiler/linker work with huge numbers of kernels on Windows\nifeq (WIN_NT, $(findstring WIN_NT, $(OSUPPER)))\n\tNVCCFLAGS += -Xcompiler /bigobj -Xcompiler /Zm500\nendif\n\n# 32/64-bit (32-bit device pointers by default) \nifeq ($(force32), 1)\n\tCPU_ARCH = -m32\n\tCPU_ARCH_SUFFIX = i386\nelse\n\tCPU_ARCH = -m64\n\tCPU_ARCH_SUFFIX = x86_64\nendif\n\n# CUDA ABI enable/disable (enabled by default) \nifneq ($(abi), 0)\n\tABI_SUFFIX = abi\nelse \n\tNVCCFLAGS += -Xptxas -abi=no\n\tABI_SUFFIX = noabi\nendif\n\n# NVVM/Open64 middle-end compiler (nvvm by default)\nifeq ($(open64), 1)\n\tNVCCFLAGS += -open64\n\tPTX_SUFFIX = open64\nelse \n\tPTX_SUFFIX = nvvm\nendif\n\n# Verbose toolchain output from nvcc\nifeq ($(verbose), 1)\n\tNVCCFLAGS += -v\nendif\n\n# Keep intermediate compilation artifacts\nifeq ($(keep), 1)\n\tNVCCFLAGS += -keep\nendif\n\n# Data type size to compile a schmoo binary for\nifdef tunesize\n TUNE_SIZE = $(tunesize)\nelse \n\tTUNE_SIZE = 4\nendif\n\n\nSUFFIX = $(TUNE_SIZE)B_sm$(SM_ARCH)_$(PTX_SUFFIX)_$(NVCC_VERSION)_$(ABI_SUFFIX)_$(CPU_ARCH_SUFFIX)\n\n#-------------------------------------------------------------------------------\n# Dependency Lists\n#-------------------------------------------------------------------------------\n\nrwildcard=$(foreach d,$(wildcard $1*),$(call rwildcard,$d/,$2) $(filter $(subst *,%,$2),$d))\n\nDEPS =\t ./Makefile \\\n\t\t../test/test_util.h \\\n\t\t$(call rwildcard,../cub/,*.cuh)\n\n\n#-------------------------------------------------------------------------------\n# make default\n#-------------------------------------------------------------------------------\n\ndefault:\n\n\n#-------------------------------------------------------------------------------\n# make clean\n#-------------------------------------------------------------------------------\n\nclean :\n\trm -f bin/*$(CPU_ARCH_SUFFIX)* \n\trm -f *.i* *.cubin *.cu.c *.cudafe* *.fatbin.c *.ptx *.hash *.cu.cpp *.o\n\n\n\n#-------------------------------------------------------------------------------\n# make tune_device_reduce\n#-------------------------------------------------------------------------------\n\ntune_device_reduce: bin/tune_device_reduce_$(SUFFIX)\n\nbin/tune_device_reduce_$(SUFFIX) : tune_device_reduce.cu $(DEPS)\n\tmkdir -p bin\n\t$(NVCC) $(DEFINES) $(SM_TARGETS) -o bin/tune_device_reduce_$(SUFFIX) tune_device_reduce.cu $(NVCCFLAGS) $(CPU_ARCH) $(INC) $(LIBS) -O3 -DTUNE_ARCH=$(SM_ARCH) -DTUNE_SIZE=$(TUNE_SIZE)\n\n" }, { "path": "external/cub/tune/tune_device_reduce.cu", "content": "/******************************************************************************\n * Copyright (c) 2011, Duane Merrill. All rights reserved.\n * Copyright (c) 2011-2017, NVIDIA CORPORATION. All rights reserved.\n *\n * Redistribution and use in source and binary forms, with or without\n * modification, are permitted provided that the following conditions are met:\n * * Redistributions of source code must retain the above copyright\n * notice, this list of conditions and the following disclaimer.\n * * Redistributions in binary form must reproduce the above copyright\n * notice, this list of conditions and the following disclaimer in the\n * documentation and/or other materials provided with the distribution.\n * * Neither the name of the NVIDIA CORPORATION nor the\n * names of its contributors may be used to endorse or promote products\n * derived from this software without specific prior written permission.\n *\n * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND\n * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED\n * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE\n * DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE FOR ANY\n * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES\n * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;\n * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND\n * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT\n * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS\n * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n *\n ******************************************************************************/\n\n/******************************************************************************\n * Evaluates different tuning configurations of DeviceReduce.\n *\n * The best way to use this program:\n * (1) Find the best all-around single-block tune for a given arch.\n * For example, 100 samples [1 ..512], 100 timing iterations per config per sample:\n * ./bin/tune_device_reduce_sm200_nvvm_5.0_abi_i386 --i=100 --s=100 --n=512 --single --device=0\n * (2) Update the single tune in device_reduce.cuh\n * (3) Find the best all-around multi-block tune for a given arch.\n * For example, 100 samples [single-block tile-size .. 50,331,648], 100 timing iterations per config per sample:\n * ./bin/tune_device_reduce_sm200_nvvm_5.0_abi_i386 --i=100 --s=100 --device=0\n * (4) Update the multi-block tune in device_reduce.cuh\n *\n ******************************************************************************/\n\n// Ensure printing of CUDA runtime errors to console\n#define CUB_STDERR\n\n#include \n#include \n#include \n#include \n#include \"../test/test_util.h\"\n\nusing namespace cub;\nusing namespace std;\n\n\n//---------------------------------------------------------------------\n// Globals, constants and typedefs\n//---------------------------------------------------------------------\n\n#ifndef TUNE_ARCH\n#define TUNE_ARCH 100\n#endif\n\nint g_max_items = 48 * 1024 * 1024;\nint g_samples = 100;\nint g_timing_iterations = 2;\nbool g_verbose = false;\nbool g_single = false;\nbool g_verify = true;\nCachingDeviceAllocator g_allocator;\n\n\n//---------------------------------------------------------------------\n// Host utility subroutines\n//---------------------------------------------------------------------\n\n/**\n * Initialize problem\n */\ntemplate \nvoid Initialize(\n GenMode gen_mode,\n T *h_in,\n int num_items)\n{\n for (int i = 0; i < num_items; ++i)\n {\n InitValue(gen_mode, h_in[i], i);\n }\n}\n\n/**\n * Sequential reduction\n */\ntemplate \nT Reduce(\n T *h_in,\n ReductionOp reduction_op,\n int num_items)\n{\n T retval = h_in[0];\n for (int i = 1; i < num_items; ++i)\n retval = reduction_op(retval, h_in[i]);\n\n return retval;\n}\n\n\n\n//---------------------------------------------------------------------\n// Full tile test generation\n//---------------------------------------------------------------------\n\n\n\n/**\n * Wrapper structure for generating and running different tuning configurations\n */\ntemplate <\n typename T,\n typename OffsetT,\n typename ReductionOp>\nstruct Schmoo\n{\n //---------------------------------------------------------------------\n // Types\n //---------------------------------------------------------------------\n\n /// Pairing of kernel function pointer and corresponding dispatch params\n template \n struct DispatchTuple\n {\n KernelPtr kernel_ptr;\n DeviceReduce::KernelDispachParams params;\n\n float avg_throughput;\n float best_avg_throughput;\n OffsetT best_size;\n float hmean_speedup;\n\n\n DispatchTuple() :\n kernel_ptr(0),\n params(DeviceReduce::KernelDispachParams()),\n avg_throughput(0.0),\n best_avg_throughput(0.0),\n hmean_speedup(0.0),\n best_size(0)\n {}\n };\n\n /**\n * Comparison operator for DispatchTuple.avg_throughput\n */\n template \n static bool MinSpeedup(const Tuple &a, const Tuple &b)\n {\n float delta = a.hmean_speedup - b.hmean_speedup;\n\n return ((delta < 0.02) && (delta > -0.02)) ?\n (a.best_avg_throughput < b.best_avg_throughput) : // Negligible average performance differences: defer to best performance\n (a.hmean_speedup < b.hmean_speedup);\n }\n\n\n\n /// Multi-block reduction kernel type and dispatch tuple type\n typedef void (*MultiBlockDeviceReduceKernelPtr)(T*, T*, OffsetT, GridEvenShare, GridQueue, ReductionOp);\n typedef DispatchTuple MultiDispatchTuple;\n\n /// Single-block reduction kernel type and dispatch tuple type\n typedef void (*SingleBlockDeviceReduceKernelPtr)(T*, T*, OffsetT, ReductionOp);\n typedef DispatchTuple SingleDispatchTuple;\n\n\n //---------------------------------------------------------------------\n // Fields\n //---------------------------------------------------------------------\n\n vector multi_kernels; // List of generated multi-block kernels\n vector single_kernels; // List of generated single-block kernels\n\n\n //---------------------------------------------------------------------\n // Kernel enumeration methods\n //---------------------------------------------------------------------\n\n /**\n * Must have smem that fits in the SM\n * Must have vector load length that divides items per thread\n */\n template \n struct SmemSize\n {\n enum\n {\n BYTES = sizeof(typename BlockReduceTiles::TempStorage),\n IS_OK = ((BYTES < ArchProps::SMEM_BYTES) &&\n (TilesReducePolicy::ITEMS_PER_THREAD % TilesReducePolicy::VECTOR_LOAD_LENGTH == 0))\n };\n };\n\n\n /**\n * Specialization that allows kernel generation with the specified TilesReducePolicy\n */\n template <\n typename TilesReducePolicy,\n bool IsOk = SmemSize::IS_OK>\n struct Ok\n {\n /// Enumerate multi-block kernel and add to the list\n template \n static void GenerateMulti(\n KernelsVector &multi_kernels,\n int subscription_factor)\n {\n MultiDispatchTuple tuple;\n tuple.params.template Init(subscription_factor);\n tuple.kernel_ptr = ReducePrivatizedKernel;\n multi_kernels.push_back(tuple);\n }\n\n\n /// Enumerate single-block kernel and add to the list\n template \n static void GenerateSingle(KernelsVector &single_kernels)\n {\n SingleDispatchTuple tuple;\n tuple.params.template Init();\n tuple.kernel_ptr = ReduceSingleKernel;\n single_kernels.push_back(tuple);\n }\n };\n\n /**\n * Specialization that rejects kernel generation with the specified TilesReducePolicy\n */\n template \n struct Ok\n {\n template \n static void GenerateMulti(KernelsVector &multi_kernels, int subscription_factor) {}\n\n template \n static void GenerateSingle(KernelsVector &single_kernels) {}\n };\n\n\n /// Enumerate block-scheduling variations\n template <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int VECTOR_LOAD_LENGTH,\n BlockReduceAlgorithm BLOCK_ALGORITHM,\n CacheLoadModifier LOAD_MODIFIER>\n void Enumerate()\n {\n // Multi-block kernels\n Ok >::GenerateMulti(multi_kernels, 1);\n Ok >::GenerateMulti(multi_kernels, 2);\n Ok >::GenerateMulti(multi_kernels, 4);\n Ok >::GenerateMulti(multi_kernels, 8);\n#if TUNE_ARCH >= 200\n Ok >::GenerateMulti(multi_kernels, 1);\n#endif\n\n // Single-block kernels\n Ok >::GenerateSingle(single_kernels);\n }\n\n\n /// Enumerate load modifier variations\n template <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int VECTOR_LOAD_LENGTH,\n BlockReduceAlgorithm BLOCK_ALGORITHM>\n void Enumerate()\n {\n Enumerate();\n#if TUNE_ARCH >= 350\n Enumerate();\n#endif\n }\n\n\n /// Enumerate block algorithms\n template <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD,\n int VECTOR_LOAD_LENGTH>\n void Enumerate()\n {\n Enumerate();\n Enumerate();\n }\n\n\n /// Enumerate vectorization variations\n template <\n int BLOCK_THREADS,\n int ITEMS_PER_THREAD>\n void Enumerate()\n {\n Enumerate();\n Enumerate();\n Enumerate();\n }\n\n\n /// Enumerate thread-granularity variations\n template \n void Enumerate()\n {\n Enumerate();\n Enumerate();\n Enumerate();\n\n Enumerate();\n Enumerate();\n Enumerate();\n\n Enumerate();\n Enumerate();\n Enumerate();\n\n Enumerate();\n Enumerate();\n Enumerate();\n\n Enumerate();\n Enumerate();\n Enumerate();\n }\n\n\n /// Enumerate block size variations\n void Enumerate()\n {\n printf(\"\\nEnumerating kernels\\n\"); fflush(stdout);\n\n Enumerate<32>();\n Enumerate<64>();\n Enumerate<96>();\n Enumerate<128>();\n Enumerate<160>();\n Enumerate<192>();\n Enumerate<256>();\n Enumerate<512>();\n }\n\n\n //---------------------------------------------------------------------\n // Test methods\n //---------------------------------------------------------------------\n\n /**\n * Test a configuration\n */\n void TestConfiguration(\n MultiDispatchTuple &multi_dispatch,\n SingleDispatchTuple &single_dispatch,\n T* d_in,\n T* d_out,\n T* h_reference,\n OffsetT num_items,\n ReductionOp reduction_op)\n {\n // Clear output\n if (g_verify) CubDebugExit(cudaMemset(d_out, 0, sizeof(T)));\n\n // Allocate temporary storage\n void *d_temp_storage = NULL;\n size_t temp_storage_bytes = 0;\n CubDebugExit(DeviceReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n multi_dispatch.kernel_ptr,\n single_dispatch.kernel_ptr,\n FillAndResetDrainKernel,\n multi_dispatch.params,\n single_dispatch.params,\n d_in,\n d_out,\n num_items,\n reduction_op));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Warmup/correctness iteration\n CubDebugExit(DeviceReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n multi_dispatch.kernel_ptr,\n single_dispatch.kernel_ptr,\n FillAndResetDrainKernel,\n multi_dispatch.params,\n single_dispatch.params,\n d_in,\n d_out,\n num_items,\n reduction_op));\n\n if (g_verify) CubDebugExit(cudaDeviceSynchronize());\n\n // Copy out and display results\n int compare = (g_verify) ?\n CompareDeviceResults(h_reference, d_out, 1, true, false) :\n 0;\n\n // Performance\n GpuTimer gpu_timer;\n float elapsed_millis = 0.0;\n for (int i = 0; i < g_timing_iterations; i++)\n {\n gpu_timer.Start();\n\n CubDebugExit(DeviceReduce::Dispatch(\n d_temp_storage,\n temp_storage_bytes,\n multi_dispatch.kernel_ptr,\n single_dispatch.kernel_ptr,\n FillAndResetDrainKernel,\n multi_dispatch.params,\n single_dispatch.params,\n d_in,\n d_out,\n num_items,\n reduction_op));\n\n gpu_timer.Stop();\n elapsed_millis += gpu_timer.ElapsedMillis();\n }\n\n // Mooch\n CubDebugExit(cudaDeviceSynchronize());\n\n float avg_elapsed = elapsed_millis / g_timing_iterations;\n float avg_throughput = float(num_items) / avg_elapsed / 1000.0 / 1000.0;\n float avg_bandwidth = avg_throughput * sizeof(T);\n\n multi_dispatch.avg_throughput = CUB_MAX(avg_throughput, multi_dispatch.avg_throughput);\n if (avg_throughput > multi_dispatch.best_avg_throughput)\n {\n multi_dispatch.best_avg_throughput = avg_throughput;\n multi_dispatch.best_size = num_items;\n }\n\n single_dispatch.avg_throughput = CUB_MAX(avg_throughput, single_dispatch.avg_throughput);\n if (avg_throughput > single_dispatch.best_avg_throughput)\n {\n single_dispatch.best_avg_throughput = avg_throughput;\n single_dispatch.best_size = num_items;\n }\n\n if (g_verbose)\n {\n printf(\"\\t%.2f GB/s, multi_dispatch( \", avg_bandwidth);\n multi_dispatch.params.Print();\n printf(\" ), single_dispatch( \");\n single_dispatch.params.Print();\n printf(\" )\\n\");\n fflush(stdout);\n }\n\n AssertEquals(0, compare);\n\n // Cleanup temporaries\n if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n }\n\n\n /**\n * Evaluate multi-block configurations\n */\n void TestMulti(\n T* h_in,\n T* d_in,\n T* d_out,\n ReductionOp reduction_op)\n {\n // Simple single kernel tuple for use with multi kernel sweep\n typedef typename DeviceReduce::TunedPolicies::SinglePolicy SimpleSinglePolicy;\n SingleDispatchTuple simple_single_tuple;\n simple_single_tuple.params.template Init();\n simple_single_tuple.kernel_ptr = ReduceSingleKernel;\n\n double max_exponent = log2(double(g_max_items));\n double min_exponent = log2(double(simple_single_tuple.params.tile_size));\n unsigned int max_int = (unsigned int) -1;\n\n for (int sample = 0; sample < g_samples; ++sample)\n {\n printf(\"\\nMulti-block sample %d, \", sample);\n\n int num_items;\n if (sample == 0)\n {\n // First sample: use max items\n num_items = g_max_items;\n printf(\"num_items: %d\", num_items); fflush(stdout);\n }\n else\n {\n // Sample a problem size from [2^g_min_exponent, g_max_items]. First 2/3 of the samples are log-distributed, the other 1/3 are uniformly-distributed.\n unsigned int bits;\n RandomBits(bits);\n double scale = double(bits) / max_int;\n\n if (sample < g_samples / 2)\n {\n // log bias\n double exponent = ((max_exponent - min_exponent) * scale) + min_exponent;\n num_items = pow(2.0, exponent);\n num_items = CUB_MIN(num_items, g_max_items);\n printf(\"num_items: %d (2^%.2f)\", num_items, exponent); fflush(stdout);\n }\n else\n {\n // uniform bias\n num_items = CUB_MAX(pow(2.0, min_exponent), scale * g_max_items);\n num_items = CUB_MIN(num_items, g_max_items);\n printf(\"num_items: %d (%.2f * %d)\", num_items, scale, g_max_items); fflush(stdout);\n }\n }\n if (g_verbose)\n printf(\"\\n\");\n else\n printf(\", \");\n\n // Compute reference\n T h_reference = Reduce(h_in, reduction_op, num_items);\n\n // Run test on each multi-kernel configuration\n float best_avg_throughput = 0.0;\n for (int j = 0; j < multi_kernels.size(); ++j)\n {\n multi_kernels[j].avg_throughput = 0.0;\n\n TestConfiguration(multi_kernels[j], simple_single_tuple, d_in, d_out, &h_reference, num_items, reduction_op);\n\n best_avg_throughput = CUB_MAX(best_avg_throughput, multi_kernels[j].avg_throughput);\n }\n\n // Print best throughput for this problem size\n printf(\"Best: %.2fe9 items/s (%.2f GB/s)\\n\", best_avg_throughput, best_avg_throughput * sizeof(T));\n\n // Accumulate speedup (inverse for harmonic mean)\n for (int j = 0; j < multi_kernels.size(); ++j)\n multi_kernels[j].hmean_speedup += best_avg_throughput / multi_kernels[j].avg_throughput;\n }\n\n // Find max overall throughput and compute hmean speedups\n float overall_max_throughput = 0.0;\n for (int j = 0; j < multi_kernels.size(); ++j)\n {\n overall_max_throughput = CUB_MAX(overall_max_throughput, multi_kernels[j].best_avg_throughput);\n multi_kernels[j].hmean_speedup = float(g_samples) / multi_kernels[j].hmean_speedup;\n }\n\n // Sort by cumulative speedup\n sort(multi_kernels.begin(), multi_kernels.end(), MinSpeedup);\n\n // Print ranked multi configurations\n printf(\"\\nRanked multi_kernels:\\n\");\n for (int j = 0; j < multi_kernels.size(); ++j)\n {\n printf(\"\\t (%d) params( \", multi_kernels.size() - j);\n multi_kernels[j].params.Print();\n printf(\" ) hmean speedup: %.3f, best throughput %.2f @ %d elements (%.2f GB/s, %.2f%%)\\n\",\n multi_kernels[j].hmean_speedup,\n multi_kernels[j].best_avg_throughput,\n (int) multi_kernels[j].best_size,\n multi_kernels[j].best_avg_throughput * sizeof(T),\n multi_kernels[j].best_avg_throughput / overall_max_throughput);\n }\n\n printf(\"\\nMax multi-block throughput %.2f (%.2f GB/s)\\n\", overall_max_throughput, overall_max_throughput * sizeof(T));\n }\n\n\n /**\n * Evaluate single-block configurations\n */\n void TestSingle(\n T* h_in,\n T* d_in,\n T* d_out,\n ReductionOp reduction_op)\n {\n // Construct a NULL-ptr multi-kernel tuple that forces a single-kernel pass\n MultiDispatchTuple multi_tuple;\n\n double max_exponent = log2(double(g_max_items));\n unsigned int max_int = (unsigned int) -1;\n\n for (int sample = 0; sample < g_samples; ++sample)\n {\n printf(\"\\nSingle-block sample %d, \", sample);\n\n int num_items;\n if (sample == 0)\n {\n // First sample: use max items\n num_items = g_max_items;\n printf(\"num_items: %d\", num_items); fflush(stdout);\n }\n else\n {\n // Sample a problem size from [2, g_max_items], log-distributed\n unsigned int bits;\n RandomBits(bits);\n double scale = double(bits) / max_int;\n double exponent = ((max_exponent - 1) * scale) + 1;\n num_items = pow(2.0, exponent);\n printf(\"num_items: %d (2^%.2f)\", num_items, exponent); fflush(stdout);\n }\n\n if (g_verbose)\n printf(\"\\n\");\n else\n printf(\", \");\n\n // Compute reference\n T h_reference = Reduce(h_in, reduction_op, num_items);\n\n // Run test on each single-kernel configuration (pick first multi-config to use, which shouldn't be\n float best_avg_throughput = 0.0;\n for (int j = 0; j < single_kernels.size(); ++j)\n {\n single_kernels[j].avg_throughput = 0.0;\n\n TestConfiguration(multi_tuple, single_kernels[j], d_in, d_out, &h_reference, num_items, reduction_op);\n\n best_avg_throughput = CUB_MAX(best_avg_throughput, single_kernels[j].avg_throughput);\n }\n\n // Print best throughput for this problem size\n printf(\"Best: %.2fe9 items/s (%.2f GB/s)\\n\", best_avg_throughput, best_avg_throughput * sizeof(T));\n\n // Accumulate speedup (inverse for harmonic mean)\n for (int j = 0; j < single_kernels.size(); ++j)\n single_kernels[j].hmean_speedup += best_avg_throughput / single_kernels[j].avg_throughput;\n }\n\n // Find max overall throughput and compute hmean speedups\n float overall_max_throughput = 0.0;\n for (int j = 0; j < single_kernels.size(); ++j)\n {\n overall_max_throughput = CUB_MAX(overall_max_throughput, single_kernels[j].best_avg_throughput);\n single_kernels[j].hmean_speedup = float(g_samples) / single_kernels[j].hmean_speedup;\n }\n\n // Sort by cumulative speedup\n sort(single_kernels.begin(), single_kernels.end(), MinSpeedup);\n\n // Print ranked single configurations\n printf(\"\\nRanked single_kernels:\\n\");\n for (int j = 0; j < single_kernels.size(); ++j)\n {\n printf(\"\\t (%d) params( \", single_kernels.size() - j);\n single_kernels[j].params.Print();\n printf(\" ) hmean speedup: %.3f, best throughput %.2f @ %d elements (%.2f GB/s, %.2f%%)\\n\",\n single_kernels[j].hmean_speedup,\n single_kernels[j].best_avg_throughput,\n (int) single_kernels[j].best_size,\n single_kernels[j].best_avg_throughput * sizeof(T),\n single_kernels[j].best_avg_throughput / overall_max_throughput);\n }\n\n printf(\"\\nMax single-block throughput %.2f (%.2f GB/s)\\n\", overall_max_throughput, overall_max_throughput * sizeof(T));\n }\n\n};\n\n\n\n//---------------------------------------------------------------------\n// Main\n//---------------------------------------------------------------------\n\n/**\n * Main\n */\nint main(int argc, char** argv)\n{\n // Initialize command line\n CommandLineArgs args(argc, argv);\n args.GetCmdLineArgument(\"n\", g_max_items);\n args.GetCmdLineArgument(\"s\", g_samples);\n args.GetCmdLineArgument(\"i\", g_timing_iterations);\n g_verbose = args.CheckCmdLineFlag(\"v\");\n g_single = args.CheckCmdLineFlag(\"single\");\n g_verify = !args.CheckCmdLineFlag(\"noverify\");\n\n // Print usage\n if (args.CheckCmdLineFlag(\"help\"))\n {\n printf(\"%s \"\n \"[--device=] \"\n \"[--n=]\"\n \"[--s=]\"\n \"[--i=]\"\n \"[--single]\"\n \"[--v]\"\n \"[--noverify]\"\n \"\\n\", argv[0]);\n exit(0);\n }\n\n // Initialize device\n CubDebugExit(args.DeviceInit());\n\n#if (TUNE_SIZE == 1)\n typedef unsigned char T;\n#elif (TUNE_SIZE == 2)\n typedef unsigned short T;\n#elif (TUNE_SIZE == 4)\n typedef unsigned int T;\n#elif (TUNE_SIZE == 8)\n typedef unsigned long long T;\n#else\n // Default\n typedef unsigned int T;\n#endif\n\n typedef unsigned int OffsetT;\n Sum reduction_op;\n\n // Enumerate kernels\n Schmoo schmoo;\n schmoo.Enumerate();\n\n // Allocate host arrays\n T *h_in = new T[g_max_items];\n\n // Initialize problem\n Initialize(UNIFORM, h_in, g_max_items);\n\n // Initialize device arrays\n T *d_in = NULL;\n T *d_out = NULL;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(T) * g_max_items));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(T) * 1));\n CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(T) * g_max_items, cudaMemcpyHostToDevice));\n\n // Test kernels\n if (g_single)\n schmoo.TestSingle(h_in, d_in, d_out, reduction_op);\n else\n schmoo.TestMulti(h_in, d_in, d_out, reduction_op);\n\n // Cleanup\n if (h_in) delete[] h_in;\n if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));\n if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));\n\n return 0;\n}\n\n\n\n" }, { "path": "src/bitonicTopK.cuh", "content": "#pragma once\n\n#include \n#include \n#include \n\n#include \"sharedmem.cuh\"\n\nusing namespace std;\nusing namespace cub;\n\n#define NUM_ELEM_PT 16\n#define NUM_ELEM_BITSHIFT 4\n\n//#define K 32\n//#define KLog2 5\n\n#define ORDERV(x,a,b) { bool swap = reverse ^ (x[a]\n__forceinline__\n__device__ T get(T* sdata, int i) {\n return sdata[i + (i>>5)];\n}\n\n#define set(a,b,c) { int tempIndex = b; a[tempIndex + (tempIndex >> 5)] = c; }\n\n#define NUM_GROUPS (NUM_ELEM_PT/2)\n#define NUM_GROUPS_BITSHIFT (NUM_ELEM_BITSHIFT-1)\n\n#define RUN_64(X) { \\\n inc >>= 5; \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 6) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define RUN_32(X) { \\\n inc >>= 4; \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 5) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define RUN_16(X) { \\\n inc >>= 3; \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 4) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define RUN_8(X) { \\\n inc >>= 2; \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 3) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define RUN_4(X) { \\\n inc >>= 1; \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 2) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define RUN_2(X) { \\\n low = t & (inc - 1); \\\n tCur = ((t - low) << 1) + low; \\\n reverse = ((dir & tCur) == 0); \\\n for (int j=0; j>= 1; \\\n}\n\n#define REDUCE(X) { \\\n tCur = ((t >> klog2) << (klog2 + 1)) + (t & (k-1)); \\\n for(int j=0; j\n__global__ void Bitonic_TopKLocalSortInPlace(T* __restrict__ in, T* __restrict__ out,\n const int k, const int klog2)\n{\n/* const int k = K;*/\n /*const int klog2 = KLog2;*/\n\n // Shared mem size is determined by the host app at run time.\n // For n elements, we have n * 33/32 shared memory.\n // We use this to break bank conflicts.\n SharedMemory smem;\n T* sdata = smem.getPointer();\n\n const int t = threadIdx.x; // index in workgroup\n const int wg = blockDim.x; // workgroup size = block size, power of 2\n const int gid = blockIdx.x;\n\n int length = min(NUM_GROUPS, k >> 1);\n int inc = length;\n inc >>= NUM_GROUPS_BITSHIFT;\n int low = t & (inc - 1);\n int dir = length << 1;\n bool reverse;\n\n T x[NUM_ELEM_PT];\n\n // Move IN, OUT to block start\n in += NUM_ELEM_PT * gid * wg;\n\n int tCur = t << NUM_ELEM_BITSHIFT;\n for (int i=0; i> 1) + 1)&1;\n B2V(x,i);\n }\n if (k > 2) {\n#if NUM_ELEM_PT > 4\n for (int i=0; i> 2) + 1)&1;\n B4V(x,i);\n }\n if (k > 4) {\n#if NUM_ELEM_PT > 8\n for (int i=0; i> 3) + 1)&1;\n B8V(x,i);\n }\n if (k > 8) {\n#if NUM_ELEM_PT > 16\n for (int i=0; i> 4) + 1)&1;\n B16V(x,i);\n }\n if (k > 16) {\n#if NUM_ELEM_PT > 32\n for (int i=0; i> 5) + 1)&1;\n B32V(x,i);\n }\n if (k > 32) {\n reverse = ((dir & tCur) == 0); B64V(x,0);\n }\n#else\n reverse = ((dir & tCur) == 0); B32V(x,0);\n#endif\n }\n#else\n reverse = ((dir & tCur) == 0); B16V(x,0);\n#endif\n }\n#else\n reverse = ((dir & tCur) == 0); B8V(x,0);\n#endif\n }\n#else\n reverse = ((dir & tCur) == 0); B4V(x,0);\n#endif\n }\n\n for (int i=0; i>= (NUM_ELEM_BITSHIFT - 0);\n\n if (mod & 1)\n {\n RUN_2(1)\n __syncthreads();\n }\n if (mod & 2)\n {\n RUN_4(1)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 4)\n {\n RUN_8(1)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 16\n if (mod & 8)\n {\n RUN_16(1)\n __syncthreads();\n }\n while (inc > 8)\n {\n RUN_32(1)\n __syncthreads();\n }\n#else\n while (inc > 4)\n {\n RUN_16(1)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 2)\n {\n RUN_8(1)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8\n }\n\n // Step 2: Reduce the size by factor 2 by pairwise comparing adjacent sequences.\n REDUCE(1)\n __syncthreads();\n // End of Step 2;\n\n // Step 3: Construct sorted sequence of length k from bitonic sequence of length k.\n // We now have n/2 elements.\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(1) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 0);\n\n if (mod & 1)\n {\n RUN_2(2)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 4\n if (mod & 2)\n {\n RUN_4(2)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 4)\n {\n RUN_8(2)\n __syncthreads();\n }\n while (inc > 4)\n {\n if (t < (wg >> 1)) {\n RUN_16(1)\n } else {\n inc >>= 4;\n }\n __syncthreads();\n }\n#else\n while (inc > 2)\n {\n RUN_8(2)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 1)\n {\n RUN_4(2)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8\n\n // Step 4: Reduce size again by 2.\n REDUCE(2)\n __syncthreads();\n // End of Step 1;\n\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(2) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 1);\n\n#if NUM_ELEM_PT > 4\n if (mod & 1)\n {\n RUN_2(4)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 2)\n {\n RUN_4(4)\n __syncthreads();\n }\n while (inc > 2)\n {\n if (t < (wg >> 1)) {\n RUN_8(2)\n } else {\n inc >>= 3;\n }\n __syncthreads();\n }\n#else\n while (inc > 1)\n {\n RUN_4(4)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 0)\n {\n RUN_2(4)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8 while (inc > 0)\n\n // Step 4: Reduce size again by 2.\n REDUCE(4)\n __syncthreads();\n // End of Step 1;\n\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(4) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 2);\n\n if (mod & 1)\n {\n RUN_2(8)\n __syncthreads();\n }\n while (inc > 0)\n {\n if (t < (wg >> 1)) {\n RUN_4(4)\n } else {\n inc >>= 2;\n }\n __syncthreads();\n }\n\n out += (NUM_ELEM_PT/16) * gid * wg;\n tCur = ((t >> klog2) << (klog2+1)) + (t&(k-1));\n for (int j=0; j> klog2) << (klog2+1)) + (t&(k-1));\n //for (int j=0; j\n__global__ void Bitonic_TopKReduce(T* __restrict__ in, T* __restrict__ out,\n const int k, const int klog2)\n{\n/* const int k = K;*/\n /*const int klog2 = KLog2;*/\n\n // Shared mem size is determined by the host app at run time.\n // For n elements, we have n * 33/32 shared memory.\n // We use this to break bank conflicts.\n SharedMemory smem;\n T* sdata = smem.getPointer();\n\n const int t = threadIdx.x; // index in workgroup\n const int wg = blockDim.x; // workgroup size = block size, power of 2\n const int gid = blockIdx.x;\n\n int length = min(NUM_GROUPS, k >> 1);\n int inc = length;\n inc >>= NUM_GROUPS_BITSHIFT;\n int low = t & (inc - 1);\n int dir = length << 1;\n bool reverse;\n\n T x[NUM_ELEM_PT];\n\n // Move IN, OUT to block start\n in += NUM_ELEM_PT * gid * wg;\n\n int tCur = t << NUM_ELEM_BITSHIFT;\n for (int i=0; i> 1);\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(1) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 0);\n\n if (mod & 1)\n {\n RUN_2(1)\n __syncthreads();\n }\n if (mod & 2)\n {\n RUN_4(1)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 4)\n {\n RUN_8(1)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 16\n if (mod & 8)\n {\n RUN_16(1)\n __syncthreads();\n }\n while (inc > 8)\n {\n RUN_32(1)\n __syncthreads();\n }\n#else\n while (inc > 4)\n {\n RUN_16(1)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 2)\n {\n RUN_8(1)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8\n\n // Step 2: Reduce the size by factor 2 by pairwise comparing adjacent sequences.\n REDUCE(1)\n __syncthreads();\n // End of Step 2;\n\n // Step 3: Construct sorted sequence of length k from bitonic sequence of length k.\n // We now have n/2 elements.\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(1) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 0);\n\n if (mod & 1)\n {\n RUN_2(2)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 4\n if (mod & 2)\n {\n RUN_4(2)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 4)\n {\n RUN_8(2)\n __syncthreads();\n }\n while (inc > 4)\n {\n if (t < (wg >> 1)) {\n RUN_16(1)\n } else {\n inc >>= 4;\n }\n __syncthreads();\n }\n#else\n while (inc > 2)\n {\n RUN_8(2)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 1)\n {\n RUN_4(2)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8\n\n // Step 4: Reduce size again by 2.\n REDUCE(2)\n __syncthreads();\n // End of Step 1;\n\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(2) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 1);\n\n#if NUM_ELEM_PT > 4\n if (mod & 1)\n {\n RUN_2(4)\n __syncthreads();\n }\n#if NUM_ELEM_PT > 8\n if (mod & 2)\n {\n RUN_4(4)\n __syncthreads();\n }\n while (inc > 2)\n {\n if (t < (wg >> 1)) {\n RUN_8(2)\n } else {\n inc >>= 3;\n }\n __syncthreads();\n }\n#else\n while (inc > 1)\n {\n RUN_4(4)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 16\n#else\n while (inc > 0)\n {\n RUN_2(4)\n __syncthreads();\n }\n#endif // NUM_ELEM_PT > 8 while (inc > 0)\n\n // Step 4: Reduce size again by 2.\n REDUCE(4)\n __syncthreads();\n // End of Step 1;\n\n length = k >> 1;\n dir = length << 1;\n // Loop on comparison distance (between keys)\n inc = length;\n mod = inc;\n mask = ~(NUM_ELEM_PT/(4) - 1);\n while ((mod & mask) != 0) mod >>= (NUM_ELEM_BITSHIFT - 2);\n\n if (mod & 1)\n {\n RUN_2(8)\n __syncthreads();\n }\n while (inc > 0)\n {\n if (t < (wg >> 1)) {\n RUN_4(4)\n } else {\n inc >>= 2;\n }\n __syncthreads();\n }\n\n out += (NUM_ELEM_PT/16) * gid * wg;\n tCur = ((t >> klog2) << (klog2+1)) + (t&(k-1));\n for (int j=0; j> 1;\n value |= value >> 2;\n value |= value >> 4;\n value |= value >> 8;\n value |= value >> 16;\n return tab32[(uint)(value*0x07C4ACDD) >> 27];\n}\n\ntemplate\ncudaError_t bitonicTopK(KeyT *d_keys_in, unsigned int num_items, unsigned int k, KeyT *d_keys_out,\n CachingDeviceAllocator& g_allocator) {\n if (k < 16) k = 16;\n\n int klog2 = log2_32(k);\n\n DoubleBuffer d_keys;\n d_keys.d_buffers[0] = d_keys_in;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(KeyT) * num_items));\n\n int current = 0;\n int numThreads = num_items;\n\n int wg_size = max(64,k);\n\n numThreads >>= 1; // Each thread processes 2 elements.\n numThreads >>= NUM_GROUPS_BITSHIFT;\n Bitonic_TopKLocalSortInPlace<<>>(d_keys.Current(), d_keys.Alternate(), k, klog2);\n current = 1-current;\n\n // Toggle the buffer index in the double buffer\n d_keys.selector = d_keys.selector ^ 1;\n\n numThreads >>= (1 + NUM_GROUPS_BITSHIFT);\n\n while (numThreads >= wg_size)\n {\n Bitonic_TopKReduce<<>>(d_keys.Current(), d_keys.Alternate(), k, klog2);\n\n // Toggle the buffer index in the double buffer\n d_keys.selector = d_keys.selector ^ 1;\n\n numThreads >>= (1 + NUM_GROUPS_BITSHIFT);\n }\n\n KeyT* res_vec = (KeyT*) malloc(sizeof(KeyT) * 2 * numThreads * NUM_GROUPS);\n cudaMemcpy(res_vec, d_keys.Current(), 2 * numThreads * NUM_GROUPS * sizeof(KeyT), cudaMemcpyDeviceToHost);\n std::sort(res_vec, res_vec + 2*numThreads*NUM_GROUPS, std::greater());\n cudaMemcpy(d_keys_out, res_vec, k * sizeof(KeyT), cudaMemcpyHostToDevice);\n\n if (d_keys.d_buffers[1])\n CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[1]));\n\n return cudaSuccess;\n}\n\n\n" }, { "path": "src/radixSelectTopK.cuh", "content": "#include \n#include \n#include \n\nusing namespace cub;\nusing namespace std;\n\n/**\n * Computes the histogram over the digit values of an array of keys that MUST have a length of an integer multiple of (KPT * blockDim.x).\n * The padding to the integer multiple can be done by adding 0's at the end and subtracting the number of padded 0's from the final result's 0 bin.\n * The 2^NUM_BITS possible counts (0..2^NUM_BITSNUM_BITS-1) will be placed in global_histo.\n * @param keys [IN] The keys for which to compute the histogram\n * @param digit [IN]\n * @param global_histo [OUT] The array of element counts, MUST be 256 in size.\n * @param per_block_histo [OUT]\n */\ntemplate<\n typename KeyT, // Data type of the keys within device memory. Data will be twiddled (if necessary) to unsigned type\n typename IndexT, // Data type used for key's offsets and counters (limits number of supported keys, uint = 2^32)\n int NUM_BITS, // Number of bits being sorted at a time\n int KPT, // Number of keys per thread\n int TPB, // Number of threads per block\n int PRE_SORT_RUNS_LENGTH // For values greater than 1, this causes to sort a thread's keys by runs of a given length to improve run-length encoded updates to shared memory.\n>\n__global__ void rdxsrt_histogram(KeyT *__restrict__ keys, const uint digit, IndexT *global_histo)\n{\n /*** TYPEDEFs***/\n typedef Traits KeyTraits;\n typedef typename KeyTraits::UnsignedBits UnsignedBits;\n /*typedef LoadUnit KeyLoader;*/\n\n /*** DECLARATIONS ***/\n UnsignedBits tloc_keys[KPT];\n uint tloc_masked[KPT];\n __shared__ uint shared_bins[0x01<(keys, tloc_keys);\n #pragma unroll\n for (int i=0; i(keys)[block_offset + threadIdx.x + blockDim.x * i];\n }\n\n#if true || USE_RLE_HISTO\n // Mask\n #pragma unroll\n for (int i=0; i>((sizeof(KeyT)*8)-(NUM_BITS*(digit+1))))&((0x01<1){\n SortingNetwork::sort_runs(tloc_masked);\n }\n#endif\n\n /*** COMPUTE HISTO ***/\n uint rle = 1;\n #pragma unroll\n for(int i=1; i>((sizeof(KeyT)*8)-(NUM_BITS*(digit+1))))&((0x01<\n__global__ void rdxsrt_histogram_with_guards(KeyT *__restrict__ keys, const uint digit, IndexT *global_histo, const IndexT total_keys, const int block_index_offset)\n{\n /*** TYPEDEFs***/\n typedef Traits KeyTraits;\n typedef typename KeyTraits::UnsignedBits UnsignedBits;\n /*typedef LoadUnit KeyLoader;*/\n\n /*** DECLARATIONS ***/\n UnsignedBits tloc_keys[KPT];\n uint tloc_masked[KPT];\n __shared__ uint shared_bins[(0x01<(keys, tloc_keys, block_max_num_keys);\n #pragma unroll\n for (int i=0; i(keys)[block_offset + threadIdx.x + blockDim.x * i];\n }\n }\n\n #pragma unroll\n for(int i=0; i>((sizeof(KeyT)*8)-(NUM_BITS*(digit+1))))&((0x01< highest digit, 3 => lowest digit for 32-bit)\n * @param digit_val [IN] Digit value.\n * @param num_items [IN] Number of entries.\n * @param d_keys_buffer [OUT] Entries with x[digit] > digit_val.\n * @param d_keys_out [OUT] Entries with x[digit] > digit_val.\n * @param d_index_buffer [OUT] Index into d_keys_buffer.\n * @param d_index_out [OUT] Index into d_keys_out.\n */\ntemplate<\n typename KeyT, // Data type of the keys within device memory. Data will be twiddled (if necessary) to unsigned type\n typename IndexT, // Data type used for key's offsets and counters (limits number of supported keys, uint = 2^32)\n int NUM_BITS, // Number of bits being sorted at a time\n int KPT, // Number of keys per thread\n int TPB // Number of threads per block\n>\n__global__ void select_kth_bucket(KeyT* d_keys_in, const uint digit, const uint digit_val, uint num_items,\n KeyT* d_keys_buffer, KeyT* d_keys_out, uint* d_index_buffer, uint* d_index_out)\n{\n typedef Traits KeyTraits;\n typedef typename KeyTraits::UnsignedBits UnsignedBits;\n\n // Specialize BlockLoad for a 1D block of TPB threads owning KPT integer items each\n typedef cub::BlockLoad BlockLoadT;\n\n // Specialize BlockScan type for our thread block\n typedef BlockScan BlockScanT;\n\n const int tile_size = TPB * KPT;\n int tile_idx = blockIdx.x; // Current tile index\n int tile_offset = tile_idx * tile_size;\n\n // Allocate shared memory for BlockLoad\n __shared__ union TempStorage\n {\n typename BlockLoadT::TempStorage load_items;\n typename BlockScanT::TempStorage scan;\n int offset[1];\n UnsignedBits raw_exchange[2 * TPB * KPT];\n } temp_storage;\n\n // Load a segment of consecutive items that are blocked across threads\n UnsignedBits key_entries[KPT];\n /*float payload_entries[KPT];*/\n int selection_flags[KPT];\n int selection_indices[KPT];\n\n int num_tiles = (num_items + tile_size - 1) / tile_size;\n int num_tile_items = tile_size;\n bool is_last_tile = false;\n if (tile_idx == num_tiles - 1) {\n num_tile_items = num_items - tile_offset;\n is_last_tile = true;\n }\n\n // Load keys\n if (is_last_tile)\n BlockLoadT(temp_storage.load_items).Load(reinterpret_cast(d_keys_in) + tile_offset, key_entries, num_tile_items);\n else\n BlockLoadT(temp_storage.load_items).Load(reinterpret_cast(d_keys_in) + tile_offset, key_entries);\n\n\n#if 0\n if (is_last_tile)\n BlockLoadT(temp_storage.load_items).Load(payload + tile_offset, payload_entries, num_tile_items);\n else\n BlockLoadT(temp_storage.load_items).Load(payload + tile_offset, payload_entries);\n#endif\n\n __syncthreads();\n\n /*** Step 1: Find keys with digit value to selected digit value ***/\n #pragma unroll\n for (int ITEM = 0; ITEM < KPT; ++ITEM)\n {\n // Out-of-bounds items are selection_flags\n selection_flags[ITEM] = 0;\n\n if (!is_last_tile || (int(threadIdx.x * KPT) + ITEM < num_tile_items)) {\n UnsignedBits key = KeyTraits::TwiddleIn(key_entries[ITEM]);\n uint masked_key = (key>>((sizeof(KeyT)*8)-(NUM_BITS*(digit+1))))&((0x01< digit_val);\n }\n }\n\n __syncthreads();\n\n // Compute exclusive prefix sum\n int num_selected;\n BlockScanT(temp_storage.scan).ExclusiveSum(selection_flags, selection_indices, num_selected);\n\n __syncthreads();\n\n if (num_selected > 0) {\n int index_out;\n if (threadIdx.x == 0) {\n // Find index into keys_out array\n index_out = atomicAdd(d_index_out, num_selected);\n temp_storage.offset[0] = index_out;\n }\n\n __syncthreads();\n\n index_out = temp_storage.offset[0];\n\n __syncthreads();\n\n // Compact and scatter items\n #pragma unroll\n for (int ITEM = 0; ITEM < KPT; ++ITEM)\n {\n int local_scatter_offset = selection_indices[ITEM];\n if (selection_flags[ITEM])\n {\n temp_storage.raw_exchange[local_scatter_offset] = key_entries[ITEM];\n /*temp_storage.raw_exchange[tile_size + local_scatter_offset] = payload_entries[ITEM];*/\n }\n }\n\n __syncthreads();\n\n // Write out matched entries to output array\n for (int item = threadIdx.x; item < num_selected; item += TPB)\n {\n reinterpret_cast(d_keys_out)[index_out + item] = temp_storage.raw_exchange[item];\n }\n\n __syncthreads();\n\n#if 0\n for (int item = threadIdx.x; item < num_selected; item += TPB)\n {\n payload_out[num_selections_prefix + item] = temp_storage.raw_exchange[tile_size + item];\n }\n#endif\n }\n\n /*** Step 2: Find entries that have digit equal to digit value ***/\n #pragma unroll\n for (int ITEM = 0; ITEM < KPT; ++ITEM)\n {\n // Out-of-bounds items are selection_flags\n selection_flags[ITEM] = 0;\n\n if (!is_last_tile || (int(threadIdx.x * KPT) + ITEM < num_tile_items)) {\n UnsignedBits key = KeyTraits::TwiddleIn(key_entries[ITEM]);\n uint masked_key = (key>>((sizeof(KeyT)*8)-(NUM_BITS*(digit+1))))&((0x01< 0) {\n int index_buffer;\n if (threadIdx.x == 0) {\n index_buffer = atomicAdd(d_index_buffer, num_selected);\n temp_storage.offset[0] = index_buffer;\n }\n\n __syncthreads();\n\n index_buffer = temp_storage.offset[0];\n\n __syncthreads();\n\n // Compact and scatter items\n #pragma unroll\n for (int ITEM = 0; ITEM < KPT; ++ITEM)\n {\n int local_scatter_offset = selection_indices[ITEM];\n if (selection_flags[ITEM])\n {\n temp_storage.raw_exchange[local_scatter_offset] = key_entries[ITEM];\n /*temp_storage.raw_exchange[tile_size + local_scatter_offset] = payload_entries[ITEM];*/\n }\n }\n\n __syncthreads();\n\n // Write out output entries\n for (int item = threadIdx.x; item < num_selected; item += TPB)\n {\n reinterpret_cast(d_keys_buffer)[index_buffer + item] = temp_storage.raw_exchange[item];\n }\n\n __syncthreads();\n }\n}\n\n#define KPT 16\n#define TPB 384\n#define DIGIT_BITS 8\ntemplate\ncudaError_t radixSelectTopK(KeyT *d_keys_in, uint num_items, uint k, KeyT *d_keys_out,\n CachingDeviceAllocator& g_allocator) {\n cudaError error = cudaSuccess;\n\n DoubleBuffer d_keys;\n d_keys.d_buffers[0] = d_keys_in;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(KeyT) * num_items));\n\n uint* d_histogram;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_histogram, sizeof(uint) * num_items));\n\n // We allocate two indices, one that maintains index into output array (this goes till K)\n // second maintains index into the output buffer containing reduced set of top-k candidates.\n uint* d_index_out;\n uint* d_index_buffer;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_index_out, sizeof(uint)));\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_index_buffer, sizeof(uint)));\n\n // Set the index into output array to 0.\n cudaMemset(d_index_out, 0, 4);\n\n uint* h_histogram = new uint[256];\n\n uint KPB = KPT * TPB;\n\n for (uint digit = 0; digit < 4; digit++) {\n uint num_blocks = num_items / KPB;// Pass-0 rough processing blocks (floor on purpose)\n uint processed_elements = num_blocks * KPB;// Pass-0 number of rough processed elements\n uint remaining_elements = num_items - processed_elements;// Do the remaining elements with a check in the inner loop\n uint remainder_blocks = (KPB-1+remaining_elements) / KPB;// Number of blocks required for remaining elements (typically 0 or 1)\n\n // Zero out the histogram\n cudaMemset(d_histogram, 0, 256 * 4);\n\n if (num_blocks > 0)\n rdxsrt_histogram<<>>(d_keys.Current(), digit, d_histogram);\n if (remaining_elements > 0)\n rdxsrt_histogram_with_guards<<>>(d_keys.Current(), digit, d_histogram, num_items, num_blocks);\n\n cudaMemcpy(h_histogram, d_histogram, 256 * sizeof(uint), cudaMemcpyDeviceToHost);\n\n // Check for failure to launch\n CubDebugExit(error = cudaPeekAtLastError());\n\n cudaMemcpy(h_histogram, d_histogram, 256 * sizeof(uint), cudaMemcpyDeviceToHost);\n uint rolling_sum = 0;\n uint digit_val;\n for (int i=255; i>=0; i--) {\n if ((rolling_sum + h_histogram[i]) > k) {\n digit_val = i;\n k -= rolling_sum;\n break;\n }\n rolling_sum += h_histogram[i];\n }\n\n cudaMemset(d_index_buffer, 0, 4);\n\n select_kth_bucket<<>>(d_keys.Current(), digit, digit_val, num_items, d_keys.Alternate(), d_keys_out, d_index_buffer, d_index_out);\n\n CubDebugExit(error = cudaPeekAtLastError());\n\n uint h_index_out;\n uint h_index_buffer;\n\n cudaMemcpy(&h_index_out, d_index_out, sizeof(uint), cudaMemcpyDeviceToHost);\n cudaMemcpy(&h_index_buffer, d_index_buffer, sizeof(uint), cudaMemcpyDeviceToHost);\n\n // Update number of items to reflect reduced number of elements.\n num_items = h_index_buffer;\n\n if (k == 0) break;\n else if (k != 0 && digit == 3) {\n // We are at last digit and k != 4 implies that kth value has repetition.\n // Copy any of the repeated values to out array to complete the array.\n cudaMemcpy(d_keys_out + h_index_out, d_keys.Alternate(), k * sizeof(KeyT), cudaMemcpyDeviceToDevice);\n k -= k;\n }\n\n // Toggle the buffer index in the double buffer\n d_keys.selector = d_keys.selector ^ 1;\n }\n\n // Cleanup\n if (d_keys.d_buffers[1])\n CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[1]));\n if (d_histogram)\n CubDebugExit(g_allocator.DeviceFree(d_histogram));\n if (d_index_buffer)\n CubDebugExit(g_allocator.DeviceFree(d_index_buffer));\n if (d_index_out)\n CubDebugExit(g_allocator.DeviceFree(d_index_out));\n\n return error;\n}\n\n" }, { "path": "src/sharedmem.cuh", "content": "/*\n * Copyright 1993-2014 NVIDIA Corporation. All rights reserved.\n *\n * Please refer to the NVIDIA end user license agreement (EULA) associated\n * with this source code for terms and conditions that govern your use of\n * this software. Any use, reproduction, disclosure, or distribution of\n * this software and related documentation outside the terms of the EULA\n * is strictly prohibited.\n *\n */\n\n#ifndef _SHAREDMEM_H_\n#define _SHAREDMEM_H_\n\n//****************************************************************************\n// Because dynamically sized shared memory arrays are declared \"extern\",\n// we can't templatize them directly. To get around this, we declare a\n// simple wrapper struct that will declare the extern array with a different\n// name depending on the type. This avoids compiler errors about duplicate\n// definitions.\n//\n// To use dynamically allocated shared memory in a templatized __global__ or\n// __device__ function, just replace code like this:\n//\n//\n// template\n// __global__ void\n// foo( T* g_idata, T* g_odata)\n// {\n// // Shared mem size is determined by the host app at run time\n// extern __shared__ T sdata[];\n// ...\n// doStuff(sdata);\n// ...\n// }\n//\n// With this\n// template\n// __global__ void\n// foo( T* g_idata, T* g_odata)\n// {\n// // Shared mem size is determined by the host app at run time\n// SharedMemory smem;\n// T* sdata = smem.getPointer();\n// ...\n// doStuff(sdata);\n// ...\n// }\n//****************************************************************************\n\n// This is the un-specialized struct. Note that we prevent instantiation of this\n// struct by putting an undefined symbol in the function body so it won't compile.\ntemplate \nstruct SharedMemory\n{\n // Ensure that we won't compile any un-specialized types\n __device__ T *getPointer()\n {\n extern __device__ void error(void);\n error();\n return NULL;\n }\n};\n\n// Following are the specializations for the following types.\n// int, uint, char, uchar, short, ushort, long, ulong, bool, float, and double\n// One could also specialize it for user-defined types.\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ int *getPointer()\n {\n extern __shared__ int s_int[];\n return s_int;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ unsigned int *getPointer()\n {\n extern __shared__ unsigned int s_uint[];\n return s_uint;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ char *getPointer()\n {\n extern __shared__ char s_char[];\n return s_char;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ unsigned char *getPointer()\n {\n extern __shared__ unsigned char s_uchar[];\n return s_uchar;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ short *getPointer()\n {\n extern __shared__ short s_short[];\n return s_short;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ unsigned short *getPointer()\n {\n extern __shared__ unsigned short s_ushort[];\n return s_ushort;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ long *getPointer()\n {\n extern __shared__ long s_long[];\n return s_long;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ unsigned long *getPointer()\n {\n extern __shared__ unsigned long s_ulong[];\n return s_ulong;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ bool *getPointer()\n {\n extern __shared__ bool s_bool[];\n return s_bool;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ float *getPointer()\n {\n extern __shared__ float s_float[];\n return s_float;\n }\n};\n\ntemplate <>\nstruct SharedMemory \n{\n __device__ double *getPointer()\n {\n extern __shared__ double s_double[];\n return s_double;\n }\n};\n\n\n#endif //_SHAREDMEM_H_\n" }, { "path": "src/sortTopK.cuh", "content": "#include \n#include \n#include \n\nusing namespace std;\nusing namespace cub;\n\ntemplate\ncudaError_t sortTopK(KeyT *d_keys_in, unsigned int num_items, unsigned int k, KeyT *d_keys_out,\n CachingDeviceAllocator& g_allocator) {\n // Allocate device memory for input/output\n DoubleBuffer d_keys;\n d_keys.d_buffers[0] = d_keys_in;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(KeyT) * num_items));\n\n // Allocate temporary storage\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n CubDebugExit(DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys, num_items));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Sort\n CubDebugExit(DeviceRadixSort::SortKeysDescending(d_temp_storage, temp_storage_bytes, d_keys, num_items));\n\n // Copy results for verification. GPU-side part is done.\n CubDebugExit(cudaMemcpy(d_keys_out, d_keys.Current(), sizeof(KeyT) * k, cudaMemcpyDeviceToDevice));\n\n // Cleanup\n if (d_keys.d_buffers[1])\n CubDebugExit(g_allocator.DeviceFree(d_keys.d_buffers[1]));\n if (d_temp_storage)\n CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n return cudaSuccess;\n}\n\n" }, { "path": "src/sortingNetwork.cuh", "content": "template\nstruct SortingNetwork\n{\n enum { NETWORK_MAX_NUM_ITEMS = 18 };\n\nprivate:\n static __device__ __forceinline__ void compare_and_swap(KeyT *keys, int low_idx, int high_idx)\n {\n if(keys[low_idx]>keys[high_idx]){\n KeyT tmp = keys[low_idx];\n keys[low_idx] = keys[high_idx];\n keys[high_idx] = tmp;\n }\n }\n\n template\n struct SortingNetworkAgent\n {\n static __device__ __forceinline__ void sort(KeyT *keys){\n#if __cplusplus >= 199711\n static_assert(NUM_ITEMS<=NETWORK_MAX_NUM_ITEMS, \"Sorting network of given size not supported.\");\n#endif\n }\n };\n\n /*** SORTING NETWORK 1 ***/\n template\n struct SortingNetworkAgent<1, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys){}\n };\n\n /*** SORTING NETWORK 2 ***/\n template\n struct SortingNetworkAgent<2, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys){\n compare_and_swap(keys, 0, 1);\n }\n };\n\n /*** SORTING NETWORK 3 ***/\n template\n struct SortingNetworkAgent<3, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 0, 1);\n }\n };\n\n /*** SORTING NETWORK 4 ***/\n template\n struct SortingNetworkAgent<4, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 1, 2);\n }\n };\n\n /*** SORTING NETWORK 5 ***/\n template\n struct SortingNetworkAgent<5, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 0, 3);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 1, 2);\n }\n };\n\n /*** SORTING NETWORK 6 ***/\n template\n struct SortingNetworkAgent<6, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 0, 3);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 2, 3);\n }\n };\n\n /*** SORTING NETWORK 7 ***/\n template\n struct SortingNetworkAgent<7, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 0, 3);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 2, 3);\n }\n };\n\n /*** SORTING NETWORK 8 ***/\n template\n struct SortingNetworkAgent<8, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 3, 6);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 3, 4);\n }\n };\n\n /*** SORTING NETWORK 9 ***/\n template\n struct SortingNetworkAgent<9, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 0, 3);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 1, 6);\n compare_and_swap(keys, 0, 5);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 4, 7);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n }\n };\n\n /*** SORTING NETWORK 10 ***/\n template\n struct SortingNetworkAgent<10, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 4, 9);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 2, 7);\n compare_and_swap(keys, 1, 6);\n compare_and_swap(keys, 0, 5);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 0, 3);\n compare_and_swap(keys, 5, 8);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 3, 6);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 4, 7);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 4, 5);\n }\n };\n\n\n /*** SORTING NETWORK 11 ***/\n template\n struct SortingNetworkAgent<11, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 6, 10);\n compare_and_swap(keys, 5, 9);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 6, 10);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 7, 10);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 3, 4);\n }\n };\n\n /*** SORTING NETWORK 12 ***/\n template\n struct SortingNetworkAgent<12, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 6, 10);\n compare_and_swap(keys, 5, 9);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 6, 10);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 7, 10);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 7, 8);\n }\n };\n\n /*** SORTING NETWORK 13 ***/\n template\n struct SortingNetworkAgent<13, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 1, 7);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 8);\n compare_and_swap(keys, 0, 12);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 11);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 5, 9);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 6, 12);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 4, 9);\n compare_and_swap(keys, 6, 10);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 1, 7);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 4, 7);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 0, 5);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n }\n };\n\n /*** SORTING NETWORK 14 ***/\n template\n struct SortingNetworkAgent<14, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 12, 13);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 8, 12);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 9, 13);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 0, 8);\n compare_and_swap(keys, 1, 9);\n compare_and_swap(keys, 2, 10);\n compare_and_swap(keys, 3, 11);\n compare_and_swap(keys, 4, 12);\n compare_and_swap(keys, 5, 13);\n compare_and_swap(keys, 5, 10);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 3, 12);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 7, 13);\n compare_and_swap(keys, 2, 8);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 13);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 10, 12);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n }\n };\n\n /*** SORTING NETWORK 15 ***/\n template\n struct SortingNetworkAgent<15, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 12, 13);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 12, 14);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 8, 12);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 9, 13);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 10, 14);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 0, 8);\n compare_and_swap(keys, 1, 9);\n compare_and_swap(keys, 2, 10);\n compare_and_swap(keys, 3, 11);\n compare_and_swap(keys, 4, 12);\n compare_and_swap(keys, 5, 13);\n compare_and_swap(keys, 6, 14);\n compare_and_swap(keys, 5, 10);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 3, 12);\n compare_and_swap(keys, 13, 14);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 7, 13);\n compare_and_swap(keys, 2, 8);\n compare_and_swap(keys, 11, 14);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 13);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 10, 12);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n }\n };\n\n /*** SORTING NETWORK 16 ***/\n template\n struct SortingNetworkAgent<16, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 12, 13);\n compare_and_swap(keys, 14, 15);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 12, 14);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 13, 15);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 8, 12);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 9, 13);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 10, 14);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 11, 15);\n compare_and_swap(keys, 0, 8);\n compare_and_swap(keys, 1, 9);\n compare_and_swap(keys, 2, 10);\n compare_and_swap(keys, 3, 11);\n compare_and_swap(keys, 4, 12);\n compare_and_swap(keys, 5, 13);\n compare_and_swap(keys, 6, 14);\n compare_and_swap(keys, 7, 15);\n compare_and_swap(keys, 5, 10);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 3, 12);\n compare_and_swap(keys, 13, 14);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 7, 13);\n compare_and_swap(keys, 2, 8);\n compare_and_swap(keys, 11, 14);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 13);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 10, 12);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 8, 9);\n }\n };\n\n /*** SORTING NETWORK 17 ***/\n template\n struct SortingNetworkAgent<17, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 2, 6);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 3, 6);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 8, 9);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 12, 13);\n compare_and_swap(keys, 15, 16);\n compare_and_swap(keys, 14, 16);\n compare_and_swap(keys, 14, 15);\n compare_and_swap(keys, 12, 15);\n compare_and_swap(keys, 12, 14);\n compare_and_swap(keys, 13, 16);\n compare_and_swap(keys, 13, 15);\n compare_and_swap(keys, 13, 14);\n compare_and_swap(keys, 8, 13);\n compare_and_swap(keys, 8, 12);\n compare_and_swap(keys, 9, 14);\n compare_and_swap(keys, 9, 13);\n compare_and_swap(keys, 9, 12);\n compare_and_swap(keys, 10, 15);\n compare_and_swap(keys, 11, 16);\n compare_and_swap(keys, 11, 15);\n compare_and_swap(keys, 10, 13);\n compare_and_swap(keys, 10, 12);\n compare_and_swap(keys, 11, 14);\n compare_and_swap(keys, 11, 13);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 0, 9);\n compare_and_swap(keys, 0, 8);\n compare_and_swap(keys, 1, 10);\n compare_and_swap(keys, 1, 9);\n compare_and_swap(keys, 1, 8);\n compare_and_swap(keys, 2, 11);\n compare_and_swap(keys, 3, 12);\n compare_and_swap(keys, 3, 11);\n compare_and_swap(keys, 2, 9);\n compare_and_swap(keys, 2, 8);\n compare_and_swap(keys, 3, 10);\n compare_and_swap(keys, 3, 9);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 4, 13);\n compare_and_swap(keys, 5, 14);\n compare_and_swap(keys, 5, 13);\n compare_and_swap(keys, 6, 15);\n compare_and_swap(keys, 7, 16);\n compare_and_swap(keys, 7, 15);\n compare_and_swap(keys, 6, 13);\n compare_and_swap(keys, 7, 14);\n compare_and_swap(keys, 7, 13);\n compare_and_swap(keys, 4, 9);\n compare_and_swap(keys, 4, 8);\n compare_and_swap(keys, 5, 10);\n compare_and_swap(keys, 5, 9);\n compare_and_swap(keys, 5, 8);\n compare_and_swap(keys, 6, 11);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 7, 10);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 7, 8);\n }\n };\n\n /*** SORTING NETWORK 18 ***/\n template\n struct SortingNetworkAgent<18, DUMMY>\n {\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n compare_and_swap(keys, 0, 1);\n compare_and_swap(keys, 2, 3);\n compare_and_swap(keys, 0, 2);\n compare_and_swap(keys, 1, 3);\n compare_and_swap(keys, 1, 2);\n compare_and_swap(keys, 4, 5);\n compare_and_swap(keys, 7, 8);\n compare_and_swap(keys, 6, 8);\n compare_and_swap(keys, 6, 7);\n compare_and_swap(keys, 4, 7);\n compare_and_swap(keys, 4, 6);\n compare_and_swap(keys, 5, 8);\n compare_and_swap(keys, 5, 7);\n compare_and_swap(keys, 5, 6);\n compare_and_swap(keys, 0, 5);\n compare_and_swap(keys, 0, 4);\n compare_and_swap(keys, 1, 6);\n compare_and_swap(keys, 1, 5);\n compare_and_swap(keys, 1, 4);\n compare_and_swap(keys, 2, 7);\n compare_and_swap(keys, 3, 8);\n compare_and_swap(keys, 3, 7);\n compare_and_swap(keys, 2, 5);\n compare_and_swap(keys, 2, 4);\n compare_and_swap(keys, 3, 6);\n compare_and_swap(keys, 3, 5);\n compare_and_swap(keys, 3, 4);\n compare_and_swap(keys, 9, 10);\n compare_and_swap(keys, 11, 12);\n compare_and_swap(keys, 9, 11);\n compare_and_swap(keys, 10, 12);\n compare_and_swap(keys, 10, 11);\n compare_and_swap(keys, 13, 14);\n compare_and_swap(keys, 16, 17);\n compare_and_swap(keys, 15, 17);\n compare_and_swap(keys, 15, 16);\n compare_and_swap(keys, 13, 16);\n compare_and_swap(keys, 13, 15);\n compare_and_swap(keys, 14, 17);\n compare_and_swap(keys, 14, 16);\n compare_and_swap(keys, 14, 15);\n compare_and_swap(keys, 9, 14);\n compare_and_swap(keys, 9, 13);\n compare_and_swap(keys, 10, 15);\n compare_and_swap(keys, 10, 14);\n compare_and_swap(keys, 10, 13);\n compare_and_swap(keys, 11, 16);\n compare_and_swap(keys, 12, 17);\n compare_and_swap(keys, 12, 16);\n compare_and_swap(keys, 11, 14);\n compare_and_swap(keys, 11, 13);\n compare_and_swap(keys, 12, 15);\n compare_and_swap(keys, 12, 14);\n compare_and_swap(keys, 12, 13);\n compare_and_swap(keys, 0, 9);\n compare_and_swap(keys, 1, 10);\n compare_and_swap(keys, 1, 9);\n compare_and_swap(keys, 2, 11);\n compare_and_swap(keys, 3, 12);\n compare_and_swap(keys, 3, 11);\n compare_and_swap(keys, 2, 9);\n compare_and_swap(keys, 3, 10);\n compare_and_swap(keys, 3, 9);\n compare_and_swap(keys, 4, 13);\n compare_and_swap(keys, 5, 14);\n compare_and_swap(keys, 5, 13);\n compare_and_swap(keys, 6, 15);\n compare_and_swap(keys, 7, 16);\n compare_and_swap(keys, 8, 17);\n compare_and_swap(keys, 8, 16);\n compare_and_swap(keys, 7, 15);\n compare_and_swap(keys, 8, 15);\n compare_and_swap(keys, 6, 13);\n compare_and_swap(keys, 7, 14);\n compare_and_swap(keys, 8, 14);\n compare_and_swap(keys, 7, 13);\n compare_and_swap(keys, 8, 13);\n compare_and_swap(keys, 4, 9);\n compare_and_swap(keys, 5, 10);\n compare_and_swap(keys, 5, 9);\n compare_and_swap(keys, 6, 11);\n compare_and_swap(keys, 7, 12);\n compare_and_swap(keys, 8, 12);\n compare_and_swap(keys, 7, 11);\n compare_and_swap(keys, 8, 11);\n compare_and_swap(keys, 6, 9);\n compare_and_swap(keys, 7, 10);\n compare_and_swap(keys, 8, 10);\n compare_and_swap(keys, 7, 9);\n compare_and_swap(keys, 8, 9);\n }\n };\n\n template<\n int RUN_LENGTH,\n int NUM_RUN_ITEMS, // Number of items to sort in this run (starting at keys)\n bool IS_LAST_RUN // If this run is the last run to be sorted\n >\n struct SortRunAgent\n {\n static __device__ __forceinline__ void sort_runs(KeyT *keys)\n {\n SortingNetwork::sort(keys);\n SortRunAgent::sort_runs(&keys[RUN_LENGTH]);\n }\n };\n\n template<\n int RUN_LENGTH,\n int NUM_RUN_ITEMS\n >\n struct SortRunAgent\n {\n static __device__ __forceinline__ void sort_runs(KeyT *keys)\n {\n SortingNetwork::sort(keys);\n }\n };\n\npublic:\n\n template\n static __device__ __forceinline__ void sort(KeyT (&keys)[NUM_ITEMS])\n {\n SortingNetworkAgent::sort(keys);\n }\n\n template\n static __device__ __forceinline__ void sort(KeyT *keys)\n {\n SortingNetworkAgent::sort(keys);\n }\n\n template<\n int RUN_LENGTH,\n int NUM_ITEMS\n >\n static __device__ __forceinline__ void sort_runs(KeyT (&keys)[NUM_ITEMS])\n {\n SortRunAgentRUN_LENGTH)>::sort_runs(keys);\n }\n\n template<\n int RUN_LENGTH,\n int NUM_ITEMS\n >\n static __device__ __forceinline__ void sort_runs(KeyT *keys)\n {\n SortRunAgentRUN_LENGTH)>::sort_runs(keys);\n }\n}\n" }, { "path": "test/compareTopKAlgorithms.cu", "content": "#include \n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \"printFunctions.cuh\"\n#include \"generateProblems.cuh\"\n#include \"sortTopK.cuh\"\n#include \"radixSelectTopK.cuh\"\n#include \"bitonicTopK.cuh\"\n\n#define SETUP_TIMING() cudaEvent_t start, stop; cudaEventCreate(&start); cudaEventCreate(&stop);\n\n#define TIME_FUNC(f,t) { \\\n cudaEventRecord(start, 0); \\\n f; \\\n cudaEventRecord(stop, 0); \\\n cudaEventSynchronize(stop); \\\n cudaEventElapsedTime(&t, start,stop); \\\n}\n\nusing namespace std;\n\nCachingDeviceAllocator g_allocator(true); // Caching allocator for device memory\n\n#define NUMBEROFALGORITHMS 3\nconst char* namesOfTimingFunctions[NUMBEROFALGORITHMS] = {\"Sort\", \"Radix Select\", \"Bitonic TopK\"};\n\ntemplate\nvoid compareAlgorithms(uint size, uint k, uint numTests,uint *algorithmsToTest, uint generateType){\n KeyT *d_vec;\n KeyT *d_vec_copy;\n KeyT *d_res;\n float timeArray[NUMBEROFALGORITHMS][numTests];\n float totalTimesPerAlgorithm[NUMBEROFALGORITHMS];\n float minTimesPerAlgorithm[NUMBEROFALGORITHMS];\n KeyT* resultsArray[NUMBEROFALGORITHMS][numTests];\n std::fill_n(minTimesPerAlgorithm, NUMBEROFALGORITHMS, 2000);\n\n uint winnerArray[numTests];\n uint timesWon[NUMBEROFALGORITHMS];\n uint i,j,m,x;\n int runOrder[NUMBEROFALGORITHMS];\n\n unsigned long long seed;\n timeval t1;\n float runtime;\n\n SETUP_TIMING()\n\n typedef cudaError_t (*ptrToTimingFunction)(KeyT*, uint, uint, KeyT*, CachingDeviceAllocator&);\n typedef void (*ptrToGeneratingFunction)(KeyT*, uint, curandGenerator_t, CachingDeviceAllocator&);\n //these are the functions that can be called\n ptrToTimingFunction arrayOfTimingFunctions[NUMBEROFALGORITHMS] = {&sortTopK , &radixSelectTopK, &bitonicTopK};\n /*ptrToTimingFunction arrayOfTimingFunctions[NUMBEROFALGORITHMS] = {NULL, NULL, NULL};*/\n\n ptrToGeneratingFunction *arrayOfGenerators;\n const char** namesOfGeneratingFunctions;\n //this is the array of names of functions that generate problems of this type, ie float, double, or uint\n namesOfGeneratingFunctions = returnNamesOfGenerators();\n arrayOfGenerators = (ptrToGeneratingFunction *) returnGenFunctions();\n\n //zero out the totals and times won\n bzero(totalTimesPerAlgorithm, NUMBEROFALGORITHMS * sizeof(uint));\n bzero(timesWon, NUMBEROFALGORITHMS * sizeof(uint));\n //allocate space for d_vec, and d_vec_copy\n cudaMalloc(&d_vec, size * sizeof(KeyT));\n cudaMalloc(&d_vec_copy, size * sizeof(KeyT));\n cudaMalloc(&d_res, k * sizeof(KeyT));\n\n //create the random generator.\n curandGenerator_t generator;\n srand(unsigned(time(NULL)));\n\n printf(\"The distribution is: %s\\n\", namesOfGeneratingFunctions[generateType]);\n for(i = 0; i < numTests; i++){\n // cudaDeviceReset();\n gettimeofday(&t1, NULL);\n seed = t1.tv_usec * t1.tv_sec;\n\n for(m = 0; m < NUMBEROFALGORITHMS;m++){\n runOrder[m] = m;\n }\n std::random_shuffle(runOrder, runOrder + NUMBEROFALGORITHMS);\n curandCreateGenerator(&generator, CURAND_RNG_PSEUDO_DEFAULT);\n curandSetPseudoRandomGeneratorSeed(generator,seed);\n curandSetPseudoRandomGeneratorSeed(generator,0);\n printf(\"Running test %u of %u for size: %u and k: %u\\n\", i + 1, numTests,size,k);\n //generate the random vector using the specified distribution\n arrayOfGenerators[generateType](d_vec, size, generator, g_allocator);\n\n //copy the vector to d_vec_copy, which will be used to restore it later\n cudaMemcpy(d_vec_copy, d_vec, size * sizeof(KeyT), cudaMemcpyDeviceToDevice);\n\n uint* h_vec = new uint[size];\n cudaMemcpy(h_vec, d_vec, size * sizeof(KeyT), cudaMemcpyDeviceToHost);\n\n/* uint *b = new uint[size];*/\n /*uint *c = new uint[size];*/\n /*uint *d = new uint[size];*/\n\n /*int x=0,y=0,z=0;*/\n /*for (int r=0; r());*/\n /*cout << \"Count \" << x << endl;*/\n /*for (int r=0; r());*/\n /*for (int r=0; r());\n\n //record the value returned\n resultsArray[j][i] = h_res;\n //update the current \"winner\" if necessary\n if(timeArray[j][i] < currentWinningTime){\n currentWinningTime = runtime;\n winnerArray[i] = j;\n }\n\n //perform clean up\n cudaMemcpy(d_vec_copy, d_vec, size * sizeof(KeyT), cudaMemcpyDeviceToDevice);\n }\n }\n curandDestroyGenerator(generator);\n/* //for(x = 0; x < NUMBEROFALGORITHMS; x++){*/\n //if(algorithmsToTest[x]){\n //fileCsv << namesOfTimingFunctions[x] << \",\"<< resultsArray[x][i] <<\",\"<< timeArray[x][i] <<\",\";\n //}\n /*}*/\n // uint flag = 0;\n/* T tempResult = resultsArray[0][i];*/\n //for(m = 1; m < NUMBEROFALGORITHMS;m++){\n //if(algorithmsToTest[m]){\n //if(resultsArray[m][i] != tempResult){\n //flag++;\n //}\n //}\n /*}*/\n //fileCsv << flag << \"\\n\";\n }\n\n //calculate the total time each algorithm took\n for(i = 0; i < numTests; i++){\n for(j = 0; j < NUMBEROFALGORITHMS;j++){\n if(algorithmsToTest[j]){\n totalTimesPerAlgorithm[j] += timeArray[j][i];\n minTimesPerAlgorithm[j] = min(minTimesPerAlgorithm[j], timeArray[j][i]);\n }\n }\n }\n\n //count the number of times each algorithm won.\n for(i = 0; i < numTests;i++){\n timesWon[winnerArray[i]]++;\n }\n\n printf(\"\\n\\n\");\n\n //print out the average times\n for(i = 0; i < NUMBEROFALGORITHMS; i++){\n if(algorithmsToTest[i]){\n printf(\"%-20s averaged: %f ms\\n\", namesOfTimingFunctions[i], totalTimesPerAlgorithm[i] / numTests);\n }\n }\n for(i = 0; i < NUMBEROFALGORITHMS; i++){\n if(algorithmsToTest[i]){\n printf(\"%-20s minimum: %f ms\\n\", namesOfTimingFunctions[i], minTimesPerAlgorithm[i]);\n }\n }\n for(i = 0; i < NUMBEROFALGORITHMS; i++){\n if(algorithmsToTest[i]){\n printf(\"%s won %u times\\n\", namesOfTimingFunctions[i], timesWon[i]);\n }\n }\n if (algorithmsToTest[0]) {\n for(i = 0; i < numTests; i++){\n for(j =1 ; j< NUMBEROFALGORITHMS; j++){\n if(algorithmsToTest[j]){\n for (int m=0; m(resultsArray[j][i], k);\n std::cout << \"Sort:\\t\";\n PrintFunctions::printArray(resultsArray[0][i], k);\n break;\n }\n }\n }\n }\n }\n\n //free d_vec and d_vec_copy\n cudaFree(d_vec);\n cudaFree(d_vec_copy);\n}\n\ntemplate\nvoid runTests(uint generateType, int K, uint startPower, uint stopPower, uint timesToTestEachK = 3) {\n // Algorithms To Run\n // timeSort, timeRadixSelect, timeBitonicTopK\n uint algorithmsToRun[NUMBEROFALGORITHMS]= {1,1,1};\n uint size;\n for (size = (1 << startPower); size <= (1 <(size, K, timesToTestEachK,algorithmsToRun,generateType);\n }\n}\n\nint main(int argc, char** argv)\n{\n uint testCount;\n int K;\n uint type,distributionType,startPower,stopPower;\n if (argc == 7) {\n type = atoi(argv[1]);\n distributionType = atoi(argv[2]);\n K = atoi(argv[3]);\n testCount = atoi(argv[4]);\n startPower = atoi(argv[5]);\n stopPower = atoi(argv[6]);\n } else {\n printf(\"Please enter the type of value you want to test:\\n1-float\\n2-double\\n3-uint\\n\");\n cin >> type;\n printf(\"Please enter distribution type: \");\n cin >> distributionType;\n printf(\"Please enter K: \");\n cin >> K;\n printf(\"Please enter number of tests to run per K: \");\n cin >> testCount;\n printf(\"Please enter start power (dataset size starts at 2^start)(max val: 29): \");\n cin >> startPower;\n printf(\"Please enter stop power (dataset size stops at 2^stop)(max val: 29): \");\n cin >> stopPower;\n }\n\n switch(type){\n case 1:\n runTests(distributionType,K,startPower,stopPower,testCount);\n break;\n case 2:\n // runTests(distributionType,K,startPower,stopPower,testCount);\n break;\n case 3:\n runTests(distributionType,K,startPower,stopPower,testCount);\n break;\n default:\n printf(\"You entered and invalid option, now exiting\\n\");\n break;\n }\n\n return 0;\n}\n\n" }, { "path": "test/generateProblems.cuh", "content": "#pragma once\n\n#include \n#include \n#include \n#include \n#include \n#include \n/*#include */\nusing namespace cub;\nusing namespace std;\n\ntemplate\nvoid sort_keys(KeyT* d_vec, const unsigned int num_keys, CachingDeviceAllocator& g_allocator) {\n // Allocate device memory for input/output\n DoubleBuffer d_keys;\n d_keys.d_buffers[0] = d_vec;\n CubDebugExit(g_allocator.DeviceAllocate((void**)&d_keys.d_buffers[1], sizeof(KeyT) * num_keys));\n\n // Allocate temporary storage\n size_t temp_storage_bytes = 0;\n void *d_temp_storage = NULL;\n CubDebugExit(DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys, num_keys));\n CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));\n\n // Sort\n CubDebugExit(DeviceRadixSort::SortKeys(d_temp_storage, temp_storage_bytes, d_keys, num_keys));\n\n // Copy results for verification. GPU-side part is done.\n if (d_keys.Current() != d_vec) {\n CubDebugExit(cudaMemcpy(d_vec, d_keys.Current(), sizeof(KeyT) * num_keys, cudaMemcpyDeviceToDevice));\n }\n\n if (d_temp_storage)\n CubDebugExit(g_allocator.DeviceFree(d_temp_storage));\n\n if (d_keys.d_buffers[1])\n CubDebugExit( g_allocator.DeviceFree(d_keys.d_buffers[1]) );\n}\n\n///////////////////////////////////////////////////////////////////\n//// FUNCTIONS TO GENERATE UINTS\n///////////////////////////////////////////////////////////////////\ntypedef void (*ptrToUintGeneratingFunction)(uint*, uint, curandGenerator_t, CachingDeviceAllocator&);\n\nvoid generateUniformUints(uint *d_vec, uint num_keys, curandGenerator_t generator, \n CachingDeviceAllocator& g_allocator) {\n curandGenerate(generator, d_vec, num_keys);\n}\n\nvoid generateSortedIncreasingUints(uint* d_vec, uint num_keys, curandGenerator_t gen, \n CachingDeviceAllocator& g_allocator) {\n curandGenerate(gen, d_vec,num_keys);\n sort_keys(d_vec, num_keys, g_allocator);\n}\n\nvoid generateSortedDecreasingUints(uint* d_vec, uint num_keys, curandGenerator_t gen, \n CachingDeviceAllocator& g_allocator) {\n curandGenerate(gen, d_vec,num_keys);\n sort_keys(d_vec, num_keys, g_allocator);\n}\n\n#define NUMBEROFUINTDISTRIBUTIONS 3\nptrToUintGeneratingFunction arrayOfUintGenerators[NUMBEROFUINTDISTRIBUTIONS] = {&generateUniformUints, &generateSortedIncreasingUints, &generateSortedDecreasingUints};\nconst char* namesOfUintGeneratingFunctions[NUMBEROFUINTDISTRIBUTIONS]={\"UNIFORM UINTS\",\"SORTED INC UINTS\", \"SORTED DEC UINTS\"};\n\n///////////////////////////////////////////////////////////////////\n//// FUNCTIONS TO GENERATE FLOATS\n///////////////////////////////////////////////////////////////////\ntypedef void (*ptrToFloatGeneratingFunction)(float*, uint, curandGenerator_t, CachingDeviceAllocator&);\n\nvoid generateUniformFloats(float *d_vec, uint num_keys, curandGenerator_t generator, \n CachingDeviceAllocator& g_allocator){\n curandGenerateUniform(generator, d_vec,num_keys);\n}\n\nvoid generateSortedIncreasingFloats(float *d_vec, uint num_keys, curandGenerator_t generator, \n CachingDeviceAllocator& g_allocator) {\n curandGenerateUniform(generator, d_vec,num_keys);\n sort_keys(d_vec, num_keys, g_allocator);\n}\n\nvoid generateSortedDecreasingFloats(float *d_vec, uint num_keys, curandGenerator_t generator, \n CachingDeviceAllocator& g_allocator) {\n curandGenerateUniform(generator, d_vec,num_keys);\n sort_keys(d_vec, num_keys, g_allocator);\n}\n\n/*void generateBucketKillerFloats(float *d_vec, uint num_keys, curandGenerator_t generator){*/\n /*int i;*/\n /*float * d_generated = d_vec;*/\n /*curandGenerateUniform(generator, d_generated,num_keys);*/\n /*thrust::device_ptr dev_ptr((uint *)d_generated);*/\n /*thrust::for_each( dev_ptr, dev_ptr + num_keys, makeSmallFloat());*/\n /*thrust::sort(dev_ptr,dev_ptr + num_keys);*/\n\n /*float* h_vec = (float*) malloc(num_keys * sizeof(float));*/\n /*cudaMemcpy(h_vec, d_generated, num_keys * sizeof(float), cudaMemcpyDeviceToHost);*/\n\n /*for(i = -126; i < 127; i++){*/\n /*h_vec[i + 126] = pow(2.0f,(float)i);*/\n /*}*/\n /*cudaMemcpy(d_generated, h_vec, num_keys * sizeof(float), cudaMemcpyHostToDevice);*/\n /*free(h_vec);*/\n/*}*/\n\n#define NUMBEROFFLOATDISTRIBUTIONS 3\nptrToFloatGeneratingFunction arrayOfFloatGenerators[NUMBEROFFLOATDISTRIBUTIONS] = {&generateUniformFloats, &generateSortedIncreasingFloats, &generateSortedDecreasingFloats};\n\nconst char* namesOfFloatGeneratingFunctions[NUMBEROFFLOATDISTRIBUTIONS] = {\"UNIFORM FLOATS\", \"SORTED INC FLOATS\", \"SORTED DEC FLOATS\"};\n\n///////////////////////////////////////////////////////////////////\n//// FUNCTIONS TO GENERATE DOUBLES\n///////////////////////////////////////////////////////////////////\n\ntypedef void (*ptrToDoubleGeneratingFunction)(double*, uint, curandGenerator_t, CachingDeviceAllocator&);\n\nvoid generateUniformDoubles(double *d_generated, uint num_keys, curandGenerator_t generator, \n CachingDeviceAllocator& g_allocator) {\n curandGenerateUniformDouble(generator, d_generated,num_keys);\n}\n\n#define NUMBEROFDOUBLEDISTRIBUTIONS 1\nptrToDoubleGeneratingFunction arrayOfDoubleGenerators[NUMBEROFDOUBLEDISTRIBUTIONS] = {&generateUniformDoubles};\nconst char* namesOfDoubleGeneratingFunctions[NUMBEROFDOUBLEDISTRIBUTIONS]={\"UNIFORM DOUBLES\"};\n\ntemplate void* returnGenFunctions(){\n if(typeid(T) == typeid(uint)){\n return arrayOfUintGenerators;\n }\n else if(typeid(T) == typeid(float)){\n return arrayOfFloatGenerators;\n }\n else{\n return arrayOfDoubleGenerators;\n }\n}\n\n\ntemplate const char** returnNamesOfGenerators(){\n if(typeid(T) == typeid(uint)){\n return &namesOfUintGeneratingFunctions[0];\n }\n else if(typeid(T) == typeid(float)){\n return &namesOfFloatGeneratingFunctions[0];\n }\n else {\n return &namesOfDoubleGeneratingFunctions[0];\n }\n}\n\n/*template void* returnGenFunctions();*/\n/*template void* returnGenFunctions();*/\n/*template void* returnGenFunctions();*/\n\n/*template const char** returnNamesOfGenerators();*/\n/*template const char** returnNamesOfGenerators();*/\n/*template const char** returnNamesOfGenerators();*/\n\n" }, { "path": "test/printFunctions.cuh", "content": "/* Copyright 2011 Russel Steinbach, Jeffrey Blanchard, Bradley Gordon,\n * and Toluwaloju Alabi\n * Licensed under the Apache License, Version 2.0 (the \"License\");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an \"AS IS\" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n */\n#pragma once\n#include \n#include \n#include \nusing namespace std;\n\nnamespace PrintFunctions{\n template\n void printArray(T *h_vec,uint size){\n printf(\"\\n\");\n for(int i = 0; i < size; i++){\n std::cout <\n void printCudaArray(T *d_vec,uint size){\n T* h_vec = (T *) std::malloc(size * sizeof(T));\n cudaMemcpy(h_vec, d_vec, size*sizeof(T), cudaMemcpyDeviceToHost);\n printArray(h_vec, size);\n free(h_vec);\n }\n}//end namespace \n\n" } ]