[ { "path": ".astylerc", "content": "# astyle -n -r \"benchmark/*.h,*.cpp\" \"src/*.h,*.cpp\" \"tests/*.h,*.cpp\" \"tools/*.h,*.cpp\" \"examples/*.h,*.cpp\"\n\n# brace style\n--style=allman\n\n# tab\n--attach-namespaces\n--attach-extern-c\n--attach-closing-while\n\n# indentation\n--indent-preproc-define\n--indent-col1-comments\n--min-conditional-indent=0\n--max-continuation-indent=120\n\n# padding\n--pad-oper\n--pad-comma\n--pad-header\n--align-pointer=type\n--align-reference=type\n\n# formatting\n--break-closing-braces\n--attach-return-type\n--attach-return-type-decl\n--keep-one-line-blocks\n--keep-one-line-statements\n--convert-tabs\n--max-code-length=200\n--mode=c\n\n# other\n--lineend=linux\n" }, { "path": ".clang-format", "content": "# find src/ tools/ tests/ examples/ benchmark/ -type f -name '*.c' -o -name '*.cpp' -o -name '*.h' | xargs -i clang-format -i {}\n\n# need clang-format >= 10.0\n\nAccessModifierOffset: -4\nAlignAfterOpenBracket: Align\nAlignConsecutiveAssignments: false\n# AlignConsecutiveBitFields: true\nAlignConsecutiveDeclarations: false\nAlignConsecutiveMacros: true\nAlignEscapedNewlines: Left\n# AlignOperands: AlignAfterOperator\nAlignTrailingComments: true\nAllowAllArgumentsOnNextLine: true\nAllowAllConstructorInitializersOnNextLine: true\nAllowAllParametersOfDeclarationOnNextLine: true\nAllowShortBlocksOnASingleLine: Always\nAllowShortCaseLabelsOnASingleLine: true\n# AllowShortEnumsOnASingleLine: true\nAllowShortFunctionsOnASingleLine: None\nAllowShortIfStatementsOnASingleLine: WithoutElse\nAllowShortLambdasOnASingleLine: All\nAllowShortLoopsOnASingleLine: true\nAlwaysBreakAfterReturnType: None\nAlwaysBreakBeforeMultilineStrings: false\nAlwaysBreakTemplateDeclarations: Yes\nBinPackArguments: true\nBinPackParameters: true\nBraceWrapping:\n AfterCaseLabel: true\n AfterClass: true\n AfterControlStatement: Always\n AfterEnum: true\n AfterFunction: true\n AfterNamespace: false\n AfterObjCDeclaration: false\n AfterStruct: true\n AfterUnion: true\n AfterExternBlock: false\n BeforeCatch: true\n BeforeElse: true\n# BeforeLambdaBody: false\n# BeforeWhile: false\n IndentBraces: false\n SplitEmptyFunction: true\n SplitEmptyRecord: true\n SplitEmptyNamespace: false\nBreakAfterJavaFieldAnnotations: true\nBreakBeforeBinaryOperators: All\nBreakBeforeBraces: Custom\nBreakBeforeTernaryOperators: true\nBreakConstructorInitializers: BeforeColon\nBreakInheritanceList: BeforeColon\nBreakStringLiterals: false\nColumnLimit: 0\n# CommentPragmas:\nCompactNamespaces: false\nConstructorInitializerAllOnOneLineOrOnePerLine: true\nConstructorInitializerIndentWidth: 4\nContinuationIndentWidth: 4\nCpp11BracedListStyle: true\nDeriveLineEnding: false\nDerivePointerAlignment: false\n# DisableFormat:\n# ExperimentalAutoDetectBinPacking:\nFixNamespaceComments: true\n# ForEachMacros:\nIncludeBlocks: Regroup\n# IncludeCategories:\n# IncludeIsMainRegex:\n# IncludeIsMainSourceRegex:\n# IndentCaseBlocks: false\nIndentCaseLabels: false\n# IndentExternBlock: NoIndent\nIndentGotoLabels: false\nIndentPPDirectives: None\nIndentWidth: 4\n# IndentWrappedFunctionNames: 4\n# InsertTrailingCommas: None\n# JavaImportGroups:\n# JavaScriptQuotes\n# JavaScriptWrapImports:\nKeepEmptyLinesAtTheStartOfBlocks: false\nLanguage: Cpp\n# MacroBlockBegin:\n# MacroBlockEnd:\nMaxEmptyLinesToKeep: 1\nNamespaceIndentation: None\n# NamespaceMacros:\n# ObjCBinPackProtocolList:\n# ObjCBlockIndentWidth:\n# ObjCBreakBeforeNestedBlockParam:\n# ObjCSpaceAfterProperty:\n# ObjCSpaceBeforeProtocolList:\n# PenaltyBreakAssignment:\n# PenaltyBreakBeforeFirstCallParameter:\n# PenaltyBreakComment:\n# PenaltyBreakFirstLessLess:\n# PenaltyBreakString:\n# PenaltyBreakTemplateDeclaration:\n# PenaltyExcessCharacter:\n# PenaltyReturnTypeOnItsOwnLine:\nPointerAlignment: Left\n# RawStringFormats:\nReflowComments: false\nSortIncludes: false\nSortUsingDeclarations: true\nSpaceAfterCStyleCast: false\nSpaceAfterLogicalNot: false\nSpaceAfterTemplateKeyword: false\nSpaceBeforeAssignmentOperators: true\nSpaceBeforeCpp11BracedList: false\nSpaceBeforeCtorInitializerColon: true\nSpaceBeforeInheritanceColon: true\nSpaceBeforeParens: ControlStatements\nSpaceBeforeRangeBasedForLoopColon: true\nSpaceBeforeSquareBrackets: false\nSpaceInEmptyBlock: false\nSpaceInEmptyParentheses: false\nSpacesBeforeTrailingComments: 1\nSpacesInAngles: false\nSpacesInCStyleCastParentheses: false\nSpacesInConditionalStatement: false\nSpacesInContainerLiterals: false\nSpacesInParentheses: false\nSpacesInSquareBrackets: false\nStandard: c++03\n#StatementMacros:\nTabWidth: 4\n# TypenameMacros:\nUseCRLF: false\nUseTab: Never\n" }, { "path": ".gitattributes", "content": "*.comp linguist-language=GLSL\n" }, { "path": ".github/ISSUE_TEMPLATE/bug.md", "content": "---\nname: \"\\U0001F41B bug issue\"\nabout: submit a bug report +_+\n---\n\n## error log | 日志或报错信息 | ログ\n\n## context | 编译/运行环境 | バックグラウンド\n\n## how to reproduce | 复现步骤 | 再現方法\n1.\n2.\n3.\n\n## more | 其他 | その他\n" }, { "path": ".github/ISSUE_TEMPLATE/model-convert.md", "content": "---\nname: \"\\U0001F6B8 model convert issue\"\nabout: \"Life is Short, Use pnnx and convertmodel.com\"\n---\n\n## error log | 日志或报错信息 | ログ\n\n## model | 模型 | モデル\n1. original model\n\n## how to reproduce | 复现步骤 | 再現方法\n1.\n2.\n3.\n" }, { "path": ".github/ISSUE_TEMPLATE/others.md", "content": "---\nname: \"\\U0001F4DD others\"\nabout: discussion, suggestion and question\n---\n\n## detail | 详细描述 | 詳細な説明\n" }, { "path": ".github/ISSUE_TEMPLATE/quantization.md", "content": "---\nname: \"\\U0001F4C8 quantization\"\nabout: best wishes for your low bit quantization has a low accuracy loss...\\(^▽^)/...2333... \n---\n\n## expectation | 诉求 | 期待する\n1. speed \n2. precision\n\n## model | 模型 | モデル\n1. model.param and model.bin\n\n## detail | 详细描述 | 詳細な説明\n" }, { "path": ".github/dependabot.yml", "content": "version: 2\nupdates:\n - package-ecosystem: \"github-actions\"\n directory: \"/\"\n schedule:\n interval: \"daily\"\n" }, { "path": ".github/labeler.yml", "content": "cmake:\n- changed-files:\n - any-glob-to-any-file: ['cmake/**', 'toolchains/**']\n\ndoc: \n- changed-files:\n - any-glob-to-any-file: docs/**\n\npython: \n- changed-files:\n - any-glob-to-any-file: python/**\n\nexample: \n- changed-files:\n - any-glob-to-any-file: examples/**\n\ntest: \n- changed-files:\n - any-glob-to-any-file: tests/**\n\ntool: \n- changed-files:\n - any-glob-to-any-file: tools/**\npnnx: \n- changed-files:\n - any-glob-to-any-file: tools/pnnx/**\n\ncore: \n- changed-files:\n - any-glob-to-any-file: src/*\nlayer: \n- changed-files:\n - any-glob-to-any-file: src/layer/*\n\narm: \n- changed-files:\n - any-glob-to-any-file: src/layer/arm/**\nloongarch: \n- changed-files:\n - any-glob-to-any-file: src/layer/loongarch/**\nmips: \n- changed-files:\n - any-glob-to-any-file: src/layer/mips/**\nriscv: \n- changed-files:\n - any-glob-to-any-file: src/layer/riscv/**\nvulkan: \n- changed-files:\n - any-glob-to-any-file: src/layer/vulkan/**\nx86: \n- changed-files:\n - any-glob-to-any-file: src/layer/x86/**\n" }, { "path": ".github/workflows/android.yml", "content": "name: android\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/android.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/riscv/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/android.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/riscv/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\nconcurrency:\n group: android-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: ubuntu-latest\n env:\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_LATEST_HOME/build/cmake/android.toolchain.cmake \\\n -DANDROID_PLATFORM=android-21 \\\n -DANDROID_SUPPORT_FLEXIBLE_PAGE_SIZES=ON \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: armeabi-v7a\n run: |\n mkdir build-armeabi-v7a && cd build-armeabi-v7a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON ..\n cmake --build . -j $(nproc)\n - name: arm64-v8a\n run: |\n mkdir build-arm64-v8a && cd build-arm64-v8a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"arm64-v8a\" ..\n cmake --build . -j $(nproc)\n - name: x86\n run: |\n mkdir build-x86 && cd build-x86\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"x86\" ..\n cmake --build . -j $(nproc)\n - name: x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"x86_64\" ..\n cmake --build . -j $(nproc)\n - name: riscv64\n run: |\n mkdir build-riscv64 && cd build-riscv64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"riscv64\" ..\n cmake --build . -j $(nproc)\n\n - name: armeabi-v7a-shared\n run: |\n mkdir build-armeabi-v7a-shared && cd build-armeabi-v7a-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: arm64-v8a-shared\n run: |\n mkdir build-arm64-v8a-shared && cd build-arm64-v8a-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"arm64-v8a\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: x86-shared\n run: |\n mkdir build-x86-shared && cd build-x86-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"x86\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: x86_64-shared\n run: |\n mkdir build-x86_64-shared && cd build-x86_64-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"x86_64\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: riscv64-shared\n run: |\n mkdir build-riscv64-shared && cd build-riscv64-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"riscv64\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n\n ndk-r16b:\n runs-on: ubuntu-latest\n env:\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=$GITHUB_WORKSPACE/android-ndk-r16b/build/cmake/android.toolchain.cmake \\\n -DANDROID_PLATFORM=android-21 \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: ndk-r16b\n env:\n DEBIAN_FRONTEND: noninteractive\n run: |\n pushd /usr/lib/x86_64-linux-gnu/\n sudo ln -s libncurses.so.6 libncurses.so.5\n sudo ln -s libtinfo.so.6 libtinfo.so.5\n popd\n wget -q https://dl.google.com/android/repository/android-ndk-r16b-linux-x86_64.zip -O $GITHUB_WORKSPACE/android-ndk-r16b-linux-x86_64.zip\n cd $GITHUB_WORKSPACE && unzip -q android-ndk-r16b-linux-x86_64.zip\n\n - name: armeabi-v7a\n run: |\n mkdir build-armeabi-v7a && cd build-armeabi-v7a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON ..\n cmake --build . -j $(nproc)\n - name: armeabi-v7a-no-neon\n run: |\n mkdir build-armeabi-v7a-no-neon && cd build-armeabi-v7a-no-neon\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=OFF ..\n cmake --build . -j $(nproc)\n - name: arm64-v8a\n run: |\n mkdir build-arm64-v8a && cd build-arm64-v8a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"arm64-v8a\" ..\n cmake --build . -j $(nproc)\n\n - name: armeabi-v7a-shared\n run: |\n mkdir build-armeabi-v7a-shared && cd build-armeabi-v7a-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: armeabi-v7a-no-neon-shared\n run: |\n mkdir build-armeabi-v7a-no-neon-shared && cd build-armeabi-v7a-no-neon-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=OFF -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: arm64-v8a-shared\n run: |\n mkdir build-arm64-v8a-shared && cd build-arm64-v8a-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"arm64-v8a\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n" }, { "path": ".github/workflows/code-format-msg.yml", "content": "name: code-format-msg\n\non:\n workflow_run:\n workflows: [code-format]\n types: [completed]\n\nconcurrency:\n group: code-format-msg-${{ github.head_ref || github.run_id }}\n\npermissions:\n contents: read\n pull-requests: write\n\njobs:\n pr-context:\n name: acquire-pr-context\n runs-on: ubuntu-latest\n outputs:\n PR_HEADSHA: ${{ steps.set-pr-context.outputs.head-sha }}\n PR_NUMBER: ${{ steps.set-pr-context.outputs.number }}\n if: ${{ github.event.workflow_run.event == 'pull_request' }}\n steps:\n - name: get-pr-context\n id: set-pr-context\n env:\n GH_TOKEN: ${{ github.token }}\n PR_TARGET_REPO: ${{ github.repository }}\n PR_BRANCH: |-\n ${{\n (github.event.workflow_run.head_repository.owner.login != github.event.workflow_run.repository.owner.login)\n && format('{0}:{1}', github.event.workflow_run.head_repository.owner.login, github.event.workflow_run.head_branch)\n || github.event.workflow_run.head_branch\n }}\n run: |\n gh pr view --repo \"${PR_TARGET_REPO}\" \"${PR_BRANCH}\" \\\n --json 'number,headRefOid' \\\n --jq '\"number=\\(.number)\\nhead-sha=\\(.headRefOid)\"' \\\n >> $GITHUB_OUTPUT\n\n remove-comment-if-success:\n if: ${{ github.event.workflow_run.conclusion == 'success' }}\n runs-on: ubuntu-latest\n needs: [pr-context]\n env:\n PR_HEADSHA: ${{ needs.pr-context.outputs.PR_HEADSHA }}\n PR_NUMBER: ${{ needs.pr-context.outputs.PR_NUMBER }}\n steps:\n - name: Remove existing \"format check failed\" comment\n uses: actions/github-script@v8\n with:\n script: |\n const owner = context.repo.owner;\n const repo = context.repo.repo;\n const { data: comments } = await github.rest.issues.listComments({\n owner,\n repo,\n issue_number: ${{ env.PR_NUMBER }},\n });\n\n const targetComment = comments.find(comment =>\n comment.body.includes(\"Please enable github action in **YOUR FORKED REPO** to make code-format workflow work\")\n );\n\n if (targetComment) {\n await github.rest.issues.deleteComment({\n owner,\n repo,\n comment_id: targetComment.id,\n });\n console.log(\"Removed existing code-format failure comment.\");\n } else {\n console.log(\"No existing format failure comment to remove.\");\n }\n\n post-comment-if-failure:\n if: ${{ github.event.workflow_run.conclusion == 'failure' }}\n runs-on: ubuntu-latest\n needs: [pr-context]\n env:\n PR_HEADSHA: ${{ needs.pr-context.outputs.PR_HEADSHA }}\n PR_NUMBER: ${{ needs.pr-context.outputs.PR_NUMBER }}\n steps:\n - name: Post comment on failed code-format if not existing\n uses: actions/github-script@v8\n with:\n script: |\n const owner = context.repo.owner;\n const repo = context.repo.repo;\n const { data: comments } = await github.rest.issues.listComments({\n owner,\n repo,\n issue_number: ${{ env.PR_NUMBER }},\n });\n\n const existingComment = comments.find(comment =>\n comment.body.includes(\"Please enable github action in **YOUR FORKED REPO** to make code-format workflow work\")\n );\n\n if (existingComment) {\n console.log(\"A code-format failure comment already exists.\");\n } else {\n await github.rest.issues.createComment({\n owner,\n repo,\n issue_number: ${{ env.PR_NUMBER }},\n body: \"Please enable github action in **YOUR FORKED REPO** to make code-format workflow work\",\n });\n console.log(\"Created code-format failure comment.\");\n }\n" }, { "path": ".github/workflows/code-format.yml", "content": "name: code-format\n\non: [push, pull_request]\n\nconcurrency:\n group: code-format-${{ github.ref }}\n cancel-in-progress: true\n\npermissions:\n contents: write\n\njobs:\n code-format:\n runs-on: ubuntu-latest\n container: ubuntu:20.04\n steps:\n - name: astyle\n run: |\n export DEBIAN_FRONTEND=noninteractive\n apt-get update -y\n apt-get install -y astyle git\n\n - uses: actions/checkout@v6\n\n - name: cache-clang-format\n id: cache-clang-format\n uses: actions/cache@v5\n with:\n path: clang-format-install\n key: clang-format-install-5\n - name: clang-format\n if: steps.cache-clang-format.outputs.cache-hit != 'true'\n run: |\n export DEBIAN_FRONTEND=noninteractive\n apt-get update -y\n apt-get install -y build-essential wget curl cmake unzip zip python3-pip\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-10.0.1/llvm-project-10.0.1.tar.xz\n tar -xf llvm-project-10.0.1.tar.xz\n cd llvm-project-10.0.1\n mkdir build\n cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DLLVM_ENABLE_PROJECTS=\"clang\" -DLLVM_TARGETS_TO_BUILD=\"\" -DLLVM_INCLUDE_EXAMPLES=OFF -DLLVM_INCLUDE_TESTS=OFF -DLLVM_INCLUDE_DOCS=OFF ../llvm/\n make -j4 clang-format\n mkdir $GITHUB_WORKSPACE/clang-format-install\n cp -r bin/clang-format $GITHUB_WORKSPACE/clang-format-install\n cd ../../\n rm -rf llvm-project-10.0.1\n rm llvm-project-10.0.1.tar.xz\n\n - name: cache-clang-format-21\n id: cache-clang-format-21\n uses: actions/cache@v5\n with:\n path: clang-format-21-install\n key: clang-format-21-install\n - name: clang-format-21\n if: steps.cache-clang-format-21.outputs.cache-hit != 'true'\n run: |\n export DEBIAN_FRONTEND=noninteractive\n apt-get update -y\n apt-get install -y build-essential wget curl cmake unzip zip python3-pip\n pip install cmake\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-21.1.8/llvm-project-21.1.8.src.tar.xz\n tar -xf llvm-project-21.1.8.src.tar.xz\n cd llvm-project-21.1.8.src\n mkdir build\n cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=OFF -DLLVM_ENABLE_PROJECTS=\"clang\" -DLLVM_TARGETS_TO_BUILD=\"\" -DLLVM_INCLUDE_EXAMPLES=OFF -DLLVM_INCLUDE_TESTS=OFF -DLLVM_INCLUDE_DOCS=OFF ../llvm/\n make -j4 clang-format\n mkdir $GITHUB_WORKSPACE/clang-format-21-install\n cp -r bin/clang-format $GITHUB_WORKSPACE/clang-format-21-install\n cd ../../\n rm -rf llvm-project-21.1.8.src\n rm llvm-project-21.1.8.src.tar.xz\n\n - name: code-format\n run: |\n mv $GITHUB_WORKSPACE/clang-format-install/clang-format /usr/local/bin/clang-format\n rm -rf $GITHUB_WORKSPACE/clang-format-install\n sh codeformat.sh\n\n - name: code-format-glsl\n run: |\n mv $GITHUB_WORKSPACE/clang-format-21-install/clang-format /usr/local/bin/clang-format-21\n rm -rf $GITHUB_WORKSPACE/clang-format-21-install\n cd src/layer/vulkan/shader\n find . -type f -name '*.comp' | xargs -i clang-format-21 -i -assume-filename=main.cpp {}\n\n - name: configure-git-safe-directory\n run: git config --global --add safe.directory /__w/ncnn/ncnn\n\n - uses: stefanzweifel/git-auto-commit-action@v7\n with:\n commit_message: apply code-format changes\n\n - name: restore-clang-format-cache\n run: |\n mkdir $GITHUB_WORKSPACE/clang-format-install\n cp -r /usr/local/bin/clang-format $GITHUB_WORKSPACE/clang-format-install\n mkdir $GITHUB_WORKSPACE/clang-format-21-install\n cp -r /usr/local/bin/clang-format-21 $GITHUB_WORKSPACE/clang-format-21-install/clang-format\n" }, { "path": ".github/workflows/codeql-analysis.yml", "content": "# For most projects, this workflow file will not need changing; you simply need\n# to commit it to your repository.\n#\n# You may wish to alter this file to override the set of languages analyzed,\n# or to provide custom queries or build logic.\nname: \"CodeQL\"\n\non:\n push:\n branches: [master]\n paths-ignore: ['**.md']\n pull_request:\n # The branches below must be a subset of the branches above\n branches: [master]\n paths-ignore: ['**.md']\n schedule:\n - cron: '0 20 * * 4'\n\nconcurrency:\n group: CodeQL-${{ github.ref }}\n cancel-in-progress: true\n\npermissions:\n contents: read\n\njobs:\n analyze:\n permissions:\n actions: read # for github/codeql-action/init to get workflow details\n contents: read # for actions/checkout to fetch code\n security-events: write # for github/codeql-action/autobuild to send a status report\n name: Analyze\n runs-on: ubuntu-latest\n\n strategy:\n fail-fast: false\n matrix:\n # Override automatic language detection by changing the below list\n # Supported options are ['csharp', 'cpp', 'go', 'java', 'javascript', 'python']\n language: ['cpp']\n # Learn more...\n # https://docs.github.com/en/github/finding-security-vulnerabilities-and-errors-in-your-code/configuring-code-scanning#overriding-automatic-language-detection\n\n steps:\n - name: Checkout repository\n uses: actions/checkout@v6\n with:\n # We must fetch at least the immediate parents so that if this is\n # a pull request then we can checkout the head.\n fetch-depth: 2\n\n # If this run was triggered by a pull request event, then checkout\n # the head of the pull request instead of the merge commit.\n - run: git checkout HEAD^2\n if: ${{ github.event_name == 'pull_request' }}\n\n # Initializes the CodeQL tools for scanning.\n - name: Initialize CodeQL\n uses: github/codeql-action/init@v4\n with:\n languages: ${{ matrix.language }}\n # If you wish to specify custom queries, you can do so here or in a config file.\n # By default, queries listed here will override any specified in a config file. \n # Prefix the list here with \"+\" to use these queries and those in the config file.\n # queries: ./path/to/local/query, your-org/your-repo/queries@main\n\n # Autobuild attempts to build any compiled languages (C/C++, C#, or Java).\n # If this step fails, then you should remove it and run the build manually (see below)\n - name: Autobuild\n uses: github/codeql-action/autobuild@v4\n\n # ℹ️ Command-line programs to run using the OS shell.\n # 📚 https://git.io/JvXDl\n\n # ✏️ If the Autobuild fails above, remove it and uncomment the following three lines\n # and modify them (or add more) to build your code if your project\n # uses a compiled language\n\n #- run: |\n # make bootstrap\n # make release\n\n - name: Perform CodeQL Analysis\n uses: github/codeql-action/analyze@v4\n" }, { "path": ".github/workflows/compare-binary-size-pr-comment.yml", "content": "name: compare-binary-size-pr-comment\non:\n workflow_run:\n workflows: [\"compare-binary-size\"]\n types:\n - completed\n\npermissions:\n actions: read\n contents: read\n pull-requests: write\n\njobs:\n pr-comment:\n runs-on: ubuntu-latest\n steps:\n - name: Setup tools\n run: |\n sudo apt-get update\n sudo apt-get install -y jq unzip\n\n - name: Ensure workflow_run is for a PR\n id: validate\n run: |\n # Use the event payload file provided by GitHub Actions directly\n echo \"Using event payload from: $GITHUB_EVENT_PATH\"\n echo \"Event file size: $(wc -c < \"$GITHUB_EVENT_PATH\") bytes\"\n\n # Safely compute number of associated PRs (use // 0 to default if missing)\n PR_COUNT=$(jq -r '.workflow_run.pull_requests | length // 0' \"$GITHUB_EVENT_PATH\")\n echo \"Associated pull_request count: $PR_COUNT\"\n\n if [ \"$PR_COUNT\" -eq 0 ]; then\n echo \"No pull_request associated with this workflow_run; nothing to do.\"\n echo \"skip=true\" >> $GITHUB_OUTPUT\n exit 0\n fi\n\n echo \"skip=false\" >> $GITHUB_OUTPUT\n\n - name: Download artifact zip for this run\n if: steps.validate.outputs.skip != 'true'\n env:\n RUN_ID: ${{ github.event.workflow_run.id }}\n OWNER: ${{ github.repository_owner }}\n REPO: ${{ github.repository }}\n TOKEN: ${{ secrets.COMMENTER_PAT }}\n ART_NAME: \"compare-binary-size.md\"\n run: |\n echo \"Listing artifacts for run $RUN_ID\"\n API=\"https://api.github.com/repos/$OWNER/${REPO#*/}/actions/runs/$RUN_ID/artifacts\"\n\n # Save artifact list to a file (avoid pipe/echo issues)\n curl -s -H \"Authorization: token $TOKEN\" \"$API\" -o /tmp/art_list.json\n echo \"Art list size: $(wc -c < /tmp/art_list.json) bytes\"\n if ! jq . /tmp/art_list.json; then\n echo \"Failed to parse /tmp/art_list.json with jq; aborting for safety.\"\n exit 1\n fi\n\n # find artifact archive_download_url by name (first match)\n ARCHIVE_URL=$(jq -r --arg name \"$ART_NAME\" '.artifacts[] | select(.name==$name) | .archive_download_url' /tmp/art_list.json | head -n1)\n if [ -z \"$ARCHIVE_URL\" ] || [ \"$ARCHIVE_URL\" = \"null\" ]; then\n echo \"Artifact named '$ART_NAME' not found for run $RUN_ID. Exiting.\"\n exit 0\n fi\n echo \"Downloading artifact from: $ARCHIVE_URL\"\n\n # download and unzip to temp dir\n mkdir -p /tmp/artifact_contents\n curl -L -H \"Authorization: token $TOKEN\" -o /tmp/artifact.zip \"$ARCHIVE_URL\"\n if ! unzip -q /tmp/artifact.zip -d /tmp/artifact_contents; then\n echo \"Failed to unzip /tmp/artifact.zip\"; exit 1\n fi\n ls -la /tmp/artifact_contents\n\n - name: Read compare-binary-size.md content\n if: steps.validate.outputs.skip != 'true'\n id: read\n run: |\n # find file inside artifact_contents\n FILE=$(find /tmp/artifact_contents -type f -name \"compare-binary-size.md\" | head -n1 || true)\n if [ -z \"$FILE\" ]; then\n # If artifact name matched but internal filename differs, try any .md\n FILE=$(find /tmp/artifact_contents -type f -name \"*.md\" | head -n1 || true)\n fi\n\n if [ -z \"$FILE\" ]; then\n echo \"compare_content<> $GITHUB_OUTPUT\n echo \"No compare-binary-size.md found in artifact.\" >> $GITHUB_OUTPUT\n echo \"EOF\" >> $GITHUB_OUTPUT\n else\n # Truncate to avoid overly long comments (adjust lines as needed)\n head -n 1000 \"$FILE\" > /tmp/compare-truncated.md || true\n echo \"compare_content<> $GITHUB_OUTPUT\n cat /tmp/compare-truncated.md >> $GITHUB_OUTPUT\n echo \"EOF\" >> $GITHUB_OUTPUT\n fi\n\n - name: Post or update PR comment via actions/github-script\n if: steps.validate.outputs.skip != 'true'\n uses: actions/github-script@v8\n with:\n github-token: ${{ secrets.COMMENTER_PAT }}\n script: |\n const pr = context.payload.workflow_run.pull_requests[0];\n if (!pr) {\n core.info(\"No pull request found in workflow_run payload; skipping.\");\n return;\n }\n\n const owner = context.repo.owner;\n const repo = context.repo.repo;\n const issue_number = pr.number;\n const marker = '';\n\n // Read the compare content from env (set in previous step outputs)\n const compare = process.env.COMPARE_CONTENT || \"\";\n\n const body = `${marker}\\n**Binary size comparison** (from artifact)\\n\\n\\`\\`\\`markdown\\n${compare}\\n\\`\\`\\``;\n\n // List existing comments and find our bot comment (by marker)\n const { data: comments } = await github.rest.issues.listComments({\n owner,\n repo,\n issue_number,\n per_page: 100\n });\n\n const existing = comments.find(c => c.body && c.body.includes(marker));\n\n if (existing) {\n await github.rest.issues.updateComment({\n owner,\n repo,\n comment_id: existing.id,\n body\n });\n core.info(`Updated comment id=${existing.id}`);\n } else {\n await github.rest.issues.createComment({\n owner,\n repo,\n issue_number,\n body\n });\n core.info(\"Created new comment\");\n }\n env:\n # pass the content from previous step into the github-script environment\n COMPARE_CONTENT: ${{ steps.read.outputs.compare_content }}\n" }, { "path": ".github/workflows/compare-binary-size.yml", "content": "name: compare-binary-size\non:\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/compare-binary-size.yml'\n - 'toolchains/**'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/**'\n - 'glslang'\n\nconcurrency:\n group: compare-binary-size-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n actions: read\n\njobs:\n compare-size:\n runs-on: ubuntu-latest\n steps:\n - name: checkout-pr-branch\n uses: actions/checkout@v6\n with:\n ref: refs/pull/${{ github.event.pull_request.number }}/merge\n submodules: true\n path: pr\n\n - name: checkout-base-branch\n uses: actions/checkout@v6\n with:\n ref: ${{ github.event.pull_request.base.ref }}\n repository: ${{ github.event.pull_request.base.repo.full_name }}\n submodules: true\n path: base\n\n - name: install-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-arm-linux-gnueabihf g++-aarch64-linux-gnu\n\n - name: compare-sizes\n env:\n COMMON_CMAKE_ARGS: -DNCNN_SHARED_LIB=ON -DNCNN_VULKAN=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF\n run: |\n # define target architectures\n archs=(\"x86_64\" \"armhf\" \"aarch64\")\n\n # generate table\n echo \"The binary size change of libncnn.so (bytes)\" >> compare-binary-size.md\n echo \"| architecture | base size | pr size | difference |\" >> compare-binary-size.md\n echo \"|--------------|-----------|---------|------------|\" >> compare-binary-size.md\n\n for arch in \"${archs[@]}\"; do\n\n mkdir -p pr/build_$arch\n pushd pr/build_$arch\n if [ \"$arch\" = \"armhf\" ]; then\n cmake ${{env.COMMON_CMAKE_ARGS}} -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..\n elif [ \"$arch\" = \"aarch64\" ]; then\n cmake ${{env.COMMON_CMAKE_ARGS}} -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..\n else\n cmake ${{env.COMMON_CMAKE_ARGS}} ..\n fi\n cmake --build . -j $(nproc)\n PR_SIZE=$(stat -c%s $(readlink -f src/libncnn.so))\n popd\n\n mkdir -p base/build_$arch\n pushd base/build_$arch\n if [ \"$arch\" = \"armhf\" ]; then\n cmake ${{env.COMMON_CMAKE_ARGS}} -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..\n elif [ \"$arch\" = \"aarch64\" ]; then\n cmake ${{env.COMMON_CMAKE_ARGS}} -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..\n else\n cmake ${{env.COMMON_CMAKE_ARGS}} ..\n fi\n cmake --build . -j $(nproc)\n BASE_SIZE=$(stat -c%s $(readlink -f src/libncnn.so))\n popd\n\n DIFF=$(($PR_SIZE - $BASE_SIZE))\n if [ $DIFF -gt 0 ]; then\n DIFF_STR=\"+$DIFF :warning:\"\n else\n DIFF_STR=\"$DIFF :kissing_heart:\"\n fi\n\n echo \"| $arch | $BASE_SIZE | $PR_SIZE | $DIFF_STR |\" >> compare-binary-size.md\n done\n\n cat compare-binary-size.md\n\n - name: upload-compare-binary-size-md\n uses: actions/upload-artifact@v6\n with:\n name: compare-binary-size.md\n path: compare-binary-size.md\n" }, { "path": ".github/workflows/elf-riscv32.yml", "content": "name: elf-riscv32\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/elf-riscv32.yml'\n - 'toolchains/riscv32-unknown-elf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/elf-riscv32.yml'\n - 'toolchains/riscv32-unknown-elf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\nconcurrency:\n group: elf-riscv32-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n rv32gc:\n runs-on: [self-hosted, linux, centos]\n steps:\n - uses: actions/checkout@v6\n\n #- name: riscv-gnu-toolchain\n #run: |\n #wget -c https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2025.01.20/riscv32-elf-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz\n #tar -xf riscv32-elf-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz\n #mv riscv riscv32-elf\n\n #- name: checkout-riscv-pk\n #uses: actions/checkout@v6\n #with:\n #repository: riscv/riscv-pk\n #path: riscv-pk\n #ref: d8659a4e8e888bdc9caf840ad17bfe83239b1d64\n #- name: riscv-pk\n #run: |\n #cd riscv-pk\n #mkdir build && cd build\n #export PATH=$GITHUB_WORKSPACE/riscv32-elf/bin:$PATH\n #export CFLAGS=\"-O3\"\n #export CXXFLAGS=\"-O3\"\n #../configure --prefix=$GITHUB_WORKSPACE/riscv32-elf --with-arch=rv32gc_zicsr_zifencei --host=riscv32-unknown-elf --with-abi=ilp32d\n #make -j4\n #make install\n\n #- name: checkout-riscv-isa-sim\n #uses: actions/checkout@v6\n #with:\n #repository: riscv-software-src/riscv-isa-sim\n #path: riscv-isa-sim\n #ref: 5ef9a61f5fecdb9bf77da155172c8018ce820308\n #- name: riscv-isa-sim\n #run: |\n #cd riscv-isa-sim\n #mkdir build && cd build\n #export PATH=$GITHUB_WORKSPACE/riscv32-elf/bin:$PATH\n #export CFLAGS=\"-O3\"\n #export CXXFLAGS=\"-O3\"\n #../configure --prefix=$GITHUB_WORKSPACE/riscv32-elf\n #make -j4\n #make install\n\n #- name: riscv-strip-install\n #run: find $GITHUB_WORKSPACE/riscv32-elf -type f | xargs -i strip -g {} || true\n\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/riscv32-elf\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv32-unknown-elf.toolchain.cmake -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=OFF -DNCNN_ZFH=OFF -DNCNN_ZVFH=OFF ..\n cmake --build . -j 4\n\n - name: test\n run: |\n export PATH=/data/action/osd/riscv32-elf/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=spike TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"--isa=rv32gc;/data/action/osd/riscv32-elf/riscv32-unknown-elf/bin/pk\" ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/elf-riscv64.yml", "content": "name: elf-riscv64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/elf-riscv64.yml'\n - 'toolchains/riscv64-unknown-elf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/elf-riscv64.yml'\n - 'toolchains/riscv64-unknown-elf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\nconcurrency:\n group: elf-riscv64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n rv64gc:\n runs-on: [self-hosted, linux, centos]\n steps:\n - uses: actions/checkout@v6\n\n #- name: riscv-gnu-toolchain\n #run: |\n #wget -c https://github.com/riscv-collab/riscv-gnu-toolchain/releases/download/2025.01.20/riscv64-elf-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz\n #tar -xf riscv64-elf-ubuntu-22.04-gcc-nightly-2025.01.20-nightly.tar.xz\n #mv riscv riscv64-elf\n\n #- name: checkout-riscv-pk\n #uses: actions/checkout@v6\n #with:\n #repository: riscv/riscv-pk\n #path: riscv-pk\n #ref: d8659a4e8e888bdc9caf840ad17bfe83239b1d64\n #- name: riscv-pk\n #run: |\n #cd riscv-pk\n #mkdir build && cd build\n #export PATH=$GITHUB_WORKSPACE/riscv64-elf/bin:$PATH\n #export CFLAGS=\"-O3\"\n #export CXXFLAGS=\"-O3\"\n #../configure --prefix=$GITHUB_WORKSPACE/riscv64-elf --with-arch=rv64gc_zicsr_zifencei --host=riscv64-unknown-elf --with-abi=lp64d\n #make -j4\n #make install\n\n #- name: checkout-riscv-isa-sim\n #uses: actions/checkout@v6\n #with:\n #repository: riscv-software-src/riscv-isa-sim\n #path: riscv-isa-sim\n #ref: 5ef9a61f5fecdb9bf77da155172c8018ce820308\n #- name: riscv-isa-sim\n #run: |\n #cd riscv-isa-sim\n #mkdir build && cd build\n #export PATH=$GITHUB_WORKSPACE/riscv64-elf/bin:$PATH\n #export CFLAGS=\"-O3\"\n #export CXXFLAGS=\"-O3\"\n #../configure --prefix=$GITHUB_WORKSPACE/riscv64-elf\n #make -j4\n #make install\n\n #- name: riscv-strip-install\n #run: find $GITHUB_WORKSPACE/riscv64-elf -type f | xargs -i strip -g {} || true\n\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/riscv64-elf\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-unknown-elf.toolchain.cmake -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_XTHEADVECTOR=OFF ..\n cmake --build . -j 4\n\n - name: test\n run: |\n export PATH=/data/action/osd/riscv64-elf/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=spike TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"--isa=rv64gc;/data/action/osd/riscv64-elf/riscv64-unknown-elf/bin/pk\" ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/esp32.yml", "content": "name: ESP32\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/esp32.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/esp32.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n\nconcurrency:\n group: esp32-${{ github.ref }}\n cancel-in-progress: true\n\npermissions:\n contents: read\n\njobs:\n build:\n name: ESP32\n runs-on: ubuntu-latest\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: Setup Python\n uses: actions/setup-python@v6\n with:\n python-version: '3.8'\n\n - name: Install dependencies\n run: |\n sudo apt-get update\n sudo apt-get install -y cmake ninja-build ccache\n \n - name: Checkout ESP-IDF\n uses: actions/checkout@v6\n with:\n repository: espressif/esp-idf\n path: esp-idf-install\n ref: release/v5.3\n \n - name: Install ESP-IDF\n run: |\n cd esp-idf-install\n git submodule update --init --recursive\n ./install.sh\n\n - name: Set environment and build NCNN for ESP32\n run: |\n source esp-idf-install/export.sh\n echo \"IDF_PATH=$IDF_PATH\" >> $GITHUB_ENV\n echo \"${IDF_PATH}/tools\" >> $GITHUB_PATH\n echo \"${IDF_PATH}/components\" >> $GITHUB_PATH\n mkdir -p build-esp32 && cd build-esp32\n cmake -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/esp32.toolchain.cmake\" -DCMAKE_BUILD_TYPE=Release -DNCNN_BUILD_EXAMPLES=OFF ..\n make -j 4\n make install\n" }, { "path": ".github/workflows/harmonyos.yml", "content": "name: harmonyos\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/harmonyos.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/harmonyos.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\nconcurrency:\n group: harmonyos-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: [self-hosted, linux, centos]\n\n env:\n OHOS_NDK_HOME: /data/action/osd/ohos-sdk/linux/native\n OHOS_NDK_CMAKE: /data/action/osd/ohos-sdk/linux/native/build-tools/cmake/bin/cmake\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=/data/action/osd/ohos-sdk/linux/native/build/cmake/ohos.toolchain.cmake \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DNCNN_SIMPLEOMP=ON \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n # - name: setup-sdk\n # run: |\n # cd /data/action/osd\n # wget -q https://repo.huaweicloud.com/harmonyos/os/4.1.1-Release/ohos-sdk-windows_linux-public.tar.gz\n # tar -xf ohos-sdk-windows_linux-public.tar.gz\n # cd ohos-sdk/linux\n # unzip -q native-linux-x64-4.1.7.8-Release.zip\n\n - name: armeabi-v7a\n run: |\n mkdir build-armeabi-v7a && cd build-armeabi-v7a\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"armeabi-v7a\" ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n - name: arm64-v8a\n run: |\n mkdir build-arm64-v8a && cd build-arm64-v8a\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"arm64-v8a\" ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n - name: x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"x86_64\" ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n\n - name: armeabi-v7a-shared\n run: |\n mkdir build-armeabi-v7a-shared && cd build-armeabi-v7a-shared\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"armeabi-v7a\" -DNCNN_SHARED_LIB=ON ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n - name: arm64-v8a-shared\n run: |\n mkdir build-arm64-v8a-shared && cd build-arm64-v8a-shared\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"arm64-v8a\" -DNCNN_SHARED_LIB=ON ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n - name: x86_64-shared\n run: |\n mkdir build-x86_64-shared && cd build-x86_64-shared\n ${{ env.OHOS_NDK_CMAKE }} ${{ env.NCNN_CMAKE_OPTIONS }} -DOHOS_ARCH=\"x86_64\" -DNCNN_SHARED_LIB=ON ..\n ${{ env.OHOS_NDK_CMAKE }} --build . -j 4\n" }, { "path": ".github/workflows/ios.yml", "content": "name: ios\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/ios.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/ios.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\nconcurrency:\n group: ios-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n IOS_DEPLOYMENT_TARGET: '13.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-ios-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=OS64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR64 -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATORARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/ios\n mkdir -p $GITHUB_WORKSPACE/openmp-install/ios-simulator\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install/ios\n mkdir -p $GITHUB_WORKSPACE/openmp-install/ios/lib\n cp openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a $GITHUB_WORKSPACE/openmp-install/ios/lib/libomp.a\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/include $GITHUB_WORKSPACE/openmp-install/ios-simulator\n mkdir -p $GITHUB_WORKSPACE/openmp-install/ios-simulator/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/ios-simulator/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/ios/include/* $DEVELOPER_DIR/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/ios/lib/libomp.a $DEVELOPER_DIR/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib\n\n sudo cp $GITHUB_WORKSPACE/openmp-install/ios-simulator/include/* $DEVELOPER_DIR/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/ios-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib\n\n - name: arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=OS64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n - name: simulator-x86_64\n run: |\n mkdir build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR64 -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n - name: simulator-arm64\n run: |\n mkdir build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATORARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n" }, { "path": ".github/workflows/labeler.yml", "content": "name: labeler\non: [pull_request_target]\n\npermissions:\n contents: read\n pull-requests: write\n\njobs:\n label:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/labeler@v6\n" }, { "path": ".github/workflows/linux-aarch64.yml", "content": "name: linux-aarch64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-aarch64.yml'\n - 'toolchains/aarch64-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-aarch64.yml'\n - 'toolchains/aarch64-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'tests/**'\nconcurrency:\n group: linux-aarch64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n aarch64-native:\n runs-on: ubuntu-24.04-arm\n steps:\n - uses: actions/checkout@v6\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test\n run: cd build && ctest --output-on-failure -j $(nproc)\n\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: cd build-noint8 && ctest --output-on-failure -j $(nproc)\n\n - name: build-simplestl-simplemath\n run: |\n mkdir build-simplestl-simplemath && cd build-simplestl-simplemath \n cmake -DNCNN_STDIO=ON -DNCNN_STRING=ON -DNCNN_SIMPLESTL=ON -DNCNN_SIMPLEMATH=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-simplestl-simplemath\n run: cd build-simplestl-simplemath && ctest --output-on-failure -j $(nproc)\n\n asan:\n runs-on: ubuntu-24.04-arm\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=relwithdebinfo -DNCNN_ASAN=ON -DNCNN_BUILD_TESTS=ON -DNCNN_SHARED_LIB=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test\n run: |\n cd build\n ctest --output-on-failure -j $(nproc)\n\n aarch64:\n runs-on: ubuntu-24.04\n steps:\n - uses: actions/checkout@v6\n\n - name: aarch64-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-aarch64-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test-a53\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;cortex-a53\" ctest --output-on-failure -j $(nproc)\n\n - name: test-a55\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;cortex-a55\" ctest --output-on-failure -j $(nproc)\n\n - name: test-a72\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;cortex-a72\" ctest --output-on-failure -j $(nproc)\n\n - name: test-a76\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;cortex-a76\" ctest --output-on-failure -j $(nproc)\n\n - name: test-a710\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;cortex-a710\" ctest --output-on-failure -j $(nproc)\n\n - name: test-max\n run: cd build && TESTS_EXECUTABLE_LOADER=qemu-aarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/aarch64-linux-gnu;-cpu;max\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-arm.yml", "content": "name: linux-arm\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-arm.yml'\n - 'toolchains/arm-linux-gnueabi.toolchain.cmake'\n - 'toolchains/arm-linux-gnueabihf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-arm.yml'\n - 'toolchains/arm-linux-gnueabi.toolchain.cmake'\n - 'toolchains/arm-linux-gnueabihf.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'tests/**'\nconcurrency:\n group: linux-arm-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n arm:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: arm-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-arm-linux-gnueabi qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabi\" ctest --output-on-failure -j $(nproc)\n\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: |\n cd build-noint8\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabi\" ctest --output-on-failure -j $(nproc)\n\n armhf:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: arm-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-arm-linux-gnueabihf qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabihf\" ctest --output-on-failure -j $(nproc)\n\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: |\n cd build-noint8\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabihf\" ctest --output-on-failure -j $(nproc)\n\n armhf-vfpv3-d16:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: arm-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-arm-linux-gnueabihf qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabihf\" ctest --output-on-failure -j $(nproc)\n\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: |\n cd build-noint8\n TESTS_EXECUTABLE_LOADER=qemu-arm-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/arm-linux-gnueabihf\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-loongarch64.yml", "content": "name: linux-loongarch64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-loongarch64.yml'\n - 'toolchains/loongarch64-linux-gnu.toolchain.cmake'\n - 'toolchains/loongarch64-unknown-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/loongarch/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-loongarch64.yml'\n - 'toolchains/loongarch64-linux-gnu.toolchain.cmake'\n - 'toolchains/loongarch64-unknown-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/loongarch/**'\n - 'tests/**'\nconcurrency:\n group: linux-loongarch64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n gcc-loongarch64:\n runs-on: [self-hosted, linux, centos]\n\n steps:\n - uses: actions/checkout@v6\n\n # - name: qemu\n # run: |\n # sudo apt-get update\n # sudo apt-get install -y qemu-user-static\n\n # - name: loongarch64-toolchain\n # run: |\n # wget https://github.com/sunhaiyong1978/CLFS-for-LoongArch/releases/download/8.0/loongarch64-clfs-8.0-cross-tools-gcc-full.tar.xz\n # tar -xf loongarch64-clfs-8.0-cross-tools-gcc-full.tar.xz\n\n - name: build\n run: |\n export LOONGARCH64_ROOT_PATH=/data/action/osd/cross-tools\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/loongarch64-unknown-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 4\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-loongarch64-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/data/action/osd/cross-tools/target\" ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/linux-mips.yml", "content": "name: linux-mips\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-mips.yml'\n - 'toolchains/mipsel-linux-gnu.toolchain.cmake'\n - 'toolchains/mipsisa32r6el-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/mips/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-mips.yml'\n - 'toolchains/mipsel-linux-gnu.toolchain.cmake'\n - 'toolchains/mipsisa32r6el-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/mips/**'\n - 'tests/**'\nconcurrency:\n group: linux-mips-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n mipsel:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: mipsel-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-mipsel-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/mipsel-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-mipsel-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/mipsel-linux-gnu\" ctest --output-on-failure -j $(nproc)\n\n mipsisa32r6el:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: mipsisa32r6el-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-mipsisa32r6el-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/mipsisa32r6el-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-mipsel-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/mipsisa32r6el-linux-gnu\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-mips64.yml", "content": "name: linux-mips64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-mips64.yml'\n - 'toolchains/mips64el-linux-gnuabi64.toolchain.cmake'\n - 'toolchains/mipsisa64r6el-linux-gnuabi64.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/mips/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-mips64.yml'\n - 'toolchains/mips64el-linux-gnuabi64.toolchain.cmake'\n - 'toolchains/mipsisa64r6el-linux-gnuabi64.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/mips/**'\n - 'tests/**'\nconcurrency:\n group: linux-mips64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n mips64el:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: mips64el-gnuabi64-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-mips64el-linux-gnuabi64 qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/mips64el-linux-gnuabi64.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-mips64el-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/mips64el-linux-gnuabi64\" ctest --output-on-failure -j $(nproc)\n\n mipsisa64r6el:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: mipsisa64r6el-gnuabi64-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-mipsisa64r6el-linux-gnuabi64 qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/mipsisa64r6el-linux-gnuabi64.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-mips64el-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/mipsisa64r6el-linux-gnuabi64\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-ppc64.yml", "content": "name: linux-ppc64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-ppc64.yml'\n - 'toolchains/powerpc64le-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/*'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-ppc64.yml'\n - 'toolchains/powerpc64le-linux-gnu.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/*'\n - 'tests/**'\nconcurrency:\n group: linux-ppc64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n ppc:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: powerpc-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-powerpc-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/powerpc-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-ppc-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/powerpc-linux-gnu\" ctest --output-on-failure -j $(nproc)\n\n ppc64le:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: powerpc64le-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-powerpc64le-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/powerpc64le-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-ppc64le-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/powerpc64le-linux-gnu\" ctest --output-on-failure -j $(nproc)\n\n power8le-vsx:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: powerpc64le-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-powerpc64le-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/power8le-linux-gnu-vsx.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-ppc64le-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/powerpc64le-linux-gnu\" ctest --output-on-failure -j $(nproc)\n\n power9le-vsx:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: powerpc64le-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-powerpc64le-linux-gnu qemu-user-static\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/power9le-linux-gnu-vsx.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-ppc64le-static TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/powerpc64le-linux-gnu;-cpu;power9_v2.0\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-riscv32.yml", "content": "name: linux-riscv32\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-riscv32.yml'\n - 'toolchains/c907-rv32-v310.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-riscv32.yml'\n - 'toolchains/c907-rv32-v310.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\nconcurrency:\n group: linux-riscv32-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n xuantie:\n name: xuantie-${{ matrix.cpu }}\n runs-on: [self-hosted, linux, ubuntu]\n strategy:\n fail-fast: false\n matrix:\n include:\n - { cpu: c907-rv32, QEMU_CPU: c907fdv-rv32, OPENMP: ON, RVV: ON, XTHEADVECTOR: OFF, ZFH: ON, ZVFH: ON }\n\n steps:\n - uses: actions/checkout@v6\n\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/Xuantie-900-gcc-linux-6.6.36-glibc-x86_64-V3.3.0\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/${{ matrix.cpu }}-v310.toolchain.cmake -DCMAKE_BUILD_TYPE=release \\\n -DNCNN_OPENMP=${{ matrix.OPENMP }} -DNCNN_THREADS=${{ matrix.OPENMP }} \\\n -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_RVV=${{ matrix.RVV }} \\\n -DNCNN_XTHEADVECTOR=${{ matrix.XTHEADVECTOR }} \\\n -DNCNN_ZFH=${{ matrix.ZFH }} \\\n -DNCNN_ZVFH=${{ matrix.ZVFH }} \\\n -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test\n run: |\n export PATH=/data/action/osd/Xuantie-qemu-x86_64-Ubuntu-20.04-V5.2.8-B20250721-0303/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv32 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;${{ matrix.QEMU_CPU }}\" ctest --output-on-failure -j 8\n" }, { "path": ".github/workflows/linux-riscv64.yml", "content": "name: linux-riscv64\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-riscv64.yml'\n - 'toolchains/riscv64-linux-gnu.toolchain.cmake'\n - 'toolchains/riscv64-unknown-linux-gnu.toolchain.cmake'\n - 'toolchains/riscv64-unknown-linux-gnu.llvm-toolchain.cmake'\n - 'toolchains/c906-v310.toolchain.cmake'\n - 'toolchains/c908-v310.toolchain.cmake'\n - 'toolchains/c910-v310.toolchain.cmake'\n - 'toolchains/k1.toolchain.cmake'\n - 'toolchains/k1.llvm.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\n - 'examples/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-riscv64.yml'\n - 'toolchains/riscv64-linux-gnu.toolchain.cmake'\n - 'toolchains/riscv64-unknown-linux-gnu.toolchain.cmake'\n - 'toolchains/riscv64-unknown-linux-gnu.llvm-toolchain.cmake'\n - 'toolchains/c906-v310.toolchain.cmake'\n - 'toolchains/c908-v310.toolchain.cmake'\n - 'toolchains/c910-v310.toolchain.cmake'\n - 'toolchains/k1.toolchain.cmake'\n - 'toolchains/k1.llvm.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/riscv/**'\n - 'tests/**'\n - 'examples/**'\nconcurrency:\n group: linux-riscv64-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n gcc-riscv64:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: cache-qemu\n id: cache-qemu\n uses: actions/cache@v5\n with:\n path: qemu-install\n key: qemu-riscv64-install-20220502-4\n - name: install-qemu-build-deps\n if: steps.cache-qemu.outputs.cache-hit != 'true'\n run: |\n sudo apt-get update\n sudo apt-get install autoconf automake autotools-dev ninja-build build-essential pkg-config libglib2.0-dev libpixman-1-dev zlib1g-dev python3\n - name: checkout-qemu\n if: steps.cache-qemu.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: qemu/qemu\n path: qemu\n ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65\n - name: qemu\n if: steps.cache-qemu.outputs.cache-hit != 'true'\n run: |\n cd qemu\n wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch\n patch -p1 -i 0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch\n ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system\n make -j$(nproc)\n make install\n\n - name: riscv64-gnu-toolchain\n run: |\n sudo apt-get update\n sudo apt-get install g++-riscv64-linux-gnu\n\n - name: configure\n run: mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n - name: build\n run: cmake --build build -j $(nproc)\n\n - name: test\n run: |\n export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-L;/usr/riscv64-linux-gnu\" ctest --output-on-failure -j $(nproc)\n\n xuantie:\n name: xuantie-${{ matrix.cpu }}\n runs-on: [self-hosted, linux, ubuntu]\n strategy:\n fail-fast: false\n matrix:\n include:\n - { cpu: c906, QEMU_CPU: c906fdv, OPENMP: OFF, RVV: OFF, XTHEADVECTOR: ON, ZFH: ON, ZVFH: OFF }\n - { cpu: c910, QEMU_CPU: c910v, OPENMP: ON, RVV: OFF, XTHEADVECTOR: ON, ZFH: ON, ZVFH: OFF }\n - { cpu: c908, QEMU_CPU: c908v, OPENMP: ON, RVV: ON, XTHEADVECTOR: OFF, ZFH: ON, ZVFH: ON }\n - { cpu: c907, QEMU_CPU: c907fdv-rv64, OPENMP: ON, RVV: ON, XTHEADVECTOR: OFF, ZFH: ON, ZVFH: ON }\n\n steps:\n - uses: actions/checkout@v6\n\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/Xuantie-900-gcc-linux-6.6.36-glibc-x86_64-V3.3.0\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/${{ matrix.cpu }}-v310.toolchain.cmake -DCMAKE_BUILD_TYPE=release \\\n -DNCNN_OPENMP=${{ matrix.OPENMP }} -DNCNN_THREADS=${{ matrix.OPENMP }} \\\n -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_RVV=${{ matrix.RVV }} \\\n -DNCNN_XTHEADVECTOR=${{ matrix.XTHEADVECTOR }} \\\n -DNCNN_ZFH=${{ matrix.ZFH }} \\\n -DNCNN_ZVFH=${{ matrix.ZVFH }} \\\n -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test\n run: |\n export PATH=/data/action/osd/Xuantie-qemu-x86_64-Ubuntu-20.04-V5.2.8-B20250721-0303/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;${{ matrix.QEMU_CPU }}\" ctest --output-on-failure -j 8\n\n spacemit:\n name: spacemit-${{ matrix.cpu }}\n runs-on: [self-hosted, linux, ubuntu]\n strategy:\n fail-fast: false\n matrix:\n include:\n - { cpu: x60, QEMU_CPU: \"max,vlen=256,elen=64,vext_spec=v1.0\", OPENMP: ON, RVV: ON, XTHEADVECTOR: OFF, ZFH: ON, ZVFH: ON }\n\n steps:\n - uses: actions/checkout@v6\n\n # https://archive.spacemit.com/toolchain/spacemit-toolchain-linux-glibc-x86_64-v1.1.2.tar.xz\n - name: build-gcc\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/spacemit-toolchain-linux-glibc-x86_64-v1.1.2\n mkdir build-gcc && cd build-gcc\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/k1.toolchain.cmake -DCMAKE_BUILD_TYPE=release \\\n -DNCNN_OPENMP=${{ matrix.OPENMP }} -DNCNN_THREADS=${{ matrix.OPENMP }} \\\n -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_RVV=${{ matrix.RVV }} \\\n -DNCNN_XTHEADVECTOR=${{ matrix.XTHEADVECTOR }} \\\n -DNCNN_ZFH=${{ matrix.ZFH }} \\\n -DNCNN_ZVFH=${{ matrix.ZVFH }} \\\n -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: build-llvm\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/spacemit-toolchain-linux-glibc-x86_64-v1.1.2\n mkdir build-llvm && cd build-llvm\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/k1.llvm.toolchain.cmake -DCMAKE_BUILD_TYPE=release \\\n -DNCNN_OPENMP=${{ matrix.OPENMP }} -DNCNN_THREADS=${{ matrix.OPENMP }} \\\n -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_RVV=${{ matrix.RVV }} \\\n -DNCNN_XTHEADVECTOR=${{ matrix.XTHEADVECTOR }} \\\n -DNCNN_ZFH=${{ matrix.ZFH }} \\\n -DNCNN_ZVFH=${{ matrix.ZVFH }} \\\n -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n # https://archive.spacemit.com/spacemit-ai/qemu/jdsk-qemu-v0.0.14.tar.gz\n - name: test-gcc\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/spacemit-toolchain-linux-glibc-x86_64-v1.1.2\n export PATH=/data/action/osd/jdsk-qemu/bin:$PATH\n cd build-gcc\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;${{ matrix.QEMU_CPU }};-L;${RISCV_ROOT_PATH}/sysroot\" ctest --output-on-failure -j 8\n\n - name: test-llvm\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/spacemit-toolchain-linux-glibc-x86_64-v1.1.2\n export PATH=/data/action/osd/jdsk-qemu/bin:$PATH\n cd build-llvm\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;${{ matrix.QEMU_CPU }};-L;${RISCV_ROOT_PATH}/sysroot\" ctest --output-on-failure -j 8\n\n gcc-rvv:\n runs-on: [self-hosted, linux, ubuntu]\n steps:\n - uses: actions/checkout@v6\n\n #- name: cache-qemu\n #id: cache-qemu\n #uses: actions/cache@v5\n #with:\n #path: qemu-install\n #key: qemu-riscv64-install-20241202\n #- name: install-qemu-build-deps\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #run: |\n #sudo apt-get update\n #sudo apt-get install autoconf automake autotools-dev ninja-build\n #- name: checkout-qemu\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #uses: actions/checkout@v6\n #with:\n #repository: qemu/qemu\n #path: qemu\n #ref: 72b88908d12ee9347d13539c7dd9a252625158d1\n #- name: qemu\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #run: |\n #cd qemu\n #./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system\n #make -j4\n #make install\n\n #- name: cache-riscv\n #id: cache-riscv\n #uses: actions/cache@v5\n #with:\n #path: riscv-install\n #key: riscv-linux-install-20241202\n\n #- name: install-riscv-build-deps\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #sudo apt-get update\n #sudo apt-get install autoconf automake autotools-dev curl python3 libmpc-dev libmpfr-dev libgmp-dev gawk build-essential bison flex texinfo gperf libtool patchutils bc zlib1g-dev libexpat-dev device-tree-compiler\n\n #- name: checkout-riscv-gnu-toolchain\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #uses: actions/checkout@v6\n #with:\n #repository: riscv-collab/riscv-gnu-toolchain\n #path: riscv-gnu-toolchain\n #ref: 20f615317e2ce888dfc11b29ccde4a649494b654\n #- name: checkout-riscv-gnu-toolchain-submodules\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #cd riscv-gnu-toolchain\n #git submodule update --init --recursive --depth 1 glibc\n #git submodule update --init --recursive --depth 1 newlib\n #git submodule update --init --recursive --depth 1 riscv-binutils\n #git submodule update --init --recursive --depth 1 riscv-gcc\n #git submodule update --init --recursive --depth 1 riscv-dejagnu\n #git submodule update --init --recursive --depth 1 riscv-gdb\n #- name: riscv-gnu-toolchain\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #cd riscv-gnu-toolchain\n #./configure --prefix=$GITHUB_WORKSPACE/riscv\n #make linux -j4\n\n #- name: riscv-strip-install\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: find $GITHUB_WORKSPACE/riscv -type f | xargs -i strip -g {} || true\n\n - name: configure\n run: export RISCV_ROOT_PATH=/data/action/osd/riscv && mkdir build && cd build && cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-unknown-linux-gnu.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n - name: build\n run: cmake --build build -j 8\n\n - name: test-vlen256\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n\n - name: test-vlen128\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n\n clang-rvv:\n runs-on: [self-hosted, linux, ubuntu]\n steps:\n - uses: actions/checkout@v6\n\n #- name: cache-qemu\n #id: cache-qemu\n #uses: actions/cache@v5\n #with:\n #path: qemu-install\n #key: qemu-riscv64-install-20241202\n #- name: install-qemu-build-deps\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #run: |\n #sudo apt-get update\n #sudo apt-get install autoconf automake autotools-dev ninja-build\n #- name: checkout-qemu\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #uses: actions/checkout@v6\n #with:\n #repository: qemu/qemu\n #path: qemu\n #ref: 72b88908d12ee9347d13539c7dd9a252625158d1\n #- name: qemu\n #if: steps.cache-qemu.outputs.cache-hit != 'true'\n #run: |\n #cd qemu\n #./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system\n #make -j4\n #make install\n\n #- name: cache-riscv\n #id: cache-riscv\n #uses: actions/cache@v5\n #with:\n #path: riscv-install\n #key: riscv-linux-install-20241202\n\n #- name: install-riscv-build-deps\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #sudo apt-get update\n #sudo apt-get install autoconf automake autotools-dev curl python3 libmpc-dev libmpfr-dev libgmp-dev gawk build-essential bison flex texinfo gperf libtool patchutils bc zlib1g-dev libexpat-dev device-tree-compiler\n\n #- name: checkout-riscv-gnu-toolchain\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #uses: actions/checkout@v6\n #with:\n #repository: riscv-collab/riscv-gnu-toolchain\n #path: riscv-gnu-toolchain\n #ref: 20f615317e2ce888dfc11b29ccde4a649494b654\n #- name: checkout-riscv-gnu-toolchain-submodules\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #cd riscv-gnu-toolchain\n #git submodule update --init --recursive --depth 1 glibc\n #git submodule update --init --recursive --depth 1 newlib\n #git submodule update --init --recursive --depth 1 riscv-binutils\n #git submodule update --init --recursive --depth 1 riscv-gcc\n #git submodule update --init --recursive --depth 1 riscv-dejagnu\n #git submodule update --init --recursive --depth 1 riscv-gdb\n #- name: riscv-gnu-toolchain\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: |\n #cd riscv-gnu-toolchain\n #./configure --prefix=$GITHUB_WORKSPACE/riscv\n #make linux -j4\n\n #- name: riscv-strip-install\n #if: steps.cache-riscv.outputs.cache-hit != 'true'\n #run: find $GITHUB_WORKSPACE/riscv -type f | xargs -i strip -g {} || true\n\n # - name: install-clang\n # run: |\n # wget https://github.com/llvm/llvm-project/releases/download/llvmorg-19.1.4/llvm-project-19.1.4.src.tar.xz\n # tar -xf llvm-project-19.1.4.src.tar.xz\n # cd llvm-project-19.1.4.src\n # mkdir build\n # cd build\n # cmake -DCMAKE_INSTALL_PREFIX=$GITHUB_WORKSPACE/riscv -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DLLVM_ENABLE_PROJECTS=\"clang\" -DLLVM_TARGETS_TO_BUILD=\"RISCV\" -DLLVM_INCLUDE_EXAMPLES=OFF -DLLVM_INCLUDE_TESTS=OFF ../llvm/\n # make -j16\n # make install\n\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/riscv\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-unknown-linux-gnu.llvm-toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test-vlen256\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n\n - name: test-vlen128\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n" }, { "path": ".github/workflows/linux-x64-cpu-clang.yml", "content": "name: linux-x64-cpu-clang\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-clang.yml'\n - 'toolchains/host-c.clang.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-clang.yml'\n - 'toolchains/host-c.clang.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\nconcurrency:\n group: linux-x64-cpu-clang-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-clang:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: update\n run: sudo apt-get update\n - name: protobuf\n run: sudo apt-get install libprotobuf-dev protobuf-compiler libopencv-dev\n - name: build-sse2\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-sse2 && cd build-sse2\n cmake -DNCNN_AVX=OFF -DNCNN_AVX2=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-sse2\n run: cd build-sse2 && ctest --output-on-failure -j $(nproc)\n - name: build-shared\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_AVX2=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: build-avx2\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-avx2 && cd build-avx2\n cmake -DNCNN_AVX2=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx2\n run: cd build-avx2 && ctest --output-on-failure -j $(nproc)\n - name: build-avx\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-avx && cd build-avx\n cmake -DNCNN_AVX2=OFF -DNCNN_AVX=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx\n run: cd build-avx && ctest --output-on-failure -j $(nproc)\n - name: build-avx1-2\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-avx1-2 && cd build-avx1-2\n cmake -DNCNN_AVX2=ON -DNCNN_AVX=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx1-2\n run: cd build-avx1-2 && ctest --output-on-failure -j $(nproc)\n - name: build-noint8\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DNCNN_INT8=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: cd build-noint8 && ctest --output-on-failure -j $(nproc)\n\n linux-clang-simplestl:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: build-simplestl\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-simplestl && cd build-simplestl\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.clang.toolchain.cmake -DNCNN_STDIO=ON -DNCNN_STRING=ON -DNCNN_SIMPLESTL=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-simplestl\n run: cd build-simplestl && ctest --output-on-failure -j $(nproc)\n - name: build-simplestl-simpleomp\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-simplestl-simpleomp && cd build-simplestl-simpleomp\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.clang.toolchain.cmake -DNCNN_STDIO=ON -DNCNN_STRING=ON -DNCNN_SIMPLESTL=ON -DNCNN_SIMPLEOMP=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-simplestl-simpleomp\n run: cd build-simplestl-simpleomp && ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-x64-cpu-gcc-musl.yml", "content": "name: linux-x64-cpu-gcc-musl\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-gcc-musl.yml'\n - 'toolchains/host-c.gcc.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-gcc-musl.yml'\n - 'toolchains/host-c.gcc.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\nconcurrency:\n group: linux-x64-cpu-gcc-musl-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-gcc-musl:\n runs-on: ubuntu-latest\n steps:\n - uses: jirutka/setup-alpine@v1\n with:\n packages: >\n cmake\n clang\n clang-dev\n make\n gcc\n g++\n libc-dev\n linux-headers\n\n - uses: actions/checkout@v6\n - name: build\n shell: alpine.sh {0}\n run: |\n mkdir build && cd build\n cmake -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test\n shell: alpine.sh {0}\n run: cd build && ctest --output-on-failure -j $(nproc)\n - name: build-shared\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n" }, { "path": ".github/workflows/linux-x64-cpu-gcc.yml", "content": "name: linux-x64-cpu-gcc\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-gcc.yml'\n - 'toolchains/host-c.gcc.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-cpu-gcc.yml'\n - 'toolchains/host-c.gcc.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\nconcurrency:\n group: linux-x64-cpu-gcc-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-gcc:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: update\n run: sudo apt-get update\n - name: protobuf\n run: sudo apt-get install libprotobuf-dev protobuf-compiler libopencv-dev\n - name: build-sse2\n run: |\n mkdir build-sse2 && cd build-sse2\n cmake -DNCNN_AVX=OFF -DNCNN_AVX2=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-sse2\n run: cd build-sse2 && ctest --output-on-failure -j $(nproc)\n - name: build-shared\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_AVX2=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: build-avx2\n run: |\n mkdir build-avx2 && cd build-avx2\n cmake -DNCNN_AVX2=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx2\n run: cd build-avx2 && ctest --output-on-failure -j $(nproc)\n - name: build-avx\n run: |\n mkdir build-avx && cd build-avx\n cmake -DNCNN_AVX2=OFF -DNCNN_AVX=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx\n run: cd build-avx && ctest --output-on-failure -j $(nproc)\n - name: build-avx1-2\n run: |\n mkdir build-avx1-2 && cd build-avx1-2\n cmake -DNCNN_AVX2=ON -DNCNN_AVX=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-avx1-2\n run: cd build-avx1-2 && ctest --output-on-failure -j $(nproc)\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DNCNN_INT8=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: cd build-noint8 && ctest --output-on-failure -j $(nproc)\n\n asan:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=relwithdebinfo -DNCNN_ASAN=ON -DNCNN_BUILD_TESTS=ON -DNCNN_SHARED_LIB=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test\n run: |\n cd build\n ctest --output-on-failure -j $(nproc)\n\n linux-gcc-cpp03-nostdio-nostring-simplestl:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: build-nostdio\n run: |\n mkdir build-nostdio && cd build-nostdio\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc-c++03.toolchain.cmake -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-nostdio\n run: cd build-nostdio && ctest --output-on-failure -j $(nproc)\n - name: build-nostdio-nostring\n run: |\n mkdir build-nostdio-nostring && cd build-nostdio-nostring\n cmake -DNCNN_STDIO=OFF -DNCNN_STRING=OFF -DNCNN_BUILD_TESTS=OFF -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: build-simplestl\n run: |\n mkdir build-simplestl && cd build-simplestl\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.gcc.toolchain.cmake -DNCNN_STDIO=ON -DNCNN_STRING=ON -DNCNN_SIMPLESTL=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-simplestl\n run: cd build-simplestl && ctest --output-on-failure -j $(nproc)\n - name: build-simplestl-simpleomp\n run: |\n mkdir build-simplestl-simpleomp && cd build-simplestl-simpleomp\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.gcc.toolchain.cmake -DNCNN_STDIO=ON -DNCNN_STRING=ON -DNCNN_SIMPLESTL=ON -DNCNN_SIMPLEOMP=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-simplestl-simpleomp\n run: cd build-simplestl-simpleomp && ctest --output-on-failure -j $(nproc)\n\n linux-gcc-avx512:\n runs-on: [self-hosted, linux, t4]\n steps:\n - uses: actions/checkout@v6\n - name: build\n env:\n CC: gcc\n CXX: g++\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: |\n mkdir build && cd build\n cmake -DNCNN_AVX2=ON -DNCNN_AVX512=ON -DNCNN_AVX512VNNI=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j 4\n - name: test\n env:\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: cd build && ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/linux-x64-gpu-clang.yml", "content": "name: linux-x64-gpu-clang\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-gpu-clang.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-gpu-clang.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\nconcurrency:\n group: linux-x64-gpu-clang-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-clang-gpu:\n runs-on: [self-hosted, linux, ubuntu25]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-swiftshader\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-linux-install-20250508\n - name: checkout-swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: 930d46d31b5d637f313fd5ef55da2bbf053c26c1\n - name: swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n mkdir -p build; cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . -j 8\n mkdir $GITHUB_WORKSPACE/swiftshader-install\n cp Linux/* $GITHUB_WORKSPACE/swiftshader-install\n\n - name: build\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build && cd build\n cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n printf \"[Processor]\\nThreadCount=1\\n\" > build/tests/SwiftShader.ini\n export VK_ICD_FILENAMES=\"$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json\"\n cd build && ctest --output-on-failure -j 8\n - name: build-shared\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j 8\n" }, { "path": ".github/workflows/linux-x64-gpu-gcc.yml", "content": "name: linux-x64-gpu-gcc\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-gpu-gcc.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-gpu-gcc.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\nconcurrency:\n group: linux-x64-gpu-gcc-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-gcc-gpu:\n runs-on: [self-hosted, linux, ubuntu25]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-swiftshader\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-linux-install-20250508\n - name: checkout-swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: 930d46d31b5d637f313fd5ef55da2bbf053c26c1\n - name: swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n mkdir -p build; cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . -j 8\n mkdir $GITHUB_WORKSPACE/swiftshader-install\n cp Linux/* $GITHUB_WORKSPACE/swiftshader-install\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n printf \"[Processor]\\nThreadCount=1\\n\" > build/tests/SwiftShader.ini\n export VK_ICD_FILENAMES=\"$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json\"\n cd build && ctest --output-on-failure -j 8\n - name: build-shared\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j 8\n\n linux-gcc-gpu-system-glslang:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n\n - name: install-deps\n run: |\n sudo apt-get update\n sudo apt-get install libprotobuf-dev protobuf-compiler libopencv-dev libvulkan-dev glslang-dev glslang-tools spirv-tools\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DGLSLANG_TARGET_DIR=/usr/lib/x86_64-linux-gnu/cmake ..\n cmake --build . -j $(nproc)\n - name: build-shared\n run: |\n mkdir build-shared && cd build-shared\n cmake -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DGLSLANG_TARGET_DIR=/usr/lib/x86_64-linux-gnu/cmake -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n\n linux-gcc-gpu-t4:\n runs-on: [self-hosted, linux, t4]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: build\n env:\n CC: gcc\n CXX: g++\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: |\n mkdir build && cd build\n cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j 4\n - name: test\n env:\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: |\n cd build && ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/linux-x64-sde.yml", "content": "name: linux-x64-sde\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-sde.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x64-sde.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\nconcurrency:\n group: linux-x64-sde-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n gcc-sde:\n runs-on: ubuntu-24.04\n steps:\n - uses: actions/checkout@v6\n - name: update\n run: sudo apt-get update\n - name: gcc14\n run: sudo apt-get install gcc-14 g++-14\n - name: Setup SDE binaries\n uses: petarpetrovt/setup-sde@v3.0\n - name: build\n env:\n CC: gcc-14\n CXX: g++-14\n run: |\n mkdir build && cd build\n cmake -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-p4p\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-p4p;--\" ctest --output-on-failure -j $(nproc)\n - name: test-snb\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-snb;--\" ctest --output-on-failure -j $(nproc)\n - name: test-hsw\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-hsw;--\" ctest --output-on-failure -j $(nproc)\n - name: test-adl\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-adl;--\" ctest --output-on-failure -j $(nproc)\n - name: test-arl\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-arl;--\" ctest --output-on-failure -j $(nproc)\n - name: test-skx\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-skx;--\" ctest --output-on-failure -j $(nproc)\n - name: test-spr\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-spr;--\" ctest --output-on-failure -j $(nproc)\n - name: test-gnr\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-gnr;--\" ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-x86-cpu-clang.yml", "content": "name: linux-x86-cpu-clang\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x86-cpu-clang.yml'\n - 'toolchains/host.clang-m32.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x86-cpu-clang.yml'\n - 'toolchains/host.clang-m32.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\nconcurrency:\n group: linux-x86-cpu-clang-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-clang:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: update\n run: sudo apt-get update\n - name: gcc-multilib\n run: sudo apt-get install gcc-multilib g++-multilib\n - name: build\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.clang-m32.toolchain.cmake -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test\n run: cd build && ctest --output-on-failure -j $(nproc)\n - name: build-shared\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-shared && cd build-shared\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.clang-m32.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: build-noint8\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.clang-m32.toolchain.cmake -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: cd build-noint8 && ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/linux-x86-cpu-gcc.yml", "content": "name: linux-x86-cpu-gcc\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/linux-x86-cpu-gcc.yml'\n - 'toolchains/host.gcc-m32.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/linux-x86-cpu-gcc.yml'\n - 'toolchains/host.gcc-m32.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\nconcurrency:\n group: linux-x86-cpu-gcc-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-gcc:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: update\n run: sudo apt-get update\n - name: gcc-multilib\n run: sudo apt-get install gcc-multilib g++-multilib\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc-m32.toolchain.cmake -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test\n run: cd build && ctest --output-on-failure -j $(nproc)\n - name: build-nosse\n run: |\n mkdir build-nosse && cd build-nosse\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc-m32.toolchain.cmake -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: test-nosse\n run: cd build-nosse && ctest --output-on-failure -j $(nproc)\n - name: build-shared\n run: |\n mkdir build-shared && cd build-shared\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc-m32.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j $(nproc)\n - name: build-noint8\n run: |\n mkdir build-noint8 && cd build-noint8\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc-m32.toolchain.cmake -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_INT8=OFF ..\n cmake --build . -j $(nproc)\n - name: test-noint8\n run: cd build-noint8 && ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/mac-catalyst.yml", "content": "name: mac-catalyst\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/mac-catalyst.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/mac-catalyst.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\nconcurrency:\n group: mac-catalyst-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n MAC_CATALYST_DEPLOYMENT_TARGET: '13.1'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_CATALYST_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_CATALYST_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-mac-catalyst-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/mac-catalyst\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install/mac-catalyst\n mkdir -p $GITHUB_WORKSPACE/openmp-install/mac-catalyst/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/mac-catalyst/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/mac-catalyst/include/* $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/mac-catalyst/lib/libomp.a $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib\n\n - name: x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n - name: arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n" }, { "path": ".github/workflows/macos.yml", "content": "name: macos\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/macos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/macos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\nconcurrency:\n group: macos-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n MAC_DEPLOYMENT_TARGET: '11.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-macos-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/include/* $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/lib/libomp.a $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib\n\n - name: cache-swiftshader\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-macos-install-20251004\n - name: checkout-swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: de870ac7518fe2b6bb651ecc22fc36647cf7b986\n - name: checkout-swiftshader-submodules\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n - name: swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n mkdir -p build; cd build\n cmake -DCMAKE_POLICY_VERSION_MINIMUM=3.5 -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . -j 4\n mkdir $GITHUB_WORKSPACE/swiftshader-install\n cp Darwin/* $GITHUB_WORKSPACE/swiftshader-install\n\n - name: arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n - name: x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 4\n\n - name: arm64-shared\n run: |\n mkdir build-arm64-shared && cd build-arm64-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j 4\n - name: x86_64-shared\n run: |\n mkdir build-x86_64-shared && cd build-x86_64-shared\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" -DNCNN_SHARED_LIB=ON ..\n cmake --build . -j 4\n\n - name: x86_64-test\n run: |\n printf \"[Processor]\\nThreadCount=1\\n\" > build-x86_64/tests/SwiftShader.ini\n export VK_ICD_FILENAMES=\"$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json\"\n cd build-x86_64 && ctest --output-on-failure -j 4\n" }, { "path": ".github/workflows/pnnx.yml", "content": "name: pnnx\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/pnnx.yml'\n - 'src/layer/*'\n - 'tools/pnnx/**'\n - '!tools/pnnx/README.md'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/pnnx.yml'\n - 'src/layer/*'\n - 'tools/pnnx/**'\n - '!tools/pnnx/README.md'\nconcurrency:\n group: pnnx-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\nenv:\n LIBTORCH_VERSION: 2.10.0\n TORCHVISION_VERSION: 0.25.0\n PROTOBUF_VERSION: 21.12\n ONNXRUNTIME_VERSION: 1.24.3\n CACHE_DATE: 20260309\n SEGMENT_DOWNLOAD_TIMEOUT_MINS: 15\n\njobs:\n quick-test:\n runs-on: ${{ matrix.os }}\n strategy:\n fail-fast: false\n matrix:\n os: [ubuntu-latest, macos-latest, windows-latest]\n\n env:\n PYTHONUSERBASE: ${{ github.workspace }}/torch\n UseMultiToolTask: true\n steps:\n - uses: actions/checkout@v6\n\n - uses: actions/setup-python@v6\n with:\n python-version: 3.12\n\n - name: setup-pytorch\n run: |\n python3 -m pip config set global.break-system-packages true\n pip3 install --user torch --index-url https://download.pytorch.org/whl/cpu\n pip3 install --user numpy packaging\n\n - name: build-pnnx\n run: |\n cd tools/pnnx\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . --config Release -j 4\n\n - name: quick-test\n if: matrix.os != 'windows-latest'\n run: |\n cd tools/pnnx\n cd build && ctest -C Release --output-on-failure -R test_nn_Conv\n\n build:\n runs-on: [self-hosted, linux, ubuntu25]\n\n steps:\n - uses: actions/checkout@v6\n\n - name: local-cache-libtorch\n id: local-cache-libtorch\n uses: maxnowack/local-cache@v2\n with:\n path: libtorch-${{ env.LIBTORCH_VERSION }}-install\n key: libtorch-${{ env.LIBTORCH_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: local-cache-torchvision\n id: local-cache-torchvision\n uses: maxnowack/local-cache@v2\n with:\n path: torchvision-${{ env.TORCHVISION_VERSION }}-install\n key: torchvision-${{ env.TORCHVISION_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: local-cache-onnxruntime\n id: local-cache-onnxruntime\n uses: maxnowack/local-cache@v2\n with:\n path: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install\n key: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: cache-libtorch\n id: cache-libtorch\n uses: actions/cache@v4\n with:\n path: libtorch-${{ env.LIBTORCH_VERSION }}-install\n key: libtorch-${{ env.LIBTORCH_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: cache-torchvision\n id: cache-torchvision\n uses: actions/cache@v4\n with:\n path: torchvision-${{ env.TORCHVISION_VERSION }}-install\n key: torchvision-${{ env.TORCHVISION_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: cache-onnxruntime\n id: cache-onnxruntime\n uses: actions/cache@v4\n with:\n path: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install\n key: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: pnnx-patches\n if: (steps.local-cache-libtorch.outputs.cache-hit != 'true' && steps.cache-libtorch.outputs.cache-hit != 'true') || (steps.local-cache-torchvision.outputs.cache-hit != 'true' && steps.cache-torchvision.outputs.cache-hit != 'true') || (steps.local-cache-onnxruntime.outputs.cache-hit != 'true' && steps.cache-onnxruntime.outputs.cache-hit != 'true')\n uses: actions/checkout@v6\n with:\n repository: pnnx/pnnx\n path: pnnx-patches\n\n - name: libtorch\n if: steps.local-cache-libtorch.outputs.cache-hit != 'true' && steps.cache-libtorch.outputs.cache-hit != 'true'\n run: |\n wget -q https://github.com/pytorch/pytorch/releases/download/v${{ env.LIBTORCH_VERSION }}/pytorch-v${{ env.LIBTORCH_VERSION }}.tar.gz\n tar -xf pytorch-v${{ env.LIBTORCH_VERSION }}.tar.gz\n cd pytorch-v${{ env.LIBTORCH_VERSION }}\n pip3 install -r requirements.txt --break-system-packages\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/pytorch-v${{ env.LIBTORCH_VERSION }}-fix-mobile-build.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/pytorch-v${{ env.LIBTORCH_VERSION }}-no-link-system-lib.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/pytorch-v${{ env.LIBTORCH_VERSION }}-fix-eigen-build.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/pytorch-v${{ env.LIBTORCH_VERSION }}-fix-link-local-sleef.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/pytorch-v${{ env.LIBTORCH_VERSION }}-revert-nativert-api.patch\n mkdir -p build && cd build\n cmake -DCMAKE_INSTALL_PREFIX=$GITHUB_WORKSPACE/libtorch-${{ env.LIBTORCH_VERSION }}-install \\\n -DCMAKE_BUILD_TYPE=MinSizeRel \\\n -DBUILD_SHARED_LIBS=OFF \\\n -DCMAKE_POLICY_VERSION_MINIMUM=3.5 \\\n -DBUILD_CUSTOM_PROTOBUF=OFF \\\n -DBUILD_LITE_INTERPRETER=OFF \\\n -DBUILD_PYTHON=OFF \\\n -DINTERN_BUILD_MOBILE=ON \\\n -DINTERN_DISABLE_AUTOGRAD=ON \\\n -DINTERN_DISABLE_ONNX=ON \\\n -DUSE_CUDA=OFF \\\n -DUSE_DISTRIBUTED=OFF \\\n -DUSE_ITT=OFF \\\n -DUSE_KINETO=OFF \\\n -DUSE_LITE_INTERPRETER_PROFILER=OFF \\\n -DUSE_MKLDNN=OFF \\\n -DUSE_MPS=OFF \\\n -DUSE_NUMPY=OFF \\\n -DUSE_OPENMP=OFF \\\n -DUSE_SOURCE_DEBUG_ON_MOBILE=OFF \\\n -DUSE_XNNPACK=OFF \\\n -DBUILD_TEST=OFF \\\n -DATEN_NO_TEST=ON \\\n ..\n cmake --build . -j 8\n cmake --build . -j 8 --target install/strip\n\n - name: torchvision\n if: steps.local-cache-torchvision.outputs.cache-hit != 'true' && steps.cache-torchvision.outputs.cache-hit != 'true'\n run: |\n wget -q https://github.com/pytorch/vision/archive/v${{ env.TORCHVISION_VERSION }}.zip -O vision-${{ env.TORCHVISION_VERSION }}.zip\n unzip -q vision-${{ env.TORCHVISION_VERSION }}.zip\n cd vision-${{ env.TORCHVISION_VERSION }}\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/vision-${{ env.TORCHVISION_VERSION }}-ops-only.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/vision-${{ env.TORCHVISION_VERSION }}-no-cuda-version.patch\n mkdir -p build && cd build\n cmake -DCMAKE_INSTALL_PREFIX=$GITHUB_WORKSPACE/torchvision-${{ env.TORCHVISION_VERSION }}-install \\\n -DTorch_DIR=$GITHUB_WORKSPACE/libtorch-${{ env.LIBTORCH_VERSION }}-install/share/cmake/Torch \\\n -DCMAKE_BUILD_TYPE=MinSizeRel \\\n -DWITH_PNG=OFF \\\n -DWITH_JPEG=OFF ..\n cmake --build . -j 8\n cmake --build . -j 8 --target install/strip\n\n - name: onnxruntime\n if: steps.local-cache-onnxruntime.outputs.cache-hit != 'true' && steps.cache-onnxruntime.outputs.cache-hit != 'true'\n run: |\n wget -q https://github.com/protocolbuffers/protobuf/archive/v${{ env.PROTOBUF_VERSION }}.zip -O protobuf-${{ env.PROTOBUF_VERSION }}.zip\n unzip -q protobuf-${{ env.PROTOBUF_VERSION }}.zip\n cd protobuf-${{ env.PROTOBUF_VERSION }}\n mkdir -p build2 && cd build2\n cmake -DCMAKE_INSTALL_PREFIX=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install \\\n -Dprotobuf_BUILD_TESTS=OFF \\\n -DCMAKE_BUILD_TYPE=MinSizeRel \\\n -DCMAKE_POSITION_INDEPENDENT_CODE=ON ..\n cmake --build . -j 8\n cmake --build . -j 8 --target install/strip\n\n cd ../../\n wget -q https://github.com/microsoft/onnxruntime/archive/v${{ env.ONNXRUNTIME_VERSION }}.zip -O onnxruntime-${{ env.ONNXRUNTIME_VERSION }}.zip\n unzip -q onnxruntime-${{ env.ONNXRUNTIME_VERSION }}.zip\n cd onnxruntime-${{ env.ONNXRUNTIME_VERSION }}\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-less-mlas-features.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-monolithic-static-library.patch\n patch -p1 -i $GITHUB_WORKSPACE/pnnx-patches/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-use-clog.patch\n mkdir -p build2 && cd build2\n cmake -DCMAKE_INSTALL_PREFIX=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install \\\n -DCMAKE_BUILD_TYPE=MinSizeRel \\\n -Donnxruntime_USE_FULL_PROTOBUF=ON \\\n -Donnxruntime_BUILD_SHARED_LIB=ON \\\n -Donnxruntime_BUILD_UNIT_TESTS=OFF \\\n -Donnxruntime_ENABLE_CPUINFO=OFF \\\n -Donnxruntime_DISABLE_CONTRIB_OPS=ON \\\n -Donnxruntime_DISABLE_ML_OPS=ON \\\n -Donnxruntime_DISABLE_SPARSE_TENSORS=ON \\\n -DCMAKE_POSITION_INDEPENDENT_CODE=ON \\\n --compile-no-warning-as-error ../cmake\n cmake --build . -j 8\n cmake --build . -j 8 --target install/strip\n\n - name: pnnx\n run: |\n cd tools/pnnx\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=MinSizeRel \\\n -DTorch_INSTALL_DIR=$GITHUB_WORKSPACE/libtorch-${{ env.LIBTORCH_VERSION }}-install \\\n -DTorchVision_INSTALL_DIR=$GITHUB_WORKSPACE/torchvision-${{ env.TORCHVISION_VERSION }}-install \\\n -Donnxruntime_INSTALL_DIR=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install \\\n -Dprotobuf_DIR=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install/lib/cmake/protobuf ..\n cmake --build . -j 8\n strip src/pnnx\n\n - name: upload-pnnx\n uses: actions/upload-artifact@v5\n with:\n name: pnnx\n path: tools/pnnx/build/src/pnnx\n compression-level: 9\n\n test:\n needs: [build]\n runs-on: [self-hosted, linux, ubuntu25]\n strategy:\n fail-fast: false\n matrix:\n include:\n - { python: '3.8', numpy: '1.24.4', opencv: '4.5.*', torch: '1.8.1', torchvision: '0.9.1', torchaudio: '0.8.1', transformers: '4.52.1' }\n - { python: '3.8', numpy: '1.24.4', opencv: '4.5.*', torch: '1.9.1', torchvision: '0.10.1', torchaudio: '0.9.1', transformers: '4.52.1' }\n - { python: '3.8', numpy: '1.24.4', opencv: '4.6.*', torch: '1.10.0', torchvision: '0.11.1', torchaudio: '0.10.0+cpu', transformers: '4.52.1' }\n - { python: '3.9', numpy: '1.26.4', opencv: '4.6.*', torch: '1.11.0', torchvision: '0.12.0', torchaudio: '0.11.0+cpu', transformers: '4.52.1' }\n - { python: '3.9', numpy: '1.26.4', opencv: '4.7.*', torch: '1.12.0', torchvision: '0.13.0', torchaudio: '0.12.0+cpu', transformers: '4.52.1' }\n - { python: '3.10', numpy: '1.26.4', opencv: '4.7.*', torch: '1.13.0', torchvision: '0.14.0', torchaudio: '0.13.0+cpu', transformers: '4.52.1' }\n - { python: '3.10', numpy: '1.26.4', opencv: '4.8.*', torch: '2.0.0', torchvision: '0.15.1', torchaudio: '2.0.0+cpu', transformers: '4.52.1' }\n - { python: '3.10', numpy: '1.26.4', opencv: '4.8.*', torch: '2.1.0', torchvision: '0.16.0', torchaudio: '2.1.0+cpu', transformers: '4.52.1' }\n - { python: '3.11', numpy: '1.26.4', opencv: '4.9.*', torch: '2.2.1', torchvision: '0.17.1', torchaudio: '2.2.1+cpu', transformers: '4.52.1' }\n - { python: '3.11', numpy: '1.26.4', opencv: '4.9.*', torch: '2.3.0', torchvision: '0.18.0', torchaudio: '2.3.0+cpu', transformers: '4.52.1' }\n - { python: '3.11', numpy: '2.2.5', opencv: '4.10.*', torch: '2.4.0', torchvision: '0.19.0', torchaudio: '2.4.0+cpu', transformers: '4.52.1' }\n - { python: '3.12', numpy: '2.2.5', opencv: '4.10.*', torch: '2.5.0', torchvision: '0.20.0', torchaudio: '2.5.0+cpu', transformers: '4.52.1' }\n - { python: '3.12', numpy: '2.2.5', opencv: '4.11.*', torch: '2.6.0', torchvision: '0.21.0', torchaudio: '2.6.0+cpu', transformers: '4.52.1' }\n - { python: '3.12', numpy: '2.2.5', opencv: '4.11.*', torch: '2.7.0', torchvision: '0.22.0', torchaudio: '2.7.0+cpu', transformers: '4.52.1' }\n - { python: '3.13', numpy: '2.2.5', opencv: '4.12.*', torch: '2.8.0', torchvision: '0.23.0', torchaudio: '2.8.0+cpu', transformers: '4.56.2' }\n - { python: '3.13', numpy: '2.2.5', opencv: '4.12.*', torch: '2.9.0', torchvision: '0.24.0', torchaudio: '2.9.0+cpu', transformers: '4.56.2' }\n - { python: '3.13', numpy: '2.2.5', opencv: '4.12.*', torch: '2.10.0', torchvision: '0.25.0', torchaudio: '2.10.0+cpu', transformers: '4.56.2' }\n\n name: test-${{ matrix.torch }}-py${{ matrix.python }}\n\n env:\n PYTHONUSERBASE: ${{ github.workspace }}/python-${{ matrix.python }}\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: local-cache-libtorch\n id: local-cache-libtorch\n uses: maxnowack/local-cache@v2\n with:\n path: libtorch-${{ env.LIBTORCH_VERSION }}-install\n key: libtorch-${{ env.LIBTORCH_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: local-cache-torchvision\n id: local-cache-torchvision\n uses: maxnowack/local-cache@v2\n with:\n path: torchvision-${{ env.TORCHVISION_VERSION }}-install\n key: torchvision-${{ env.TORCHVISION_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: local-cache-onnxruntime\n id: local-cache-onnxruntime\n uses: maxnowack/local-cache@v2\n with:\n path: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install\n key: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n\n - name: cache-libtorch\n if: steps.local-cache-libtorch.outputs.cache-hit != 'true'\n id: cache-libtorch\n uses: actions/cache/restore@v5\n with:\n path: libtorch-${{ env.LIBTORCH_VERSION }}-install\n key: libtorch-${{ env.LIBTORCH_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n fail-on-cache-miss: true\n\n - name: cache-torchvision\n if: steps.local-cache-torchvision.outputs.cache-hit != 'true'\n id: cache-torchvision\n uses: actions/cache/restore@v5\n with:\n path: torchvision-${{ env.TORCHVISION_VERSION }}-install\n key: torchvision-${{ env.TORCHVISION_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n fail-on-cache-miss: true\n\n - name: cache-onnxruntime\n if: steps.local-cache-onnxruntime.outputs.cache-hit != 'true'\n id: cache-onnxruntime\n uses: actions/cache/restore@v5\n with:\n path: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install\n key: onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-linux-install-${{ env.CACHE_DATE }}\n fail-on-cache-miss: true\n\n - uses: actions/setup-python@v6\n with:\n python-version: ${{ matrix.python }}\n\n - name: setup-pytorch\n run: |\n export PATH=${{ env.PYTHONUSERBASE }}/bin:$PATH\n pip3 install --user pytest wheel twine requests einops numpy==${{ matrix.numpy }} opencv-python==${{ matrix.opencv }}\n pip3 install --user torch==${{ matrix.torch }}+cpu torchvision==${{ matrix.torchvision }}+cpu torchaudio==${{ matrix.torchaudio }} --index-url https://download.pytorch.org/whl/cpu\n pip3 install --user onnx onnxscript onnxruntime\n pip3 install --user \"transformers<=${{ matrix.transformers }}\" diffusers \"safetensors<=0.6.2\"\n\n - name: setup-pytorch-execstack-or-patchelf\n if: ${{ matrix.python }} == '3.8' || ${{ matrix.python }} == '3.9'\n run: |\n execstack -c ${{ env.PYTHONUSERBASE }}/lib/python${{ matrix.python }}/site-packages/torch/lib/libtorch_cpu.so || true\n patchelf --clear-execstack ${{ env.PYTHONUSERBASE }}/lib/python${{ matrix.python }}/site-packages/torch/lib/libtorch_cpu.so || true\n\n - name: python-ncnn\n run: |\n export CMAKE_BUILD_PARALLEL_LEVEL=8\n pip3 install --user . --verbose\n\n - name: pnnx\n run: |\n cd tools/pnnx\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=Release \\\n -DTorch_INSTALL_DIR=$GITHUB_WORKSPACE/libtorch-${{ env.LIBTORCH_VERSION }}-install \\\n -DTorchVision_INSTALL_DIR=$GITHUB_WORKSPACE/torchvision-${{ env.TORCHVISION_VERSION }}-install \\\n -Donnxruntime_INSTALL_DIR=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install \\\n -Dprotobuf_DIR=$GITHUB_WORKSPACE/onnxruntime-${{ env.ONNXRUNTIME_VERSION }}-install/lib/cmake/protobuf ..\n\n - name: download-pnnx\n uses: actions/download-artifact@v8\n with:\n name: pnnx\n path: tools/pnnx/build/src\n\n - name: test\n run: |\n export PATH=${{ env.PYTHONUSERBASE }}/bin:$PATH\n chmod +x tools/pnnx/build/src/pnnx\n export OMP_THREAD_LIMIT=1\n export OMP_NUM_THREADS=1\n export MKL_NUM_THREADS=1\n export MKL_ENABLE_INSTRUCTIONS=SSE4_2\n cd tools/pnnx/build\n ctest --output-on-failure -j 8\n\n - name: python-pnnx\n run: |\n export PATH=${{ env.PYTHONUSERBASE }}/bin:$PATH\n export PNNX_WHEEL_WITHOUT_BUILD=ON\n cd tools/pnnx/python\n cp ../build/src/pnnx pnnx/\n python3 setup.py install --user\n pytest tests\n" }, { "path": ".github/workflows/python.yml", "content": "name: python\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/python.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'python/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/python.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'python/**'\n - 'glslang'\nconcurrency:\n group: python-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n MAC_DEPLOYMENT_TARGET: '11.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\n CMAKE_BUILD_PARALLEL_LEVEL: 4\n UseMultiToolTask: true\npermissions:\n contents: read\n\njobs:\n build:\n strategy:\n matrix:\n os: [ubuntu-latest, macos-15-intel, windows-latest]\n python-version: [3.9, 3.12]\n\n runs-on: ${{ matrix.os }}\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-swiftshader\n if: matrix.os == 'ubuntu-latest'\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-linux-install-20240622\n - name: checkout-swiftshader\n if: matrix.os == 'ubuntu-latest' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: de870ac7518fe2b6bb651ecc22fc36647cf7b986\n - name: checkout-swiftshader-submodules\n if: matrix.os == 'ubuntu-latest' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n - name: swiftshader\n if: matrix.os == 'ubuntu-latest' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n mkdir -p build; cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . -j $(nproc)\n mkdir $GITHUB_WORKSPACE/swiftshader-install\n cp Linux/* $GITHUB_WORKSPACE/swiftshader-install\n\n - name: setup-python\n uses: actions/setup-python@v6\n with:\n python-version: ${{ matrix.python-version }}\n - name: install-deps\n run: |\n python -m pip install --upgrade pip\n pip install pytest setuptools wheel twine importlib-metadata\n\n - name: build\n if: matrix.os == 'ubuntu-latest'\n env:\n CC: clang\n CXX: clang++\n run: |\n mkdir build && cd build\n cmake -DNCNN_VULKAN=ON -DNCNN_PYTHON=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j $(nproc)\n - name: build\n if: matrix.os == 'macos-15-intel'\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DPLATFORM=MAC -DARCHS=\"x86_64\" \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET -DENABLE_BITCODE=$ENABLE_BITCODE -DENABLE_ARC=$ENABLE_ARC -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DNCNN_VULKAN=OFF -DNCNN_PYTHON=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . -j 4\n - name: build\n if: matrix.os == 'windows-latest'\n run: |\n mkdir build; cd build\n cmake -T v142,host=x64 -A x64 -DNCNN_VULKAN=OFF -DNCNN_PYTHON=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . --config Release -j 4\n - name: build-python\n run: cd python && pip install .\n - name: test\n if: matrix.os == 'ubuntu-latest'\n run: |\n export VK_ICD_FILENAMES=\"$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json\"\n cd python && pytest tests\n - name: test\n if: matrix.os != 'ubuntu-latest'\n run: |\n cd python && pytest tests\n" }, { "path": ".github/workflows/release-python.yml", "content": "name: release-python\non:\n push:\n tags:\n - '*'\n workflow_dispatch:\n\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n MAC_DEPLOYMENT_TARGET: '11.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\n CIBW_SKIP: \"cp3??t-*\"\n\njobs:\n build_sdist:\n name: Build SDist\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - uses: actions/setup-python@v6\n with:\n python-version: '3.x'\n\n - name: Install deps\n run: python -m pip install twine build\n\n - name: Build SDist\n run: python -m build -s\n\n - name: Check metadata\n run: twine check dist/*\n\n - uses: actions/upload-artifact@v6\n with:\n name: sdist\n path: dist/*.tar.gz\n\n build_wheels:\n name: ${{ matrix.arch }} ${{ matrix.build_id }} on ${{ matrix.os }}\n runs-on: ${{ matrix.os }}\n strategy:\n fail-fast: false\n matrix:\n include:\n - { os: ubuntu-24.04, arch: x86_64, build: 'cp*-manylinux*', build_id: cp-manylinux }\n - { os: ubuntu-24.04, arch: x86_64, build: 'cp*-musllinux*', build_id: cp-musllinux }\n - { os: ubuntu-24.04, arch: x86_64, build: 'pp*', build_id: pp }\n - { os: ubuntu-24.04, arch: i686, build: 'cp*-manylinux*', build_id: cp-manylinux }\n - { os: ubuntu-24.04, arch: i686, build: 'cp*-musllinux*', build_id: cp-musllinux }\n - { os: ubuntu-24.04, arch: i686, build: 'pp*', build_id: pp }\n - { os: windows-2025, arch: x86, build: 'cp*', build_id: cp }\n - { os: windows-2025, arch: AMD64, build: 'cp*', build_id: cp }\n - { os: windows-2025, arch: AMD64, build: 'pp*', build_id: pp }\n - { os: windows-11-arm, arch: ARM64, build: 'cp*', build_id: cp }\n - { os: macos-15-intel, arch: x86_64, build: 'cp*', build_id: cp }\n - { os: macos-15, arch: arm64, build: 'cp*', build_id: cp }\n - { os: ubuntu-24.04-arm, arch: armv7l, build: 'cp*-manylinux*', build_id: cp-manylinux }\n - { os: ubuntu-24.04-arm, arch: armv7l, build: 'cp*-musllinux*', build_id: cp-musllinux }\n - { os: ubuntu-24.04-arm, arch: aarch64, build: 'cp*-manylinux*', build_id: cp-manylinux }\n - { os: ubuntu-24.04-arm, arch: aarch64, build: 'cp*-musllinux*', build_id: cp-musllinux }\n - { os: ubuntu-24.04-arm, arch: aarch64, build: 'pp*', build_id: pp }\n\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n # build wheels for ubuntu\n - name: Build wheels for ubuntu\n if: matrix.os == 'ubuntu-24.04'\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_LINUX: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4\n with:\n output-dir: wheelhouse\n\n # build wheels for ubuntu armv7l\n - name: Build wheels for ubuntu armv7l\n if: matrix.os == 'ubuntu-24.04-arm' && (matrix.arch == 'armv7l')\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_LINUX: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4\n CFLAGS=\"-mfpu=neon\" CXXFLAGS=\"-mfpu=neon\"\n with:\n output-dir: wheelhouse\n\n # build wheels for ubuntu aarch64\n - name: Build wheels for ubuntu aarch64\n if: matrix.os == 'ubuntu-24.04-arm' && (matrix.arch == 'aarch64')\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_LINUX: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4\n with:\n output-dir: wheelhouse\n\n # build wheels for windows\n - name: Build wheels for windows\n if: matrix.os == 'windows-2025' && (matrix.arch == 'AMD64' || matrix.arch == 'x86')\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_WINDOWS: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT_WINDOWS: CMAKE_BUILD_PARALLEL_LEVEL=4\n CIBW_BEFORE_BUILD: pip install delvewheel\n CIBW_REPAIR_WHEEL_COMMAND: delvewheel repair -w {dest_dir} {wheel}\n with:\n output-dir: wheelhouse\n\n - name: Build wheels for windows ARM64\n if: matrix.os == 'windows-11-arm' && matrix.arch == 'ARM64'\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_WINDOWS: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT_WINDOWS: CMAKE_BUILD_PARALLEL_LEVEL=4\n CIBW_BEFORE_BUILD: pip install delvewheel\n CIBW_REPAIR_WHEEL_COMMAND: delvewheel repair -w {dest_dir} {wheel} --no-dll \"msvcp140.dll;vcomp140.dll\"\n with:\n output-dir: wheelhouse\n\n # build wheels for macos\n - name: cache-openmp for macos\n if: matrix.os == 'macos-15-intel' || matrix.os == 'macos-15'\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-macos-install-20251004\n\n - name: openmp for macos\n if: (matrix.os == 'macos-15-intel' || matrix.os == 'macos-15') && steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n\n - name: openmp-build-x86_64 for macos\n if: (matrix.os == 'macos-15-intel' || matrix.os == 'macos-15') && steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n\n - name: openmp-build-arm64 for macos\n if: (matrix.os == 'macos-15-intel' || matrix.os == 'macos-15') && steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n\n - name: openmp-merge-fat-library for macos\n if: (matrix.os == 'macos-15-intel' || matrix.os == 'macos-15') && steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n\n - name: install-openmp for macos\n if: matrix.os == 'macos-15-intel' || matrix.os == 'macos-15'\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/include/* $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/lib/libomp.a $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib\n\n - name: vulkansdk for macos\n if: matrix.os == 'macos-15-intel' || matrix.os == 'macos-15'\n run: |\n wget -q https://sdk.lunarg.com/sdk/download/1.4.335.1/mac/vulkansdk-macos-1.4.335.1.zip?Human=true -O vulkansdk-macos-1.4.335.1.zip\n unzip vulkansdk-macos-1.4.335.1.zip\n sudo vulkansdk-macOS-1.4.335.1.app/Contents/MacOS/vulkansdk-macOS-1.4.335.1 --root $GITHUB_WORKSPACE/vulkansdk-macos-1.4.335.1 --accept-licenses --default-answer --confirm-command install\n\n - name: Build wheels for macos x86_64\n if: matrix.os == 'macos-15-intel' && matrix.arch == 'x86_64'\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_MACOS: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4\n CMAKE_TOOLCHAIN_FILE=$GITHUB_WORKSPACE/toolchains/ios.toolchain.cmake PLATFORM=MAC ARCHS=\"x86_64\"\n DEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET ENABLE_BITCODE=OFF ENABLE_ARC=OFF ENABLE_VISIBILITY=OFF\n OpenMP_C_FLAGS=\"-Xclang -fopenmp\" OpenMP_CXX_FLAGS=\"-Xclang -fopenmp\"\n OpenMP_C_LIB_NAMES=\"libomp\" OpenMP_CXX_LIB_NAMES=\"libomp\"\n OpenMP_libomp_LIBRARY=\"libomp.a\"\n Vulkan_LIBRARY=$GITHUB_WORKSPACE/vulkansdk-macos-1.4.335.1/macOS/lib/libMoltenVK.dylib\n MACOSX_DEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET\n with:\n output-dir: wheelhouse\n\n - name: Build wheels for macos arm64\n if: matrix.os == 'macos-15' && matrix.arch == 'arm64'\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_MACOS: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build }}\n CIBW_ENABLE: pypy\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4\n CMAKE_TOOLCHAIN_FILE=$GITHUB_WORKSPACE/toolchains/ios.toolchain.cmake PLATFORM=MAC_ARM64 ARCHS=\"arm64\"\n DEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET ENABLE_BITCODE=OFF ENABLE_ARC=OFF ENABLE_VISIBILITY=OFF\n OpenMP_C_FLAGS=\"-Xclang -fopenmp\" OpenMP_CXX_FLAGS=\"-Xclang -fopenmp\"\n OpenMP_C_LIB_NAMES=\"libomp\" OpenMP_CXX_LIB_NAMES=\"libomp\"\n OpenMP_libomp_LIBRARY=\"libomp.a\"\n Vulkan_LIBRARY=$GITHUB_WORKSPACE/vulkansdk-macos-1.4.335.1/macOS/lib/libMoltenVK.dylib\n MACOSX_DEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET\n with:\n output-dir: wheelhouse\n\n - name: Show files\n run: ls -lh wheelhouse\n shell: bash\n\n - name: Verify clean directory\n run: git diff --exit-code\n shell: bash\n\n - name: Upload wheels\n uses: actions/upload-artifact@v6\n with:\n name: wheels-${{ matrix.os }}-${{ matrix.arch }}-${{ matrix.build_id }}\n path: wheelhouse/*.whl\n\n build_wheels_qemu_cp:\n name: ${{ matrix.arch }} ${{ matrix.build_cp }} ${{ matrix.build_sub }}\n runs-on: ubuntu-24.04\n\n strategy:\n fail-fast: false\n matrix:\n arch: [riscv64]\n build_cp: [cp38, cp39, cp310, cp311, cp312, cp313, cp314]\n build_sub: [manylinux, musllinux]\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: Set up QEMU\n uses: docker/setup-qemu-action@v3\n with:\n platforms: all\n\n - name: Build wheels with qemu\n uses: pypa/cibuildwheel@v3.3.1\n env:\n CIBW_ARCHS_LINUX: ${{ matrix.arch }}\n CIBW_BUILD: ${{ matrix.build_cp }}-${{ matrix.build_sub }}*\n CIBW_BUILD_VERBOSITY: 1\n CIBW_ENVIRONMENT: CMAKE_BUILD_PARALLEL_LEVEL=4 EXTRA_CMAKE_ARGS=\"-DNCNN_XTHEADVECTOR=OFF\"\n with:\n output-dir: wheelhouse\n\n - name: Show files\n run: ls -lh wheelhouse\n shell: bash\n\n - name: Verify clean directory\n run: git diff --exit-code\n shell: bash\n\n - name: Upload wheels\n uses: actions/upload-artifact@v6\n with:\n name: wheels_qemu_cp-${{ matrix.arch }}-${{ matrix.build_cp }}-${{ matrix.build_sub }}\n path: wheelhouse/*.whl\n\n upload_all:\n permissions:\n contents: none\n name: Upload\n needs: [build_wheels, build_wheels_qemu_cp, build_sdist]\n runs-on: ubuntu-latest\n\n steps:\n - uses: actions/download-artifact@v8\n with:\n path: dist\n merge-multiple: true\n\n - uses: pypa/gh-action-pypi-publish@release/v1\n with:\n user: __token__\n password: ${{ secrets.PYPI_API_TOKEN }}\n" }, { "path": ".github/workflows/release.yml", "content": "name: release\non:\n push:\n tags:\n - '*'\n\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n IOS_DEPLOYMENT_TARGET: '13.0'\n MAC_DEPLOYMENT_TARGET: '11.0'\n MAC_CATALYST_DEPLOYMENT_TARGET: '13.1'\n WATCHOS_DEPLOYMENT_TARGET: '6.0'\n TVOS_DEPLOYMENT_TARGET: '11.0'\n VISIONOS_DEPLOYMENT_TARGET: '1.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\n EMSCRIPTEN_VERSION: 3.1.28\n\npermissions:\n contents: read\n\njobs:\n\n setup:\n permissions:\n contents: none\n runs-on: ubuntu-latest\n outputs:\n VERSION: ${{ steps.get_version.outputs.VERSION }}\n steps:\n - name: get-version\n id: get_version\n run: echo \"VERSION=${GITHUB_REF/refs\\/tags\\//}\" >> $GITHUB_OUTPUT\n\n full-source:\n needs: [setup]\n runs-on: ubuntu-latest\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-full-source\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: package\n run: |\n rm -rf .git\n rm -f /tmp/${{ env.PACKAGENAME }}.zip\n zip -9 -y -r /tmp/${{ env.PACKAGENAME }}.zip .\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: /tmp/${{ env.PACKAGENAME }}.zip\n\n ubuntu:\n needs: [setup]\n strategy:\n matrix:\n opt:\n - { shared-lib: OFF, os: ubuntu-22.04, id: ubuntu-2204 }\n - { shared-lib: OFF, os: ubuntu-24.04, id: ubuntu-2404 }\n - { shared-lib: ON, os: ubuntu-22.04, id: ubuntu-2204-shared }\n - { shared-lib: ON, os: ubuntu-24.04, id: ubuntu-2404-shared }\n runs-on: ${{ matrix.opt.os }}\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: apt\n run: |\n sudo apt-get install -y libprotobuf-dev protobuf-compiler\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TOOLS=ON -DNCNN_BUILD_BENCHMARK=OFF -DNCNN_SHARED_LIB=${{ matrix.opt.shared-lib }} ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: package\n run: |\n rm -rf ${{ env.PACKAGENAME }}\n mkdir -p ${{ env.PACKAGENAME }}\n cp -a build/install/* ${{ env.PACKAGENAME }}\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip ${{ env.PACKAGENAME }}\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-macos:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-macos-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-macos\n path: openmp-install\n\n macos:\n needs: [setup, openmp-macos]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: macos }\n - { vulkan: ON, id: macos-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_TOOLS=OFF \\\n -DNCNN_BUILD_EXAMPLES=OFF \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-macos\n uses: actions/download-artifact@v8\n with:\n name: openmp-macos\n path: openmp-macos\n - name: install-openmp\n run: |\n sudo cp openmp-macos/include/* $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include\n sudo cp openmp-macos/lib/libomp.a $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-macos/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-macos/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-x86_64/install/lib/libglslang.a \\\n build-x86_64/install/lib/libSPIRV.a \\\n -o build-x86_64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n lipo -create build-x86_64/install/lib/libglslang_combined.a build-arm64/install/lib/libglslang_combined.a -o glslang.framework/Versions/A/glslang\n cp -a build-x86_64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create build-x86_64/install/lib/libncnn.a build-arm64/install/lib/libncnn.a -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-ios:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-ios-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=OS64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n cp openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-ios\n path: openmp-install\n\n ios:\n needs: [setup, openmp-ios]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: ios }\n - { vulkan: ON, id: ios-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-ios\n uses: actions/download-artifact@v8\n with:\n name: openmp-ios\n path: openmp-ios\n - name: install-openmp\n run: |\n sudo cp openmp-ios/include/* $DEVELOPER_DIR/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/include\n sudo cp openmp-ios/lib/libomp.a $DEVELOPER_DIR/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=OS64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-ios/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-ios/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n cp build-arm64/install/lib/libglslang_combined.a glslang.framework/Versions/A/glslang\n cp -a build-arm64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n cp build-arm64/install/lib/libncnn.a ncnn.framework/Versions/A/ncnn\n cp -a build-arm64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-ios-simulator:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-ios-simulator-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR64 -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATORARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-ios-simulator\n path: openmp-install\n\n ios-simulator:\n needs: [setup, openmp-ios-simulator]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: ios-simulator }\n - { vulkan: ON, id: ios-simulator-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$IOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-ios-simulator\n uses: actions/download-artifact@v8\n with:\n name: openmp-ios-simulator\n path: openmp-ios-simulator\n - name: install-openmp\n run: |\n sudo cp openmp-ios-simulator/include/* $DEVELOPER_DIR/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/include\n sudo cp openmp-ios-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR64 -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATORARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-ios-simulator/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-ios-simulator/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-x86_64/install/lib/libglslang.a \\\n build-x86_64/install/lib/libSPIRV.a \\\n -o build-x86_64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n lipo -create \\\n build-x86_64/install/lib/libglslang_combined.a \\\n build-arm64/install/lib/libglslang_combined.a \\\n -o glslang.framework/Versions/A/glslang\n cp -a build-x86_64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-x86_64/install/lib/libncnn.a \\\n build-arm64/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-mac-catalyst:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_CATALYST_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-mac-catalyst-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST_ARM64 -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-mac-catalyst\n path: openmp-install\n\n mac-catalyst:\n needs: [setup, openmp-mac-catalyst]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: mac-catalyst }\n - { vulkan: ON, id: mac-catalyst-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$MAC_CATALYST_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-mac-catalyst\n uses: actions/download-artifact@v8\n with:\n name: openmp-mac-catalyst\n path: openmp-mac-catalyst\n - name: install-openmp\n run: |\n sudo cp openmp-mac-catalyst/include/* $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include\n sudo cp openmp-mac-catalyst/lib/libomp.a $DEVELOPER_DIR/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=MAC_CATALYST -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-mac-catalyst/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-mac-catalyst/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-x86_64/install/lib/libglslang.a \\\n build-x86_64/install/lib/libSPIRV.a \\\n -o build-x86_64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n lipo -create \\\n build-x86_64/install/lib/libglslang_combined.a \\\n build-arm64/install/lib/libglslang_combined.a \\\n -o glslang.framework/Versions/A/glslang\n cp -a build-x86_64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-x86_64/install/lib/libncnn.a \\\n build-arm64/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-watchos:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-watchos-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-armv7k\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-armv7k && cd build-armv7k\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"armv7k\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64_32\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64_32 && cd build-arm64_32\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"arm64_32\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64_32/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-armv7k/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64_32/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-watchos\n path: openmp-install\n\n watchos:\n needs: [setup, openmp-watchos]\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-watchos\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n\n steps:\n - uses: actions/checkout@v6\n - name: download-openmp-watchos\n uses: actions/download-artifact@v8\n with:\n name: openmp-watchos\n path: openmp-watchos\n - name: install-openmp\n run: |\n sudo cp openmp-watchos/include/* $DEVELOPER_DIR/Platforms/WatchOS.platform/Developer/SDKs/WatchOS.sdk/usr/include\n sudo cp openmp-watchos/lib/libomp.a $DEVELOPER_DIR/Platforms/WatchOS.platform/Developer/SDKs/WatchOS.sdk/usr/lib\n - name: build-armv7k\n run: |\n mkdir build-armv7k && cd build-armv7k\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"armv7k\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64_32\n run: |\n mkdir build-arm64_32 && cd build-arm64_32\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"arm64_32\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-watchos/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-watchos/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-armv7k/install/lib/libncnn.a \\\n build-arm64_32/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-arm64_32/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-watchos-simulator:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-watchos-simulator-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-watchos-simulator\n path: openmp-install\n\n watchos-simulator:\n needs: [setup, openmp-watchos-simulator]\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-watchos-simulator\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n\n steps:\n - uses: actions/checkout@v6\n - name: download-openmp-watchos-simulator\n uses: actions/download-artifact@v8\n with:\n name: openmp-watchos-simulator\n path: openmp-watchos-simulator\n - name: install-openmp\n run: |\n sudo cp openmp-watchos-simulator/include/* $DEVELOPER_DIR/Platforms/WatchSimulator.platform/Developer/SDKs/WatchSimulator.sdk/usr/include\n sudo cp openmp-watchos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/WatchSimulator.platform/Developer/SDKs/WatchSimulator.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-watchos-simulator/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-watchos-simulator/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-x86_64/install/lib/libncnn.a \\\n build-arm64/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-tvos:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-tvos-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64e\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64e && cd build-arm64e\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64e\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64e/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-tvos\n path: openmp-install\n\n tvos:\n needs: [setup, openmp-tvos]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: tvos }\n - { vulkan: ON, id: tvos-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-tvos\n uses: actions/download-artifact@v8\n with:\n name: openmp-tvos\n path: openmp-tvos\n - name: install-openmp\n run: |\n sudo cp openmp-tvos/include/* $DEVELOPER_DIR/Platforms/AppleTVOS.platform/Developer/SDKs/AppleTVOS.sdk/usr/include\n sudo cp openmp-tvos/lib/libomp.a $DEVELOPER_DIR/Platforms/AppleTVOS.platform/Developer/SDKs/AppleTVOS.sdk/usr/lib\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64e\n run: |\n mkdir build-arm64e && cd build-arm64e\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64e\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-tvos/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-tvos/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64e/install/lib/libglslang.a \\\n build-arm64e/install/lib/libSPIRV.a \\\n -o build-arm64e/install/lib/libglslang_combined.a\n lipo -create \\\n build-arm64/install/lib/libglslang_combined.a \\\n build-arm64e/install/lib/libglslang_combined.a \\\n -o glslang.framework/Versions/A/glslang\n cp -a build-arm64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-arm64/install/lib/libncnn.a \\\n build-arm64e/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-arm64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-tvos-simulator:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-tvos-simulator-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-tvos-simulator\n path: openmp-install\n\n tvos-simulator:\n needs: [setup, openmp-tvos-simulator]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: tvos-simulator }\n - { vulkan: ON, id: tvos-simulator-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-tvos-simulator\n uses: actions/download-artifact@v8\n with:\n name: openmp-tvos-simulator\n path: openmp-tvos-simulator\n - name: install-openmp\n run: |\n sudo cp openmp-tvos-simulator/include/* $DEVELOPER_DIR/Platforms/AppleTVSimulator.platform/Developer/SDKs/AppleTVSimulator.sdk/usr/include\n sudo cp openmp-tvos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/AppleTVSimulator.platform/Developer/SDKs/AppleTVSimulator.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-tvos-simulator/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-tvos-simulator/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-x86_64/install/lib/libglslang.a \\\n build-x86_64/install/lib/libSPIRV.a \\\n -o build-x86_64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n lipo -create \\\n build-x86_64/install/lib/libglslang_combined.a \\\n build-arm64/install/lib/libglslang_combined.a \\\n -o glslang.framework/Versions/A/glslang\n cp -a build-x86_64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-x86_64/install/lib/libncnn.a \\\n build-arm64/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-visionos:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-visionos-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n cp openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-visionos\n path: openmp-install\n\n visionos:\n needs: [setup, openmp-visionos]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: visionos }\n - { vulkan: ON, id: visionos-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-visionos\n uses: actions/download-artifact@v8\n with:\n name: openmp-visionos\n path: openmp-visionos\n - name: install-openmp\n run: |\n sudo cp openmp-visionos/include/* $DEVELOPER_DIR/Platforms/XROS.platform/Developer/SDKs/XROS.sdk/usr/include\n sudo cp openmp-visionos/lib/libomp.a $DEVELOPER_DIR/Platforms/XROS.platform/Developer/SDKs/XROS.sdk/usr/lib\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-visionos/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-visionos/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n cp build-arm64/install/lib/libglslang_combined.a glslang.framework/Versions/A/glslang\n cp -a build-arm64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n cp build-arm64/install/lib/libncnn.a ncnn.framework/Versions/A/ncnn\n cp -a build-arm64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n openmp-visionos-simulator:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n steps:\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-visionos-simulator-release-18.1.2-20251004\n - name: checkout\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: build-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-x86_64 && cd build-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n rm -rf $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/include $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/lib/libomp.a\n - name: upload\n uses: actions/upload-artifact@v6\n with:\n name: openmp-visionos-simulator\n path: openmp-install\n\n visionos-simulator:\n needs: [setup, openmp-visionos-simulator]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, id: visionos-simulator }\n - { vulkan: ON, id: visionos-simulator-vulkan }\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: download-openmp-visionos-simulator\n uses: actions/download-artifact@v8\n with:\n name: openmp-visionos-simulator\n path: openmp-visionos-simulator\n - name: install-openmp\n run: |\n sudo cp openmp-visionos-simulator/include/* $DEVELOPER_DIR/Platforms/XRSimulator.platform/Developer/SDKs/XRSimulator.sdk/usr/include\n sudo cp openmp-visionos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/XRSimulator.platform/Developer/SDKs/XRSimulator.sdk/usr/lib\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: build-arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install/strip\n - name: package-openmp\n run: |\n rm -rf openmp.framework\n mkdir -p openmp.framework/Versions/A/Headers\n mkdir -p openmp.framework/Versions/A/Resources\n ln -s A openmp.framework/Versions/Current\n ln -s Versions/Current/Headers openmp.framework/Headers\n ln -s Versions/Current/Resources openmp.framework/Resources\n ln -s Versions/Current/openmp openmp.framework/openmp\n cp openmp-visionos-simulator/lib/libomp.a openmp.framework/Versions/A/openmp\n cp -a openmp-visionos-simulator/include/* openmp.framework/Versions/A/Headers/\n sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/18.1/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n - name: package-glslang\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -rf glslang.framework\n mkdir -p glslang.framework/Versions/A/Headers\n mkdir -p glslang.framework/Versions/A/Resources\n ln -s A glslang.framework/Versions/Current\n ln -s Versions/Current/Headers glslang.framework/Headers\n ln -s Versions/Current/Resources glslang.framework/Resources\n ln -s Versions/Current/glslang glslang.framework/glslang\n libtool -static \\\n build-x86_64/install/lib/libglslang.a \\\n build-x86_64/install/lib/libSPIRV.a \\\n -o build-x86_64/install/lib/libglslang_combined.a\n libtool -static \\\n build-arm64/install/lib/libglslang.a \\\n build-arm64/install/lib/libSPIRV.a \\\n -o build-arm64/install/lib/libglslang_combined.a\n lipo -create \\\n build-x86_64/install/lib/libglslang_combined.a \\\n build-arm64/install/lib/libglslang_combined.a \\\n -o glslang.framework/Versions/A/glslang\n cp -a build-x86_64/install/include/glslang glslang.framework/Versions/A/Headers/\n sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n - name: package-ncnn\n run: |\n rm -rf ncnn.framework\n mkdir -p ncnn.framework/Versions/A/Headers\n mkdir -p ncnn.framework/Versions/A/Resources\n ln -s A ncnn.framework/Versions/Current\n ln -s Versions/Current/Headers ncnn.framework/Headers\n ln -s Versions/Current/Resources ncnn.framework/Resources\n ln -s Versions/Current/ncnn ncnn.framework/ncnn\n lipo -create \\\n build-x86_64/install/lib/libncnn.a \\\n build-arm64/install/lib/libncnn.a \\\n -o ncnn.framework/Versions/A/ncnn\n cp -a build-x86_64/install/include/* ncnn.framework/Versions/A/Headers/\n sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n - name: package\n if: matrix.opt.vulkan == 'OFF'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework ncnn.framework\n - name: package\n if: matrix.opt.vulkan == 'ON'\n run: |\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.framework glslang.framework ncnn.framework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n android:\n needs: [setup]\n strategy:\n matrix:\n opt:\n - { vulkan: OFF, shared-lib: OFF, id: android }\n - { vulkan: OFF, shared-lib: ON, id: android-shared }\n - { vulkan: ON, shared-lib: OFF, id: android-vulkan }\n - { vulkan: ON, shared-lib: ON, id: android-vulkan-shared }\n runs-on: ubuntu-latest\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-${{ matrix.opt.id }}\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_LATEST_HOME/build/cmake/android.toolchain.cmake \\\n -DANDROID_PLATFORM=android-21 \\\n -DANDROID_USE_LEGACY_TOOLCHAIN_FILE=False \\\n -DANDROID_SUPPORT_FLEXIBLE_PAGE_SIZES=ON \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_BUILD_BENCHMARK=OFF \\\n -DNCNN_VULKAN=${{ matrix.opt.vulkan }} \\\n -DNCNN_SHARED_LIB=${{ matrix.opt.shared-lib }} \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: ndk-fix-debug\n run: sed -i -e '/^ -g$/d' $ANDROID_NDK_LATEST_HOME/build/cmake/android-legacy.toolchain.cmake\n - name: build-armeabi-v7a\n run: |\n mkdir build-armeabi-v7a && cd build-armeabi-v7a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-arm64-v8a\n run: |\n mkdir build-arm64-v8a && cd build-arm64-v8a\n cmake ${{ env.NCNN_CMAKE_OPTIONS }}-DANDROID_ABI=\"arm64-v8a\" ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-x86\n run: |\n mkdir build-x86 && cd build-x86\n cmake ${{ env.NCNN_CMAKE_OPTIONS }}-DANDROID_ABI=\"x86\" ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-x86_64\n run: |\n mkdir build-x86_64 && cd build-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }}-DANDROID_ABI=\"x86_64\" ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-riscv64\n run: |\n mkdir build-riscv64 && cd build-riscv64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }}-DANDROID_ABI=\"riscv64\" ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: package\n run: |\n rm -rf ${{ env.PACKAGENAME }}\n mkdir -p ${{ env.PACKAGENAME }}\n cp -a build-armeabi-v7a/install ${{ env.PACKAGENAME }}/armeabi-v7a\n cp -a build-arm64-v8a/install ${{ env.PACKAGENAME }}/arm64-v8a\n cp -a build-x86/install ${{ env.PACKAGENAME }}/x86\n cp -a build-x86_64/install ${{ env.PACKAGENAME }}/x86_64\n cp -a build-riscv64/install ${{ env.PACKAGENAME }}/riscv64\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip ${{ env.PACKAGENAME }}\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n webassembly:\n needs: [setup]\n runs-on: ubuntu-latest\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-webassembly\n steps:\n - uses: actions/checkout@v6\n - name: emsdk\n run: |\n git clone https://github.com/emscripten-core/emsdk.git\n cd emsdk\n ./emsdk install $EMSCRIPTEN_VERSION\n ./emsdk activate $EMSCRIPTEN_VERSION\n - name: build\n run: |\n source emsdk/emsdk_env.sh\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-simd\n run: |\n source emsdk/emsdk_env.sh\n mkdir build-simd && cd build-simd\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-threads\n run: |\n source emsdk/emsdk_env.sh\n mkdir build-threads && cd build-threads\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: build-simd-threads\n run: |\n source emsdk/emsdk_env.sh\n mkdir build-simd-threads && cd build-simd-threads\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" \\\n -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\n cmake --build . -j $(nproc)\n cmake --build . --target install/strip\n - name: package\n run: |\n rm -rf ${{ env.PACKAGENAME }}\n mkdir -p ${{ env.PACKAGENAME }}\n cp -a build/install ${{ env.PACKAGENAME }}/basic\n cp -a build-simd/install ${{ env.PACKAGENAME }}/simd\n cp -a build-threads/install ${{ env.PACKAGENAME }}/threads\n cp -a build-simd-threads/install ${{ env.PACKAGENAME }}/simd-threads\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip ${{ env.PACKAGENAME }}\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n windows:\n needs: [setup]\n strategy:\n matrix:\n opt:\n - { shared-lib: OFF, os: windows-2022, toolset-version: v140, windows-sdk-version: 22621, id: vs2015 }\n - { shared-lib: OFF, os: windows-2022, toolset-version: v141, windows-sdk-version: 22621, id: vs2017 }\n - { shared-lib: OFF, os: windows-2022, toolset-version: v142, windows-sdk-version: 22621, id: vs2019 }\n - { shared-lib: OFF, os: windows-2022, toolset-version: v143, windows-sdk-version: 26100, id: vs2022 }\n - { shared-lib: ON, os: windows-2022, toolset-version: v140, windows-sdk-version: 22621, id: vs2015-shared }\n - { shared-lib: ON, os: windows-2022, toolset-version: v141, windows-sdk-version: 22621, id: vs2017-shared }\n - { shared-lib: ON, os: windows-2022, toolset-version: v142, windows-sdk-version: 22621, id: vs2019-shared }\n - { shared-lib: ON, os: windows-2022, toolset-version: v143, windows-sdk-version: 26100, id: vs2022-shared }\n runs-on: ${{ matrix.opt.os }}\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-windows-${{ matrix.opt.id }}\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: |\n -T ${{ matrix.opt.toolset-version }},host=x64 `\n -DCMAKE_BUILD_TYPE=Release `\n -DCMAKE_INSTALL_PREFIX=install `\n -DNCNN_VERSION_STRING=\"${{ needs.setup.outputs.VERSION }}\" `\n -DNCNN_BUILD_EXAMPLES=OFF `\n -DNCNN_BUILD_TOOLS=ON `\n -DNCNN_BUILD_BENCHMARK=OFF `\n -DNCNN_VULKAN=ON `\n -DNCNN_SHARED_LIB=${{ matrix.opt.shared-lib }} `\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: Install VS 2017 (v141) Build Tools\n if: matrix.opt.toolset-version == 'v141'\n run: |\n $vsInstallPath = & \"${env:ProgramFiles(x86)}\\Microsoft Visual Studio\\Installer\\vswhere.exe\" -latest -property installationPath\n Start-Process -FilePath \"${env:ProgramFiles(x86)}\\Microsoft Visual Studio\\Installer\\vs_installer.exe\" -ArgumentList \"modify --installPath `\"$vsInstallPath`\" --add Microsoft.VisualStudio.Component.VC.v141.x86.x64 --quiet --norestart --nocache\" -Wait\n - name: Install and Setup VS 2015 (v140) Build Tools\n if: matrix.opt.toolset-version == 'v140'\n run: |\n $vs140Path = \"C:/vs140_build_tools\"\n Invoke-WebRequest -Uri \"https://aka.ms/vs/15/release/vs_buildtools.exe\" -OutFile vs_buildtools.exe\n Start-Process -FilePath \"vs_buildtools.exe\" -ArgumentList \"--installPath `\"$vs140Path`\" --add Microsoft.VisualStudio.Workload.VCTools --add Microsoft.VisualStudio.Component.VC.140 --quiet --wait --norestart --nocache\" -Wait\n\n $vcvarsPath = (Get-ChildItem -Path $vs140Path -Filter \"vcvars64.bat\" -Recurse | Select-Object -First 1).FullName\n $cmd = \"`\"$vcvarsPath`\" && powershell -Command `\"`$env:PATH;`$env:INCLUDE;`$env:LIB`\"\"\n $output = cmd.exe /c $cmd\n $lines = $output -split \"`r`n\"\n\n echo \"PATH=$($lines[0]);$($env:PATH)\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n echo \"INCLUDE=$($lines[1])\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n echo \"LIB=$($lines[2])\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n\n - uses: GuillaumeFalourd/setup-windows10-sdk-action@v2.4\n with:\n sdk-version: ${{ matrix.opt.windows-sdk-version }}\n\n - name: cache-protobuf\n id: cache-protobuf\n uses: actions/cache@v5\n with:\n path: \"protobuf-install\"\n key: protobuf-${{ matrix.opt.toolset-version }}-x86-x64-install\n - name: protobuf\n if: steps.cache-protobuf.outputs.cache-hit != 'true'\n run: |\n Invoke-WebRequest -Uri https://github.com/protocolbuffers/protobuf/archive/v3.11.2.zip -OutFile protobuf-3.11.2.zip\n 7z x ./protobuf-3.11.2.zip\n cd protobuf-3.11.2\n mkdir build-x86; cd build-x86;\n cmake -T ${{ matrix.opt.toolset-version }},host=x64 -A Win32,version=10.0.${{ matrix.opt.windows-sdk-version }}.0 -DCMAKE_INSTALL_PREFIX=\"$env:GITHUB_WORKSPACE\\protobuf-install\\x86\" -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n cd ..\n mkdir build-x64; cd build-x64;\n cmake -T ${{ matrix.opt.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.opt.windows-sdk-version }}.0 -DCMAKE_INSTALL_PREFIX=\"$env:GITHUB_WORKSPACE\\protobuf-install\\x64\" -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n - name: build-x86\n run: |\n mkdir build-x86; cd build-x86\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -A Win32,version=10.0.${{ matrix.opt.windows-sdk-version }}.0 -Dprotobuf_DIR=\"$env:GITHUB_WORKSPACE\\protobuf-install\\x86\\cmake\" ..\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n - name: build-x64\n run: |\n mkdir build-x64; cd build-x64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -A x64,version=10.0.${{ matrix.opt.windows-sdk-version }}.0 -Dprotobuf_DIR=\"$env:GITHUB_WORKSPACE\\protobuf-install\\x64\\cmake\" ..\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n - name: build-arm64\n if: matrix.opt.toolset-version == 'v142' || matrix.opt.toolset-version == 'v143'\n run: |\n mkdir build-arm64; cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -A arm64,version=10.0.${{ matrix.opt.windows-sdk-version }}.0 ..\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n - name: package\n if: matrix.opt.toolset-version == 'v140' || matrix.opt.toolset-version == 'v141'\n run: |\n mkdir ${{ env.PACKAGENAME }}\n mkdir ${{ env.PACKAGENAME }}/x86\n mkdir ${{ env.PACKAGENAME }}/x64\n Copy-Item -Verbose -Recurse -Path \"build-x86\\install\\*\" -Destination \"${{ env.PACKAGENAME }}\\x86\"\n Copy-Item -Verbose -Recurse -Path \"build-x64\\install\\*\" -Destination \"${{ env.PACKAGENAME }}\\x64\"\n 7z a -r ${{ env.PACKAGENAME }}.zip ${{ env.PACKAGENAME }}\n - name: package\n if: matrix.opt.toolset-version == 'v142' || matrix.opt.toolset-version == 'v143'\n run: |\n mkdir ${{ env.PACKAGENAME }}\n mkdir ${{ env.PACKAGENAME }}/x86\n mkdir ${{ env.PACKAGENAME }}/x64\n mkdir ${{ env.PACKAGENAME }}/arm64\n Copy-Item -Verbose -Recurse -Path \"build-x86\\install\\*\" -Destination \"${{ env.PACKAGENAME }}\\x86\"\n Copy-Item -Verbose -Recurse -Path \"build-x64\\install\\*\" -Destination \"${{ env.PACKAGENAME }}\\x64\"\n Copy-Item -Verbose -Recurse -Path \"build-arm64\\install\\*\" -Destination \"${{ env.PACKAGENAME }}\\arm64\"\n 7z a -r ${{ env.PACKAGENAME }}.zip ${{ env.PACKAGENAME }}\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n\n apple:\n needs: [setup, macos, ios, ios-simulator, mac-catalyst, watchos, watchos-simulator, tvos, tvos-simulator, visionos, visionos-simulator]\n runs-on: macos-15-intel\n env:\n PACKAGENAME: ncnn-${{ needs.setup.outputs.VERSION }}-apple\n steps:\n - run: sudo xcode-select --switch /Applications/Xcode_16.4.0.app\n - name: download\n uses: actions/download-artifact@v8\n with:\n path: artifacts\n\n - name: unzip\n run: |\n mkdir -p ncnn-ios\n mkdir -p ncnn-ios-vulkan\n mkdir -p ncnn-ios-simulator\n mkdir -p ncnn-ios-simulator-vulkan\n mkdir -p ncnn-mac-catalyst\n mkdir -p ncnn-mac-catalyst-vulkan\n mkdir -p ncnn-macos\n mkdir -p ncnn-macos-vulkan\n mkdir -p ncnn-tvos\n mkdir -p ncnn-tvos-vulkan\n mkdir -p ncnn-tvos-simulator\n mkdir -p ncnn-tvos-simulator-vulkan\n mkdir -p ncnn-visionos\n mkdir -p ncnn-visionos-vulkan\n mkdir -p ncnn-visionos-simulator\n mkdir -p ncnn-visionos-simulator-vulkan\n mkdir -p ncnn-watchos\n mkdir -p ncnn-watchos-simulator\n\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-ios/ncnn-${{ needs.setup.outputs.VERSION }}-ios.zip -d ncnn-ios\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-ios-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-ios-vulkan.zip -d ncnn-ios-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-ios-simulator/ncnn-${{ needs.setup.outputs.VERSION }}-ios-simulator.zip -d ncnn-ios-simulator\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-ios-simulator-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-ios-simulator-vulkan.zip -d ncnn-ios-simulator-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-mac-catalyst/ncnn-${{ needs.setup.outputs.VERSION }}-mac-catalyst.zip -d ncnn-mac-catalyst\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-mac-catalyst-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-mac-catalyst-vulkan.zip -d ncnn-mac-catalyst-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-macos/ncnn-${{ needs.setup.outputs.VERSION }}-macos.zip -d ncnn-macos\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-macos-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-macos-vulkan.zip -d ncnn-macos-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-tvos/ncnn-${{ needs.setup.outputs.VERSION }}-tvos.zip -d ncnn-tvos\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-vulkan.zip -d ncnn-tvos-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-simulator/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-simulator.zip -d ncnn-tvos-simulator\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-simulator-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-tvos-simulator-vulkan.zip -d ncnn-tvos-simulator-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-visionos/ncnn-${{ needs.setup.outputs.VERSION }}-visionos.zip -d ncnn-visionos\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-vulkan.zip -d ncnn-visionos-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-simulator/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-simulator.zip -d ncnn-visionos-simulator\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-simulator-vulkan/ncnn-${{ needs.setup.outputs.VERSION }}-visionos-simulator-vulkan.zip -d ncnn-visionos-simulator-vulkan\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-watchos/ncnn-${{ needs.setup.outputs.VERSION }}-watchos.zip -d ncnn-watchos\n unzip -q artifacts/ncnn-${{ needs.setup.outputs.VERSION }}-watchos-simulator/ncnn-${{ needs.setup.outputs.VERSION }}-watchos-simulator.zip -d ncnn-watchos-simulator\n\n - name: create-xcframwork\n run: |\n rm -rf openmp.xcframework\n xcodebuild -create-xcframework \\\n -framework ncnn-macos/openmp.framework \\\n -framework ncnn-ios/openmp.framework \\\n -framework ncnn-ios-simulator/openmp.framework \\\n -framework ncnn-mac-catalyst/openmp.framework \\\n -framework ncnn-watchos/openmp.framework \\\n -framework ncnn-watchos-simulator/openmp.framework \\\n -framework ncnn-tvos/openmp.framework \\\n -framework ncnn-tvos-simulator/openmp.framework \\\n -framework ncnn-visionos/openmp.framework \\\n -framework ncnn-visionos-simulator/openmp.framework \\\n -output openmp.xcframework\n\n rm -rf ncnn.xcframework\n xcodebuild -create-xcframework \\\n -framework ncnn-macos/ncnn.framework \\\n -framework ncnn-ios/ncnn.framework \\\n -framework ncnn-ios-simulator/ncnn.framework \\\n -framework ncnn-mac-catalyst/ncnn.framework \\\n -framework ncnn-watchos/ncnn.framework \\\n -framework ncnn-watchos-simulator/ncnn.framework \\\n -framework ncnn-tvos/ncnn.framework \\\n -framework ncnn-tvos-simulator/ncnn.framework \\\n -framework ncnn-visionos/ncnn.framework \\\n -framework ncnn-visionos-simulator/ncnn.framework \\\n -output ncnn.xcframework\n\n rm -f ${{ env.PACKAGENAME }}.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}.zip openmp.xcframework ncnn.xcframework\n - name: create-xcframwork-vulkan\n run: |\n rm -rf openmp.xcframework\n xcodebuild -create-xcframework \\\n -framework ncnn-macos-vulkan/openmp.framework \\\n -framework ncnn-ios-vulkan/openmp.framework \\\n -framework ncnn-ios-simulator-vulkan/openmp.framework \\\n -framework ncnn-mac-catalyst-vulkan/openmp.framework \\\n -framework ncnn-watchos/openmp.framework \\\n -framework ncnn-watchos-simulator/openmp.framework \\\n -framework ncnn-tvos-vulkan/openmp.framework \\\n -framework ncnn-tvos-simulator-vulkan/openmp.framework \\\n -framework ncnn-visionos/openmp.framework \\\n -framework ncnn-visionos-simulator/openmp.framework \\\n -output openmp.xcframework\n\n rm -rf glslang.xcframework\n xcodebuild -create-xcframework \\\n -framework ncnn-macos-vulkan/glslang.framework \\\n -framework ncnn-ios-vulkan/glslang.framework \\\n -framework ncnn-ios-simulator-vulkan/glslang.framework \\\n -framework ncnn-mac-catalyst-vulkan/glslang.framework \\\n -framework ncnn-tvos-vulkan/glslang.framework \\\n -framework ncnn-tvos-simulator-vulkan/glslang.framework \\\n -framework ncnn-visionos-vulkan/glslang.framework \\\n -framework ncnn-visionos-simulator-vulkan/glslang.framework \\\n -output glslang.xcframework\n\n rm -rf ncnn.xcframework\n xcodebuild -create-xcframework \\\n -framework ncnn-macos-vulkan/ncnn.framework \\\n -framework ncnn-ios-vulkan/ncnn.framework \\\n -framework ncnn-ios-simulator-vulkan/ncnn.framework \\\n -framework ncnn-mac-catalyst-vulkan/ncnn.framework \\\n -framework ncnn-watchos/ncnn.framework \\\n -framework ncnn-watchos-simulator/ncnn.framework \\\n -framework ncnn-tvos-vulkan/ncnn.framework \\\n -framework ncnn-tvos-simulator-vulkan/ncnn.framework \\\n -framework ncnn-visionos-vulkan/ncnn.framework \\\n -framework ncnn-visionos-simulator-vulkan/ncnn.framework \\\n -output ncnn.xcframework\n\n rm -f ${{ env.PACKAGENAME }}-vulkan.zip\n zip -9 -y -r ${{ env.PACKAGENAME }}-vulkan.zip openmp.xcframework glslang.xcframework ncnn.xcframework\n - name: upload-zip\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}\n path: ${{ env.PACKAGENAME }}.zip\n - name: upload-zip-vulkan\n uses: actions/upload-artifact@v6\n with:\n name: ${{ env.PACKAGENAME }}-vulkan\n path: ${{ env.PACKAGENAME }}-vulkan.zip\n\n release:\n permissions:\n contents: write # for softprops/action-gh-release to create a release\n needs: [setup, full-source, ubuntu, macos, ios, ios-simulator, mac-catalyst, watchos, watchos-simulator, tvos, tvos-simulator, android, webassembly, windows, apple]\n runs-on: ubuntu-latest\n steps:\n - name: download\n uses: actions/download-artifact@v8\n with:\n path: artifacts\n\n - name: create-release\n uses: softprops/action-gh-release@v2\n with:\n token: ${{ secrets.GITHUB_TOKEN }}\n tag_name: ${{ needs.setup.outputs.VERSION }}\n name: Release ${{ needs.setup.outputs.VERSION }}\n files: artifacts/*/*.zip\n" }, { "path": ".github/workflows/sync-wiki.yml", "content": "name: sync-wiki\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/sync-wiki.yml'\n - 'docs/**'\nconcurrency:\n group: sync-wiki-${{ github.ref }}\n cancel-in-progress: true\n\npermissions:\n contents: read\n\njobs:\n sync-wiki:\n permissions:\n contents: write # for Git to git push\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: sync\n run: |\n cp -r docs $GITHUB_WORKSPACE/ncnn.wiki\n cd $GITHUB_WORKSPACE/ncnn.wiki\n git config --global user.name \"wiki-sync-bot\"\n git config --global user.email \"wiki-sync-bot@qq.com\"\n git init\n git add .\n git commit -m \"sync\"\n git remote add upstream https://${{ secrets.WIKI_SYNC_BOT_TOKEN }}@github.com/Tencent/ncnn.wiki.git\n git push upstream master -f\n" }, { "path": ".github/workflows/test-coverage.yml", "content": "name: test-coverage\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/test-coverage.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/**'\n - 'tests/**'\n - 'toolchains/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/test-coverage.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/**'\n - 'tests/**'\n - 'toolchains/**'\n - 'glslang'\nconcurrency:\n group: test-coverage-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n linux-gcc-gpu-t4:\n runs-on: [self-hosted, linux, t4]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: build\n env:\n CC: gcc\n CXX: g++\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_VULKAN=ON -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX2=ON -DNCNN_XOP=OFF -DNCNN_AVXVNNI=OFF -DNCNN_AVXNECONVERT=OFF -DNCNN_AVX512=OFF -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 4\n - name: test\n env:\n LD_LIBRARY_PATH: /data/action/install/lib64\n run: cd build && ctest --output-on-failure -j 4\n - name: lcov-collect\n run: |\n cd build\n lcov -d ./src -c -o lcov.info\n lcov -r lcov.info '/usr/*' -o lcov.info\n lcov -r lcov.info '*/install/*' -o lcov.info\n lcov -r lcov.info '*/build/*' -o lcov.info\n lcov --list lcov.info\n\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/.local/bin/codecov\n files: build/lcov.info\n\n linux-gcc-x64:\n name: x64-${{ matrix.name }}\n runs-on: [self-hosted, linux, ubuntu25]\n strategy:\n fail-fast: false\n matrix:\n include:\n - { name: 'none', SSE2: OFF, AVX: OFF, F16C: OFF, FMA: OFF, AVX2: OFF, AVX512: OFF, AVX512VNNI: OFF }\n - { name: 'sse2', SSE2: ON, AVX: OFF, F16C: OFF, FMA: OFF, AVX2: OFF, AVX512: OFF, AVX512VNNI: OFF }\n - { name: 'avx', SSE2: ON, AVX: ON, F16C: OFF, FMA: OFF, AVX2: OFF, AVX512: OFF, AVX512VNNI: OFF }\n - { name: 'avx2', SSE2: ON, AVX: ON, F16C: ON, FMA: ON, AVX2: ON, AVX512: OFF, AVX512VNNI: OFF }\n - { name: 'avx512', SSE2: ON, AVX: ON, F16C: ON, FMA: ON, AVX2: ON, AVX512: ON, AVX512VNNI: OFF }\n - { name: 'avx512vnni', SSE2: ON, AVX: ON, F16C: ON, FMA: ON, AVX2: ON, AVX512: ON, AVX512VNNI: ON }\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_SSE2=${{ matrix.SSE2 }} \\\n -DNCNN_AVX=${{ matrix.AVX }} \\\n -DNCNN_F16C=${{ matrix.F16C }} \\\n -DNCNN_FMA=${{ matrix.FMA }} \\\n -DNCNN_AVX2=${{ matrix.AVX2 }} \\\n -DNCNN_AVX512=${{ matrix.AVX512 }} \\\n -DNCNN_AVX512VNNI=${{ matrix.AVX512VNNI }} \\\n -DNCNN_XOP=OFF \\\n -DNCNN_AVXVNNI=OFF \\\n -DNCNN_AVX512BF16=OFF \\\n -DNCNN_AVX512FP16=OFF \\\n -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n cd build\n ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n - name: build-openmp\n run: |\n mkdir build-openmp && cd build-openmp\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_SSE2=${{ matrix.SSE2 }} \\\n -DNCNN_AVX=${{ matrix.AVX }} \\\n -DNCNN_F16C=${{ matrix.F16C }} \\\n -DNCNN_FMA=${{ matrix.FMA }} \\\n -DNCNN_AVX2=${{ matrix.AVX2 }} \\\n -DNCNN_AVX512=${{ matrix.AVX512 }} \\\n -DNCNN_AVX512VNNI=${{ matrix.AVX512VNNI }} \\\n -DNCNN_XOP=OFF \\\n -DNCNN_AVXVNNI=OFF \\\n -DNCNN_AVX512BF16=OFF \\\n -DNCNN_AVX512FP16=OFF \\\n -DNCNN_OPENMP=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n export OMP_THREAD_LIMIT=1\n export OMP_NUM_THREADS=1\n cd build-openmp\n ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build-openmp\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build-openmp/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info,build-openmp/lcov.info\n\n linux-gcc-x64-simplestl-simplemath:\n name: simplestl-simplemath\n runs-on: [self-hosted, linux, ubuntu25]\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host-c.gcc.toolchain.cmake \\\n -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_SIMPLESTL=ON -DNCNN_SIMPLEMATH=ON \\\n -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n cd build\n ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info\n\n linux-gcc-x64-sde:\n name: sde-${{ matrix.cpu }}\n runs-on: [self-hosted, linux, ubuntu25]\n env:\n SDE_PATH: /data/action/osd/sde-external-9.33.0-2024-01-07-lin\n strategy:\n fail-fast: false\n matrix:\n include:\n - { cpu: hsw, AVX2: ON, AVXVNNI: OFF, AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }\n - { cpu: adl, AVX2: ON, AVXVNNI: ON, AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }\n - { cpu: arl, AVX2: ON, AVXVNNI: ON, AVXVNNIINT8: ON, AVXNECONVERT: ON, AVX512: OFF, AVX512VNNI: OFF, AVX512BF16: OFF, AVX512FP16: OFF }\n - { cpu: spr, AVX2: ON, AVXVNNI: OFF, AVXVNNIINT8: OFF, AVXNECONVERT: OFF, AVX512: ON, AVX512VNNI: ON, AVX512BF16: ON, AVX512FP16: ON }\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_AVX=ON \\\n -DNCNN_F16C=ON \\\n -DNCNN_XOP=OFF \\\n -DNCNN_AVX2=${{ matrix.AVX2 }} \\\n -DNCNN_AVXVNNI=${{ matrix.AVXVNNI }} \\\n -DNCNN_AVXVNNIINT8=${{ matrix.AVXVNNIINT8 }} \\\n -DNCNN_AVXNECONVERT=${{ matrix.AVXNECONVERT }} \\\n -DNCNN_AVX512=${{ matrix.AVX512 }} \\\n -DNCNN_AVX512VNNI=${{ matrix.AVX512VNNI }} \\\n -DNCNN_AVX512BF16=${{ matrix.AVX512BF16 }} \\\n -DNCNN_AVX512FP16=${{ matrix.AVX512FP16 }} \\\n -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-${{ matrix.cpu }};--\" ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info\n\n linux-gcc-x64-sde-combined:\n name: sde-combined\n runs-on: [self-hosted, linux, ubuntu25]\n env:\n SDE_PATH: /data/action/osd/sde-external-9.33.0-2024-01-07-lin\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_OPENMP=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test-p4p\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-p4p;--\" ctest --output-on-failure -j 8\n - name: test-snb\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-snb;--\" ctest --output-on-failure -j 8\n - name: test-hsw\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-hsw;--\" ctest --output-on-failure -j 8\n - name: test-adl\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-adl;--\" ctest --output-on-failure -j 8\n - name: test-arl\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-arl;--\" ctest --output-on-failure -j 8\n - name: test-skx\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-skx;--\" ctest --output-on-failure -j 8\n - name: test-spr\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-spr;--\" ctest --output-on-failure -j 8\n - name: test-gnr\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=$SDE_PATH/sde64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-gnr;--\" ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info\n\n linux-gcc-riscv64-rvv:\n strategy:\n matrix:\n openmp: [ON, OFF]\n runs-on: [self-hosted, linux, ubuntu]\n steps:\n - uses: actions/checkout@v6\n - name: build\n run: |\n export RISCV_ROOT_PATH=/data/action/osd/riscv\n mkdir build\n cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-unknown-linux-gnu.toolchain.cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=ON -DNCNN_ZFH=ON -DNCNN_ZVFH=ON -DNCNN_OPENMP=${{ matrix.openmp }} -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test-vlen256\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n\n - name: test-vlen128\n run: |\n export PATH=/data/action/osd/qemu-install/bin:$PATH\n cd build\n TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"-cpu;rv64,v=true,zfh=true,zvfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;/data/action/osd/riscv/sysroot\" ctest --output-on-failure -j 8\n\n - name: lcov-collect\n run: |\n cd build\n lcov --gcov-tool /data/action/osd/riscv/bin/riscv64-unknown-linux-gnu-gcov -d ./src -c -o lcov.info\n lcov -r lcov.info '/usr/*' -o lcov.info\n lcov -r lcov.info '*/install/*' -o lcov.info\n lcov -r lcov.info '*/build/*' -o lcov.info\n lcov --list lcov.info\n\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n files: build/lcov.info\n\n linux-gpu-llvmpipe:\n runs-on: [self-hosted, linux, ubuntu25]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX2=ON -DNCNN_AVXVNNI=OFF -DNCNN_AVXNECONVERT=OFF -DNCNN_AVX512=ON -DNCNN_AVX512VNNI=ON -DNCNN_AVX512BF16=OFF -DNCNN_AVX512FP16=OFF -DNCNN_XOP=OFF -DNCNN_OPENMP=OFF -DNCNN_VULKAN=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n export LP_NUM_THREADS=4\n cd build && ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info\n\n linux-gpu-swiftshader:\n runs-on: [self-hosted, linux, ubuntu25]\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-swiftshader\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-linux-install-20250508\n - name: checkout-swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: 930d46d31b5d637f313fd5ef55da2bbf053c26c1\n - name: swiftshader\n if: steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n mkdir -p build; cd build\n cmake -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . -j 8\n mkdir $GITHUB_WORKSPACE/swiftshader-install\n cp Linux/* $GITHUB_WORKSPACE/swiftshader-install\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX2=ON -DNCNN_AVXVNNI=OFF -DNCNN_AVXNECONVERT=OFF -DNCNN_AVX512=ON -DNCNN_AVX512VNNI=ON -DNCNN_AVX512BF16=OFF -DNCNN_AVX512FP16=OFF -DNCNN_XOP=OFF -DNCNN_OPENMP=OFF -DNCNN_VULKAN=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n - name: test\n run: |\n printf \"[Processor]\\nThreadCount=1\\n\" > build/tests/SwiftShader.ini\n export VK_ICD_FILENAMES=\"$GITHUB_WORKSPACE/swiftshader-install/vk_swiftshader_icd.json\"\n cd build && ctest --output-on-failure -j 8\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info\n\n linux-gcc-cross:\n name: ${{ matrix.arch }}\n runs-on: [self-hosted, linux, ubuntu25]\n strategy:\n fail-fast: false\n matrix:\n include:\n - arch: arm\n toolchain: arm-linux-gnueabi.toolchain.cmake\n extra-cmake-args: -DNCNN_VFPV4=ON\n qemu: qemu-arm-static\n qemu-args: \"-L;/usr/arm-linux-gnueabi\"\n\n - arch: arm-noinlineasm\n toolchain: arm-linux-gnueabi.toolchain.cmake\n extra-cmake-args: -DNCNN_GNU_INLINE_ASM=OFF -DNCNN_VFPV4=ON\n qemu: qemu-arm-static\n qemu-args: \"-L;/usr/arm-linux-gnueabi\"\n\n - arch: armhf-vfpv3-d16\n toolchain: arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake\n extra-cmake-args: -DNCNN_VFPV4=OFF\n qemu: qemu-arm-static\n qemu-args: \"-L;/usr/arm-linux-gnueabihf\"\n\n - arch: armhf-vfpv3-d16-noinlineasm\n toolchain: arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake\n extra-cmake-args: -DNCNN_GNU_INLINE_ASM=OFF -DNCNN_VFPV4=OFF\n qemu: qemu-arm-static\n qemu-args: \"-L;/usr/arm-linux-gnueabihf\"\n\n - arch: aarch64-armv8.0\n toolchain: aarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_ARM82=OFF\n qemu: qemu-aarch64-static\n qemu-args: \"-L;/usr/aarch64-linux-gnu\"\n\n - arch: aarch64-armv8.2\n toolchain: aarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_ARM82DOT=OFF -DNCNN_ARM82FP16FML=OFF\n qemu: qemu-aarch64-static\n qemu-args: \"-L;/usr/aarch64-linux-gnu\"\n\n - arch: aarch64-armv8.4\n toolchain: aarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_ARM84BF16=OFF -DNCNN_ARM84I8MM=OFF\n qemu: qemu-aarch64-static\n qemu-args: \"-L;/usr/aarch64-linux-gnu\"\n\n - arch: aarch64-armv8.6\n toolchain: aarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_ARM86SVE=OFF\n qemu: qemu-aarch64-static\n qemu-args: \"-L;/usr/aarch64-linux-gnu\"\n\n - arch: aarch64-armv8.6-noinlineasm\n toolchain: aarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_GNU_INLINE_ASM=OFF -DNCNN_ARM86SVE=OFF\n qemu: qemu-aarch64-static\n qemu-args: \"-L;/usr/aarch64-linux-gnu\"\n\n - arch: mipsisa32r6el\n toolchain: mipsisa32r6el-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_MSA=OFF -DNCNN_MMI=OFF\n qemu: qemu-mipsel-static\n qemu-args: \"-L;/usr/mipsisa32r6el-linux-gnu\"\n\n - arch: mipsisa64r6el\n toolchain: mipsisa64r6el-linux-gnuabi64.toolchain.cmake\n extra-cmake-args: -DNCNN_MSA=ON -DNCNN_MMI=OFF\n qemu: qemu-mips64el-static\n qemu-args: \"-L;/usr/mipsisa64r6el-linux-gnuabi64\"\n\n - arch: powerpc\n toolchain: powerpc-linux-gnu.toolchain.cmake\n extra-cmake-args:\n qemu: qemu-ppc-static\n qemu-args: \"-L;/usr/powerpc-linux-gnu\"\n\n - arch: powerpc64le\n toolchain: powerpc64le-linux-gnu.toolchain.cmake\n extra-cmake-args:\n qemu: qemu-ppc64le-static\n qemu-args: \"-L;/usr/powerpc64le-linux-gnu\"\n\n - arch: riscv64\n toolchain: riscv64-linux-gnu.toolchain.cmake\n extra-cmake-args:\n qemu: qemu-riscv64-static\n qemu-args: \"-L;/usr/riscv64-linux-gnu\"\n\n - arch: loongarch64-la264\n toolchain: loongarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_LSX=ON -DNCNN_LASX=OFF\n qemu: qemu-loongarch64-static\n qemu-args: \"-L;/usr/loongarch64-linux-gnu\"\n\n - arch: loongarch64-la664\n toolchain: loongarch64-linux-gnu.toolchain.cmake\n extra-cmake-args: -DNCNN_LSX=ON -DNCNN_LASX=ON\n qemu: qemu-loongarch64-static\n qemu-args: \"-L;/usr/loongarch64-linux-gnu\"\n\n steps:\n - uses: actions/checkout@v6\n\n - name: build\n run: |\n mkdir build && cd build\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/${{ matrix.toolchain }} ${{ matrix.extra-cmake-args }} -DNCNN_OPENMP=OFF \\\n -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=${{ matrix.qemu }} TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"${{ matrix.qemu-args }}\" ctest --output-on-failure -j 8\n\n - name: lcov-collect\n run: |\n cd build\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n\n - name: build-openmp\n run: |\n mkdir build-openmp && cd build-openmp\n cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/${{ matrix.toolchain }} ${{ matrix.extra-cmake-args }} -DNCNN_OPENMP=ON \\\n -DCMAKE_BUILD_TYPE=debug -DNCNN_COVERAGE=ON -DNCNN_RUNTIME_CPU=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j 8\n\n - name: test-openmp\n run: |\n export OMP_THREAD_LIMIT=1\n export OMP_NUM_THREADS=1\n cd build-openmp\n TESTS_EXECUTABLE_LOADER=${{ matrix.qemu }} TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"${{ matrix.qemu-args }}\" ctest --output-on-failure -j 8\n\n - name: lcov-collect-openmp\n run: |\n cd build-openmp\n lcov --ignore-errors inconsistent -d ./src -c -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '/usr/*' -o lcov.info\n lcov --ignore-errors inconsistent -r lcov.info '*/build-openmp/*' -o lcov.info\n lcov --ignore-errors inconsistent --list lcov.info\n\n - name: codecov\n uses: codecov/codecov-action@v5\n with:\n token: ${{ secrets.CODECOV_TOKEN }}\n disable_search: true\n plugins: noop\n binary: /data/action/osd/codecov\n files: build/lcov.info,build-openmp/lcov.info\n" }, { "path": ".github/workflows/tvos.yml", "content": "name: tvos\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/tvos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/tvos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'glslang'\nconcurrency:\n group: tvos-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n TVOS_DEPLOYMENT_TARGET: '11.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$TVOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-tvos-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-arm64e\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64e && cd build-arm64e\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64e\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/tvos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/tvos-simulator\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install/tvos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/tvos/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64e/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/tvos/lib/libomp.a\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/include $GITHUB_WORKSPACE/openmp-install/tvos-simulator\n mkdir -p $GITHUB_WORKSPACE/openmp-install/tvos-simulator/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/tvos-simulator/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/tvos/include/* $DEVELOPER_DIR/Platforms/AppleTVOS.platform/Developer/SDKs/AppleTVOS.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/tvos/lib/libomp.a $DEVELOPER_DIR/Platforms/AppleTVOS.platform/Developer/SDKs/AppleTVOS.sdk/usr/lib\n\n sudo cp $GITHUB_WORKSPACE/openmp-install/tvos-simulator/include/* $DEVELOPER_DIR/Platforms/AppleTVSimulator.platform/Developer/SDKs/AppleTVSimulator.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/tvos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/AppleTVSimulator.platform/Developer/SDKs/AppleTVSimulator.sdk/usr/lib\n\n - name: arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n - name: arm64e\n run: |\n mkdir build-arm64e && cd build-arm64e\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=TVOS -DARCHS=\"arm64e\" ..\n cmake --build . -j 4\n - name: simulator-x86_64\n run: |\n mkdir build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_TVOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n - name: simulator-arm64\n run: |\n mkdir build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATORARM64_TVOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n" }, { "path": ".github/workflows/visionos.yml", "content": "name: visionos\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/visionos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/visionos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\nconcurrency:\n group: visionos-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n VISIONOS_DEPLOYMENT_TARGET: '1.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$VISIONOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n -DNCNN_VULKAN=ON \\\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-visionos-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64 && cd build-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/visionos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/visionos-simulator\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/include $GITHUB_WORKSPACE/openmp-install/visionos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/visionos/lib\n cp openmp-${{ env.OPENMP_VERSION }}.src/build-arm64/install/lib/libomp.a $GITHUB_WORKSPACE/openmp-install/visionos/lib/libomp.a\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/include $GITHUB_WORKSPACE/openmp-install/visionos-simulator\n mkdir -p $GITHUB_WORKSPACE/openmp-install/visionos-simulator/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/visionos-simulator/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/visionos/include/* $DEVELOPER_DIR/Platforms/XROS.platform/Developer/SDKs/XROS.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/visionos/lib/libomp.a $DEVELOPER_DIR/Platforms/XROS.platform/Developer/SDKs/XROS.sdk/usr/lib\n\n sudo cp $GITHUB_WORKSPACE/openmp-install/visionos-simulator/include/* $DEVELOPER_DIR/Platforms/XRSimulator.platform/Developer/SDKs/XRSimulator.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/visionos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/XRSimulator.platform/Developer/SDKs/XRSimulator.sdk/usr/lib\n\n - name: arm64\n run: |\n mkdir build-arm64 && cd build-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n - name: simulator-x86_64\n run: |\n mkdir build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n - name: simulator-arm64\n run: |\n mkdir build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_VISIONOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n" }, { "path": ".github/workflows/watchos.yml", "content": "name: watchos\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/watchos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/watchos.yml'\n - 'toolchains/ios.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\nconcurrency:\n group: watchos-${{ github.ref }}\n cancel-in-progress: true\nenv:\n DEVELOPER_DIR: /Applications/Xcode_16.4.0.app/Contents/Developer\n WATCHOS_DEPLOYMENT_TARGET: '6.0'\n ENABLE_BITCODE: OFF\n ENABLE_ARC: OFF\n ENABLE_VISIBILITY: OFF\npermissions:\n contents: read\n\njobs:\n build:\n runs-on: macos-15-intel\n env:\n OPENMP_VERSION: '18.1.2'\n OPENMP_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DLIBOMP_ENABLE_SHARED=OFF \\\n -DLIBOMP_OMPT_SUPPORT=OFF \\\n -DLIBOMP_USE_HWLOC=OFF \\\n\n NCNN_CMAKE_OPTIONS: |\n -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake \\\n -DDEPLOYMENT_TARGET=$WATCHOS_DEPLOYMENT_TARGET \\\n -DENABLE_BITCODE=$ENABLE_BITCODE \\\n -DENABLE_ARC=$ENABLE_ARC \\\n -DENABLE_VISIBILITY=$ENABLE_VISIBILITY \\\n -DCMAKE_INSTALL_PREFIX=install \\\n -DCMAKE_BUILD_TYPE=Release \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"libomp.a\" \\\n\n steps:\n - uses: actions/checkout@v6\n\n - name: cache-openmp\n id: cache-openmp\n uses: actions/cache@v5\n with:\n path: openmp-install\n key: openmp-watchos-install-20251004\n - name: openmp\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf cmake-${{ env.OPENMP_VERSION }}.src.tar.xz\n wget https://github.com/llvm/llvm-project/releases/download/llvmorg-${{ env.OPENMP_VERSION }}/openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n tar -xf openmp-${{ env.OPENMP_VERSION }}.src.tar.xz\n mv cmake-${{ env.OPENMP_VERSION }}.src/Modules/* openmp-${{ env.OPENMP_VERSION }}.src/cmake/\n cd openmp-${{ env.OPENMP_VERSION }}.src\n wget https://github.com/nihui/llvm-project/commit/ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n patch -p2 -i ef8c35bcf5d9cfdb0764ffde6a63c04ec715bc37.patch\n wget https://github.com/nihui/llvm-project/commit/5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n patch -p2 -i 5c12711f9a21f41bea70566bf15a4026804d6b20.patch\n - name: openmp-armv7k\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-armv7k && cd build-armv7k\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"armv7k\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-arm64_32\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-arm64_32 && cd build-arm64_32\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"arm64_32\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-x86_64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-simulator-arm64\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n cd openmp-${{ env.OPENMP_VERSION }}.src\n mkdir -p build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.OPENMP_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n cmake --build . --target install\n - name: openmp-merge-fat-library\n if: steps.cache-openmp.outputs.cache-hit != 'true'\n run: |\n mkdir -p $GITHUB_WORKSPACE/openmp-install\n mkdir -p $GITHUB_WORKSPACE/openmp-install/watchos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/watchos-simulator\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-arm64_32/install/include $GITHUB_WORKSPACE/openmp-install/watchos\n mkdir -p $GITHUB_WORKSPACE/openmp-install/watchos/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-armv7k/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-arm64_32/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/watchos/lib/libomp.a\n\n cp -a openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/include $GITHUB_WORKSPACE/openmp-install/watchos-simulator\n mkdir -p $GITHUB_WORKSPACE/openmp-install/watchos-simulator/lib\n lipo -create \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-x86_64/install/lib/libomp.a \\\n openmp-${{ env.OPENMP_VERSION }}.src/build-simulator-arm64/install/lib/libomp.a \\\n -o $GITHUB_WORKSPACE/openmp-install/watchos-simulator/lib/libomp.a\n\n - name: install-openmp\n run: |\n sudo cp $GITHUB_WORKSPACE/openmp-install/watchos/include/* $DEVELOPER_DIR/Platforms/WatchOS.platform/Developer/SDKs/WatchOS.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/watchos/lib/libomp.a $DEVELOPER_DIR/Platforms/WatchOS.platform/Developer/SDKs/WatchOS.sdk/usr/lib\n\n sudo cp $GITHUB_WORKSPACE/openmp-install/watchos-simulator/include/* $DEVELOPER_DIR/Platforms/WatchSimulator.platform/Developer/SDKs/WatchSimulator.sdk/usr/include\n sudo cp $GITHUB_WORKSPACE/openmp-install/watchos-simulator/lib/libomp.a $DEVELOPER_DIR/Platforms/WatchSimulator.platform/Developer/SDKs/WatchSimulator.sdk/usr/lib\n\n - name: armv7k\n run: |\n mkdir build-armv7k && cd build-armv7k\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"armv7k\" ..\n cmake --build . -j 4\n - name: arm64_32\n run: |\n mkdir build-arm64_32 && cd build-arm64_32\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=WATCHOS -DARCHS=\"arm64_32\" ..\n cmake --build . -j 4\n\n - name: simulator-x86_64\n run: |\n mkdir build-simulator-x86_64 && cd build-simulator-x86_64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"x86_64\" ..\n cmake --build . -j 4\n - name: simulator-arm64\n run: |\n mkdir build-simulator-arm64 && cd build-simulator-arm64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DPLATFORM=SIMULATOR_WATCHOS -DARCHS=\"arm64\" ..\n cmake --build . -j 4\n" }, { "path": ".github/workflows/web-assembly.yml", "content": "name: web-assembly\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/web-assembly.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/web-assembly.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n\nenv:\n EMSCRIPTEN_VERSION: 3.1.28\n\nconcurrency:\n group: web-assembly-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n webassembly:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v6\n - name: emsdk\n run: |\n git clone https://github.com/emscripten-core/emsdk.git\n cd emsdk\n ./emsdk install $EMSCRIPTEN_VERSION\n ./emsdk activate $EMSCRIPTEN_VERSION\n - name: build-basic\n run: |\n source emsdk/emsdk_env.sh\n export LDFLAGS=\"-sERROR_ON_WASM_CHANGES_AFTER_LINK -sWASM_BIGINT -O1\"\n mkdir build-basic && cd build-basic\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-basic\n run: |\n cd build-basic\n TESTS_EXECUTABLE_LOADER=node ctest --output-on-failure -j $(nproc)\n - name: build-simd\n run: |\n source emsdk/emsdk_env.sh\n export LDFLAGS=\"-sERROR_ON_WASM_CHANGES_AFTER_LINK -sWASM_BIGINT -O1\"\n mkdir build-simd && cd build-simd\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-simd\n run: |\n cd build-simd\n TESTS_EXECUTABLE_LOADER=node ctest --output-on-failure -j $(nproc)\n - name: build-simd-omp\n run: |\n source emsdk/emsdk_env.sh\n export LDFLAGS=\"-sERROR_ON_WASM_CHANGES_AFTER_LINK -sWASM_BIGINT -O1\"\n mkdir build-simd-omp && cd build-simd-omp\n cmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_TESTS=ON ..\n cmake --build . -j $(nproc)\n - name: test-simd-omp\n run: |\n cd build-simd-omp\n TESTS_EXECUTABLE_LOADER=node ctest --output-on-failure -j $(nproc)\n" }, { "path": ".github/workflows/windows-arm.yml", "content": "name: windows-arm\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/windows-arm.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/windows-arm.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\nconcurrency:\n group: windows-arm-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n windows:\n name: ${{ matrix.vs-version }}\n runs-on: windows-2022\n strategy:\n matrix:\n include:\n - vs-version: vs2019\n toolset-version: v142\n windows-sdk-version: 22621\n\n - vs-version: vs2022\n toolset-version: v143\n windows-sdk-version: 26100\n\n env:\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TESTS=OFF -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_VULKAN=ON\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - uses: GuillaumeFalourd/setup-windows10-sdk-action@v2.4\n with:\n sdk-version: ${{ matrix.windows-sdk-version }}\n - name: arm64\n run: |\n mkdir build-arm64; cd build-arm64\n cmake -T ${{ matrix.toolset-version }},host=x64 -A arm64,version=10.0.${{ matrix.windows-sdk-version }}.0 ${{ env.NCNN_CMAKE_OPTIONS }} ..\n cmake --build . --config Release -j 4\n - name: arm64-shared\n run: |\n mkdir build-arm64-shared; cd build-arm64-shared\n cmake -T ${{ matrix.toolset-version }},host=x64 -A arm64,version=10.0.${{ matrix.windows-sdk-version }}.0 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_SHARED_LIB=ON ..\n cmake --build . --config Release -j 4\n\n woa-linux:\n name: woa-linux\n runs-on: ubuntu-latest\n container: linaro/wine-arm64\n steps:\n - uses: actions/checkout@v6\n - name: msvc-wine\n env:\n WINEPREFIX: /tmp/wine-x64-prefix/\n run: |\n apt-get update\n apt-get install -y wine64 python3 msitools python3-simplejson python3-six ca-certificates winbind cmake ninja-build meson\n ln -s /usr/bin/wine /usr/bin/wine64\n xvfb-run winecfg &\n git clone --depth 1 https://github.com/mstorsjo/msvc-wine\n msvc-wine/vsdownload.py --accept-license --dest /msvc\n msvc-wine/install.sh /msvc\n - name: build\n env:\n WINEPREFIX: /tmp/wine-x64-prefix/\n CC: cl\n CXX: cl\n run: |\n export PATH=/msvc/bin/arm64:$PATH\n mkdir build && cd build\n cmake -GNinja -DCMAKE_BUILD_TYPE=Release -DCMAKE_SYSTEM_NAME=Windows -DNCNN_BUILD_TESTS=ON ..\n cmake --build . --config Release -j $(nproc)\n - name: test\n run: |\n cd build\n TESTS_EXECUTABLE_LOADER=wine-arm64 TESTS_EXECUTABLE_LOADER_ARGUMENTS=\"\" ctest --output-on-failure -j $(nproc)\n\n windows-arm:\n runs-on: windows-11-arm\n env:\n UseMultiToolTask: true\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: build\n run: |\n mkdir build; cd build\n cmake -A arm64 -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_VULKAN=OFF -DNCNN_ARM82=OFF ..\n cmake --build . --config Release -j 4\n - name: test\n run: cd build; ctest -C Release --output-on-failure -j 4\n" }, { "path": ".github/workflows/windows-clang.yml", "content": "name: windows-clang\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/windows-clang.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/windows-clang.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/arm/**'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\nconcurrency:\n group: windows-clang-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n windows:\n name: ClangCL\n runs-on: windows-2022\n\n env:\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: arm64\n run: |\n mkdir build-arm64; cd build-arm64\n cmake -T ClangCL -A arm64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=OFF ..\n cmake --build . --config Release -j 4\n\n - name: arm64-vulkan\n run: |\n mkdir build-arm64-vulkan; cd build-arm64-vulkan\n cmake -T ClangCL -A arm64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . --config Release -j 4\n\n - name: x86\n run: |\n mkdir build-x86; cd build-x86\n cmake -T ClangCL -A Win32 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_BUILD_TESTS=ON -DNCNN_VULKAN=OFF ..\n cmake --build . --config Release -j 4\n - name: x86-test\n run: cd build-x86; ctest -C Release --output-on-failure -j 4\n\n - name: x86-vulkan\n run: |\n mkdir build-x86-vulkan; cd build-x86-vulkan\n cmake -T ClangCL -A Win32 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . --config Release -j 4\n\n - name: x64\n run: |\n mkdir build-x64; cd build-x64\n cmake -T ClangCL -A x64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_BUILD_TESTS=ON -DNCNN_VULKAN=OFF ..\n cmake --build . --config Release -j 4\n - name: x64-test\n run: cd build-x64; ctest -C Release --output-on-failure -j 4\n\n - name: x64-vulkan\n run: |\n mkdir build-x64-vulkan; cd build-x64-vulkan\n cmake -T ClangCL -A x64 ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON ..\n cmake --build . --config Release -j 4\n" }, { "path": ".github/workflows/windows-mingw.yml", "content": "name: windows-mingw\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/windows-mingw.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/windows-mingw.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'glslang'\nconcurrency:\n group: windows-mingw-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n windows:\n name: MinGW-w64\n runs-on: windows-2022\n\n env:\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: x64\n run: |\n mkdir build-x64; cd build-x64\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_BUILD_TESTS=ON -DNCNN_VULKAN=OFF -G \"MinGW Makefiles\" ..\n cmake --build . --config Release -j 4\n - name: x64-test\n run: cd build-x64; ctest -C Release --output-on-failure -j 4\n\n - name: x64-vulkan\n run: |\n mkdir build-x64-vulkan; cd build-x64-vulkan\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DNCNN_VULKAN=ON -DNCNN_SHARED_LIB=ON -G \"MinGW Makefiles\" ..\n cmake --build . --config Release -j 4\n" }, { "path": ".github/workflows/windows-xp.yml", "content": "name: windows-xp\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/windows-xp.yml'\n - 'toolchains/windows-xp-msvc.toolchain.cmake'\n - 'toolchains/windows-xp-mingw.toolchain.cmake'\n - 'toolchains/windows-xp-clang.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/windows-xp.yml'\n - 'toolchains/windows-xp-msvc.toolchain.cmake'\n - 'toolchains/windows-xp-mingw.toolchain.cmake'\n - 'toolchains/windows-xp-clang.toolchain.cmake'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'tests/**'\nconcurrency:\n group: windows-xp-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n MSVC:\n runs-on: windows-2025\n\n env:\n VS_INSTALL_DIR: C:\\Program Files\\Microsoft Visual Studio\\2022\\Enterprise\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: config\n shell: cmd\n run: |\n \"C:\\Program Files (x86)\\Microsoft Visual Studio\\Installer\\setup.exe\" modify --installPath \"${{ env.VS_INSTALL_DIR }}\" --channelId VisualStudio.17.Release --add Microsoft.VisualStudio.Component.WinXP --add Microsoft.VisualStudio.Component.VC.Tools.X86.X64.Spectre --add Microsoft.VisualStudio.Component.VC.Tools.X86.X64 --add Microsoft.VisualStudio.Component.VC.Tools.X86.X64 --add Microsoft.VisualStudio.Component.VC.v141.xp --nocache --quiet\n call \"${{ env.VS_INSTALL_DIR }}\\VC\\Auxiliary\\Build\\vcvarsall.bat\" x86\n - name: build\n run: |\n mkdir build; cd build\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -A WIN32 -G \"Visual Studio 17 2022\" -T v141_xp -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_OPENMP=OFF -DNCNN_BUILD_WITH_STATIC_CRT=ON -DNCNN_AVX=OFF -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-msvc.toolchain.cmake\" ..\n cmake --build . --config Release -j 4\n - name: test\n run: cd build; ctest -C Release --output-on-failure -j 4\n\n MinGW-w32:\n runs-on: windows-2025\n\n env:\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n\n - name: config\n run: |\n Invoke-WebRequest -Uri https://github.com/nihui/ncnn-assets/releases/download/toolchain/i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z -OutFile i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z\n 7z x ./i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z\n Add-Content -Path $env:GITHUB_ENV -Value \"MINGW32_ROOT_PATH=${{ github.workspace }}\\mingw32\"\n Add-Content -Path $env:GITHUB_PATH -Value \"${{ github.workspace }}\\mingw32\\bin\"\n - name: build\n run: |\n mkdir build; cd build\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-mingw.toolchain.cmake\" -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_AVX=OFF .. -G \"MinGW Makefiles\"\n cmake --build . --config Release -j 4\n - name: test\n run: cd build; ctest -C Release --output-on-failure -j 4\n\n Clang:\n runs-on: windows-2022\n\n env:\n UseMultiToolTask: true\n NCNN_CMAKE_OPTIONS: -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON\n\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n - name: Set up Clang\n run: choco install llvm --version=6.0.0 --allow-downgrade\n - name: Verify Clang\n run: |\n clang --version\n clang++ --version\n - name: config\n run: |\n Invoke-WebRequest -Uri https://github.com/nihui/ncnn-assets/releases/download/toolchain/i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z -OutFile i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z\n 7z x ./i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z\n Add-Content -Path $env:GITHUB_ENV -Value \"MINGW32_ROOT_PATH=${{ github.workspace }}\\mingw32\"\n echo \"${{ github.workspace }}\\mingw32\\bin;$env:PATH\" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8\n - name: build\n run: |\n mkdir build; cd build\n cmake ${{ env.NCNN_CMAKE_OPTIONS }} -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-clang.toolchain.cmake\" -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_AVX=OFF .. -G \"MinGW Makefiles\"\n cmake --build . --config Release -j 4\n - name: test\n run: cd build; ctest -C Release --output-on-failure -j 4\n" }, { "path": ".github/workflows/windows.yml", "content": "name: windows\non:\n push:\n branches: [master]\n paths:\n - '.github/workflows/windows.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\n pull_request:\n branches: [master]\n paths:\n - '.github/workflows/windows.yml'\n - 'CMakeLists.txt'\n - 'cmake/**'\n - 'src/*'\n - 'src/layer/*'\n - 'src/layer/x86/**'\n - 'src/layer/vulkan/**'\n - 'tests/**'\n - 'tools/**'\n - '!tools/pnnx/**'\n - 'examples/**'\n - 'glslang'\nconcurrency:\n group: windows-${{ github.ref }}\n cancel-in-progress: true\npermissions:\n contents: read\n\njobs:\n msvc:\n name: ${{ matrix.vs-version }}\n runs-on: windows-2022\n strategy:\n matrix:\n include:\n - vs-version: vs2015\n toolset-version: v140\n windows-sdk-version: 22621\n - vs-version: vs2017\n toolset-version: v141\n windows-sdk-version: 22621\n - vs-version: vs2019\n toolset-version: v142\n windows-sdk-version: 26100\n - vs-version: vs2022\n toolset-version: v143\n windows-sdk-version: 26100\n\n env:\n UseMultiToolTask: true\n steps:\n - uses: actions/checkout@v6\n with:\n submodules: true\n \n - name: Install VS 2017 (v141) Build Tools\n if: matrix.vs-version == 'vs2017'\n run: |\n $vsInstallPath = & \"${env:ProgramFiles(x86)}\\Microsoft Visual Studio\\Installer\\vswhere.exe\" -latest -property installationPath\n Start-Process -FilePath \"${env:ProgramFiles(x86)}\\Microsoft Visual Studio\\Installer\\vs_installer.exe\" -ArgumentList \"modify --installPath `\"$vsInstallPath`\" --add Microsoft.VisualStudio.Component.VC.v141.x86.x64 --quiet --norestart --nocache\" -Wait\n - name: Install and Setup VS 2015 (v140) Build Tools\n if: matrix.vs-version == 'vs2015'\n run: |\n $vs140Path = \"C:/vs140_build_tools\"\n Invoke-WebRequest -Uri \"https://aka.ms/vs/15/release/vs_buildtools.exe\" -OutFile vs_buildtools.exe\n Start-Process -FilePath \"vs_buildtools.exe\" -ArgumentList \"--installPath `\"$vs140Path`\" --add Microsoft.VisualStudio.Workload.VCTools --add Microsoft.VisualStudio.Component.VC.140 --quiet --wait --norestart --nocache\" -Wait\n\n $vcvarsPath = (Get-ChildItem -Path $vs140Path -Filter \"vcvars64.bat\" -Recurse | Select-Object -First 1).FullName\n $cmd = \"`\"$vcvarsPath`\" && powershell -Command `\"`$env:PATH;`$env:INCLUDE;`$env:LIB`\"\"\n $output = cmd.exe /c $cmd\n $lines = $output -split \"`r`n\"\n \n echo \"PATH=$($lines[0]);$($env:PATH)\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n echo \"INCLUDE=$($lines[1])\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n echo \"LIB=$($lines[2])\" | Out-File -FilePath $env:GITHUB_ENV -Encoding utf8 -Append\n\n - uses: GuillaumeFalourd/setup-windows10-sdk-action@v2.4\n with:\n sdk-version: ${{ matrix.windows-sdk-version }}\n\n - name: cache-protobuf\n id: cache-protobuf\n uses: actions/cache@v5\n with:\n path: \"protobuf-install\"\n key: protobuf-${{ matrix.vs-version }}-x64-install-3\n - name: protobuf\n if: steps.cache-protobuf.outputs.cache-hit != 'true'\n run: |\n Invoke-WebRequest -Uri https://github.com/protocolbuffers/protobuf/archive/v3.11.2.zip -OutFile protobuf-3.11.2.zip\n 7z x ./protobuf-3.11.2.zip\n cd protobuf-3.11.2\n mkdir build-${{ matrix.vs-version }}; cd build-${{ matrix.vs-version }}\n cmake -T ${{ matrix.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.windows-sdk-version }}.0 -DCMAKE_INSTALL_PREFIX=\"$env:GITHUB_WORKSPACE\\protobuf-install\" -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF -DNCNN_BUILD_TESTS=ON ../cmake\n cmake --build . --config Release -j 4\n cmake --build . --config Release --target install\n - name: cache-swiftshader\n if: matrix.vs-version != 'vs2015' && matrix.vs-version != 'vs2017'\n id: cache-swiftshader\n uses: actions/cache@v5\n with:\n path: swiftshader-install\n key: swiftshader-${{ matrix.vs-version }}-x64-install-20251010\n - name: checkout-swiftshader\n if: matrix.vs-version != 'vs2015' && matrix.vs-version != 'vs2017' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n uses: actions/checkout@v6\n with:\n repository: google/swiftshader\n path: swiftshader\n ref: de870ac7518fe2b6bb651ecc22fc36647cf7b986\n - name: checkout-swiftshader-submodules\n if: matrix.vs-version != 'vs2015' && matrix.vs-version != 'vs2017' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n git -c submodule.\"third_party/git-hooks\".update=none submodule update --init --recursive\n - name: swiftshader\n if: matrix.vs-version != 'vs2015' && matrix.vs-version != 'vs2017' && steps.cache-swiftshader.outputs.cache-hit != 'true'\n run: |\n cd swiftshader\n mkdir build-${{ matrix.vs-version }}; cd build-${{ matrix.vs-version }}\n cmake -T ${{ matrix.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.windows-sdk-version }}.0 -DCMAKE_INSTALL_PREFIX=install -DSWIFTSHADER_BUILD_EGL=FALSE -DSWIFTSHADER_BUILD_GLESv2=FALSE -DSWIFTSHADER_BUILD_GLES_CM=FALSE -DSWIFTSHADER_BUILD_VULKAN=TRUE -DSWIFTSHADER_BUILD_PVR=FALSE -DSWIFTSHADER_BUILD_TESTS=FALSE -DSWIFTSHADER_ENABLE_ASTC=FALSE -DSWIFTSHADER_WARNINGS_AS_ERRORS=FALSE -DREACTOR_BACKEND=Subzero -DREACTOR_DEFAULT_OPT_LEVEL=Default -DCMAKE_BUILD_TYPE=Release ..\n cmake --build . --config Release -j 4\n mkdir \"$env:GITHUB_WORKSPACE/swiftshader-install\"\n Copy-Item -Path \"Windows\\*\" -Destination \"$env:GITHUB_WORKSPACE\\swiftshader-install\"\n\n - name: x64\n run: |\n mkdir build-x64; cd build-x64\n cmake -T ${{ matrix.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.windows-sdk-version }}.0 -Dprotobuf_DIR=\"$env:GITHUB_WORKSPACE\\protobuf-install\\cmake\" -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON ..\n cmake --build . --config Release -j 4\n - name: x64-test\n if: matrix.vs-version != 'vs2015' && matrix.vs-version != 'vs2017'\n run: |\n echo \"[Processor]`nThreadCount=1`n\" > build-x64/tests/Release/SwiftShader.ini\n Copy-Item -Path \"$env:GITHUB_WORKSPACE\\swiftshader-install\\vulkan-1.dll\" -Destination 'build-x64\\tests'\n cd build-x64; ctest -C Release --output-on-failure -j 4\n\n - name: x64-sse2\n run: |\n mkdir build-x64-sse2; cd build-x64-sse2\n cmake -T ${{ matrix.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.windows-sdk-version }}.0 -DNCNN_RUNTIME_CPU=OFF -DNCNN_XOP=OFF -DNCNN_AVX=OFF -DNCNN_AVX2=OFF -DNCNN_AVX512=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . --config Release -j 4\n - name: x64-sse2-test\n run: cd build-x64-sse2; ctest -C Release --output-on-failure -j 4\n\n - name: x64-avx\n run: |\n mkdir build-x64-avx; cd build-x64-avx\n cmake -T ${{ matrix.toolset-version }},host=x64 -A x64,version=10.0.${{ matrix.windows-sdk-version }}.0 -DNCNN_RUNTIME_CPU=OFF -DNCNN_XOP=OFF -DNCNN_AVX=ON -DNCNN_AVX2=OFF -DNCNN_AVX512=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . --config Release -j 4\n - name: x64-avx-test\n run: cd build-x64-avx; ctest -C Release --output-on-failure -j 4\n\n - name: x86\n run: |\n mkdir build-x86; cd build-x86\n cmake -T ${{ matrix.toolset-version }},host=x64 -A Win32,version=10.0.${{ matrix.windows-sdk-version }}.0 -DNCNN_SHARED_LIB=ON -DNCNN_BUILD_TESTS=ON -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF ..\n cmake --build . --config Release -j 4\n - name: x86-test\n run: |\n Copy-Item -Path \"build-x86\\src\\Release\\ncnn.dll\" -Destination 'build-x86\\tests'\n cd build-x86; ctest -C Release --output-on-failure -j 4\n" }, { "path": ".gitignore", "content": "# CMake build directory\nbuild*/\n\n# Backup files.\n*~\n\n# Prerequisites\n*.d\n\n# Compiled Object files\n*.slo\n*.lo\n*.o\n*.obj\n\n# Precompiled Headers\n*.gch\n*.pch\n\n# Compiled Dynamic libraries\n*.so\n*.dylib\n*.dll\n\n# Fortran module files\n*.mod\n*.smod\n\n# Compiled Static libraries\n*.lai\n*.la\n*.a\n*.lib\n\n# Executables\n*.exe\n*.out\n*.app\n\n# MACOSX\n.DS_Store\n\n# IDE\n.vs\n.vscode\n.idea\ncmake-build-debug\ncmake-build-release\nCMakeSettings.json\n\n# Compiled python\n__pycache__\n*.pyc\n*.pyd\n*.egg-info/\npython/setup.py\n\n# Clangd\n.cache/\n\n# Xmake\n.xmake/\n" }, { "path": ".gitmodules", "content": "[submodule \"glslang\"]\n\tpath = glslang\n\turl = https://github.com/nihui/glslang\n[submodule \"python/pybind11\"]\n\tpath = python/pybind11\n\turl = https://github.com/pybind/pybind11.git\n" }, { "path": "CITATION.cff", "content": "cff-version: 1.2.0\ntitle: ncnn\nmessage: >-\n If you use this software, please cite it using the\n metadata from this file.\ntype: software\nauthors:\n - family-names: \"Ni\"\n given-names: \"Hui\"\n - name: \"The ncnn contributors\"\nabstract: >-\n ncnn is a high-performance neural network inference\n computing framework optimized for mobile platforms. \ndate-released: 2017-06-30\nkeywords:\n - \"neural network\"\n - \"artificial intelligence\"\n - \"deep learning\"\n - android\n - ios\n - windows\n - linux\n - macos\n - pnnx\n - simd\n - vulkan\n - riscv\n - x86\n - arm\n - mips\n - loongarch\nlicense: BSD-3-Clause\nrepository-code: \"https://github.com/Tencent/ncnn\"\n" }, { "path": "CMakeLists.txt", "content": "if(CMAKE_TOOLCHAIN_FILE)\n set(LIBRARY_OUTPUT_PATH_ROOT ${CMAKE_BINARY_DIR} CACHE PATH \"root for library output, set this to change where android libs are compiled to\")\n # get absolute path, but get_filename_component ABSOLUTE only refer with source dir, so find_file here :(\n get_filename_component(CMAKE_TOOLCHAIN_FILE_NAME ${CMAKE_TOOLCHAIN_FILE} NAME)\n find_file(CMAKE_TOOLCHAIN_FILE ${CMAKE_TOOLCHAIN_FILE_NAME} PATHS ${CMAKE_SOURCE_DIR} NO_DEFAULT_PATH)\n message(STATUS \"CMAKE_TOOLCHAIN_FILE = ${CMAKE_TOOLCHAIN_FILE}\")\nendif()\n\nif(NOT DEFINED CMAKE_INSTALL_PREFIX)\n set(CMAKE_INSTALL_PREFIX \"${CMAKE_BINARY_DIR}/install\" CACHE PATH \"Installation Directory\")\nendif()\nmessage(STATUS \"CMAKE_INSTALL_PREFIX = ${CMAKE_INSTALL_PREFIX}\")\n\nif(NOT DEFINED NCNN_VERSION)\n string(TIMESTAMP NCNN_VERSION \"%Y%m%d\")\nendif()\n\nset(NCNN_VERSION_MAJOR 1)\nset(NCNN_VERSION_MINOR 0)\nset(NCNN_VERSION_PATCH ${NCNN_VERSION})\nset(NCNN_VERSION_STRING ${NCNN_VERSION_MAJOR}.${NCNN_VERSION_MINOR}.${NCNN_VERSION_PATCH})\nset(NCNN_VERSION_NUMBER ${NCNN_VERSION})\nmessage(STATUS \"NCNN_VERSION_STRING = ${NCNN_VERSION_STRING}\")\n\ncmake_minimum_required(VERSION 2.8.12...3.10)\n\nif(NOT CMAKE_BUILD_TYPE)\n set(CMAKE_BUILD_TYPE release CACHE STRING \"Choose the type of build\" FORCE)\nendif()\n\nif(NOT CMAKE_VERSION VERSION_LESS \"3.15\")\n # enable CMAKE_MSVC_RUNTIME_LIBRARY\n cmake_policy(SET CMP0091 NEW)\nendif()\n\nif(POLICY CMP0025)\n # reference from https://cmake.org/cmake/help/latest/policy/CMP0025.html\n cmake_policy(SET CMP0025 NEW)\nendif()\n\nif(POLICY CMP0057)\n # reference from https://cmake.org/cmake/help/latest/policy/CMP0057.html\n cmake_policy(SET CMP0057 NEW)\nendif()\n\nproject(ncnn)\n\nif(MSVC AND NOT CMAKE_VERSION VERSION_LESS \"3.15\")\n option(NCNN_BUILD_WITH_STATIC_CRT \"Enables use of statically linked CRT for statically linked ncnn\" OFF)\n if(NCNN_BUILD_WITH_STATIC_CRT)\n # cmake before version 3.15 not work\n set(CMAKE_MSVC_RUNTIME_LIBRARY \"MultiThreaded$<$:Debug>\")\n endif()\nendif()\n\nif(CMAKE_FIND_LIBRARY_SUFFIXES_INIT)\n # project() overwrite CMAKE_FIND_LIBRARY_SUFFIXES in toolchain, restore it\n set(CMAKE_FIND_LIBRARY_SUFFIXES ${CMAKE_FIND_LIBRARY_SUFFIXES_INIT})\nendif()\n\noption(NCNN_SHARED_LIB \"shared library support\" OFF)\noption(NCNN_ENABLE_LTO \"enable link-time optimization\" OFF)\noption(NCNN_OPENMP \"openmp support\" ON)\noption(NCNN_STDIO \"load model from external file\" ON)\noption(NCNN_STRING \"plain and verbose string\" ON)\noption(NCNN_INSTALL_SDK \"install ncnn library and headers\" ON)\noption(NCNN_SIMPLEOCV \"minimal opencv structure emulation\" OFF)\noption(NCNN_SIMPLEOMP \"minimal openmp runtime emulation\" OFF)\noption(NCNN_SIMPLESTL \"minimal cpp stl structure emulation\" OFF)\noption(NCNN_SIMPLEMATH \"minimal cmath\" OFF)\noption(NCNN_THREADS \"build with threads\" ON)\noption(NCNN_BENCHMARK \"print benchmark information for every layer\" OFF)\noption(NCNN_C_API \"build with C api\" ON)\noption(NCNN_PLATFORM_API \"build with platform api candy\" ON)\noption(NCNN_WINXP \"build with windows xp compatibility\" OFF)\noption(NCNN_PIXEL \"convert and resize from/to image pixel\" ON)\noption(NCNN_PIXEL_ROTATE \"rotate image pixel orientation\" ON)\noption(NCNN_PIXEL_AFFINE \"warp affine image pixel\" ON)\noption(NCNN_PIXEL_DRAWING \"draw basic figure and text\" ON)\noption(NCNN_CMAKE_VERBOSE \"print verbose cmake messages\" OFF)\noption(NCNN_VULKAN \"vulkan compute support\" OFF)\noption(NCNN_SIMPLEVK \"minimal in-house vulkan loader\" ON)\noption(NCNN_SYSTEM_GLSLANG \"use system glslang library\" OFF)\noption(NCNN_RUNTIME_CPU \"runtime dispatch cpu routines\" ON)\noption(NCNN_DISABLE_PIC \"disable position-independent code\" OFF)\noption(NCNN_BUILD_TESTS \"build tests\" OFF)\noption(NCNN_COVERAGE \"build for coverage\" OFF)\noption(NCNN_ASAN \"build for address sanitizer\" OFF)\noption(NCNN_BUILD_BENCHMARK \"build benchmark\" ON)\noption(NCNN_PYTHON \"build python api\" OFF)\noption(NCNN_INT8 \"int8 inference\" ON)\noption(NCNN_BF16 \"bf16 inference\" ON)\noption(NCNN_FORCE_INLINE \"force inline some function\" ON)\n\nif(ANDROID OR IOS OR NCNN_SIMPLESTL)\n option(NCNN_DISABLE_RTTI \"disable rtti\" ON)\n option(NCNN_DISABLE_EXCEPTION \"disable exception\" ON)\nelse()\n option(NCNN_DISABLE_RTTI \"disable rtti\" OFF)\n option(NCNN_DISABLE_EXCEPTION \"disable exception\" OFF)\nendif()\n\nif(ANDROID OR IOS OR NCNN_SIMPLESTL OR CMAKE_CROSSCOMPILING)\n option(NCNN_BUILD_TOOLS \"build tools\" OFF)\n option(NCNN_BUILD_EXAMPLES \"build examples\" OFF)\nelse()\n option(NCNN_BUILD_TOOLS \"build tools\" ON)\n option(NCNN_BUILD_EXAMPLES \"build examples\" ON)\nendif()\n\nif(NCNN_SHARED_LIB)\n if(NCNN_ENABLE_LTO)\n # enable global link time optimization\n cmake_policy(SET CMP0069 NEW)\n set(CMAKE_POLICY_DEFAULT_CMP0069 NEW)\n include(CheckIPOSupported)\n check_ipo_supported(RESULT ipo_supported OUTPUT ipo_supported_output)\n if(ipo_supported)\n set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)\n else()\n message(WARNING \"IPO is not supported: ${ipo_supported_output}\")\n set(NCNN_ENABLE_LTO OFF)\n endif()\n endif()\nendif()\n\nif(NOT NCNN_STDIO OR NOT NCNN_STRING)\n if(NCNN_BUILD_TOOLS)\n message(WARNING \"NCNN_STDIO or NCNN_STRING disabled, NCNN_BUILD_TOOLS will be turned off.\")\n set(NCNN_BUILD_TOOLS OFF)\n endif()\n if(NCNN_BUILD_EXAMPLES)\n message(WARNING \"NCNN_STDIO or NCNN_STRING disabled, NCNN_BUILD_EXAMPLES will be turned off.\")\n set(NCNN_BUILD_EXAMPLES OFF)\n endif()\n if(NCNN_BUILD_BENCHMARK)\n message(WARNING \"NCNN_STDIO or NCNN_STRING disabled, NCNN_BUILD_BENCHMARK will be turned off.\")\n set(NCNN_BUILD_BENCHMARK OFF)\n endif()\n if(NCNN_BUILD_TESTS)\n message(WARNING \"NCNN_STDIO or NCNN_STRING disabled, NCNN_BUILD_TESTS will be turned off.\")\n set(NCNN_BUILD_TESTS OFF)\n endif()\nendif()\n\n##############################################\n\ninclude(CheckCXXCompilerFlag)\nset(CMAKE_TRY_COMPILE_CONFIGURATION release)\nset(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)\n\n# gnu inline assembly in clang msvc does not work actually\nif(NOT (CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\" OR (CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")))\n check_cxx_source_compiles(\"int test(int a) { asm volatile(\\\"\\\" : \\\"=r\\\"(a) : \\\"0\\\"(a) : \\\"memory\\\"); return a; }\" NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)\n if(NCNN_COMPILER_SUPPORT_GNU_INLINE_ASM)\n option(NCNN_GNU_INLINE_ASM \"optimize platform with gnu style inline assembly\" ON)\n else()\n message(WARNING \"The compiler does not support gnu style inline assembly. NCNN_GNU_INLINE_ASM will be OFF.\")\n endif()\nendif()\n\nif((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES \"arm\")\n OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES \"arm64\")\n OR (CMAKE_SYSTEM_PROCESSOR MATCHES \"^(arm|aarch64)\")\n OR (CMAKE_CXX_COMPILER_ARCHITECTURE_ID MATCHES \"(ARMV7|ARM64)\")\n OR ((CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\" OR (CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")) AND (${CMAKE_GENERATOR_PLATFORM} MATCHES \"^(arm|arm64)\")))\n set(NCNN_TARGET_ARCH arm)\n\n if(APPLE AND CMAKE_OSX_ARCHITECTURES STREQUAL \"arm64_32\")\n set(NCNN_TARGET_ILP32 TRUE)\n endif()\n\n if(CMAKE_SIZEOF_VOID_P EQUAL 4 AND NOT NCNN_TARGET_ILP32)\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, float32x4_t a, float32x4_t b) { return vmlaq_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM_NEON)\n\n if(NCNN_COMPILER_SUPPORT_ARM_NEON)\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\" OR (CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\"))\n set(CMAKE_REQUIRED_FLAGS \"/arch:VFPv4\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n\n unset(CMAKE_REQUIRED_FLAGS)\n else()\n set(CMAKE_REQUIRED_FLAGS \"-mfpu=neon-vfpv4\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n\n if(NOT NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n set(CMAKE_REQUIRED_FLAGS \"-mfpu=neon-vfpv4 -mfp16-format=ieee\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4_FP16)\n endif()\n\n unset(CMAKE_REQUIRED_FLAGS)\n endif()\n endif()\n\n if(NCNN_COMPILER_SUPPORT_ARM_VFPV4 OR NCNN_COMPILER_SUPPORT_ARM_VFPV4_FP16)\n option(NCNN_VFPV4 \"optimize armv7 platform with vfpv4\" ON)\n else()\n message(WARNING \"The compiler does not support arm vfpv4. NCNN_VFPV4 will be OFF.\")\n endif()\n endif()\n\n if(CMAKE_SIZEOF_VOID_P EQUAL 8 OR NCNN_TARGET_ILP32)\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\")\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.0\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2\")\n check_cxx_source_compiles(\"#include \\nfloat16x8_t test(float16x8_t s, float16x8_t a, float16x8_t b) { return vfmaq_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vdotq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_DOTPROD)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, float16x8_t a, float16x8_t b) { return vfmlalq_low_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16FML)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.4\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, bfloat16x8_t a, bfloat16x8_t b) { return vcvt_f32_bf16(vcvt_bf16_f32(vbfmmlaq_f32(s, a, b))); }\" NCNN_COMPILER_SUPPORT_ARM84_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.4\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vmmlaq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM84_I8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6\")\n check_cxx_source_compiles(\"#include \\nsvfloat16_t test(svfloat16_t s, svfloat16_t a, svfloat16_t b, svbool_t bp) { return svmla_f16_z(bp, s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6\")\n check_cxx_source_compiles(\"#include \\nsvint16_t test(svint16_t s, svint8_t a, svint8_t b) { return svmlslb_s16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE2)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svbfloat16_t a, svbfloat16_t b) { return svbfmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEBF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6\")\n check_cxx_source_compiles(\"#include \\nsvint32_t test(svint32_t s, svint8_t a, svint8_t b) { return svmmla_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEI8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svfloat32_t a, svfloat32_t b) { return svmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)\n\n unset(CMAKE_REQUIRED_FLAGS)\n elseif(CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.0\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2 -march=armv8.2-a+fp16\")\n check_cxx_source_compiles(\"#include \\nfloat16x8_t test(float16x8_t s, float16x8_t a, float16x8_t b) { return vfmaq_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2 -march=armv8.2-a+dotprod\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vdotq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_DOTPROD)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.2 -march=armv8.2-a+fp16fml\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, float16x8_t a, float16x8_t b) { return vfmlalq_low_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16FML)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.4 -march=armv8.4-a+bf16\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, bfloat16x8_t a, bfloat16x8_t b) { return vcvt_f32_bf16(vcvt_bf16_f32(vbfmmlaq_f32(s, a, b))); }\" NCNN_COMPILER_SUPPORT_ARM84_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.4 -march=armv8.4-a+i8mm\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vmmlaq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM84_I8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6 -march=armv8.6-a+sve\")\n check_cxx_source_compiles(\"#include \\nsvfloat16_t test(svfloat16_t s, svfloat16_t a, svfloat16_t b, svbool_t bp) { return svmla_f16_z(bp, s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6 -march=armv8.6-a+sve2\")\n check_cxx_source_compiles(\"#include \\nsvint16_t test(svint16_t s, svint8_t a, svint8_t b) { return svmlslb_s16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE2)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6 -march=armv8.6-a+sve+bf16\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svbfloat16_t a, svbfloat16_t b) { return svbfmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEBF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6 -march=armv8.6-a+sve+i8mm\")\n check_cxx_source_compiles(\"#include \\nsvint32_t test(svint32_t s, svint8_t a, svint8_t b) { return svmmla_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEI8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:armv8.6 -march=armv8.6-a+sve+f32mm\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svfloat32_t a, svfloat32_t b) { return svmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)\n\n unset(CMAKE_REQUIRED_FLAGS)\n else()\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8-a\")\n check_cxx_source_compiles(\"#include \\nfloat16x4_t test(float32x4_t a) { return vcvt_f16_f32(a); }\" NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.2-a+fp16\")\n check_cxx_source_compiles(\"#include \\nfloat16x8_t test(float16x8_t s, float16x8_t a, float16x8_t b) { return vfmaq_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.2-a+dotprod\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vdotq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_DOTPROD)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.2-a+fp16fml\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, float16x8_t a, float16x8_t b) { return vfmlalq_low_f16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM82_FP16FML)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.4-a+bf16\")\n check_cxx_source_compiles(\"#include \\nfloat32x4_t test(float32x4_t s, bfloat16x8_t a, bfloat16x8_t b) { return vcvt_f32_bf16(vcvt_bf16_f32(vbfmmlaq_f32(s, a, b))); }\" NCNN_COMPILER_SUPPORT_ARM84_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.4-a+i8mm\")\n check_cxx_source_compiles(\"#include \\nint32x4_t test(int32x4_t s, int8x16_t a, int8x16_t b) { return vmmlaq_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM84_I8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.6-a+sve\")\n check_cxx_source_compiles(\"#include \\nsvfloat16_t test(svfloat16_t s, svfloat16_t a, svfloat16_t b, svbool_t bp) { return svmla_f16_z(bp, s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.6-a+sve2\")\n check_cxx_source_compiles(\"#include \\nsvint16_t test(svint16_t s, svint8_t a, svint8_t b) { return svmlslb_s16(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVE2)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.6-a+sve+bf16\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svbfloat16_t a, svbfloat16_t b) { return svbfmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEBF16)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.6-a+sve+i8mm\")\n check_cxx_source_compiles(\"#include \\nsvint32_t test(svint32_t s, svint8_t a, svint8_t b) { return svmmla_s32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEI8MM)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=armv8.6-a+sve+f32mm\")\n check_cxx_source_compiles(\"#include \\nsvfloat32_t test(svfloat32_t s, svfloat32_t a, svfloat32_t b) { return svmmla_f32(s, a, b); }\" NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)\n\n unset(CMAKE_REQUIRED_FLAGS)\n endif()\n\n if(NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n option(NCNN_VFPV4 \"optimize aarch64 platform with vfpv4\" ON)\n else()\n message(WARNING \"The compiler does not support arm vfpv4. NCNN_VFPV4 will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_ARM82_FP16)\n option(NCNN_ARM82 \"optimize aarch64 platform with armv8.2 fp16\" ON)\n if(NCNN_COMPILER_SUPPORT_ARM82_DOTPROD)\n if(NCNN_ARM82)\n option(NCNN_ARM82DOT \"optimize aarch64 platform with armv8.2 dotprod\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.2 dotprod. NCNN_ARM82DOT will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM82_FP16FML)\n if(NCNN_ARM82)\n option(NCNN_ARM82FP16FML \"optimize aarch64 platform with armv8.2 fp16fml\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.2 fp16fml. NCNN_ARM82FP16FML will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM84_BF16)\n if(NCNN_ARM82DOT AND NCNN_ARM82FP16FML)\n option(NCNN_ARM84BF16 \"optimize aarch64 platform with armv8.4 bf16\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.4 bf16. NCNN_ARM86BF16 will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM84_I8MM)\n if(NCNN_ARM82DOT AND NCNN_ARM82FP16FML)\n option(NCNN_ARM84I8MM \"optimize aarch64 platform with armv8.4 i8mm\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.4 i8mm. NCNN_ARM84I8MM will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM86_SVE)\n if(NCNN_ARM84BF16 AND NCNN_ARM84I8MM)\n option(NCNN_ARM86SVE \"optimize aarch64 platform with armv8.6 sve\" ON)\n if(NCNN_COMPILER_SUPPORT_ARM86_SVE2)\n if(NCNN_ARM86SVE)\n option(NCNN_ARM86SVE2 \"optimize aarch64 platform with armv8.6 sve2\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.6 sve2. NCNN_ARM86SVE2 will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM86_SVEBF16)\n if(NCNN_ARM86SVE)\n option(NCNN_ARM86SVEBF16 \"optimize aarch64 platform with armv8.6 sve bf16\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.6 sve bf16. NCNN_ARM86SVEBF16 will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM86_SVEI8MM)\n if(NCNN_ARM86SVE)\n option(NCNN_ARM86SVEI8MM \"optimize aarch64 platform with armv8.6 sve i8mm\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.6 sve i8mm. NCNN_ARM86SVEI8MM will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_ARM86_SVEF32MM)\n if(NCNN_ARM86SVE)\n option(NCNN_ARM86SVEF32MM \"optimize aarch64 platform with armv8.6 sve f32mm\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.6 sve f32mm. NCNN_ARM86SVEF32MM will be OFF.\")\n endif()\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.6 sve. NCNN_ARM86SVE will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support armv8.2 fp16. NCNN_ARM82 will be OFF.\")\n endif()\n endif()\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(mips)\")\n set(NCNN_TARGET_ARCH mips)\n\n check_cxx_compiler_flag(\"-mmsa\" NCNN_COMPILER_SUPPORT_MIPS_MSA)\n\n set(CMAKE_REQUIRED_FLAGS \"-mloongson-mmi -I${CMAKE_CURRENT_SOURCE_DIR}/src/layer/mips\")\n check_cxx_source_compiles(\"#include \\\"loongson_mmi.h\\\"\\nint32x2_t test(int16x4_t a, int16x4_t b) { return __mmi_pmaddhw(a, b); }\" NCNN_COMPILER_SUPPORT_LOONGSON_MMI)\n\n unset(CMAKE_REQUIRED_FLAGS)\n\n if(NCNN_COMPILER_SUPPORT_MIPS_MSA)\n option(NCNN_MSA \"optimize mips platform with msa extension\" ON)\n else()\n message(WARNING \"The compiler does not support msa extension. NCNN_MSA will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_LOONGSON_MMI)\n option(NCNN_MMI \"optimize mips platform with loongson mmi extension\" ON)\n else()\n message(WARNING \"The compiler does not support loongson mmi extension. NCNN_MMI will be OFF.\")\n endif()\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(loongarch64|loongarch32)\")\n set(NCNN_TARGET_ARCH loongarch)\n\n set(CMAKE_REQUIRED_FLAGS \"-mlsx\")\n check_cxx_source_compiles(\"#include \\n__m128 test(__m128 a, __m128 b, __m128 c) { return __lsx_vfmadd_s(a, b, c); }\" NCNN_COMPILER_SUPPORT_LOONGARCH_LSX)\n\n set(CMAKE_REQUIRED_FLAGS \"-mlasx\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 a, __m256 b, __m256 c) { return __lasx_xvfmadd_s(a, b, c); }\" NCNN_COMPILER_SUPPORT_LOONGARCH_LASX)\n\n unset(CMAKE_REQUIRED_FLAGS)\n\n if(NCNN_COMPILER_SUPPORT_LOONGARCH_LSX)\n option(NCNN_LSX \"optimize loongarch platform with lsx extension\" ON)\n if(NCNN_COMPILER_SUPPORT_LOONGARCH_LASX)\n option(NCNN_LASX \"optimize loongarch platform with lasx extension\" ON)\n else()\n message(WARNING \"The compiler does not support lasx extension. NCNN_LASX will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support lsx extension. NCNN_LSX will be OFF.\")\n endif()\n\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(riscv)\")\n set(NCNN_TARGET_ARCH riscv)\n\n if(CMAKE_SIZEOF_VOID_P EQUAL 8)\n set(CMAKE_REQUIRED_FLAGS \"-march=rv64gcv\")\n check_cxx_source_compiles(\"#include \\nvfloat32m8_t test(vfloat32m8_t s, vfloat32m8_t w, float v, size_t vl) { return __riscv_vfmacc_vf_f32m8(s, v, w, vl); }\\nvfloat32m1x2_t test2(vfloat32m1_t x) { return __riscv_vcreate_v_f32m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_V)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv64gc_zfh -D__fp16=_Float16\")\n check_cxx_source_compiles(\"__fp16 test(__fp16 a) { return a * a; }\" NCNN_COMPILER_SUPPORT_RISCV_ZFH)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv64gcv_zfh_zvfh -D__fp16=_Float16\")\n check_cxx_source_compiles(\"#include \\nvfloat16m8_t test(vfloat16m8_t s, vfloat16m8_t w, __fp16 v, size_t vl) { return __riscv_vfmacc_vf_f16m8(s, v, w, vl); }\\nvfloat16m1x2_t test2(vfloat16m1_t x){ return __riscv_vcreate_v_f16m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_ZVFH)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv64gc_zfh_xtheadvector -D__fp16=_Float16\")\n check_cxx_source_compiles(\"#include \\nvfloat16m8_t test(vfloat16m8_t s, vfloat16m8_t w, __fp16 v, size_t vl) { return __riscv_vfmacc_vf_f16m8(s, v, w, vl); }\\nvfloat16m1x2_t test2(vfloat16m1_t x){ return __riscv_vcreate_v_f16m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n\n unset(CMAKE_REQUIRED_FLAGS)\n\n if(NCNN_COMPILER_SUPPORT_RISCV_V OR NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n option(NCNN_RVV \"optimize risc-v platform with v extension\" ON)\n else()\n message(WARNING \"The compiler does not support risc-v v or xtheadvector extension. NCNN_RVV will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n option(NCNN_XTHEADVECTOR \"optimize risc-v platform with xtheadvector extension\" ON)\n else()\n message(WARNING \"The compiler does not support risc-v xtheadvector extension. NCNN_XTHEADVECTOR will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_RISCV_ZFH)\n option(NCNN_ZFH \"optimize risc-v platform with zfh extension\" ON)\n if(NCNN_COMPILER_SUPPORT_RISCV_ZVFH OR NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n if(NCNN_RVV AND NCNN_ZFH)\n option(NCNN_ZVFH \"optimize risc-v platform with zvfh extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support zvfh extension. NCNN_ZVFH will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support risc-v zfh extension. NCNN_ZFH will be OFF.\")\n endif()\n\n elseif(CMAKE_SIZEOF_VOID_P EQUAL 4)\n set(CMAKE_REQUIRED_FLAGS \"-march=rv32gcv\")\n check_cxx_source_compiles(\"#include \\nvfloat32m8_t test(vfloat32m8_t s, vfloat32m8_t w, float v, size_t vl) { return __riscv_vfmacc_vf_f32m8(s, v, w, vl); }\\nvfloat32m1x2_t test2(vfloat32m1_t x) { return __riscv_vcreate_v_f32m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_V)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv32gc_zfh -D__fp16=_Float16\")\n check_cxx_source_compiles(\"__fp16 test(__fp16 a) { return a * a; }\" NCNN_COMPILER_SUPPORT_RISCV_ZFH)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv32gcv_zfh_zvfh -D__fp16=_Float16\")\n check_cxx_source_compiles(\"#include \\nvfloat16m8_t test(vfloat16m8_t s, vfloat16m8_t w, __fp16 v, size_t vl) { return __riscv_vfmacc_vf_f16m8(s, v, w, vl); }\\nvfloat16m1x2_t test2(vfloat16m1_t x){ return __riscv_vcreate_v_f16m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_ZVFH)\n\n set(CMAKE_REQUIRED_FLAGS \"-march=rv32gc_zfh_xtheadvector -D__fp16=_Float16\")\n check_cxx_source_compiles(\"#include \\nvfloat16m8_t test(vfloat16m8_t s, vfloat16m8_t w, __fp16 v, size_t vl) { return __riscv_vfmacc_vf_f16m8(s, v, w, vl); }\\nvfloat16m1x2_t test2(vfloat16m1_t x){ return __riscv_vcreate_v_f16m1x2(x, x); }\" NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n\n unset(CMAKE_REQUIRED_FLAGS)\n\n if(NCNN_COMPILER_SUPPORT_RISCV_V OR NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n option(NCNN_RVV \"optimize risc-v platform with v extension\" ON)\n else()\n message(WARNING \"The compiler does not support risc-v v or xtheadvector extension. NCNN_RVV will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n option(NCNN_XTHEADVECTOR \"optimize risc-v platform with xtheadvector extension\" ON)\n else()\n message(WARNING \"The compiler does not support risc-v xtheadvector extension. NCNN_XTHEADVECTOR will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_RISCV_ZFH)\n option(NCNN_ZFH \"optimize risc-v platform with zfh extension\" ON)\n if(NCNN_COMPILER_SUPPORT_RISCV_ZVFH OR NCNN_COMPILER_SUPPORT_RISCV_XTHEADVECTOR)\n if(NCNN_RVV AND NCNN_ZFH)\n option(NCNN_ZVFH \"optimize risc-v platform with zvfh extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support zvfh extension. NCNN_ZVFH will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support risc-v zfh extension. NCNN_ZFH will be OFF.\")\n endif()\n\n endif()\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(powerpc|ppc)\")\n set(NCNN_TARGET_ARCH powerpc)\n\n if(NCNN_PPC64LE_VSX)\n set(NCNN_TARGET_ARCH x86)\n\n set(CMAKE_REQUIRED_FLAGS \"-DNO_WARN_X86_INTRINSICS -D__SSE2__\")\n check_cxx_source_compiles(\"#include \\n__m128i test(__m128i a, __m128i b) { return _mm_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_PPC64LE_SSE2)\n unset(CMAKE_REQUIRED_FLAGS)\n\n set(CMAKE_REQUIRED_FLAGS \"-DNO_WARN_X86_INTRINSICS -D__SSE4_1__\")\n check_cxx_source_compiles(\"#include \\n__m128i test(__m128i a, __m128i b) { return _mm_packus_epi32(a, b); }\" NCNN_COMPILER_SUPPORT_PPC64LE_SSE41)\n unset(CMAKE_REQUIRED_FLAGS)\n\n if(NCNN_COMPILER_SUPPORT_PPC64LE_SSE2)\n option(NCNN_VSX_SSE2 \"optimize ppc64le platform with sse2 extension\" ON)\n else()\n message(WARNING \"The compiler does not support sse2 extension. NCNN_VSX_SSE2 will be OFF.\")\n endif()\n\n if(NCNN_COMPILER_SUPPORT_PPC64LE_SSE41)\n option(NCNN_VSX_SSE41 \"optimize ppc64le platform with sse4.1 extension\" ON)\n else()\n message(WARNING \"The compiler does not support sse4.1 extension. NCNN_VSX_SSE41 will be OFF.\")\n endif()\n endif()\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(xtensa)\")\n set(NCNN_TARGET_ARCH xtensa)\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(s390x)\")\n set(NCNN_TARGET_ARCH s390x)\nelseif(CMAKE_SYSTEM_PROCESSOR MATCHES \"^(sw_64)\")\n set(NCNN_TARGET_ARCH sw_64)\n #sw_64 is alpha-like platform\n set(CMAKE_C_FLAGS \"${CMAKE_C_FLAGS} -mieee\")\n set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -mieee\")\nelse()\n set(NCNN_TARGET_ARCH x86)\n\n option(NCNN_SSE2 \"optimize x86 platform with sse2 extension\" ON)\n\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\")\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 a, __m256 b) { return _mm256_mul_ps(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 s, __m256 a, __m256 b) { return _mm256_fmadd_ps(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_FMA)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX\")\n check_cxx_source_compiles(\"#include \\n#include \\n__m128i test(__m128i s, __m128i a, __m128i b) { return _mm_maddd_epi16(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_XOP)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m128i a) { return _mm256_cvtph_ps(a); }\" NCNN_COMPILER_SUPPORT_X86_F16C)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i a, __m256i b) { return _mm256_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX2)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i a, __m512i b) { return _mm512_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwssd_avx_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpbssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT8)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwsud_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2\")\n check_cxx_source_compiles(\"#include \\n__m128bh test(__m256 a) { return _mm256_cvtneps_avx_pbh(a); }\" NCNN_COMPILER_SUPPORT_X86_AVX_NE_CONVERT)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i s, __m512i a, __m512i b) { return _mm512_dpwssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512\")\n check_cxx_source_compiles(\"#include \\n__m256bh test(__m256bh s, __m512bh a, __m512bh b) { return _mm512_cvtneps_pbh(_mm512_dpbf16_ps(_mm512_cvtpbh_ps(s), a, b)); }\\n__m512i test2(__m512 a) { __m256i _a = (__m256i)_mm512_cvtneps_pbh(a); return _mm512_inserti32x8(_mm512_castsi256_si512(_a), _a, 1); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512\")\n check_cxx_source_compiles(\"#include \\n__m512h test(__m512h s, __m512h a, __m512h b) { return _mm512_fmadd_ph(s, a, b); }\\n__m512 test2(__m512 a) { return _mm512_cvtxph_ps(_mm512_cvtxps_ph(a)); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_FP16)\n\n unset(CMAKE_REQUIRED_FLAGS)\n elseif(CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")\n check_cxx_compiler_flag(\"-mrecip=none\" NCNN_COMPILER_SUPPORT_X86_RECIP_NONE)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 a, __m256 b) { return _mm256_mul_ps(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX -mfma -mf16c\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 s, __m256 a, __m256 b) { return _mm256_fmadd_ps(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_FMA)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX -mxop\")\n check_cxx_source_compiles(\"#include \\n__m128i test(__m128i s, __m128i a, __m128i b) { return _mm_maddd_epi16(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_XOP)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX -mf16c\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m128i a) { return _mm256_cvtph_ps(a); }\" NCNN_COMPILER_SUPPORT_X86_F16C)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2 -mfma -mf16c\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i a, __m256i b) { return _mm256_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX2)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i a, __m512i b) { return _mm512_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2 -mfma -mf16c -mavxvnni\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwssd_avx_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2 -mfma -mf16c -mavxvnni -mavxvnniint8\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpbssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT8)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2 -mfma -mf16c -mavxvnni -mavxvnniint16\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwsud_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX2 -mfma -mf16c -mavxneconvert\")\n check_cxx_source_compiles(\"#include \\n__m128bh test(__m256 a) { return _mm256_cvtneps_avx_pbh(a); }\" NCNN_COMPILER_SUPPORT_X86_AVX_NE_CONVERT)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i s, __m512i a, __m512i b) { return _mm512_dpwssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16\")\n check_cxx_source_compiles(\"#include \\n__m256bh test(__m256bh s, __m512bh a, __m512bh b) { return _mm512_cvtneps_pbh(_mm512_dpbf16_ps(_mm512_cvtpbh_ps(s), a, b)); }\\n__m512i test2(__m512 a) { __m256i _a = (__m256i)_mm512_cvtneps_pbh(a); return _mm512_inserti32x8(_mm512_castsi256_si512(_a), _a, 1); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"/arch:AVX512 -mfma -mf16c -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16\")\n check_cxx_source_compiles(\"#include \\n__m512h test(__m512h s, __m512h a, __m512h b) { return _mm512_fmadd_ph(s, a, b); }\\n__m512 test2(__m512 a) { return _mm512_cvtxph_ps(_mm512_cvtxps_ph(a)); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_FP16)\n\n unset(CMAKE_REQUIRED_FLAGS)\n else()\n check_cxx_compiler_flag(\"-mrecip=none\" NCNN_COMPILER_SUPPORT_X86_RECIP_NONE)\n\n set(CMAKE_REQUIRED_FLAGS \"-mavx\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 a, __m256 b) { return _mm256_mul_ps(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m256 s, __m256 a, __m256 b) { return _mm256_fmadd_ps(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_FMA)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mxop\")\n check_cxx_source_compiles(\"#include \\n__m128i test(__m128i s, __m128i a, __m128i b) { return _mm_maddd_epi16(a, b, s); }\" NCNN_COMPILER_SUPPORT_X86_XOP)\n\n set(CMAKE_REQUIRED_FLAGS \"-mf16c\")\n check_cxx_source_compiles(\"#include \\n__m256 test(__m128i a) { return _mm256_cvtph_ps(a); }\" NCNN_COMPILER_SUPPORT_X86_F16C)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx2\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i a, __m256i b) { return _mm256_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX2)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i a, __m512i b) { return _mm512_madd_epi16(a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx2 -mavxvnni\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwssd_avx_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx2 -mavxvnni -mavxvnniint8\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpbssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT8)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx2 -mavxvnni -mavxvnniint16\")\n check_cxx_source_compiles(\"#include \\n__m256i test(__m256i s, __m256i a, __m256i b) { return _mm256_dpwsud_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT16)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx2 -mavxneconvert\")\n check_cxx_source_compiles(\"#include \\n__m128bh test(__m256 a) { return _mm256_cvtneps_avx_pbh(a); }\" NCNN_COMPILER_SUPPORT_X86_AVX_NE_CONVERT)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512vnni\")\n check_cxx_source_compiles(\"#include \\n__m512i test(__m512i s, __m512i a, __m512i b) { return _mm512_dpwssd_epi32(s, a, b); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512bf16\")\n check_cxx_source_compiles(\"#include \\n__m256bh test(__m256bh s, __m512bh a, __m512bh b) { return _mm512_cvtneps_pbh(_mm512_dpbf16_ps(_mm512_cvtpbh_ps(s), a, b)); }\\n__m512i test2(__m512 a) { __m256i _a = (__m256i)_mm512_cvtneps_pbh(a); return _mm512_inserti32x8(_mm512_castsi256_si512(_a), _a, 1); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_BF16)\n\n set(CMAKE_REQUIRED_FLAGS \"-mfma -mf16c -mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mavx512fp16\")\n check_cxx_source_compiles(\"#include \\n__m512h test(__m512h s, __m512h a, __m512h b) { return _mm512_fmadd_ph(s, a, b); }\\n__m512 test2(__m512 a) { return _mm512_cvtxph_ps(_mm512_cvtxps_ph(a)); }\" NCNN_COMPILER_SUPPORT_X86_AVX512_FP16)\n\n unset(CMAKE_REQUIRED_FLAGS)\n endif()\n\n if(NOT CMAKE_SYSTEM_NAME MATCHES \"Emscripten|WASI\" AND NCNN_COMPILER_SUPPORT_X86_AVX)\n option(NCNN_AVX \"optimize x86 platform with avx extension\" ON)\n if(NCNN_COMPILER_SUPPORT_X86_FMA)\n if(NCNN_AVX)\n option(NCNN_FMA \"optimize x86 platform with fma extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support fma extension. NCNN_FMA will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_XOP)\n if(NCNN_AVX)\n option(NCNN_XOP \"optimize x86 platform with xop extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support xop extension. NCNN_XOP will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_F16C)\n if(NCNN_AVX)\n option(NCNN_F16C \"optimize x86 platform with f16c extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support f16c extension. NCNN_F16C will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX2)\n if(NCNN_AVX)\n option(NCNN_AVX2 \"optimize x86 platform with avx2 extension\" ON)\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)\n if(NCNN_AVX2)\n option(NCNN_AVXVNNI \"optimize x86 platform with avx vnni extension\" ON)\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT8)\n if(NCNN_AVXVNNI)\n option(NCNN_AVXVNNIINT8 \"optimize x86 platform with avx vnni int8 extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx vnni int8 extension. NCNN_AVXVNNIINT8 will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX_VNNI_INT16)\n if(NCNN_AVXVNNI)\n option(NCNN_AVXVNNIINT16 \"optimize x86 platform with avx vnni int16 extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx vnni int16 extension. NCNN_AVXVNNIINT16 will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support avx vnni extension. NCNN_AVXVNNI will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX_NE_CONVERT)\n if(NCNN_AVX2)\n option(NCNN_AVXNECONVERT \"optimize x86 platform with avx ne convert extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx ne convert extension. NCNN_AVXNECONVERT will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX512)\n if(NCNN_AVX2)\n option(NCNN_AVX512 \"optimize x86 platform with avx512 extension\" ON)\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)\n if(NCNN_AVX512)\n option(NCNN_AVX512VNNI \"optimize x86 platform with avx512 vnni extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx512 vnni extension. NCNN_AVX512VNNI will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX512_BF16)\n if(NCNN_AVX512)\n option(NCNN_AVX512BF16 \"optimize x86 platform with avx512 bf16 extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx512 bf16 extension. NCNN_AVX512BF16 will be OFF.\")\n endif()\n if(NCNN_COMPILER_SUPPORT_X86_AVX512_FP16)\n if(NCNN_AVX512)\n option(NCNN_AVX512FP16 \"optimize x86 platform with avx512 fp16 extension\" ON)\n endif()\n else()\n message(WARNING \"The compiler does not support avx512 fp16 extension. NCNN_AVX512FP16 will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support avx512 extension. NCNN_AVX512 will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support avx2 extension. NCNN_AVX2 will be OFF.\")\n endif()\n else()\n message(WARNING \"The compiler does not support avx extension. NCNN_AVX will be OFF.\")\n endif()\nendif()\n\nunset(CMAKE_TRY_COMPILE_CONFIGURATION)\nunset(CMAKE_TRY_COMPILE_TARGET_TYPE)\n\nif(NCNN_TARGET_ILP32)\n message(STATUS \"Target arch: ${NCNN_TARGET_ARCH} 64bit ilp32\")\nelseif(CMAKE_SIZEOF_VOID_P EQUAL 8)\n message(STATUS \"Target arch: ${NCNN_TARGET_ARCH} 64bit\")\nelse()\n message(STATUS \"Target arch: ${NCNN_TARGET_ARCH} 32bit\")\nendif()\n\n##############################################\n\n# set cmake default folder name\nset_property(GLOBAL PROPERTY USE_FOLDERS ON)\nset_property(GLOBAL PROPERTY PREDEFINED_TARGETS_FOLDER \"cmake\")\n\nif(CMAKE_SYSTEM_NAME STREQUAL \"Emscripten\")\n set(CMAKE_C_FLAGS \"${CMAKE_C_FLAGS} -s FORCE_FILESYSTEM=1 -s INITIAL_MEMORY=256MB -s EXIT_RUNTIME=1\")\n set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -s FORCE_FILESYSTEM=1 -s INITIAL_MEMORY=256MB -s EXIT_RUNTIME=1\")\n set(CMAKE_EXECUTBLE_LINKER_FLAGS \"${CMAKE_EXECUTBLE_LINKER_FLAGS} -s FORCE_FILESYSTEM=1 -s INITIAL_MEMORY=256MB -s EXIT_RUNTIME=1\")\n\n if(NCNN_OPENMP AND NCNN_SIMPLEOMP)\n # TODO better flags for emscripten\n # node --experimental-wasm-threads xxx.js\n set(CMAKE_C_FLAGS \"${CMAKE_C_FLAGS} -s USE_PTHREADS=1 -s PTHREAD_POOL_SIZE=15\")\n set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -s USE_PTHREADS=1 -s PTHREAD_POOL_SIZE=15\")\n set(CMAKE_EXECUTBLE_LINKER_FLAGS \"${CMAKE_EXECUTBLE_LINKER_FLAGS} -s USE_PTHREADS=1 -s PTHREAD_POOL_SIZE=15\")\n endif()\nendif()\n\nif(NCNN_VULKAN)\n if(NCNN_SYSTEM_GLSLANG)\n find_package(Threads)\n find_package(SPIRV-Tools QUIET)\n find_package(SPIRV-Tools-opt QUIET)\n find_package(glslang QUIET)\n if(glslang_FOUND)\n add_library(glslang ALIAS glslang::glslang)\n add_library(SPIRV ALIAS glslang::SPIRV)\n else()\n set(GLSLANG_TARGET_DIR \"GLSLANG-NOTFOUND\" CACHE PATH \"Absolute path to glslangTargets.cmake directory\")\n if(NOT GLSLANG_TARGET_DIR AND NOT DEFINED ENV{GLSLANG_TARGET_DIR})\n message(WARNING \"set glslang_DIR to glslang-config.cmake directory for using system glslang.\")\n message(WARNING \"GLSLANG_TARGET_DIR must be defined! NCNN_SYSTEM_GLSLANG will be turned off.\")\n set(NCNN_SYSTEM_GLSLANG OFF)\n else()\n include(\"${GLSLANG_TARGET_DIR}/OSDependentTargets.cmake\")\n include(\"${GLSLANG_TARGET_DIR}/OGLCompilerTargets.cmake\")\n if(EXISTS \"${GLSLANG_TARGET_DIR}/HLSLTargets.cmake\")\n # hlsl support can be optional\n include(\"${GLSLANG_TARGET_DIR}/HLSLTargets.cmake\")\n endif()\n include(\"${GLSLANG_TARGET_DIR}/glslangTargets.cmake\")\n include(\"${GLSLANG_TARGET_DIR}/SPIRVTargets.cmake\")\n endif()\n endif()\n\n if(TARGET glslang AND TARGET SPIRV)\n get_property(glslang_location TARGET glslang PROPERTY LOCATION)\n get_property(SPIRV_location TARGET SPIRV PROPERTY LOCATION)\n message(STATUS \"Found glslang: ${glslang_location} (found version \\\"${glslang_VERSION}\\\")\")\n message(STATUS \"Found SPIRV: ${SPIRV_location} (found version \\\"${glslang_VERSION}\\\")\")\n else()\n message(WARNING \"glslang or SPIRV target not found! NCNN_SYSTEM_GLSLANG will be turned off.\")\n set(NCNN_SYSTEM_GLSLANG OFF)\n endif()\n endif()\n\n if(NOT NCNN_SYSTEM_GLSLANG)\n if(NOT EXISTS \"${CMAKE_CURRENT_SOURCE_DIR}/glslang/CMakeLists.txt\")\n message(FATAL_ERROR \"The submodules were not downloaded! Please update submodules with \\\"git submodule update --init\\\" and try again.\")\n else()\n # glslang requires c++11\n set(CMAKE_CXX_STANDARD 11)\n\n option(BUILD_EXTERNAL \"\" OFF)\n option(ENABLE_SPVREMAPPER \"\" OFF)\n option(ENABLE_GLSLANG_BINARIES \"\" OFF)\n option(ENABLE_HLSL \"\" OFF)\n option(ENABLE_RTTI \"\" OFF)\n option(ENABLE_EXCEPTIONS \"\" OFF)\n option(ENABLE_OPT \"\" OFF)\n option(ENABLE_PCH \"\" OFF)\n option(GLSLANG_TESTS \"\" OFF)\n if(NCNN_SHARED_LIB)\n option(GLSLANG_ENABLE_INSTALL \"\" OFF)\n else()\n option(GLSLANG_ENABLE_INSTALL \"\" ON)\n endif()\n add_subdirectory(glslang)\n if(NCNN_SHARED_LIB)\n if(CMAKE_CXX_COMPILER_ID MATCHES \"GNU\" OR (CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND NOT CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\"))\n target_compile_options(glslang PRIVATE -fvisibility=hidden -fvisibility-inlines-hidden)\n target_compile_options(glslang-default-resource-limits PRIVATE -fvisibility=hidden -fvisibility-inlines-hidden)\n endif()\n if(NCNN_ENABLE_LTO)\n set_target_properties(glslang PROPERTIES INTERPROCEDURAL_OPTIMIZATION ON)\n set_target_properties(glslang-default-resource-limits PROPERTIES INTERPROCEDURAL_OPTIMIZATION ON)\n endif()\n endif()\n endif()\n endif()\nendif()\n\nadd_subdirectory(src)\nif(NCNN_BUILD_BENCHMARK)\n add_subdirectory(benchmark)\nendif()\nif(NCNN_BUILD_EXAMPLES)\n add_subdirectory(examples)\nendif()\nif(NCNN_BUILD_TOOLS)\n add_subdirectory(tools)\nendif()\nif(NCNN_BUILD_TESTS)\n enable_testing()\n add_subdirectory(tests)\n add_subdirectory(tests/perf)\nendif()\nif(NCNN_PYTHON)\n add_subdirectory(python)\nendif()\n" }, { "path": "CONTRIBUTING.md", "content": "\n# Acknowledgements\n\n- Thanks to bug1989 [https://github.com/bug1989] for contributing the initial quantized int8 inference code and a large variety of device benchmark\n- Thanks to zhiliu6 [https://github.com/zhiliu6] for contributing the darknet conversion tool, operators and YOLO examples\n- Thanks to Tijmen Verhulsdonck [https://github.com/Timen] for contributing the massive AVX optimization for x86 platform\n" }, { "path": "Info.plist", "content": "\n\n\n\n CFBundleName\n __NAME__\n CFBundleIdentifier\n __IDENTIFIER__\n CFBundleVersion\n __VERSION__\n CFBundleShortVersionString\n __VERSION__\n CFBundleSignature\n ????\n CFBundlePackageType\n FMWK\n\n\n" }, { "path": "LICENSE.txt", "content": "Tencent is pleased to support the open source community by making ncnn available.\nCopyright (C) 2017 Tencent. All rights reserved.\nIf you have downloaded a copy of the ncnn binary from Tencent, please note that the ncnn binary is licensed under the BSD 3-Clause License.\nIf you have downloaded a copy of the ncnn source code from Tencent, please note that ncnn source code is licensed under the BSD 3-Clause License, except for the third-party components listed below which are subject to different license terms. Your integration of ncnn into your own projects may require compliance with the BSD 3-Clause License, as well as the other licenses applicable to the third-party components included within ncnn.\nA copy of the BSD 3-Clause License is included in this file.\n\nOther dependencies and licenses:\n\nOpen Source Software Licensed Under the zlib License:\nThe below software in this distribution may have been modified by Tencent (“Tencent Modifications”). All Tencent Modifications are Copyright (C) 2017 Tencent.\n----------------------------------------------------------------------------------------\n1. neon_mathfun.h\nCopyright (C) 2011 Julien Pommier\n\n2. sse_mathfun.h\nCopyright (C) 2007 Julien Pommier\n\n3. avx_mathfun.h\nCopyright (C) 2012 Giovanni Garberoglio\nInterdisciplinary Laboratory for Computational Science (LISC)\nFondazione Bruno Kessler and University of Trento\nvia Sommarive, 18\nI-38123 Trento (Italy)\n\n\nTerms of the zlib License:\n---------------------------------------------------\nCopyright (c) \n\nThis software is provided 'as-is', without any express or implied warranty. In no event will the authors be held liable for any damages arising from the use of this software.\n\nPermission is granted to anyone to use this software for any purpose, including commercial applications, and to alter it and redistribute it freely, subject to the following restrictions:\n\n1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n3. This notice may not be removed or altered from any source distribution.\n\n\n\nOpen Source Software Licensed Under the BSD 2-Clause License:\nThe below software in this distribution may have been modified by Tencent (“Tencent Modifications”). All Tencent Modifications are Copyright (C) 2017 Tencent.\n----------------------------------------------------------------------------------------\n1. squeezenet 1.1\nCopyright (c) 2016 Forrest N. Iandola and Matthew W. Moskewicz and Khalid Ashraf and Song Han and William J. Dally and Kurt Keutzer\nAll rights reserved.\n\n2. caffe.proto master\nAll contributions by the University of California:\nCopyright (c) 2014-2017 The Regents of the University of California (Regents)\nAll rights reserved.\n\nAll other contributions:\nCopyright (c) 2014-2017, the respective contributors\nAll rights reserved.\n\n\nTerms of the BSD 2-Clause License:\n--------------------------------------------------------------------\nRedistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:\n\nRedistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.\nRedistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n\n\n\nOpen Source Software Licensed Under the BSD 3-Clause License:\nThe below software in this distribution may have been modified by Tencent (“Tencent Modifications”). All Tencent Modifications are Copyright (C) 2017 Tencent.\n----------------------------------------------------------------------------------------\n1. android.toolchain.cmake master\nCopyright (c) 2010-2011, Ethan Rublee\nCopyright (c) 2011-2014, Andrey Kamaev\nAll rights reserved.\n\n\nTerms of the BSD 3-Clause License:\n--------------------------------------------------------------------\n\nRedistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:\n\nRedistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.\nRedistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.\nNeither the name of [copyright holder] nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.\n\nTHIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\n" }, { "path": "MANIFEST.in", "content": "recursive-include cmake *\n\nrecursive-include glslang *\nprune glslang/Test\n\nrecursive-include src *\n\nrecursive-include python *\nprune python/pybind11/tests\n\ninclude CMakeLists.txt\n" }, { "path": "README.md", "content": "![ncnn](https://raw.githubusercontent.com/Tencent/ncnn/master/images/256-ncnn.png)\n\n# ncnn\n\n[![License](https://img.shields.io/badge/license-BSD_3_Clause-blue.svg?style=for-the-badge)](LICENSE.txt)\n[![Download Total Count](https://img.shields.io/github/downloads/Tencent/ncnn/total.svg?style=for-the-badge)](https://github.com/Tencent/ncnn/releases)\n[![codecov](https://img.shields.io/codecov/c/github/Tencent/ncnn/master?style=for-the-badge)](https://codecov.io/gh/Tencent/ncnn)\n\nncnn is a high-performance neural network inference computing framework optimized for mobile platforms.\nncnn is deeply considerate about deployment and uses on mobile phones from the beginning of design.\nncnn does not have third-party dependencies.\nIt is cross-platform and runs faster than all known open-source frameworks on mobile phone cpu.\nDevelopers can easily deploy deep learning algorithm models to the mobile platform by using efficient ncnn implementation, creating intelligent APPs, and bringing artificial intelligence to your fingertips.\nncnn is currently being used in many Tencent applications, such as QQ, Qzone, WeChat, Pitu, and so on.\n\nncnn 是一个为手机端极致优化的高性能神经网络前向计算框架。\nncnn 从设计之初深刻考虑手机端的部署和使用。\n无第三方依赖，跨平台，手机端 cpu 的速度快于目前所有已知的开源框架。\n基于 ncnn，开发者能够将深度学习算法轻松移植到手机端高效执行，\n开发出人工智能 APP，将 AI 带到你的指尖。\nncnn 目前已在腾讯多款应用中使用，如：QQ，Qzone，微信，天天 P 图等。\n\n---\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n技术交流 QQ 群
\n637093648 (超多大佬)
\n答案：卷卷卷卷卷（已满）\n		\nTelegram Group\n\n\n\nDiscord Channel\n\n\n
\nPocky QQ 群（MLIR YES!）
\n677104663 (超多大佬)
\n答案：multi-level intermediate representation\n
\n他们都不知道 pnnx 有多好用群
\n818998520 (新群！)\n
\n\n---\n\n## Download & Build status\n\nhttps://github.com/Tencent/ncnn/releases/latest\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\n \n\n\n **[how to build ncnn library](https://github.com/Tencent/ncnn/wiki/how-to-build) on Linux / Windows / macOS / Raspberry Pi3, Pi4 / POWER / Android / NVIDIA Jetson / iOS / WebAssembly / AllWinner D1 / Loongson 2K1000**\n\n
Source\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-full-source.zip)\n\n
\n \n\n\n- [Build for Android](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-android)\n- [Build for Termux on Android](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-termux-on-android)\n\n
Android\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-android-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-android.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Aandroid)\n\n
Android shared\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-android-vulkan-shared.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-android-shared.zip)\n\n
\n \n\n\n- [Build for HarmonyOS with cross-compiling](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-harmonyos-with-cross-compiling)\n\n
HarmonyOS\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Aharmonyos)\n\n
HarmonyOS shared\n\n
\n \n\n\n- [Build for iOS on macOS with xcode](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-ios-on-macos-with-xcode)\n\n
iOS\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ios-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ios.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Aios)\n\n
iOS-Simulator\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ios-simulator-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ios-simulator.zip)\n\n
\n \n\n\n- [Build for macOS](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-macos)\n\n
macOS\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-macos-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-macos.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Amacos)\n\n
Mac-Catalyst\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-mac-catalyst-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-mac-catalyst.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Amac-catalyst)\n\n
watchOS\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-watchos.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Awatchos)\n\n
watchOS-Simulator\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-watchos-simulator.zip)\n\n
tvOS\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-tvos-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-tvos.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Atvos)\n\n
tvOS-Simulator\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-tvos-simulator-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-tvos-simulator.zip)\n\n
visionOS\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-visionos-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-visionos.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Avisionos)\n\n
visionOS-Simulator\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-visionos-simulator-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-visionos-simulator.zip)\n\n
Apple xcframework\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-apple-vulkan.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-apple.zip)\n\n\n\n
\n \n\n\n- [Build for Linux / NVIDIA Jetson / Raspberry Pi3, Pi4 / POWER](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-linux)\n\n
Ubuntu 22.04\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ubuntu-2204.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ubuntu-2204-shared.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-x64-gpu-gcc)\n\n
Ubuntu 24.04\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ubuntu-2404.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-ubuntu-2404-shared.zip)\n\n
\n \"windows\"\n\n\n- [Build for Windows x64 using VS2017](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-windows-x64-using-visual-studio-community-2017)\n- [Build for Windows x64 using MinGW-w64](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-windows-x64-using-mingw-w64)\n\n
VS2015\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2015.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2015-shared.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Awindows)\n\n
VS2017\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2017.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2017-shared.zip)\n\n
VS2019\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2019.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2019-shared.zip)\n\n
VS2022\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2022.zip)\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-windows-vs2022-shared.zip)\n\n
\n \n\n\n- [Build for WebAssembly](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-webassembly)\n\n
WebAssembly\n\n [](https://github.com/Tencent/ncnn/releases/latest/download/ncnn-20260113-webassembly.zip)\n\n\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Aweb-assembly)\n\n
\n \n\n\n- [Build for ARM Cortex-A family with cross-compiling](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-arm-cortex-a-family-with-cross-compiling)\n- [Build for Hisilicon platform with cross-compiling](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-hisilicon-platform-with-cross-compiling)\n- [Build for AllWinner D1](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-allwinner-d1)\n- [Build for Loongson 2K1000](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-loongson-2k1000)\n- [Build for QNX](https://github.com/Tencent/ncnn/wiki/how-to-build#build-for-qnx)\n\n
Linux (arm)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-arm)\n\n
Linux (aarch64)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-aarch64)\n\n
Linux (mips)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-mips)\n\n
Linux (mips64)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-mips64)\n\n
Linux (ppc64)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-ppc64)\n\n
Linux (riscv64)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-riscv64)\n\n
Linux (loongarch64)\n\n [](https://github.com/Tencent/ncnn/actions?query=workflow%3Alinux-loongarch64)\n\n
\n\n\n---\n\n## Support most commonly used CNN network\n\n## 支持大部分常用的 CNN 网络\n\n- Classical CNN:\n [VGG](https://github.com/BVLC/caffe/wiki/Model-Zoo#models-used-by-the-vgg-team-in-ilsvrc-2014)\n [AlexNet](https://github.com/BVLC/caffe/tree/9b891540183ddc834a02b2bd81b31afae71b2153/models/bvlc_alexnet)\n [GoogleNet](https://github.com/BVLC/caffe/tree/9b891540183ddc834a02b2bd81b31afae71b2153/models/bvlc_googlenet)\n Inception\n ...\n- Practical CNN:\n [ResNet](https://github.com/tornadomeet/ResNet)\n [DenseNet](https://github.com/liuzhuang13/DenseNet)\n [SENet](https://github.com/hujie-frank/SENet)\n [FPN](https://github.com/unsky/FPN)\n ...\n- Light-weight CNN:\n [SqueezeNet](https://github.com/forresti/SqueezeNet)\n [MobileNetV1](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md)\n [MobileNetV2/V3](https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/README.md)\n [ShuffleNetV1](https://github.com/farmingyard/ShuffleNet)\n [ShuffleNetV2](https://github.com/opconty/keras-shufflenetV2)\n [MNasNet](https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet)\n ...\n- Face Detection:\n [MTCNN](https://github.com/ipazc/mtcnn)\n [RetinaFace](https://github.com/biubug6/Pytorch_Retinaface)\n [scrfd](https://github.com/nihui/ncnn-android-scrfd)\n ...\n- Detection:\n [VGG-SSD](https://github.com/lzx1413/CAFFE_SSD)\n [MobileNet-SSD](https://github.com/chuanqi305/MobileNet-SSD)\n [SqueezeNet-SSD](https://github.com/chuanqi305/SqueezeNet-SSD)\n [MobileNetV2-SSDLite](https://github.com/chuanqi305/MobileNetv2-SSDLite)\n [MobileNetV3-SSDLite](https://github.com/XiaoyuHuang96/MobilenetV3SSDLite-tfkeras)\n ...\n- Detection:\n [Faster-RCNN](https://github.com/rbgirshick/py-faster-rcnn)\n [R-FCN](https://github.com/daijifeng001/R-FCN)\n ...\n- Detection:\n [YOLOv2](https://github.com/longcw/yolo2-pytorch)\n [YOLOv3](https://github.com/ultralytics/yolov3)\n [MobileNet-YOLOv3](https://github.com/eric612/MobileNet-YOLO)\n [YOLOv4](https://github.com/Tianxiaomo/pytorch-YOLOv4)\n [YOLOv5](https://github.com/ultralytics/yolov5)\n [YOLOv7](https://github.com/WongKinYiu/yolov7)\n [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX)\n [YOLOv8](https://github.com/nihui/ncnn-android-yolov8)\n ...\n- Detection:\n [NanoDet](https://github.com/RangiLyu/nanodet)\n- Segmentation:\n [FCN](https://github.com/unsky/FPN)\n [PSPNet](https://github.com/hszhao/PSPNet)\n [UNet](https://github.com/zhixuhao/unet)\n [YOLACT](https://github.com/dbolya/yolact)\n ...\n- Pose Estimation:\n [SimplePose](https://github.com/dog-qiuqiu/Ultralight-SimplePose)\n ...\n\n---\n\n## HowTo\n\n**[use ncnn with alexnet](https://github.com/Tencent/ncnn/wiki/use-ncnn-with-alexnet) with detailed steps, recommended for beginners :)**\n\n**[ncnn 组件使用指北 alexnet](https://github.com/Tencent/ncnn/wiki/use-ncnn-with-alexnet.zh) 附带详细步骤，新人强烈推荐 :)**\n\n**[use netron for ncnn model visualization](https://netron.app)**\n\n**[use ncnn with pytorch or onnx](https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx)**\n\n[ncnn low-level operation api](https://github.com/Tencent/ncnn/wiki/low-level-operation-api)\n\n[ncnn param and model file spec](https://github.com/Tencent/ncnn/wiki/param-and-model-file-structure)\n\n[ncnn operation param weight table](https://github.com/Tencent/ncnn/wiki/operation-param-weight-table)\n\n[how to implement custom layer step by step](https://github.com/Tencent/ncnn/wiki/how-to-implement-custom-layer-step-by-step)\n\n---\n\n## FAQ\n\n**[ncnn deepwiki](https://deepwiki.com/Tencent/ncnn) LLM Answering Questions ;)** \n\n**[ncnn throw error](https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-throw-error)**\n\n**[ncnn produce wrong result](https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result)**\n\n**[ncnn vulkan](https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-vulkan)**\n\n---\n\n## Features\n\n- Supports convolutional neural networks, supports multiple input and multi-branch structure, can calculate part of the branch\n- No third-party library dependencies, does not rely on BLAS / NNPACK or any other computing framework\n- Pure C++ implementation, cross-platform, supports Android, iOS and so on\n- ARM NEON assembly level of careful optimization, calculation speed is extremely high\n- Sophisticated memory management and data structure design, very low memory footprint\n- Supports multi-core parallel computing acceleration, ARM big.LITTLE CPU scheduling optimization\n- Supports GPU acceleration via the next-generation low-overhead Vulkan API\n- Extensible model design, supports 8bit [quantization](https://github.com/Tencent/ncnn/wiki/quantized-int8-inference) and half-precision floating point storage, can import caffe/pytorch/mxnet/onnx/darknet/keras/tensorflow(mlir) models\n- Support direct memory zero copy reference load network model\n- Can be registered with custom layer implementation and extended\n- Well, it is strong, not afraid of being stuffed with 卷 QvQ\n\n## 功能概述\n\n- 支持卷积神经网络，支持多输入和多分支结构，可计算部分分支\n- 无任何第三方库依赖，不依赖 BLAS/NNPACK 等计算框架\n- 纯 C++ 实现，跨平台，支持 Android / iOS 等\n- ARM Neon 汇编级良心优化，计算速度极快\n- 精细的内存管理和数据结构设计，内存占用极低\n- 支持多核并行计算加速，ARM big.LITTLE CPU 调度优化\n- 支持基于全新低消耗的 Vulkan API GPU 加速\n- 可扩展的模型设计，支持 8bit [量化](tools/quantize) 和半精度浮点存储，可导入 caffe/pytorch/mxnet/onnx/darknet/keras/tensorflow(mlir) 模型\n- 支持直接内存零拷贝引用加载网络模型\n- 可注册自定义层实现并扩展\n- 恩，很强就是了，不怕被塞卷 QvQ\n\n---\n\n## supported platform matrix\n\n- ✅ = known work and runs fast with good optimization\n- ✔️ = known work, but speed may not be fast enough\n- ❔ = shall work, not confirmed\n- / = not applied\n\n| | Windows | Linux | Android | macOS | iOS |\n| ---------- | ------- | ----- | ------- | ----- | --- |\n| intel-cpu | ✔️ | ✔️ | ✔️ | ✔️ | / |\n| intel-gpu | ✔️ | ✔️ | ✔️ | ✔️ | / |\n| amd-cpu | ✔️ | ✔️ | ✔️ | ✔️ | / |\n| amd-gpu | ✔️ | ✔️ | ✔️ | ✔️ | / |\n| nvidia-gpu | ✔️ | ✔️ | ✔️ | ✔️ | / |\n| qcom-cpu | ✅ | ✅ | ✅ | / | / |\n| qcom-gpu | ✔️ | ✔️ | ✔️ | / | / |\n| arm-cpu | ✅ | ✅ | ✅ | / | / |\n| arm-gpu | ❔ | ✔️ | ✔️ | / | / |\n| apple-cpu | / | / | / | ✔️ | ✅ |\n| apple-gpu | / | / | / | ✔️ | ✔️ |\n| ibm-cpu | / | ✔️ | / | / | / |\n\n---\n\n## Project examples\n\n- \n- \n- \n- \n- \n- \n- 🤩\n- \n\n
\n\n-
\n\n\n- Call ncnn from Fortran\n\n- Use ncnn for real-time speech\n recognition (i.e., speech-to-text); also support embedded devices and provide\n mobile Apps (e.g., Android App)\n\n---\n\n## License\n\n[BSD 3 Clause](LICENSE.txt)\n" }, { "path": "benchmark/CMakeLists.txt", "content": "\nif(MSVC)\n # warning C4996: 'fopen': This function or variable may be unsafe. Consider using fopen_s instead. To disable deprecation, use _CRT_SECURE_NO_WARNINGS. See online help for details.\n add_definitions(/wd4996)\nendif()\n\n# ncnn macro\ninclude(${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_add_param.cmake)\n\nset(benchncnn_PARAMS\n alexnet.param\n blazeface.param\n efficientnet_b0.param\n efficientnetv2_b0.param\n FastestDet.param\n googlenet_int8.param\n googlenet.param\n mnasnet.param\n mobilenet_int8.param\n mobilenet_ssd_int8.param\n mobilenet_ssd.param\n mobilenet_v2.param\n mobilenet_v3.param\n mobilenet_yolo.param\n mobilenet.param\n mobilenetv2_yolov3.param\n nanodet_m.param\n proxylessnasnet.param\n regnety_400m.param\n resnet18_int8.param\n resnet18.param\n resnet50_int8.param\n resnet50.param\n shufflenet_v2.param\n shufflenet.param\n squeezenet_int8.param\n squeezenet_ssd_int8.param\n squeezenet_ssd.param\n squeezenet.param\n vgg16_int8.param\n vgg16.param\n vision_transformer.param\n yolo-fastest-1.1.param\n yolo-fastestv2.param\n yolov4-tiny.param\n)\n\nforeach(PARAM_FILE ${benchncnn_PARAMS})\n ncnn_add_param(\"${CMAKE_CURRENT_SOURCE_DIR}/${PARAM_FILE}\")\nendforeach()\n\nadd_custom_target(ncnn-generate-param DEPENDS ${NCNN_PARAM_HEX_FILES})\n\nconfigure_file(benchncnn_param_data.h.in ${CMAKE_CURRENT_BINARY_DIR}/benchncnn_param_data.h)\n\nadd_executable(benchncnn benchncnn.cpp)\ntarget_link_libraries(benchncnn PRIVATE ncnn)\n\ntarget_include_directories(benchncnn PRIVATE ${CMAKE_CURRENT_BINARY_DIR})\n\nif(CMAKE_SYSTEM_NAME STREQUAL \"Emscripten\")\n target_link_libraries(benchncnn PRIVATE nodefs.js)\nendif()\n\nadd_dependencies(benchncnn ncnn-generate-param)\n\n# add benchncnn to a virtual project group\nset_property(TARGET benchncnn PROPERTY FOLDER \"benchmark\")\n" }, { "path": "benchmark/FastestDet.param", "content": "7767517\n127 150\nInput in0 0 1 in0\nConvolution convrelu_0 1 1 in0 1 0=24 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=648 9=1\nPooling maxpool2d_43 1 1 1 2 0=0 1=3 11=3 12=2 13=1 2=2 3=1 5=1\nSplit splitncnn_0 1 2 2 3 4\nConvolutionDepthWise convdw_95 1 1 4 5 0=24 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=216 7=24\nConvolution convrelu_1 1 1 3 6 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConvolutionDepthWise convdw_96 1 1 6 7 0=24 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=216 7=24\nConvolution convrelu_3 1 1 5 8 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConvolution convrelu_2 1 1 7 9 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConcat cat_0 2 1 8 9 10 0=0\nShuffleChannel shufflechannel_0 1 1 10 11 0=2 1=1\nSlice shufflechannel_0_slice 1 2 11 12 13 -23300=2,-233,-233 1=0\nConvolution convrelu_4 1 1 13 14 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConvolutionDepthWise convdw_97 1 1 14 15 0=24 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=216 7=24\nConvolution convrelu_5 1 1 15 16 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConcat cat_1 2 1 12 16 17 0=0\nShuffleChannel shufflechannel_1 1 1 17 18 0=2 1=1\nSlice shufflechannel_1_slice 1 2 18 19 20 -23300=2,-233,-233 1=0\nConvolution convrelu_6 1 1 20 21 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConvolutionDepthWise convdw_98 1 1 21 22 0=24 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=216 7=24\nConvolution convrelu_7 1 1 22 23 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConcat cat_2 2 1 19 23 24 0=0\nShuffleChannel shufflechannel_2 1 1 24 25 0=2 1=1\nSlice shufflechannel_2_slice 1 2 25 26 27 -23300=2,-233,-233 1=0\nConvolution convrelu_8 1 1 27 28 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConvolutionDepthWise convdw_99 1 1 28 29 0=24 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=216 7=24\nConvolution convrelu_9 1 1 29 30 0=24 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=576 9=1\nConcat cat_3 2 1 26 30 31 0=0\nSplit splitncnn_1 1 3 31 32 33 34\nConvolutionDepthWise convdw_100 1 1 34 35 0=48 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=432 7=48\nConvolution convrelu_10 1 1 33 36 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_101 1 1 36 37 0=48 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=432 7=48\nConvolution convrelu_12 1 1 35 38 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolution convrelu_11 1 1 37 39 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_4 2 1 38 39 40 0=0\nShuffleChannel shufflechannel_3 1 1 40 41 0=2 1=1\nSlice shufflechannel_3_slice 1 2 41 42 43 -23300=2,-233,-233 1=0\nConvolution convrelu_13 1 1 43 44 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_102 1 1 44 45 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_14 1 1 45 46 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_5 2 1 42 46 47 0=0\nShuffleChannel shufflechannel_4 1 1 47 48 0=2 1=1\nSlice shufflechannel_4_slice 1 2 48 49 50 -23300=2,-233,-233 1=0\nConvolution convrelu_15 1 1 50 51 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_103 1 1 51 52 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_16 1 1 52 53 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_6 2 1 49 53 54 0=0\nShuffleChannel shufflechannel_5 1 1 54 55 0=2 1=1\nSlice shufflechannel_5_slice 1 2 55 56 57 -23300=2,-233,-233 1=0\nConvolution convrelu_17 1 1 57 58 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_104 1 1 58 59 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_18 1 1 59 60 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_7 2 1 56 60 61 0=0\nShuffleChannel shufflechannel_6 1 1 61 62 0=2 1=1\nSlice shufflechannel_6_slice 1 2 62 63 64 -23300=2,-233,-233 1=0\nConvolution convrelu_19 1 1 64 65 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_105 1 1 65 66 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_20 1 1 66 67 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_8 2 1 63 67 68 0=0\nShuffleChannel shufflechannel_7 1 1 68 69 0=2 1=1\nSlice shufflechannel_7_slice 1 2 69 70 71 -23300=2,-233,-233 1=0\nConvolution convrelu_21 1 1 71 72 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_106 1 1 72 73 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_22 1 1 73 74 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_9 2 1 70 74 75 0=0\nShuffleChannel shufflechannel_8 1 1 75 76 0=2 1=1\nSlice shufflechannel_8_slice 1 2 76 77 78 -23300=2,-233,-233 1=0\nConvolution convrelu_23 1 1 78 79 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_107 1 1 79 80 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_24 1 1 80 81 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_10 2 1 77 81 82 0=0\nShuffleChannel shufflechannel_9 1 1 82 83 0=2 1=1\nSlice shufflechannel_9_slice 1 2 83 84 85 -23300=2,-233,-233 1=0\nConvolution convrelu_25 1 1 85 86 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConvolutionDepthWise convdw_108 1 1 86 87 0=48 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=432 7=48\nConvolution convrelu_26 1 1 87 88 0=48 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=2304 9=1\nConcat cat_11 2 1 84 88 89 0=0\nSplit splitncnn_2 1 3 89 90 91 92\nConvolutionDepthWise convdw_109 1 1 92 93 0=96 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=864 7=96\nConvolution convrelu_27 1 1 91 94 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConvolutionDepthWise convdw_110 1 1 94 95 0=96 1=3 11=3 12=1 13=2 14=1 2=1 3=2 4=1 5=1 6=864 7=96\nConvolution convrelu_29 1 1 93 96 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConvolution convrelu_28 1 1 95 97 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConcat cat_12 2 1 96 97 98 0=0\nShuffleChannel shufflechannel_10 1 1 98 99 0=2 1=1\nSlice shufflechannel_10_slice 1 2 99 100 101 -23300=2,-233,-233 1=0\nConvolution convrelu_30 1 1 101 102 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConvolutionDepthWise convdw_111 1 1 102 103 0=96 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=864 7=96\nConvolution convrelu_31 1 1 103 104 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConcat cat_13 2 1 100 104 105 0=0\nShuffleChannel shufflechannel_11 1 1 105 106 0=2 1=1\nSlice shufflechannel_11_slice 1 2 106 107 108 -23300=2,-233,-233 1=0\nConvolution convrelu_32 1 1 108 109 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConvolutionDepthWise convdw_112 1 1 109 110 0=96 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=864 7=96\nConvolution convrelu_33 1 1 110 111 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConcat cat_14 2 1 107 111 112 0=0\nShuffleChannel shufflechannel_12 1 1 112 113 0=2 1=1\nSlice shufflechannel_12_slice 1 2 113 114 115 -23300=2,-233,-233 1=0\nConvolution convrelu_34 1 1 115 116 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConvolutionDepthWise convdw_113 1 1 116 117 0=96 1=3 11=3 12=1 13=1 14=1 2=1 3=1 4=1 5=1 6=864 7=96\nConvolution convrelu_35 1 1 117 118 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nConcat cat_15 2 1 114 118 119 0=0\nPooling avgpool2d_0 1 1 32 120 0=1 1=3 11=3 12=2 13=1 2=2 3=1 5=1 6=1\nInterp upsample_94 1 1 119 121 0=1 1=2.000000e+00 2=2.000000e+00 6=0\nConcat cat_16 3 1 120 90 121 122 0=0\nConvolution convrelu_36 1 1 122 123 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=32256 9=1\nSplit splitncnn_3 1 4 123 124 125 126 127\nConvolutionDepthWise convdwrelu_5 1 1 127 128 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolutionDepthWise convdwrelu_0 1 1 126 129 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolutionDepthWise convdwrelu_4 1 1 129 130 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolutionDepthWise convdwrelu_1 1 1 125 131 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolutionDepthWise convdwrelu_2 1 1 131 132 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolutionDepthWise convdwrelu_3 1 1 132 133 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConcat cat_17 3 1 128 130 133 134 0=0\nConvolution conv_38 1 1 134 135 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=27648\nBinaryOp add_0 2 1 124 135 136 0=0\nReLU relu_87 1 1 136 137\nConvolution convrelu_37 1 1 137 138 0=96 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=9216 9=1\nSplit splitncnn_4 1 3 138 139 140 141\nConvolutionDepthWise convdwrelu_7 1 1 139 142 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolution conv_41 1 1 142 143 0=80 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=7680\nConvolutionDepthWise convdwrelu_8 1 1 140 144 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolution conv_42 1 1 144 145 0=4 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=384\nSoftmax softmax_93 1 1 143 146 0=0 1=1\nConvolutionDepthWise convdwrelu_6 1 1 141 147 0=96 1=5 11=5 12=1 13=1 14=2 2=1 3=1 4=2 5=1 6=2400 7=96 9=1\nConvolution convsigmoid_38 1 1 147 148 0=1 1=1 11=1 12=1 13=1 14=0 2=1 3=1 4=0 5=1 6=96 9=4\nConcat cat_18 3 1 148 145 146 out0 0=0\n" }, { "path": "benchmark/README.md", "content": "benchncnn can be used to test neural network inference performance\n\nOnly the network definition files (ncnn param) are required.\n\nThe large model binary files (ncnn bin) are not loaded but generated randomly for speed test.\n\nIf no model specified, it would benchmark default built-in models. More model networks may be added later.\n\n---\nBuild\n```shell\n# assume you have already build ncnn library successfully\n# uncomment the following line in /CMakeLists.txt with your favorite editor\n\n# add_subdirectory(benchmark)\n\ncd /\nmake -j4\n\n# you can find benchncnn binary in //benchmark\n```\n\nUsage\n```shell\n# copy all param files to the current directory\n./benchncnn [loop count] [num threads] [powersave] [gpu device] [cooling down] [(key=value)...]\n param=model.param\n shape=[227,227,3],..\n```\nrun benchncnn on android device\n```shell\n# for running on android device, upload to /data/local/tmp/ folder\nadb push benchncnn /data/local/tmp/\n\n# (optional) upload your ncnn model param to /data/local/tmp/ folder\nadb push model.param /data/local/tmp/\n\n# executed in android adb shell\nadb shell\ncd /data/local/tmp/\n\n# sample: benchmark built-in models on cpu, with 4 threads on big core, 4 loops and cooling_down\n./benchncnn 4 4 2 -1 1\n\n# sample: benchmark built-in models on gpu id 0, with 1 thread on big core, 8 loops, without cooling_down\n./benchncnn 8 1 2 0 0\n\n./benchncnn [loop count] [num threads] [powersave] [gpu device] [cooling down] [(key=value)...]\n param=model.param\n shape=[227,227,3],..\n```\n\nParameter\n\n|param|options|default|\n|---|---|---|\n|loop count|1~N|4|\n|num threads|1~N|max_cpu_count|\n|powersave|0=all cores, 1=little cores only, 2=big cores only|0|\n|gpu device|-1=cpu-only, 0=gpu0, 1=gpu1 ...|-1|\n|cooling down|0=disable, 1=enable|1|\n|param|ncnn model.param filepath|-|\n|shape|model input shapes with, whc format|-|\n\nTips: Disable android UI server and set CPU and GPU to max frequency\n```shell\n# stopping android ui server, can be retarted later via adb shell start\nadb root\nadb shell stop\n\n# executed in android adb shell\n# set cpu performance mode\necho \"performance\" > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor\necho \"performance\" > /sys/devices/system/cpu/cpu1/cpufreq/scaling_governor\necho \"performance\" > /sys/devices/system/cpu/cpu2/cpufreq/scaling_governor\necho \"performance\" > /sys/devices/system/cpu/cpu3/cpufreq/scaling_governor\necho \"performance\" > /sys/devices/system/cpu/cpu4/cpufreq/scaling_governor\necho \"performance\" > /sys/devices/system/cpu/cpu5/cpufreq/scaling_governor\n\n# set gpu performance mode (eg. RK3399)\necho \"performance\" > /sys/class/misc/mali0/device/devfreq/ff9a0000.gpu/governor\n\n# set gpu performance mode (eg. Android Adreno)\necho 1 > /sys/class/kgsl/kgsl-3d0/force_clk_on\necho 10000000 > /sys/class/kgsl/kgsl-3d0/idle_timer\necho \"performance\" > /sys/class/kgsl/kgsl-3d0/devfreq/governor\necho > /sys/class/kgsl/kgsl-3d0/gpuclk\n```\n\n---\n\nTypical output (executed in android adb shell)\n\n### NVIDIA Jetson AGX Orin (Cortex-A78AE 2.2 GHz x 12 + Ampere@1.3 GHz Tensor Cores 64)\n```\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 64 1 0 -1 0\nloop_count = 64\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 11.66 max = 11.80 avg = 11.74\n squeezenet_int8 min = 12.24 max = 12.39 avg = 12.31\n mobilenet min = 19.56 max = 19.73 avg = 19.65\n mobilenet_int8 min = 16.06 max = 16.25 avg = 16.14\n mobilenet_v2 min = 13.20 max = 13.41 avg = 13.29\n mobilenet_v3 min = 11.39 max = 11.57 avg = 11.48\n shufflenet min = 8.07 max = 8.18 avg = 8.11\n shufflenet_v2 min = 8.41 max = 8.51 avg = 8.45\n mnasnet min = 12.74 max = 12.91 avg = 12.79\n proxylessnasnet min = 15.18 max = 15.32 avg = 15.25\n efficientnet_b0 min = 26.86 max = 26.96 avg = 26.90\n efficientnetv2_b0 min = 35.99 max = 36.15 avg = 36.07\n regnety_400m min = 16.81 max = 16.98 avg = 16.87\n blazeface min = 4.25 max = 4.37 avg = 4.29\n googlenet min = 48.73 max = 48.98 avg = 48.87\n googlenet_int8 min = 47.39 max = 47.60 avg = 47.49\n resnet18 min = 30.93 max = 31.24 avg = 31.08\n resnet18_int8 min = 55.44 max = 55.70 avg = 55.56\n alexnet min = 44.19 max = 44.43 avg = 44.33\n vgg16 min = 173.94 max = 174.97 avg = 174.46\n vgg16_int8 min = 475.10 max = 479.37 avg = 477.33\n resnet50 min = 89.50 max = 90.11 avg = 89.80\n resnet50_int8 min = 106.77 max = 107.14 avg = 106.96\n squeezenet_ssd min = 37.78 max = 38.35 avg = 37.93\n squeezenet_ssd_int8 min = 50.48 max = 50.88 avg = 50.74\n mobilenet_ssd min = 45.62 max = 46.12 avg = 45.74\n mobilenet_ssd_int8 min = 37.77 max = 38.00 avg = 37.88\n mobilenet_yolo min = 90.23 max = 90.49 avg = 90.35\n mobilenetv2_yolov3 min = 47.27 max = 47.48 avg = 47.33\n yolov4-tiny min = 60.41 max = 60.75 avg = 60.57\n nanodet_m min = 19.26 max = 19.43 avg = 19.35\n yolo-fastest-1.1 min = 8.16 max = 8.31 avg = 8.20\n yolo-fastestv2 min = 8.26 max = 8.39 avg = 8.32\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 64 2 0 -1 0\nloop_count = 64\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.83 max = 6.98 avg = 6.90\n squeezenet_int8 min = 7.39 max = 7.50 avg = 7.45\n mobilenet min = 10.40 max = 10.50 avg = 10.45\n mobilenet_int8 min = 8.92 max = 9.09 avg = 8.99\n mobilenet_v2 min = 7.67 max = 7.80 avg = 7.74\n mobilenet_v3 min = 6.86 max = 7.01 avg = 6.93\n shufflenet min = 6.34 max = 6.44 avg = 6.39\n shufflenet_v2 min = 5.71 max = 5.83 avg = 5.76\n mnasnet min = 7.47 max = 7.58 avg = 7.53\n proxylessnasnet min = 8.73 max = 8.83 avg = 8.78\n efficientnet_b0 min = 14.93 max = 15.13 avg = 15.03\n efficientnetv2_b0 min = 20.17 max = 20.70 avg = 20.29\n regnety_400m min = 12.50 max = 12.62 avg = 12.57\n blazeface min = 2.95 max = 3.06 avg = 3.00\n googlenet min = 26.25 max = 26.53 avg = 26.37\n googlenet_int8 min = 26.54 max = 26.79 avg = 26.66\n resnet18 min = 16.69 max = 16.90 avg = 16.80\n resnet18_int8 min = 29.70 max = 29.93 avg = 29.81\n alexnet min = 22.96 max = 23.12 avg = 23.03\n vgg16 min = 88.39 max = 89.16 avg = 88.79\n vgg16_int8 min = 245.86 max = 247.55 avg = 246.62\n resnet50 min = 46.55 max = 46.86 avg = 46.70\n resnet50_int8 min = 56.28 max = 56.63 avg = 56.43\n squeezenet_ssd min = 23.65 max = 24.29 avg = 23.81\n squeezenet_ssd_int8 min = 30.86 max = 31.27 avg = 30.99\n mobilenet_ssd min = 25.17 max = 25.31 avg = 25.24\n mobilenet_ssd_int8 min = 21.77 max = 21.97 avg = 21.84\n mobilenet_yolo min = 48.03 max = 48.33 avg = 48.14\n mobilenetv2_yolov3 min = 26.58 max = 26.81 avg = 26.66\n yolov4-tiny min = 35.31 max = 35.53 avg = 35.41\n nanodet_m min = 12.93 max = 13.08 avg = 13.01\n yolo-fastest-1.1 min = 6.00 max = 6.10 avg = 6.04\n yolo-fastestv2 min = 6.46 max = 6.61 avg = 6.52\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 64 4 0 -1 0\nloop_count = 64\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 4.54 max = 4.84 avg = 4.61\n squeezenet_int8 min = 4.96 max = 5.41 avg = 5.05\n mobilenet min = 5.96 max = 6.23 avg = 6.04\n mobilenet_int8 min = 5.21 max = 5.50 avg = 5.30\n mobilenet_v2 min = 5.05 max = 5.26 avg = 5.15\n mobilenet_v3 min = 4.83 max = 5.14 avg = 4.90\n shufflenet min = 5.11 max = 5.34 avg = 5.18\n shufflenet_v2 min = 4.13 max = 4.44 avg = 4.18\n mnasnet min = 4.93 max = 5.27 avg = 5.01\n proxylessnasnet min = 5.64 max = 5.89 avg = 5.72\n efficientnet_b0 min = 9.47 max = 10.60 avg = 9.60\n efficientnetv2_b0 min = 12.67 max = 13.06 avg = 12.82\n regnety_400m min = 10.27 max = 10.58 avg = 10.38\n blazeface min = 2.05 max = 2.27 avg = 2.10\n googlenet min = 15.57 max = 15.96 avg = 15.68\n googlenet_int8 min = 16.19 max = 16.65 avg = 16.32\n resnet18 min = 10.20 max = 11.76 avg = 10.35\n resnet18_int8 min = 16.89 max = 17.31 avg = 17.03\n alexnet min = 13.13 max = 13.70 avg = 13.32\n vgg16 min = 51.03 max = 52.46 avg = 51.35\n vgg16_int8 min = 131.08 max = 139.44 avg = 133.78\n resnet50 min = 26.74 max = 28.32 avg = 26.91\n resnet50_int8 min = 32.15 max = 32.74 avg = 32.38\n squeezenet_ssd min = 16.58 max = 16.99 avg = 16.70\n squeezenet_ssd_int8 min = 20.22 max = 21.67 avg = 20.51\n mobilenet_ssd min = 14.68 max = 16.07 avg = 14.83\n mobilenet_ssd_int8 min = 12.89 max = 13.27 avg = 13.01\n mobilenet_yolo min = 28.44 max = 28.85 avg = 28.58\n mobilenetv2_yolov3 min = 17.21 max = 21.31 avg = 17.44\n yolov4-tiny min = 23.68 max = 24.38 avg = 23.88\n nanodet_m min = 8.76 max = 9.17 avg = 8.86\n yolo-fastest-1.1 min = 4.83 max = 5.04 avg = 4.88\n yolo-fastestv2 min = 4.93 max = 5.17 avg = 5.00\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 64 8 0 -1 0\nloop_count = 64\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.52 max = 4.28 avg = 3.65\n squeezenet_int8 min = 3.85 max = 4.11 avg = 3.93\n mobilenet min = 3.78 max = 4.12 avg = 3.85\n mobilenet_int8 min = 3.57 max = 3.85 avg = 3.63\n mobilenet_v2 min = 4.14 max = 4.44 avg = 4.22\n mobilenet_v3 min = 3.89 max = 4.26 avg = 3.97\n shufflenet min = 4.78 max = 4.95 avg = 4.84\n shufflenet_v2 min = 3.49 max = 3.84 avg = 3.54\n mnasnet min = 3.94 max = 4.09 avg = 3.99\n proxylessnasnet min = 4.41 max = 4.68 avg = 4.47\n efficientnet_b0 min = 7.01 max = 7.85 avg = 7.13\n efficientnetv2_b0 min = 9.22 max = 9.46 avg = 9.32\n regnety_400m min = 9.34 max = 9.66 avg = 9.44\n blazeface min = 1.86 max = 1.98 avg = 1.89\n googlenet min = 10.37 max = 10.76 avg = 10.48\n googlenet_int8 min = 11.03 max = 11.34 avg = 11.16\n resnet18 min = 6.83 max = 7.12 avg = 6.93\n resnet18_int8 min = 10.25 max = 11.50 avg = 10.42\n alexnet min = 8.88 max = 9.71 avg = 9.01\n vgg16 min = 31.26 max = 31.97 avg = 31.44\n vgg16_int8 min = 71.31 max = 74.53 avg = 72.18\n resnet50 min = 16.43 max = 16.84 avg = 16.52\n resnet50_int8 min = 19.07 max = 20.28 avg = 19.42\n squeezenet_ssd min = 13.50 max = 13.69 avg = 13.56\n squeezenet_ssd_int8 min = 15.16 max = 16.06 avg = 15.30\n mobilenet_ssd min = 9.73 max = 10.85 avg = 9.90\n mobilenet_ssd_int8 min = 9.27 max = 9.46 avg = 9.36\n mobilenet_yolo min = 17.58 max = 17.79 avg = 17.67\n mobilenetv2_yolov3 min = 12.80 max = 13.50 avg = 12.90\n yolov4-tiny min = 17.98 max = 21.31 avg = 18.24\n nanodet_m min = 7.01 max = 7.18 avg = 7.09\n yolo-fastest-1.1 min = 4.76 max = 4.86 avg = 4.80\n yolo-fastestv2 min = 4.76 max = 4.88 avg = 4.82\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 64 12 0 -1 0\nloop_count = 64\nnum_threads = 12\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.50 max = 5.21 avg = 3.65\n squeezenet_int8 min = 3.97 max = 4.44 avg = 4.12\n mobilenet min = 3.49 max = 7.73 avg = 3.78\n mobilenet_int8 min = 3.40 max = 3.86 avg = 3.49\n mobilenet_v2 min = 4.07 max = 4.39 avg = 4.17\n mobilenet_v3 min = 3.92 max = 4.17 avg = 4.03\n shufflenet min = 5.08 max = 6.63 avg = 5.18\n shufflenet_v2 min = 3.64 max = 5.11 avg = 3.75\n mnasnet min = 3.86 max = 4.16 avg = 3.95\n proxylessnasnet min = 4.30 max = 5.39 avg = 4.38\n efficientnet_b0 min = 6.42 max = 9.19 avg = 6.61\n efficientnetv2_b0 min = 8.96 max = 9.43 avg = 9.12\n regnety_400m min = 10.11 max = 10.89 avg = 10.27\n blazeface min = 1.93 max = 2.16 avg = 1.99\n googlenet min = 9.72 max = 10.84 avg = 10.01\n googlenet_int8 min = 10.91 max = 13.03 avg = 11.17\n resnet18 min = 6.70 max = 7.27 avg = 6.92\n resnet18_int8 min = 9.62 max = 12.93 avg = 10.14\n alexnet min = 7.21 max = 7.47 avg = 7.32\n vgg16 min = 29.61 max = 63.73 avg = 30.86\n vgg16_int8 min = 64.91 max = 75.06 avg = 68.72\n resnet50 min = 15.35 max = 16.28 avg = 15.73\n resnet50_int8 min = 17.47 max = 18.98 avg = 18.09\n squeezenet_ssd min = 13.40 max = 28.74 avg = 14.07\n squeezenet_ssd_int8 min = 15.35 max = 16.77 avg = 15.67\n mobilenet_ssd min = 9.51 max = 11.49 avg = 9.88\n mobilenet_ssd_int8 min = 9.43 max = 10.08 avg = 9.58\n mobilenet_yolo min = 16.88 max = 17.45 avg = 17.09\n mobilenetv2_yolov3 min = 11.91 max = 31.90 avg = 12.50\n yolov4-tiny min = 17.85 max = 18.87 avg = 18.36\n nanodet_m min = 6.88 max = 7.64 avg = 7.06\n yolo-fastest-1.1 min = 5.02 max = 5.53 avg = 5.12\n yolo-fastestv2 min = 4.95 max = 5.60 avg = 5.05\ni@orin:~/projects/ncnn/benchmark$ ./benchncnn 128 1 0 0 0\n[0 NVIDIA Tegra Orin (nvgpu)] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA Tegra Orin (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra Orin (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra Orin (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 128\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 2.13 max = 3.37 avg = 2.31\n squeezenet_int8 min = 12.31 max = 12.51 avg = 12.42\n mobilenet min = 2.03 max = 2.73 avg = 2.23\n mobilenet_int8 min = 16.86 max = 17.91 avg = 16.99\n mobilenet_v2 min = 2.59 max = 3.59 avg = 2.91\n mobilenet_v3 min = 3.22 max = 4.23 avg = 3.71\n shufflenet min = 2.57 max = 3.27 avg = 2.80\n shufflenet_v2 min = 3.20 max = 4.03 avg = 3.47\n mnasnet min = 2.45 max = 3.06 avg = 2.69\n proxylessnasnet min = 2.50 max = 3.14 avg = 2.72\n efficientnet_b0 min = 4.23 max = 8.73 avg = 4.85\n efficientnetv2_b0 min = 8.15 max = 8.60 avg = 8.41\n regnety_400m min = 3.25 max = 4.17 avg = 3.54\n blazeface min = 1.29 max = 1.48 avg = 1.33\n googlenet min = 4.95 max = 12.34 avg = 6.36\n googlenet_int8 min = 47.49 max = 47.78 avg = 47.61\n resnet18 min = 3.18 max = 9.49 avg = 4.04\n resnet18_int8 min = 55.57 max = 55.88 avg = 55.73\n alexnet min = 3.22 max = 14.56 avg = 4.25\n vgg16 min = 6.82 max = 14.75 avg = 8.18\n vgg16_int8 min = 473.55 max = 479.07 avg = 476.22\n resnet50 min = 4.75 max = 15.06 avg = 6.08\n resnet50_int8 min = 106.99 max = 107.48 avg = 107.22\n squeezenet_ssd min = 6.87 max = 9.12 avg = 7.76\n squeezenet_ssd_int8 min = 50.87 max = 51.17 avg = 51.01\n mobilenet_ssd min = 4.44 max = 6.22 avg = 5.23\n mobilenet_ssd_int8 min = 37.80 max = 38.03 avg = 37.92\n mobilenet_yolo min = 5.41 max = 7.36 avg = 6.29\n mobilenetv2_yolov3 min = 7.20 max = 9.96 avg = 7.30\n yolov4-tiny min = 16.48 max = 28.81 avg = 18.40\n nanodet_m min = 5.75 max = 8.54 avg = 6.85\n yolo-fastest-1.1 min = 4.03 max = 4.75 avg = 4.35\n yolo-fastestv2 min = 4.27 max = 5.23 avg = 4.71\n```\n\n### AMD Ryzen Threadripper 3970X (Zen2 3.7 GHz ~ 4.5 GHz x 32)\n```\ni@s:~/qtang/ncnn/benchmark$ ../build-vulkan/benchmark/benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 11.73 max = 11.88 avg = 11.78\n mobilenet min = 21.63 max = 21.73 avg = 21.68\n mobilenet_v2 min = 14.70 max = 14.95 avg = 14.82\n mobilenet_v3 min = 12.12 max = 12.17 avg = 12.15\n shufflenet min = 14.08 max = 14.16 avg = 14.12\n shufflenet_v2 min = 25.99 max = 26.13 avg = 26.06\n mnasnet min = 14.12 max = 14.17 avg = 14.14\n proxylessnasnet min = 16.51 max = 16.71 avg = 16.61\n efficientnet_b0 min = 22.88 max = 22.97 avg = 22.93\n regnety_400m min = 18.50 max = 18.61 avg = 18.56\n blazeface min = 6.18 max = 6.27 avg = 6.21\n googlenet min = 58.42 max = 58.60 avg = 58.49\n resnet18 min = 61.13 max = 61.84 avg = 61.40\n alexnet min = 50.82 max = 50.98 avg = 50.92\n vgg16 min = 217.19 max = 218.40 avg = 217.87\n resnet50 min = 126.84 max = 137.46 avg = 128.21\n squeezenet_ssd min = 114.24 max = 114.57 avg = 114.47\n mobilenet_ssd min = 51.60 max = 51.89 avg = 51.77\n mobilenet_yolo min = 125.09 max = 126.33 avg = 125.83\n mobilenetv2_yolov3 min = 57.51 max = 57.79 avg = 57.65\n yolov4-tiny min = 85.65 max = 85.97 avg = 85.79\n```\n\n### NVIDIA Quadro RTX 8000 (TU102 SM x 72 + Tensor Core x 576)\n```\ni@s:~/qtang/ncnn/benchmark$ ../build-vulkan/benchmark/benchncnn 256 1 0 1 0\n[0 Quadro RTX 8000] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 Quadro RTX 8000] bugsbn1=0 bugcopc=0 bugihfa=0\n[0 Quadro RTX 8000] fp16p=1 fp16s=1 fp16a=1 int8s=1 int8a=1\n[0 Quadro RTX 8000] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\n[1 Quadro RTX 8000] queueC=2[8] queueG=0[16] queueT=1[2]\n[1 Quadro RTX 8000] bugsbn1=0 bugcopc=0 bugihfa=0\n[1 Quadro RTX 8000] fp16p=1 fp16s=1 fp16a=1 int8s=1 int8a=1\n[1 Quadro RTX 8000] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 256\nnum_threads = 1\npowersave = 0\ngpu_device = 1\ncooling_down = 0\n squeezenet min = 0.84 max = 1.39 avg = 0.93\n mobilenet min = 0.90 max = 2.30 avg = 0.91\n mobilenet_v2 min = 1.35 max = 9.59 avg = 1.46\n mobilenet_v3 min = 1.60 max = 77.94 avg = 2.12\n shufflenet min = 0.86 max = 2.27 avg = 0.88\n shufflenet_v2 min = 1.25 max = 1.47 avg = 1.27\n mnasnet min = 1.42 max = 20.77 avg = 1.72\n proxylessnasnet min = 1.48 max = 1.67 avg = 1.49\n efficientnet_b0 min = 2.56 max = 12.86 avg = 2.77\n regnety_400m min = 1.84 max = 14.98 avg = 2.42\n blazeface min = 0.64 max = 0.90 avg = 0.65\n googlenet min = 2.94 max = 76.82 avg = 3.45\n resnet18 min = 1.27 max = 10.56 avg = 1.56\n alexnet min = 1.53 max = 71.76 avg = 1.96\n vgg16 min = 4.90 max = 78.12 avg = 5.80\n resnet50 min = 3.00 max = 12.51 avg = 3.07\n squeezenet_ssd min = 5.60 max = 97.09 avg = 6.50\n mobilenet_ssd min = 2.40 max = 93.64 avg = 3.30\n mobilenet_yolo min = 2.96 max = 19.15 avg = 3.25\n mobilenetv2_yolov3 min = 4.52 max = 66.96 avg = 5.32\n yolov4-tiny min = 9.32 max = 72.92 avg = 14.01\n\n```\n\n### NVIDIA RTX3090 (GA102 SM x 82 + Tensor Core 328)\n```\n(base) i@t:~/wls/ncnn/benchmark$ ../build/benchmark/benchncnn 32 1 0 0 0\n[0 GeForce RTX 3090] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 GeForce RTX 3090] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 GeForce RTX 3090] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 GeForce RTX 3090] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\n[1 GeForce RTX 3090] queueC=2[8] queueG=0[16] queueT=1[2]\n[1 GeForce RTX 3090] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 GeForce RTX 3090] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 GeForce RTX 3090] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 32\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 1.76 max = 2.74 avg = 1.80\n squeezenet_int8 min = 47.10 max = 47.75 avg = 47.21\n mobilenet min = 4.77 max = 5.79 avg = 5.20\n mobilenet_int8 min = 64.19 max = 67.05 avg = 64.39\n mobilenet_v2 min = 2.44 max = 20.89 avg = 6.98\n mobilenet_v3 min = 2.75 max = 2.87 avg = 2.77\n shufflenet min = 2.20 max = 2.62 avg = 2.46\n shufflenet_v2 min = 5.10 max = 7.43 avg = 5.75\n mnasnet min = 3.47 max = 3.50 avg = 3.48\n proxylessnasnet min = 2.59 max = 9.08 avg = 7.28\n efficientnet_b0 min = 3.87 max = 4.65 avg = 3.91\n efficientnetv2_b0 min = 29.48 max = 41.90 avg = 30.14\n regnety_400m min = 2.89 max = 2.99 avg = 2.91\n blazeface min = 1.55 max = 2.14 avg = 1.60\n googlenet min = 4.33 max = 17.89 avg = 6.05\n googlenet_int8 min = 174.46 max = 178.19 avg = 174.74\n resnet18 min = 2.14 max = 11.04 avg = 5.33\n resnet18_int8 min = 193.37 max = 193.83 avg = 193.55\n alexnet min = 2.37 max = 15.99 avg = 4.50\n vgg16 min = 4.55 max = 16.65 avg = 5.22\n vgg16_int8 min = 1538.76 max = 1544.81 avg = 1540.79\n resnet50 min = 4.13 max = 25.86 avg = 5.80\n resnet50_int8 min = 400.89 max = 401.72 avg = 401.29\n squeezenet_ssd min = 6.95 max = 7.81 avg = 7.07\n squeezenet_ssd_int8 min = 158.51 max = 159.04 avg = 158.68\n mobilenet_ssd min = 4.36 max = 18.98 avg = 9.40\n mobilenet_ssd_int8 min = 130.74 max = 130.92 avg = 130.83\n mobilenet_yolo min = 3.96 max = 11.94 avg = 6.48\n mobilenetv2_yolov3 min = 6.07 max = 6.21 avg = 6.13\n yolov4-tiny min = 13.01 max = 26.78 avg = 14.87\n\nroot@3090:~/Desktop/ncnn-20221128/build/benchmark$ ./benchncnn 100 10 2 0 0\n[0 NVIDIA GeForce RTX 3090] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA GeForce RTX 3090] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA GeForce RTX 3090] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA GeForce RTX 3090] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 100\nnum_threads = 10\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 0.64 max = 0.66 avg = 0.65\n squeezenet_int8 min = 4.30 max = 4.93 avg = 4.45\n mobilenet min = 0.60 max = 1.85 avg = 1.32\n mobilenet_int8 min = 3.08 max = 3.17 avg = 3.12\n mobilenet_v2 min = 1.40 max = 1.46 avg = 1.42\n mobilenet_v3 min = 1.22 max = 6.10 avg = 3.02\n shufflenet min = 0.90 max = 0.97 avg = 0.92\n shufflenet_v2 min = 1.06 max = 1.13 avg = 1.09\n mnasnet min = 0.84 max = 0.98 avg = 0.91\n proxylessnasnet min = 0.99 max = 3.01 avg = 2.45\n efficientnet_b0 min = 2.11 max = 2.85 avg = 2.16\n efficientnetv2_b0 min = 7.46 max = 28.58 avg = 8.55\n regnety_400m min = 1.53 max = 1.75 avg = 1.59\n blazeface min = 0.59 max = 0.94 avg = 0.63\n googlenet min = 1.90 max = 12.22 avg = 2.63\n googlenet_int8 min = 17.45 max = 18.69 avg = 17.81\n resnet18 min = 0.90 max = 13.14 avg = 3.09\n resnet18_int8 min = 16.25 max = 17.34 avg = 16.50\n alexnet min = 0.86 max = 4.77 avg = 2.59\n vgg16 min = 1.38 max = 11.20 avg = 2.91\n vgg16_int8 min = 47.17 max = 49.02 avg = 47.57\n resnet50 min = 1.54 max = 2.16 avg = 1.64\n resnet50_int8 min = 22.90 max = 24.46 avg = 23.23\n squeezenet_ssd min = 2.25 max = 10.91 avg = 4.12\n squeezenet_ssd_int8 min = 11.98 max = 14.54 avg = 12.31\n mobilenet_ssd min = 1.46 max = 8.98 avg = 3.38\n mobilenet_ssd_int8 min = 6.13 max = 6.65 avg = 6.23\n mobilenet_yolo min = 1.29 max = 1.43 avg = 1.34\n mobilenetv2_yolov3 min = 3.64 max = 6.66 avg = 3.77\n yolov4-tiny min = 9.04 max = 11.65 avg = 9.54\n nanodet_m min = 1.43 max = 11.90 avg = 3.16\n yolo-fastest-1.1 min = 1.40 max = 1.82 avg = 1.57\n yolo-fastestv2 min = 1.36 max = 2.30 avg = 1.42\n vision_transformer min = 202.71 max = 244.47 avg = 218.69\n FastestDet min = 1.37 max = 5.37 avg = 2.77\n```\n\n### AMD Ryzen Embedded V1605B (Zen 2.0 GHz ~ 3.6 GHz x 4 + Radeon Vega 8 1.1GHz 8CU)\n```\nC:\\Users\\i\\Desktop\\benchmark>benchncnn.exe 32 1 0 -1 0\nloop_count = 32\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 22.13 max = 24.07 avg = 22.88\n squeezenet_int8 min = 58.54 max = 62.21 avg = 59.55\n mobilenet min = 40.99 max = 43.67 avg = 41.70\n mobilenet_int8 min = 98.06 max = 111.37 avg = 101.15\n mobilenet_v2 min = 26.53 max = 28.96 avg = 27.81\n mobilenet_v3 min = 22.96 max = 25.25 avg = 23.30\n shufflenet min = 20.17 max = 28.78 avg = 21.09\n shufflenet_v2 min = 19.06 max = 19.72 avg = 19.47\n mnasnet min = 25.11 max = 39.53 avg = 27.54\n proxylessnasnet min = 28.84 max = 35.16 avg = 30.03\n efficientnet_b0 min = 43.16 max = 46.03 avg = 43.65\n efficientnetv2_b0 min = 48.64 max = 52.07 avg = 49.62\n regnety_400m min = 33.43 max = 35.87 avg = 33.97\n blazeface min = 5.43 max = 6.04 avg = 5.56\n googlenet min = 85.80 max = 90.93 avg = 87.65\n googlenet_int8 min = 214.37 max = 230.75 avg = 219.50\n resnet18 min = 76.58 max = 80.38 avg = 77.34\n resnet18_int8 min = 231.16 max = 255.22 avg = 236.65\n alexnet min = 60.69 max = 64.06 avg = 61.34\n vgg16 min = 286.45 max = 307.04 avg = 290.86\n vgg16_int8 min = 1797.58 max = 2079.73 avg = 1844.78\n resnet50 min = 198.27 max = 215.03 avg = 201.37\n resnet50_int8 min = 493.52 max = 499.67 avg = 496.95\n squeezenet_ssd min = 189.97 max = 198.53 avg = 192.10\n squeezenet_ssd_int8 min = 198.81 max = 214.55 avg = 203.59\n mobilenet_ssd min = 87.56 max = 92.72 avg = 89.03\n mobilenet_ssd_int8 min = 196.97 max = 209.51 avg = 201.95\n mobilenet_yolo min = 206.87 max = 218.48 avg = 210.84\n mobilenetv2_yolov3 min = 102.72 max = 108.18 avg = 104.62\n yolov4-tiny min = 117.97 max = 134.73 avg = 121.26\n\nC:\\Users\\i\\Desktop\\benchmark>benchncnn.exe 32 2 0 -1 0\nloop_count = 32\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 13.43 max = 14.35 avg = 13.62\n squeezenet_int8 min = 32.29 max = 50.76 avg = 33.56\n mobilenet min = 23.42 max = 25.10 avg = 24.09\n mobilenet_int8 min = 51.99 max = 55.42 avg = 53.01\n mobilenet_v2 min = 15.45 max = 15.75 avg = 15.59\n mobilenet_v3 min = 14.32 max = 14.75 avg = 14.39\n shufflenet min = 12.64 max = 12.83 avg = 12.69\n shufflenet_v2 min = 11.45 max = 12.44 avg = 11.60\n mnasnet min = 14.43 max = 20.45 avg = 15.11\n proxylessnasnet min = 16.18 max = 16.38 avg = 16.24\n efficientnet_b0 min = 25.25 max = 28.42 avg = 26.59\n efficientnetv2_b0 min = 27.57 max = 32.05 avg = 30.04\n regnety_400m min = 22.74 max = 24.75 avg = 23.31\n blazeface min = 3.44 max = 3.83 avg = 3.62\n googlenet min = 49.39 max = 66.76 avg = 53.76\n googlenet_int8 min = 113.89 max = 136.75 avg = 119.29\n resnet18 min = 43.77 max = 67.24 avg = 46.14\n resnet18_int8 min = 121.44 max = 148.01 avg = 126.95\n alexnet min = 34.46 max = 37.38 avg = 35.50\n vgg16 min = 177.16 max = 207.25 avg = 184.19\n vgg16_int8 min = 951.86 max = 1155.60 avg = 990.51\n resnet50 min = 112.28 max = 137.18 avg = 115.64\n resnet50_int8 min = 260.69 max = 272.26 avg = 265.89\n squeezenet_ssd min = 108.07 max = 121.66 avg = 110.35\n squeezenet_ssd_int8 min = 109.01 max = 126.86 avg = 111.96\n mobilenet_ssd min = 49.60 max = 52.62 avg = 50.46\n mobilenet_ssd_int8 min = 104.22 max = 111.07 avg = 106.33\n mobilenet_yolo min = 117.42 max = 136.73 avg = 122.92\n mobilenetv2_yolov3 min = 61.66 max = 65.22 avg = 63.01\n yolov4-tiny min = 72.64 max = 77.09 avg = 74.30\n\nC:\\Users\\i\\Desktop\\benchmark>benchncnn.exe 32 4 0 -1 0\nloop_count = 32\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 9.19 max = 14.82 avg = 11.15\n squeezenet_int8 min = 19.00 max = 40.30 avg = 24.80\n mobilenet min = 18.02 max = 39.84 avg = 27.38\n mobilenet_int8 min = 28.04 max = 57.59 avg = 34.15\n mobilenet_v2 min = 10.26 max = 17.79 avg = 13.36\n mobilenet_v3 min = 8.87 max = 10.87 avg = 9.11\n shufflenet min = 8.93 max = 11.96 avg = 9.34\n shufflenet_v2 min = 7.37 max = 13.10 avg = 8.72\n mnasnet min = 9.24 max = 14.90 avg = 11.32\n proxylessnasnet min = 10.21 max = 11.89 avg = 10.39\n efficientnet_b0 min = 16.22 max = 23.71 avg = 16.59\n efficientnetv2_b0 min = 17.44 max = 31.42 avg = 22.85\n regnety_400m min = 18.32 max = 24.02 avg = 18.90\n blazeface min = 2.22 max = 2.81 avg = 2.30\n googlenet min = 31.52 max = 51.80 avg = 42.11\n googlenet_int8 min = 65.47 max = 114.41 avg = 75.98\n resnet18 min = 28.90 max = 64.62 avg = 37.58\n resnet18_int8 min = 71.29 max = 136.67 avg = 103.03\n alexnet min = 23.67 max = 34.01 avg = 29.78\n vgg16 min = 142.18 max = 211.00 avg = 170.46\n vgg16_int8 min = 531.36 max = 871.25 avg = 625.60\n resnet50 min = 69.23 max = 108.67 avg = 73.68\n resnet50_int8 min = 149.18 max = 309.88 avg = 168.68\n squeezenet_ssd min = 68.83 max = 81.70 avg = 71.01\n squeezenet_ssd_int8 min = 66.34 max = 118.16 avg = 74.34\n mobilenet_ssd min = 29.96 max = 34.32 avg = 30.74\n mobilenet_ssd_int8 min = 56.87 max = 92.24 avg = 65.57\n mobilenet_yolo min = 74.26 max = 113.91 avg = 81.28\n mobilenetv2_yolov3 min = 42.16 max = 63.49 avg = 45.34\n yolov4-tiny min = 53.06 max = 69.84 avg = 55.81\n\nC:\\Users\\i\\Desktop\\benchmark>benchncnn.exe 32 1 0 0 0\n[0 AMD Radeon(TM) Vega 8 Graphics] queueC=1[2] queueG=0[1] queueT=2[1]\n[0 AMD Radeon(TM) Vega 8 Graphics] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 AMD Radeon(TM) Vega 8 Graphics] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 AMD Radeon(TM) Vega 8 Graphics] subgroup=64 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 32\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 6.78 max = 7.09 avg = 6.91\n squeezenet_int8 min = 58.93 max = 62.53 avg = 60.11\n mobilenet min = 8.08 max = 8.39 avg = 8.25\n mobilenet_int8 min = 97.74 max = 116.77 avg = 100.17\n mobilenet_v2 min = 7.95 max = 8.27 avg = 8.14\n mobilenet_v3 min = 8.70 max = 9.70 avg = 9.02\n shufflenet min = 6.36 max = 7.64 avg = 7.01\n shufflenet_v2 min = 7.04 max = 8.12 avg = 7.50\n mnasnet min = 8.07 max = 9.08 avg = 8.38\n proxylessnasnet min = 8.56 max = 9.66 avg = 8.81\n efficientnet_b0 min = 16.68 max = 18.00 avg = 17.30\n efficientnetv2_b0 min = 394.82 max = 404.88 avg = 401.05\n regnety_400m min = 11.92 max = 12.17 avg = 12.03\n blazeface min = 4.82 max = 6.50 avg = 5.42\n googlenet min = 18.44 max = 19.66 avg = 19.18\n googlenet_int8 min = 213.41 max = 231.79 avg = 218.31\n resnet18 min = 14.27 max = 14.72 avg = 14.44\n resnet18_int8 min = 228.79 max = 249.65 avg = 236.06\n alexnet min = 17.31 max = 18.31 avg = 17.69\n vgg16 min = 111.85 max = 123.35 avg = 112.98\n vgg16_int8 min = 1789.64 max = 1838.84 avg = 1826.05\n resnet50 min = 31.61 max = 32.86 avg = 32.12\n resnet50_int8 min = 483.57 max = 505.72 avg = 491.76\n squeezenet_ssd min = 99.66 max = 105.68 avg = 104.57\n squeezenet_ssd_int8 min = 200.48 max = 208.71 avg = 203.02\n mobilenet_ssd min = 33.45 max = 35.64 avg = 34.75\n mobilenet_ssd_int8 min = 195.14 max = 205.35 avg = 200.18\n mobilenet_yolo min = 59.20 max = 61.06 avg = 60.47\n mobilenetv2_yolov3 min = 31.48 max = 33.25 avg = 32.84\n yolov4-tiny min = 93.75 max = 97.45 avg = 96.00\n```\n\n### Qualcomm SM8150-AC Snapdragon 855+ (Kyro485 2.96 GHz + 2.42 GHz x 3 + 1.80 GHz x 4 + Adreno 640)\n```\nOnePlus7T:/data/local/tmp # ./benchncnn 8 4 2 -1 1 \nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 3.60 max = 3.70 avg = 3.64\n squeezenet_int8 min = 3.67 max = 3.78 avg = 3.71\n mobilenet min = 5.32 max = 5.42 avg = 5.38\n mobilenet_int8 min = 4.20 max = 4.28 avg = 4.23\n mobilenet_v2 min = 4.64 max = 4.73 avg = 4.68\n mobilenet_v3 min = 4.13 max = 4.25 avg = 4.18\n shufflenet min = 3.29 max = 3.40 avg = 3.33\n shufflenet_v2 min = 2.98 max = 3.07 avg = 3.01\n mnasnet min = 4.26 max = 4.37 avg = 4.31\n proxylessnasnet min = 4.67 max = 4.78 avg = 4.72\n efficientnet_b0 min = 7.23 max = 7.34 avg = 7.30\n efficientnetv2_b0 min = 8.74 max = 8.87 avg = 8.81\n regnety_400m min = 7.88 max = 7.99 avg = 7.95\n blazeface min = 1.19 max = 1.30 avg = 1.22\n googlenet min = 13.07 max = 13.20 avg = 13.12\n googlenet_int8 min = 12.86 max = 12.98 avg = 12.93\n resnet18 min = 10.33 max = 10.36 avg = 10.35\n resnet18_int8 min = 9.42 max = 9.45 avg = 9.43\n alexnet min = 11.88 max = 11.95 avg = 11.91\n vgg16 min = 59.34 max = 60.69 avg = 60.19\n vgg16_int8 min = 68.78 max = 69.07 avg = 68.93\n resnet50 min = 26.18 max = 26.28 avg = 26.24\n resnet50_int8 min = 20.86 max = 20.95 avg = 20.91\n squeezenet_ssd min = 12.00 max = 12.76 avg = 12.19\n squeezenet_ssd_int8 min = 11.67 max = 13.13 avg = 12.03\n mobilenet_ssd min = 11.88 max = 12.68 avg = 12.03\n mobilenet_ssd_int8 min = 9.28 max = 9.68 avg = 9.35\n mobilenet_yolo min = 27.89 max = 28.06 avg = 27.96\n mobilenetv2_yolov3 min = 18.00 max = 18.13 avg = 18.06\n yolov4-tiny min = 25.25 max = 25.36 avg = 25.29\n nanodet_m min = 8.93 max = 9.00 avg = 8.96\n yolo-fastest-1.1 min = 3.73 max = 3.83 avg = 3.77\n yolo-fastestv2 min = 3.38 max = 3.47 avg = 3.41\n vision_transformer min = 567.94 max = 572.31 avg = 569.66\n FastestDet min = 3.28 max = 3.37 avg = 3.32\n\nOnePlus7T:/data/local/tmp # ./benchncnn 8 1 2 -1 1 \nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 8.24 max = 8.34 avg = 8.31\n squeezenet_int8 min = 8.23 max = 8.34 avg = 8.30\n mobilenet min = 14.38 max = 14.56 avg = 14.45\n mobilenet_int8 min = 11.12 max = 11.24 avg = 11.17\n mobilenet_v2 min = 9.82 max = 9.88 avg = 9.84\n mobilenet_v3 min = 8.15 max = 8.24 avg = 8.21\n shufflenet min = 5.32 max = 5.44 avg = 5.37\n shufflenet_v2 min = 5.38 max = 5.51 avg = 5.44\n mnasnet min = 9.25 max = 9.36 avg = 9.31\n proxylessnasnet min = 10.95 max = 11.01 avg = 10.98\n efficientnet_b0 min = 17.67 max = 17.79 avg = 17.73\n efficientnetv2_b0 min = 20.56 max = 20.70 avg = 20.60\n regnety_400m min = 11.96 max = 12.07 avg = 12.00\n blazeface min = 2.19 max = 2.87 avg = 2.47\n googlenet min = 32.10 max = 32.20 avg = 32.15\n googlenet_int8 min = 32.00 max = 32.15 avg = 32.07\n resnet18 min = 22.02 max = 22.28 avg = 22.12\n resnet18_int8 min = 26.17 max = 26.26 avg = 26.22\n alexnet min = 24.83 max = 24.99 avg = 24.92\n vgg16 min = 129.57 max = 129.95 avg = 129.78\n vgg16_int8 min = 202.08 max = 202.34 avg = 202.19\n resnet50 min = 65.85 max = 66.01 avg = 65.93\n resnet50_int8 min = 56.33 max = 56.49 avg = 56.42\n squeezenet_ssd min = 22.52 max = 24.50 avg = 22.93\n squeezenet_ssd_int8 min = 24.51 max = 26.83 avg = 24.98\n mobilenet_ssd min = 30.55 max = 32.68 avg = 30.85\n mobilenet_ssd_int8 min = 22.96 max = 23.75 avg = 23.09\n mobilenet_yolo min = 68.74 max = 69.01 avg = 68.88\n mobilenetv2_yolov3 min = 36.98 max = 37.16 avg = 37.06\n yolov4-tiny min = 47.36 max = 47.45 avg = 47.41\n nanodet_m min = 15.08 max = 15.30 avg = 15.17\n yolo-fastest-1.1 min = 5.51 max = 5.61 avg = 5.55\n yolo-fastestv2 min = 4.92 max = 5.02 avg = 4.97\n vision_transformer min = 990.13 max = 994.45 avg = 991.95\n FastestDet min = 5.06 max = 5.17 avg = 5.11\n\nOnePlus7T:/data/local/tmp $ ./benchncnn 8 1 2 0 1\n[0 Adreno (TM) 640] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 640] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=1\n[0 Adreno (TM) 640] fp16-p/s/a=1/0/1 int8-p/s/a=1/0/0\n[0 Adreno (TM) 640] subgroup=64 basic=1 vote=1 ballot=0 shuffle=0\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 8.59 max = 9.51 avg = 9.09\n mobilenet min = 13.04 max = 13.45 avg = 13.22\n mobilenet_v2 min = 10.68 max = 11.38 avg = 10.85\n mobilenet_v3 min = 11.86 max = 12.37 avg = 12.08\n shufflenet min = 8.21 max = 8.40 avg = 8.25\n shufflenet_v2 min = 8.84 max = 9.13 avg = 8.97\n mnasnet min = 11.32 max = 11.72 avg = 11.45\n proxylessnasnet min = 12.27 max = 12.86 avg = 12.55\n efficientnet_b0 min = 22.64 max = 22.82 avg = 22.75\n efficientnetv2_b0 min = 32.32 max = 38.20 avg = 35.79\n regnety_400m min = 15.35 max = 15.86 avg = 15.64\n blazeface min = 2.82 max = 2.93 avg = 2.86\n googlenet min = 28.22 max = 28.34 avg = 28.26\n resnet18 min = 24.71 max = 24.96 avg = 24.82\n alexnet min = 27.94 max = 28.10 avg = 28.01\n vgg16 min = 106.08 max = 106.53 avg = 106.30\n resnet50 min = 55.28 max = 56.03 avg = 55.68\n squeezenet_ssd min = 29.77 max = 30.65 avg = 30.05\n mobilenet_ssd min = 29.14 max = 29.39 avg = 29.25\n mobilenet_yolo min = 49.78 max = 50.09 avg = 49.94\n mobilenetv2_yolov3 min = 31.11 max = 31.97 avg = 31.60\n yolov4-tiny min = 46.22 max = 46.90 avg = 46.63\n nanodet_m min = 15.96 max = 16.52 avg = 16.13\n yolo-fastest-1.1 min = 9.59 max = 9.66 avg = 9.61\n yolo-fastestv2 min = 7.99 max = 8.23 avg = 8.13\n```\n\n### Qualcomm MSM6150 Snapdragon 675 (Kyro460 2.0GHz x 2 + Kyro460 1.7GHz x 6 + Adreno 612)\n```\nviolet:/data/local/tmp/ncnn $ ./benchncnn 8 2 0\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = -1\n squeezenet min = 23.29 max = 24.65 avg = 23.95\n squeezenet_int8 min = 23.24 max = 61.55 avg = 31.20\n mobilenet min = 31.60 max = 32.10 avg = 31.80\n mobilenet_int8 min = 30.35 max = 32.03 avg = 30.95\n mobilenet_v2 min = 25.92 max = 26.45 avg = 26.08\n shufflenet min = 11.91 max = 12.11 avg = 12.00\n mnasnet min = 21.38 max = 21.71 avg = 21.51\n proxylessnasnet min = 25.53 max = 25.78 avg = 25.62\n googlenet min = 93.62 max = 100.67 avg = 94.86\n googlenet_int8 min = 90.74 max = 91.06 avg = 90.87\n resnet18 min = 85.84 max = 87.37 avg = 86.50\n resnet18_int8 min = 77.88 max = 78.11 avg = 78.00\n alexnet min = 196.33 max = 201.73 avg = 200.19\n vgg16 min = 560.71 max = 571.75 avg = 564.84\n vgg16_int8 min = 651.51 max = 652.68 avg = 652.12\n resnet50 min = 178.25 max = 179.86 avg = 178.77\n resnet50_int8 min = 181.07 max = 183.26 avg = 181.64\n squeezenet_ssd min = 64.86 max = 68.39 avg = 66.05\n squeezenet_ssd_int8 min = 69.61 max = 70.37 avg = 69.93\n mobilenet_ssd min = 65.92 max = 67.03 avg = 66.41\n mobilenet_ssd_int8 min = 61.54 max = 63.38 avg = 62.27\n mobilenet_yolo min = 143.42 max = 146.69 avg = 144.33\n mobilenet_yolov3 min = 150.45 max = 152.30 avg = 151.36\n\nviolet:/data/local/tmp/ncnn $ ./benchncnn 8 1 0\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = -1\n squeezenet min = 36.04 max = 37.25 avg = 36.48\n squeezenet_int8 min = 37.82 max = 79.20 avg = 43.13\n mobilenet min = 54.29 max = 54.73 avg = 54.41\n mobilenet_int8 min = 58.90 max = 60.11 avg = 59.39\n mobilenet_v2 min = 38.64 max = 40.22 avg = 38.97\n shufflenet min = 18.05 max = 18.39 avg = 18.19\n mnasnet min = 34.65 max = 34.98 avg = 34.79\n proxylessnasnet min = 42.61 max = 43.12 avg = 42.80\n googlenet min = 164.74 max = 165.89 avg = 165.34\n googlenet_int8 min = 159.93 max = 160.38 avg = 160.12\n resnet18 min = 135.76 max = 137.93 avg = 136.98\n resnet18_int8 min = 140.22 max = 144.06 avg = 141.92\n alexnet min = 391.01 max = 396.85 avg = 392.74\n vgg16 min = 1019.35 max = 1022.75 avg = 1021.26\n vgg16_int8 min = 1122.25 max = 1137.99 avg = 1124.78\n resnet50 min = 302.16 max = 304.22 avg = 303.05\n resnet50_int8 min = 318.35 max = 319.50 avg = 318.84\n squeezenet_ssd min = 91.26 max = 94.86 avg = 92.39\n squeezenet_ssd_int8 min = 105.06 max = 106.17 avg = 105.56\n mobilenet_ssd min = 105.01 max = 105.95 avg = 105.40\n mobilenet_ssd_int8 min = 119.93 max = 120.50 avg = 120.19\n mobilenet_yolo min = 229.87 max = 230.76 avg = 230.21\n mobilenet_yolov3 min = 242.10 max = 242.91 avg = 242.47\n```\n\n### Kirin 970 (Cortex-A73 2.4GHz x 4 + Cortex-A53 1.8GHz x 4)\n```\nHWEML:/data/local/tmp/ncnnbench $ ./benchncnn 8 4 2 -1 1\n[0 Mali-G72] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G72] buglssc=0 bugsbn1=0 buglbia=0 bugihfa=1\n[0 Mali-G72] fp16p=1 fp16s=0 fp16a=1 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 24.38 max = 28.03 avg = 25.83\n squeezenet_int8 min = 21.79 max = 24.80 avg = 22.60\n mobilenet min = 34.09 max = 36.88 avg = 35.93\n mobilenet_int8 min = 52.62 max = 61.70 avg = 55.38\n mobilenet_v2 min = 23.71 max = 25.70 avg = 24.49\n mobilenet_v3 min = 20.66 max = 25.68 avg = 23.07\n shufflenet min = 17.89 max = 19.91 avg = 18.53\n shufflenet_v2 min = 13.73 max = 16.54 avg = 15.37\n mnasnet min = 24.36 max = 27.14 avg = 25.58\n proxylessnasnet min = 27.19 max = 29.70 avg = 28.59\n efficientnet_b0 min = 49.31 max = 50.26 avg = 49.70\n regnety_400m min = 42.54 max = 51.22 avg = 46.71\n blazeface min = 5.49 max = 7.67 avg = 6.27\n googlenet min = 72.67 max = 81.22 avg = 75.92\n googlenet_int8 min = 67.60 max = 74.50 avg = 71.21\n resnet18 min = 69.32 max = 81.59 avg = 73.45\n resnet18_int8 min = 60.92 max = 68.11 avg = 64.18\n alexnet min = 60.90 max = 79.28 avg = 66.72\n vgg16 min = 337.01 max = 378.89 avg = 352.37\n vgg16_int8 min = 465.88 max = 505.19 avg = 489.76\n resnet50 min = 207.75 max = 220.74 avg = 214.42\n resnet50_int8 min = 165.67 max = 183.80 avg = 171.27\n squeezenet_ssd min = 72.77 max = 84.45 avg = 79.09\n squeezenet_ssd_int8 min = 75.37 max = 86.58 avg = 78.70\n mobilenet_ssd min = 88.88 max = 96.43 avg = 92.02\n mobilenet_ssd_int8 min = 89.04 max = 101.35 avg = 92.23\n mobilenet_yolo min = 189.73 max = 206.55 avg = 193.64\n mobilenetv2_yolov3 min = 99.08 max = 111.64 avg = 104.23\n\nHWEML:/data/local/tmp/ncnnbench $ ./benchncnn 8 1 2 -1 1\n[0 Mali-G72] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G72] buglssc=0 bugsbn1=0 buglbia=0 bugihfa=1\n[0 Mali-G72] fp16p=1 fp16s=0 fp16a=1 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 73.47 max = 81.39 avg = 76.06\n squeezenet_int8 min = 62.63 max = 73.66 avg = 66.52\n mobilenet min = 103.85 max = 112.83 avg = 108.98\n mobilenet_int8 min = 152.27 max = 161.26 avg = 157.17\n mobilenet_v2 min = 70.53 max = 87.26 avg = 76.67\n mobilenet_v3 min = 59.87 max = 68.59 avg = 63.08\n shufflenet min = 36.69 max = 41.45 avg = 39.24\n shufflenet_v2 min = 33.97 max = 37.84 avg = 35.03\n mnasnet min = 69.24 max = 79.73 avg = 74.20\n proxylessnasnet min = 78.63 max = 88.57 avg = 81.83\n efficientnet_b0 min = 147.45 max = 159.07 avg = 152.09\n regnety_400m min = 90.83 max = 98.51 avg = 93.82\n blazeface min = 10.05 max = 11.59 avg = 10.78\n googlenet min = 240.26 max = 277.71 avg = 259.61\n googlenet_int8 min = 214.64 max = 233.56 avg = 225.01\n resnet18 min = 245.62 max = 268.49 avg = 260.37\n resnet18_int8 min = 184.85 max = 194.91 avg = 190.60\n alexnet min = 202.52 max = 241.12 avg = 211.51\n vgg16 min = 1632.98 max = 1769.05 avg = 1710.89\n vgg16_int8 min = 1237.01 max = 1316.40 avg = 1273.44\n resnet50 min = 558.41 max = 601.59 avg = 581.26\n resnet50_int8 min = 425.26 max = 445.19 avg = 436.22\n squeezenet_ssd min = 228.50 max = 255.89 avg = 244.63\n squeezenet_ssd_int8 min = 166.97 max = 193.77 avg = 180.22\n mobilenet_ssd min = 226.54 max = 246.62 avg = 235.75\n mobilenet_ssd_int8 min = 231.35 max = 249.63 avg = 241.29\n mobilenet_yolo min = 469.71 max = 508.79 avg = 497.50\n mobilenetv2_yolov3 min = 242.88 max = 265.30 avg = 254.68\n\nHWEML:/data/local/tmp/ncnnbench $ ./benchncnn 4 1 2 0 1\n[0 Mali-G72] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G72] buglssc=0 bugsbn1=0 buglbia=0 bugihfa=1\n[0 Mali-G72] fp16p=1 fp16s=0 fp16a=1 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 24.54 max = 25.75 avg = 25.16\n mobilenet min = 22.03 max = 29.61 avg = 27.31\n mobilenet_v2 min = 20.15 max = 28.05 avg = 25.35\n mobilenet_v3 min = 34.26 max = 37.49 avg = 35.51\n shufflenet min = 26.29 max = 27.68 avg = 26.86\n shufflenet_v2 min = 29.60 max = 32.08 avg = 31.27\n mnasnet min = 25.85 max = 29.38 avg = 27.98\n proxylessnasnet min = 23.64 max = 30.09 avg = 26.36\n efficientnet_b0 min = 52.55 max = 58.51 avg = 55.56\n regnety_400m min = 37.81 max = 43.22 avg = 40.30\n blazeface min = 9.14 max = 10.93 avg = 10.08\n googlenet min = 60.19 max = 62.84 avg = 61.51\n resnet18 min = 50.42 max = 52.93 avg = 51.70\n alexnet min = 195.34 max = 196.98 avg = 196.14\n vgg16 min = 725.88 max = 751.20 avg = 739.99\n resnet50 min = 124.47 max = 125.93 avg = 125.02\n squeezenet_ssd min = 91.79 max = 97.04 avg = 93.56\n mobilenet_ssd min = 51.81 max = 59.31 avg = 54.09\n mobilenet_yolo min = 124.67 max = 127.62 avg = 126.53\n mobilenetv2_yolov3 min = 53.11 max = 54.81 avg = 54.11\n```\n\n### Qualcomm MSM8998 Snapdragon 835 (Kyro 2.45GHz x 4 + Kyro 1.9GHz x 4 + Adreno 540)\n```\ntaimen:/data/local/tmp/ncnnbench $ ./benchncnn 8 4 2 -1 0\n[0 Adreno (TM) 540] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 540] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 540] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 28.46 max = 30.89 avg = 29.77\n squeezenet_int8 min = 30.32 max = 32.92 avg = 31.68\n mobilenet min = 36.65 max = 38.37 avg = 37.32\n mobilenet_int8 min = 62.91 max = 66.71 avg = 64.49\n mobilenet_v2 min = 27.85 max = 31.21 avg = 29.41\n mobilenet_v3 min = 23.83 max = 26.40 avg = 24.79\n shufflenet min = 15.65 max = 16.88 avg = 16.27\n shufflenet_v2 min = 13.70 max = 14.49 avg = 14.08\n mnasnet min = 25.04 max = 28.35 avg = 26.45\n proxylessnasnet min = 27.49 max = 29.58 avg = 28.62\n efficientnet_b0 min = 48.43 max = 49.41 avg = 48.85\n regnety_400m min = 42.48 max = 43.78 avg = 43.18\n blazeface min = 4.39 max = 4.68 avg = 4.51\n googlenet min = 75.98 max = 78.40 avg = 77.37\n googlenet_int8 min = 79.26 max = 83.20 avg = 80.55\n resnet18 min = 73.60 max = 76.97 avg = 75.63\n resnet18_int8 min = 62.93 max = 65.94 avg = 64.50\n alexnet min = 64.18 max = 67.02 avg = 65.49\n vgg16 min = 389.39 max = 399.13 avg = 394.09\n vgg16_int8 min = 509.06 max = 524.41 avg = 514.76\n resnet50 min = 188.21 max = 194.58 avg = 191.98\n resnet50_int8 min = 182.84 max = 187.22 avg = 184.23\n squeezenet_ssd min = 77.69 max = 81.17 avg = 79.24\n squeezenet_ssd_int8 min = 81.71 max = 84.12 avg = 82.90\n mobilenet_ssd min = 78.35 max = 81.50 avg = 79.82\n mobilenet_ssd_int8 min = 96.84 max = 100.97 avg = 98.42\n mobilenet_yolo min = 167.32 max = 170.71 avg = 168.87\n mobilenetv2_yolov3 min = 97.00 max = 102.11 avg = 99.01\n\ntaimen:/data/local/tmp/ncnnbench $ ./benchncnn 8 1 2 -1 1\n[0 Adreno (TM) 540] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 540] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 540] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 67.25 max = 71.39 avg = 69.35\n squeezenet_int8 min = 62.12 max = 66.35 avg = 63.73\n mobilenet min = 103.30 max = 110.39 avg = 107.13\n mobilenet_int8 min = 155.24 max = 161.42 avg = 157.82\n mobilenet_v2 min = 71.89 max = 74.73 avg = 73.48\n mobilenet_v3 min = 58.35 max = 63.43 avg = 60.68\n shufflenet min = 35.96 max = 39.43 avg = 36.94\n shufflenet_v2 min = 35.53 max = 39.86 avg = 37.10\n mnasnet min = 66.71 max = 74.00 avg = 68.65\n proxylessnasnet min = 76.50 max = 82.20 avg = 78.57\n efficientnet_b0 min = 142.32 max = 152.17 avg = 146.14\n regnety_400m min = 89.60 max = 98.27 avg = 92.62\n blazeface min = 10.45 max = 12.81 avg = 11.07\n googlenet min = 222.75 max = 233.61 avg = 228.38\n googlenet_int8 min = 206.70 max = 212.20 avg = 209.24\n resnet18 min = 210.86 max = 220.25 avg = 213.65\n resnet18_int8 min = 176.04 max = 183.58 avg = 178.71\n alexnet min = 185.97 max = 195.91 avg = 191.40\n vgg16 min = 1176.82 max = 1200.64 avg = 1187.88\n vgg16_int8 min = 1086.52 max = 1105.00 avg = 1095.53\n resnet50 min = 517.48 max = 533.99 avg = 526.04\n resnet50_int8 min = 417.30 max = 435.81 avg = 422.36\n squeezenet_ssd min = 164.88 max = 171.21 avg = 167.51\n squeezenet_ssd_int8 min = 164.78 max = 171.77 avg = 168.36\n mobilenet_ssd min = 221.41 max = 229.13 avg = 226.18\n mobilenet_ssd_int8 min = 234.15 max = 245.91 avg = 239.01\n mobilenet_yolo min = 471.34 max = 484.99 avg = 477.15\n mobilenetv2_yolov3 min = 249.14 max = 257.61 avg = 252.54\n\ntaimen:/data/local/tmp/ncnnbench $ ./benchncnn 8 1 2 0 1\n[0 Adreno (TM) 540] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 540] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 540] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 18.74 max = 19.89 avg = 19.22\n mobilenet min = 21.19 max = 25.61 avg = 22.94\n mobilenet_v2 min = 24.15 max = 34.68 avg = 30.12\n mobilenet_v3 min = 25.94 max = 33.15 avg = 30.09\n shufflenet min = 25.05 max = 31.41 avg = 27.85\n shufflenet_v2 min = 28.82 max = 32.04 avg = 30.95\n mnasnet min = 21.34 max = 27.69 avg = 24.17\n proxylessnasnet min = 25.51 max = 30.03 avg = 28.01\n efficientnet_b0 min = 42.94 max = 47.44 avg = 45.28\n regnety_400m min = 36.36 max = 55.73 avg = 41.82\n blazeface min = 11.14 max = 13.11 avg = 12.20\n googlenet min = 49.72 max = 56.92 avg = 51.79\n resnet18 min = 44.63 max = 47.37 avg = 45.86\n alexnet min = 42.83 max = 46.34 avg = 44.63\n vgg16 min = 568.82 max = 586.75 avg = 578.60\n resnet50 min = 108.63 max = 115.76 avg = 110.38\n squeezenet_ssd min = 85.22 max = 104.73 avg = 93.14\n mobilenet_ssd min = 49.91 max = 56.86 avg = 52.33\n mobilenet_yolo min = 98.76 max = 109.37 avg = 102.27\n mobilenetv2_yolov3 min = 57.49 max = 61.15 avg = 58.74\n```\n\n### Qualcomm SDM765G Snapdragon 765G (Kyro 1.8GHz x 6 + Kyro 2.2GHz x 2 + Adreno 620)\n```\n130|bramble:/data/local/tmp $ ./benchncnn 8 4 2 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 9.84 max = 11.72 avg = 10.36\n squeezenet_int8 min = 10.80 max = 11.13 avg = 10.96\n mobilenet min = 14.04 max = 14.37 avg = 14.20\n mobilenet_int8 min = 13.39 max = 13.75 avg = 13.59\n mobilenet_v2 min = 13.04 max = 13.51 avg = 13.27\n mobilenet_v3 min = 11.00 max = 13.21 avg = 12.54\n shufflenet min = 11.08 max = 11.22 avg = 11.16\n shufflenet_v2 min = 8.45 max = 8.50 avg = 8.47\n mnasnet min = 14.15 max = 14.69 avg = 14.38\n proxylessnasnet min = 14.49 max = 15.07 avg = 14.83\n efficientnet_b0 min = 28.99 max = 29.53 avg = 29.24\n efficientnetv2_b0 min = 38.92 max = 39.34 avg = 39.14\n regnety_400m min = 33.46 max = 33.81 avg = 33.62\n blazeface min = 4.22 max = 4.30 avg = 4.27\n googlenet min = 35.24 max = 36.94 avg = 35.57\n googlenet_int8 min = 45.26 max = 46.46 avg = 45.78\n resnet18 min = 33.14 max = 33.75 avg = 33.31\n resnet18_int8 min = 43.26 max = 43.50 avg = 43.35\n alexnet min = 25.40 max = 26.19 avg = 25.74\n vgg16 min = 121.39 max = 122.35 avg = 121.78\n vgg16_int8 min = 243.47 max = 249.94 avg = 245.56\n resnet50 min = 67.05 max = 70.16 avg = 68.20\n resnet50_int8 min = 76.95 max = 80.23 avg = 78.18\n squeezenet_ssd min = 32.02 max = 33.27 avg = 32.51\n squeezenet_ssd_int8 min = 36.31 max = 38.35 avg = 37.09\n mobilenet_ssd min = 32.02 max = 34.55 avg = 32.99\n mobilenet_ssd_int8 min = 32.31 max = 33.92 avg = 32.77\n mobilenet_yolo min = 99.12 max = 109.81 avg = 103.00\n mobilenetv2_yolov3 min = 59.74 max = 60.95 avg = 60.21\n yolov4-tiny min = 57.83 max = 72.15 avg = 68.75\n nanodet_m min = 22.76 max = 22.97 avg = 22.85\n yolo-fastest-1.1 min = 13.58 max = 13.93 avg = 13.80\n yolo-fastestv2 min = 12.06 max = 12.27 avg = 12.15\n vision_transformer min = 1274.67 max = 1597.52 avg = 1363.14\n FastestDet min = 9.75 max = 9.86 avg = 9.81\n\n130|bramble:/data/local/tmp $ ./benchncnn 8 4 2 0 1\n[0 Adreno (TM) 620] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 620] bugsbn1=1 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Adreno (TM) 620] fp16-p/s/u/a=1/1/0/1 int8-p/s/u/a=1/0/0/1\n[0 Adreno (TM) 620] subgroup=64 basic/vote/ballot/shuffle=1/1/1/1\n[0 Adreno (TM) 620] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 25.06 max = 25.80 avg = 25.53\n squeezenet_int8 min = 9.75 max = 9.82 avg = 9.78\n mobilenet min = 43.43 max = 44.04 avg = 43.71\n mobilenet_int8 min = 11.12 max = 11.59 avg = 11.34\n mobilenet_v2 min = 32.14 max = 32.58 avg = 32.40\n mobilenet_v3 min = 32.75 max = 32.98 avg = 32.87\n shufflenet min = 29.29 max = 29.63 avg = 29.40\n shufflenet_v2 min = 32.43 max = 33.18 avg = 32.69\n mnasnet min = 34.58 max = 35.24 avg = 35.00\n proxylessnasnet min = 40.61 max = 41.40 avg = 40.98\n efficientnet_b0 min = 49.44 max = 50.46 avg = 49.95\n efficientnetv2_b0 min = 185.31 max = 187.37 avg = 186.24\n regnety_400m min = 41.43 max = 42.75 avg = 41.84\n blazeface min = 13.47 max = 14.07 avg = 13.72\n googlenet min = 78.12 max = 79.06 avg = 78.56\n googlenet_int8 min = 48.73 max = 50.13 avg = 49.20\n resnet18 min = 73.61 max = 74.05 avg = 73.75\n resnet18_int8 min = 21.87 max = 22.05 avg = 21.95\n alexnet min = 128.58 max = 129.51 avg = 128.97\n vgg16 min = 437.64 max = 439.12 avg = 438.28\n vgg16_int8 min = 232.77 max = 243.06 avg = 239.54\n resnet50 min = 187.36 max = 188.47 avg = 188.01\n resnet50_int8 min = 75.79 max = 77.33 avg = 76.64\n squeezenet_ssd min = 80.68 max = 84.50 avg = 81.93\n squeezenet_ssd_int8 min = 29.88 max = 30.77 avg = 30.30\n mobilenet_ssd min = 94.77 max = 96.46 avg = 95.79\n mobilenet_ssd_int8 min = 29.03 max = 30.07 avg = 29.53\n mobilenet_yolo min = 185.97 max = 188.11 avg = 186.59\n mobilenetv2_yolov3 min = 108.43 max = 164.75 avg = 121.55\n yolov4-tiny min = 149.38 max = 158.39 avg = 153.92\n nanodet_m min = 46.73 max = 48.85 avg = 47.73\n yolo-fastest-1.1 min = 26.32 max = 26.77 avg = 26.54\n yolo-fastestv2 min = 38.87 max = 39.31 avg = 39.13\n vision_transformer min = 3392.80 max = 3397.79 avg = 3396.09\n FastestDet min = 43.05 max = 43.81 avg = 43.45\n```\n\n### Qualcomm SDM660 Snapdragon 660 (Kyro260 2.2GHz x 4 + Kyro260 1.84GHz x 4 + Adreno 512)\n```\nlavender:/data/local/tmp/ncnnbench $ ./benchncnn 8 8 0 -1 1\n[0 Adreno (TM) 512] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 512] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 512] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 29.05 max = 44.86 avg = 33.26\n squeezenet_int8 min = 35.47 max = 37.10 avg = 36.09\n mobilenet min = 31.59 max = 33.47 avg = 32.33\n mobilenet_int8 min = 77.50 max = 91.15 avg = 82.98\n mobilenet_v2 min = 33.63 max = 35.43 avg = 34.54\n mobilenet_v3 min = 29.97 max = 49.80 avg = 34.81\n shufflenet min = 28.52 max = 30.09 avg = 29.09\n shufflenet_v2 min = 19.15 max = 21.15 avg = 19.99\n mnasnet min = 29.91 max = 35.11 avg = 31.46\n proxylessnasnet min = 33.28 max = 117.09 avg = 55.22\n efficientnet_b0 min = 52.29 max = 57.93 avg = 55.04\n regnety_400m min = 96.05 max = 116.42 avg = 102.07\n blazeface min = 7.98 max = 11.83 avg = 8.89\n googlenet min = 76.88 max = 103.99 avg = 84.54\n googlenet_int8 min = 97.68 max = 118.56 avg = 104.92\n resnet18 min = 75.93 max = 89.31 avg = 80.00\n resnet18_int8 min = 73.27 max = 80.84 avg = 76.19\n alexnet min = 90.94 max = 114.57 avg = 96.42\n vgg16 min = 381.30 max = 615.62 avg = 555.96\n vgg16_int8 min = 803.75 max = 1126.53 avg = 886.03\n resnet50 min = 257.38 max = 285.19 avg = 266.59\n resnet50_int8 min = 304.81 max = 338.01 avg = 314.84\n squeezenet_ssd min = 117.59 max = 145.79 avg = 123.79\n squeezenet_ssd_int8 min = 132.80 max = 163.00 avg = 149.99\n mobilenet_ssd min = 103.98 max = 126.90 avg = 113.10\n mobilenet_ssd_int8 min = 167.86 max = 188.46 avg = 180.56\n mobilenet_yolo min = 201.75 max = 263.92 avg = 240.17\n mobilenetv2_yolov3 min = 143.76 max = 167.77 avg = 151.94\n\nlavender:/data/local/tmp/ncnnbench $ ./benchncnn 4 1 2 -1 1\n[0 Adreno (TM) 512] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 512] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 512] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 69.75 max = 71.33 avg = 70.38\n squeezenet_int8 min = 67.12 max = 68.07 avg = 67.59\n mobilenet min = 107.65 max = 110.48 avg = 108.82\n mobilenet_int8 min = 163.13 max = 164.74 avg = 164.24\n mobilenet_v2 min = 75.50 max = 77.36 avg = 76.38\n mobilenet_v3 min = 59.05 max = 59.36 avg = 59.23\n shufflenet min = 38.33 max = 38.74 avg = 38.57\n shufflenet_v2 min = 37.43 max = 38.97 avg = 38.32\n mnasnet min = 69.29 max = 73.20 avg = 70.73\n proxylessnasnet min = 80.81 max = 82.66 avg = 81.52\n efficientnet_b0 min = 151.20 max = 152.38 avg = 151.72\n regnety_400m min = 93.53 max = 94.53 avg = 94.19\n blazeface min = 12.15 max = 12.82 avg = 12.46\n googlenet min = 239.63 max = 242.64 avg = 241.06\n googlenet_int8 min = 214.71 max = 216.53 avg = 215.79\n resnet18 min = 234.20 max = 238.74 avg = 236.90\n resnet18_int8 min = 181.57 max = 183.97 avg = 182.66\n alexnet min = 205.94 max = 207.44 avg = 206.63\n vgg16 min = 1188.14 max = 1201.95 avg = 1196.93\n vgg16_int8 min = 1081.21 max = 1087.84 avg = 1085.17\n resnet50 min = 556.54 max = 566.68 avg = 561.21\n resnet50_int8 min = 433.19 max = 433.93 avg = 433.48\n squeezenet_ssd min = 169.02 max = 170.54 avg = 169.73\n squeezenet_ssd_int8 min = 176.28 max = 177.90 avg = 176.87\n mobilenet_ssd min = 228.15 max = 232.69 avg = 230.38\n mobilenet_ssd_int8 min = 236.97 max = 239.69 avg = 238.35\n mobilenet_yolo min = 493.33 max = 506.34 avg = 499.79\n mobilenetv2_yolov3 min = 252.53 max = 261.58 avg = 256.30\n\nlavender:/data/local/tmp/ncnnbench $ ./benchncnn 4 1 2 0 1\n[0 Adreno (TM) 512] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 512] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 512] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 34.49 max = 34.65 avg = 34.55\n mobilenet min = 54.45 max = 55.52 avg = 54.75\n mobilenet_v2 min = 39.32 max = 39.58 avg = 39.50\n mobilenet_v3 min = 36.13 max = 36.28 avg = 36.19\n shufflenet min = 35.25 max = 35.42 avg = 35.31\n shufflenet_v2 min = 31.38 max = 31.70 avg = 31.53\n mnasnet min = 40.95 max = 41.32 avg = 41.13\n proxylessnasnet min = 43.81 max = 44.05 avg = 43.90\n efficientnet_b0 min = 68.34 max = 68.56 avg = 68.47\n regnety_400m min = 53.89 max = 54.23 avg = 54.02\n blazeface min = 19.82 max = 27.74 avg = 22.01\n googlenet min = 119.46 max = 119.98 avg = 119.80\n resnet18 min = 115.56 max = 120.28 avg = 116.88\n alexnet min = 102.06 max = 105.56 avg = 102.97\n vgg16 min = 1192.29 max = 1202.17 avg = 1197.03\n resnet50 min = 294.87 max = 298.79 avg = 296.05\n squeezenet_ssd min = 167.85 max = 168.42 avg = 168.09\n mobilenet_ssd min = 120.30 max = 120.37 avg = 120.34\n mobilenet_yolo min = 256.60 max = 260.21 avg = 257.54\n mobilenetv2_yolov3 min = 121.48 max = 125.22 avg = 122.53\n```\n\n### Qualcomm MSM8996 Pro Snapdragon 821 (Kyro 2.35GHz x 2 + Kyro 2.19GHz x 2)\n```\nnatrium:/data/local/tmp # ./benchncnn 8 4 0 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 18.46 max = 19.12 avg = 18.78\n squeezenet_int8 min = 16.69 max = 17.22 avg = 16.95\n mobilenet min = 27.33 max = 28.74 avg = 27.88\n mobilenet_int8 min = 20.14 max = 20.71 avg = 20.46\n mobilenet_v2 min = 21.94 max = 23.09 avg = 22.38\n mobilenet_v3 min = 18.81 max = 19.45 avg = 19.04\n shufflenet min = 14.07 max = 14.75 avg = 14.29\n shufflenet_v2 min = 11.52 max = 11.92 avg = 11.71\n mnasnet min = 20.41 max = 21.75 avg = 20.74\n proxylessnasnet min = 22.99 max = 23.63 avg = 23.13\n efficientnet_b0 min = 34.74 max = 35.26 avg = 34.91\n efficientnetv2_b0 min = 41.16 max = 41.60 avg = 41.39\n regnety_400m min = 44.27 max = 45.01 avg = 44.69\n blazeface min = 4.25 max = 4.71 avg = 4.43\n googlenet min = 54.88 max = 55.55 avg = 55.12\n googlenet_int8 min = 51.88 max = 52.72 avg = 52.25\n resnet18 min = 44.33 max = 45.44 avg = 44.88\n resnet18_int8 min = 51.24 max = 51.94 avg = 51.54\n alexnet min = 38.62 max = 39.31 avg = 38.88\n vgg16 min = 242.53 max = 244.23 avg = 243.16\n vgg16_int8 min = 183.15 max = 204.96 avg = 192.16\n resnet50 min = 122.14 max = 124.29 avg = 122.94\n resnet50_int8 min = 116.61 max = 118.47 avg = 117.56\n squeezenet_ssd min = 47.92 max = 49.01 avg = 48.45\n squeezenet_ssd_int8 min = 43.21 max = 44.45 avg = 43.76\n mobilenet_ssd min = 56.92 max = 58.21 avg = 57.56\n mobilenet_ssd_int8 min = 42.26 max = 42.92 avg = 42.48\n mobilenet_yolo min = 126.20 max = 128.50 avg = 127.10\n mobilenetv2_yolov3 min = 75.49 max = 76.50 avg = 76.01\n yolov4-tiny min = 94.24 max = 95.75 avg = 94.83\n nanodet_m min = 31.30 max = 31.93 avg = 31.62\n yolo-fastest-1.1 min = 16.89 max = 17.56 avg = 17.23\n yolo-fastestv2 min = 12.97 max = 13.50 avg = 13.15\n\nnatrium:/data/local/tmp # ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 46.27 max = 46.60 avg = 46.45\n squeezenet_int8 min = 41.33 max = 41.73 avg = 41.56\n mobilenet min = 80.89 max = 81.16 avg = 81.00\n mobilenet_int8 min = 60.33 max = 62.29 avg = 61.33\n mobilenet_v2 min = 51.78 max = 52.02 avg = 51.88\n mobilenet_v3 min = 43.71 max = 44.17 avg = 43.91\n shufflenet min = 24.96 max = 25.08 avg = 25.02\n shufflenet_v2 min = 24.09 max = 24.26 avg = 24.17\n mnasnet min = 51.28 max = 51.42 avg = 51.35\n proxylessnasnet min = 59.25 max = 59.66 avg = 59.48\n efficientnet_b0 min = 92.16 max = 92.34 avg = 92.22\n efficientnetv2_b0 min = 112.27 max = 113.63 avg = 113.17\n regnety_400m min = 68.59 max = 68.85 avg = 68.75\n blazeface min = 7.36 max = 7.83 avg = 7.59\n googlenet min = 151.15 max = 151.53 avg = 151.37\n googlenet_int8 min = 152.01 max = 158.63 avg = 154.18\n resnet18 min = 121.49 max = 121.90 avg = 121.77\n resnet18_int8 min = 154.54 max = 166.73 avg = 161.30\n alexnet min = 97.41 max = 97.74 avg = 97.62\n vgg16 min = 674.80 max = 675.86 avg = 675.38\n vgg16_int8 min = 593.42 max = 602.98 avg = 596.93\n resnet50 min = 360.44 max = 364.31 avg = 362.01\n resnet50_int8 min = 371.21 max = 386.24 avg = 381.53\n squeezenet_ssd min = 97.72 max = 98.32 avg = 98.01\n squeezenet_ssd_int8 min = 98.33 max = 99.15 avg = 98.63\n mobilenet_ssd min = 161.72 max = 161.89 avg = 161.79\n mobilenet_ssd_int8 min = 122.44 max = 123.38 avg = 123.00\n mobilenet_yolo min = 367.34 max = 369.59 avg = 368.97\n mobilenetv2_yolov3 min = 190.09 max = 190.77 avg = 190.31\n yolov4-tiny min = 241.59 max = 242.29 avg = 241.81\n nanodet_m min = 63.03 max = 63.22 avg = 63.12\n yolo-fastest-1.1 min = 29.06 max = 29.22 avg = 29.12\n yolo-fastestv2 min = 22.72 max = 22.80 avg = 22.77\n```\n\n### Qualcomm MSM8994 Snapdragon 810 (Cortex-A57 2.0GHz x 4 + Cortex-A53 1.55GHz x 4)\n```\nangler:/data/local/tmp $ ./benchncnn 8 8 0 -1 1\nloop_count = 8\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 25.83 max = 29.17 avg = 27.69\n squeezenet_int8 min = 24.18 max = 26.31 avg = 25.18\n mobilenet min = 33.94 max = 35.29 avg = 34.44\n mobilenet_int8 min = 24.99 max = 26.12 avg = 25.46\n mobilenet_v2 min = 32.63 max = 34.44 avg = 33.56\n mobilenet_v3 min = 27.72 max = 30.14 avg = 29.35\n shufflenet min = 23.23 max = 26.78 avg = 24.58\n shufflenet_v2 min = 21.04 max = 22.25 avg = 21.68\n mnasnet min = 29.51 max = 31.26 avg = 30.27\n proxylessnasnet min = 34.21 max = 37.55 avg = 35.20\n efficientnet_b0 min = 54.75 max = 60.45 avg = 56.38\n efficientnetv2_b0 min = 63.60 max = 67.51 avg = 64.81\n regnety_400m min = 60.80 max = 72.33 avg = 68.27\n blazeface min = 5.96 max = 7.22 avg = 6.41\n googlenet min = 80.62 max = 94.46 avg = 86.50\n googlenet_int8 min = 69.05 max = 75.75 avg = 71.47\n resnet18 min = 63.90 max = 75.96 avg = 69.64\n resnet18_int8 min = 46.43 max = 62.23 avg = 53.22\n alexnet min = 82.67 max = 90.25 avg = 87.03\n vgg16 min = 562.23 max = 636.26 avg = 594.82\n vgg16_int8 min = 303.42 max = 358.03 avg = 325.60\n resnet50 min = 233.47 max = 279.99 avg = 248.49\n resnet50_int8 min = 170.11 max = 198.27 avg = 183.35\n squeezenet_ssd min = 86.97 max = 112.21 avg = 96.84\n squeezenet_ssd_int8 min = 66.09 max = 77.00 avg = 70.57\n mobilenet_ssd min = 76.95 max = 101.74 avg = 87.73\n mobilenet_ssd_int8 min = 53.27 max = 60.50 avg = 57.46\n mobilenet_yolo min = 206.42 max = 260.06 avg = 227.84\n mobilenetv2_yolov3 min = 129.32 max = 147.76 avg = 138.90\n yolov4-tiny min = 184.85 max = 213.03 avg = 203.52\n nanodet_m min = 47.66 max = 60.55 avg = 53.00\n\nangler:/data/local/tmp # ./benchncnn 4 4 2 -1 1\nloop_count = 4\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 41.39 max = 47.64 avg = 43.08\n squeezenet_int8 min = 36.92 max = 37.59 avg = 37.24\n mobilenet min = 59.04 max = 59.43 avg = 59.22\n mobilenet_int8 min = 44.67 max = 46.60 avg = 45.58\n mobilenet_v2 min = 43.38 max = 43.71 avg = 43.62\n mobilenet_v3 min = 37.57 max = 37.82 avg = 37.65\n shufflenet min = 30.67 max = 30.86 avg = 30.76\n shufflenet_v2 min = 27.80 max = 28.12 avg = 27.97\n mnasnet min = 42.99 max = 46.41 avg = 44.21\n proxylessnasnet min = 51.26 max = 53.52 avg = 52.04\n efficientnet_b0 min = 81.58 max = 82.30 avg = 82.03\n efficientnetv2_b0 min = 94.01 max = 94.48 avg = 94.27\n regnety_400m min = 82.38 max = 83.86 avg = 82.95\n blazeface min = 10.02 max = 10.42 avg = 10.18\n googlenet min = 125.47 max = 126.72 avg = 125.92\n googlenet_int8 min = 109.92 max = 111.65 avg = 110.44\n resnet18 min = 110.14 max = 111.95 avg = 110.76\n resnet18_int8 min = 78.21 max = 79.65 avg = 79.07\n alexnet min = 78.09 max = 80.34 avg = 78.87\n vgg16 min = 486.69 max = 494.97 avg = 490.35\n vgg16_int8 min = 370.66 max = 377.64 avg = 373.78\n resnet50 min = 272.31 max = 278.64 avg = 274.10\n resnet50_int8 min = 215.57 max = 218.55 avg = 217.27\n squeezenet_ssd min = 112.98 max = 114.75 avg = 113.60\n squeezenet_ssd_int8 min = 91.85 max = 94.82 avg = 93.13\n mobilenet_ssd min = 115.18 max = 116.56 avg = 115.95\n mobilenet_ssd_int8 min = 90.95 max = 92.21 avg = 91.39\n mobilenet_yolo min = 255.07 max = 259.01 avg = 256.18\n mobilenetv2_yolov3 min = 155.52 max = 156.58 avg = 156.09\n yolov4-tiny min = 231.89 max = 234.14 avg = 232.97\n nanodet_m min = 72.74 max = 74.71 avg = 73.52\n yolo-fastest-1.1 min = 35.25 max = 36.51 avg = 35.77\n yolo-fastestv2 min = 29.94 max = 31.09 avg = 30.75\n\nangler:/data/local/tmp # ./benchncnn 4 1 2 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 70.83 max = 72.68 avg = 71.77\n squeezenet_int8 min = 59.27 max = 59.60 avg = 59.51\n mobilenet min = 110.70 max = 112.72 avg = 111.48\n mobilenet_int8 min = 79.69 max = 80.01 avg = 79.81\n mobilenet_v2 min = 77.85 max = 78.19 avg = 78.03\n mobilenet_v3 min = 63.49 max = 63.92 avg = 63.73\n shufflenet min = 41.43 max = 41.60 avg = 41.49\n shufflenet_v2 min = 37.49 max = 38.26 avg = 37.97\n mnasnet min = 73.91 max = 75.91 avg = 74.59\n proxylessnasnet min = 94.13 max = 94.53 avg = 94.37\n efficientnet_b0 min = 161.91 max = 162.38 avg = 162.10\n efficientnetv2_b0 min = 179.33 max = 180.26 avg = 179.67\n regnety_400m min = 100.35 max = 100.76 avg = 100.53\n blazeface min = 12.57 max = 12.76 avg = 12.66\n googlenet min = 232.77 max = 233.08 avg = 232.91\n googlenet_int8 min = 203.39 max = 205.25 avg = 204.77\n resnet18 min = 182.58 max = 183.17 avg = 182.91\n resnet18_int8 min = 150.40 max = 152.07 avg = 151.35\n alexnet min = 147.27 max = 149.00 avg = 148.06\n vgg16 min = 986.93 max = 988.35 avg = 987.47\n vgg16_int8 min = 816.37 max = 819.93 avg = 817.79\n resnet50 min = 502.77 max = 510.88 avg = 508.53\n resnet50_int8 min = 393.33 max = 398.07 avg = 395.86\n squeezenet_ssd min = 175.01 max = 175.61 avg = 175.32\n squeezenet_ssd_int8 min = 145.19 max = 145.94 avg = 145.66\n mobilenet_ssd min = 231.04 max = 231.25 avg = 231.13\n mobilenet_ssd_int8 min = 159.81 max = 160.52 avg = 160.13\n mobilenet_yolo min = 517.86 max = 523.71 avg = 521.85\n mobilenetv2_yolov3 min = 275.84 max = 279.16 avg = 277.13\n yolov4-tiny min = 363.71 max = 366.14 avg = 364.56\n nanodet_m min = 93.90 max = 95.09 avg = 94.40\n yolo-fastest-1.1 min = 45.94 max = 46.09 avg = 46.01\n yolo-fastestv2 min = 38.23 max = 38.33 avg = 38.29\n\nangler:/data/local/tmp $ ./benchncnn 4 1 2 0 1\n[0 Adreno (TM) 430] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 430] buglssc=0 bugsbn1=1 buglbia=0 bugihfa=0\n[0 Adreno (TM) 430] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 39.49 max = 41.93 avg = 40.62\n mobilenet min = 60.30 max = 61.81 avg = 60.88\n mobilenet_v2 min = 45.38 max = 47.10 avg = 45.88\n mobilenet_v3 min = 45.97 max = 47.39 avg = 46.69\n shufflenet min = 29.12 max = 31.02 avg = 29.91\n shufflenet_v2 min = 47.58 max = 50.06 avg = 48.26\n mnasnet min = 47.84 max = 49.17 avg = 48.26\n proxylessnasnet min = 49.51 max = 51.03 avg = 49.97\n efficientnet_b0 min = 100.56 max = 105.60 avg = 102.45\n regnety_400m min = 59.67 max = 61.24 avg = 60.56\n blazeface min = 13.87 max = 13.98 avg = 13.93\n googlenet min = 131.26 max = 136.33 avg = 133.40\n resnet18 min = 116.38 max = 117.92 avg = 116.93\n alexnet min = 72.59 max = 73.94 avg = 73.29\n vgg16 min = 1090.07 max = 1101.71 avg = 1096.34\n resnet50 min = 299.76 max = 300.78 avg = 300.40\n squeezenet_ssd min = 181.95 max = 182.83 avg = 182.39\n mobilenet_ssd min = 148.44 max = 151.07 avg = 149.75\n mobilenet_yolo min = 284.46 max = 285.74 avg = 285.39\n mobilenetv2_yolov3 min = 140.28 max = 148.62 avg = 144.83\n```\n\n### Qualcomm MSM8916 Snapdragon 410 (Cortex-A53 1.2GHz x 4)\n```\nHM2014812:/data/local/tmp # ./benchncnn 8 4 0 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 65.45 max = 73.59 avg = 68.10\n squeezenet_int8 min = 59.39 max = 65.54 avg = 61.14\n mobilenet min = 86.69 max = 94.10 avg = 90.03\n mobilenet_int8 min = 62.22 max = 69.67 avg = 64.13\n mobilenet_v2 min = 77.98 max = 89.53 avg = 82.00\n mobilenet_v3 min = 62.17 max = 68.31 avg = 63.90\n shufflenet min = 47.52 max = 53.76 avg = 49.92\n shufflenet_v2 min = 39.77 max = 46.08 avg = 40.66\n mnasnet min = 69.27 max = 75.73 avg = 71.73\n proxylessnasnet min = 78.72 max = 85.37 avg = 81.33\n efficientnet_b0 min = 126.62 max = 136.67 avg = 130.69\n efficientnetv2_b0 min = 143.24 max = 150.97 avg = 146.89\n regnety_400m min = 108.79 max = 116.22 avg = 112.99\n blazeface min = 14.85 max = 15.02 avg = 14.94\n googlenet min = 180.91 max = 190.37 avg = 186.36\n googlenet_int8 min = 160.07 max = 170.86 avg = 165.05\n resnet18 min = 137.91 max = 155.37 avg = 144.99\n resnet18_int8 min = 104.34 max = 110.20 avg = 106.76\n alexnet min = 105.30 max = 114.73 avg = 109.53\n vgg16 min = 829.16 max = 942.94 avg = 853.28\n vgg16_int8 min = 515.61 max = 547.32 avg = 526.50\n resnet50 min = 380.46 max = 443.90 avg = 393.71\n resnet50_int8 min = 318.06 max = 327.13 avg = 323.23\n squeezenet_ssd min = 178.22 max = 189.02 avg = 184.51\n squeezenet_ssd_int8 min = 153.75 max = 163.44 avg = 158.05\n mobilenet_ssd min = 189.45 max = 195.17 avg = 193.10\n mobilenet_ssd_int8 min = 132.59 max = 139.63 avg = 137.23\n mobilenet_yolo min = 404.52 max = 414.20 avg = 409.97\n mobilenetv2_yolov3 min = 271.33 max = 279.98 avg = 275.08\n yolov4-tiny min = 349.36 max = 372.54 avg = 357.98\n nanodet_m min = 103.01 max = 111.71 avg = 105.82\n\nHM2014812:/data/local/tmp # ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 147.48 max = 149.35 avg = 148.40\n squeezenet_int8 min = 143.20 max = 144.55 avg = 143.98\n mobilenet min = 243.78 max = 244.33 avg = 244.08\n mobilenet_int8 min = 206.23 max = 207.13 avg = 206.55\n mobilenet_v2 min = 168.04 max = 170.37 avg = 169.06\n mobilenet_v3 min = 147.10 max = 147.91 avg = 147.55\n shufflenet min = 88.47 max = 89.31 avg = 88.85\n shufflenet_v2 min = 84.47 max = 84.80 avg = 84.60\n mnasnet min = 162.81 max = 163.93 avg = 163.22\n proxylessnasnet min = 208.18 max = 209.15 avg = 208.61\n efficientnet_b0 min = 370.06 max = 371.14 avg = 370.64\n efficientnetv2_b0 min = 418.28 max = 429.68 avg = 423.01\n regnety_400m min = 216.42 max = 217.19 avg = 216.71\n blazeface min = 27.63 max = 28.67 avg = 28.00\n googlenet min = 525.25 max = 528.83 avg = 526.23\n googlenet_int8 min = 469.78 max = 472.51 avg = 470.76\n resnet18 min = 396.46 max = 399.66 avg = 397.57\n resnet18_int8 min = 324.07 max = 326.64 avg = 325.34\n alexnet min = 362.44 max = 363.02 avg = 362.68\n vgg16 min = 2174.86 max = 2252.92 avg = 2215.62\n vgg16_int8 min = 1726.07 max = 1732.69 avg = 1729.18\n resnet50 min = 1136.96 max = 1142.94 avg = 1139.91\n resnet50_int8 min = 977.73 max = 983.64 avg = 980.71\n squeezenet_ssd min = 350.46 max = 353.35 avg = 351.37\n squeezenet_ssd_int8 min = 333.91 max = 336.59 avg = 334.77\n mobilenet_ssd min = 513.18 max = 519.05 avg = 516.22\n mobilenet_ssd_int8 min = 424.37 max = 426.89 avg = 426.03\n mobilenet_yolo min = 1143.20 max = 1145.04 avg = 1144.31\n mobilenetv2_yolov3 min = 617.45 max = 619.30 avg = 618.37\n yolov4-tiny min = 839.32 max = 847.57 avg = 844.61\n nanodet_m min = 208.41 max = 211.31 avg = 210.03\n```\n\n### Qualcomm Snapdragon 888 (Cortex-X1 2.84GHz x1 + Cortex-A78 2.4GHz x3 + Cortex-A55 1.8GHz x4 + Adreno 660)\n```\nvenus:/data/local/tmp $ ./benchncnn 8 8 2 -1 1\nloop_count = 8\nnum_threads = 8\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 5.89 max = 6.04 avg = 5.98\n squeezenet_int8 min = 6.09 max = 6.29 avg = 6.25\n mobilenet min = 9.27 max = 10.22 avg = 9.64\n mobilenet_int8 min = 5.90 max = 6.05 avg = 5.97\n mobilenet_v2 min = 6.87 max = 8.42 avg = 7.63\n mobilenet_v3 min = 8.93 max = 12.22 avg = 9.55\n shufflenet min = 8.72 max = 11.44 avg = 9.20\n shufflenet_v2 min = 6.05 max = 8.24 avg = 7.40\n mnasnet min = 7.83 max = 9.03 avg = 8.53\n proxylessnasnet min = 7.03 max = 9.62 avg = 7.88\n efficientnet_b0 min = 12.62 max = 18.01 avg = 15.51\n efficientnetv2_b0 min = 14.96 max = 23.75 avg = 19.61\n regnety_400m min = 23.58 max = 23.87 avg = 23.72\n blazeface min = 4.62 max = 4.87 avg = 4.73\n googlenet min = 17.23 max = 25.41 avg = 19.83\n googlenet_int8 min = 16.91 max = 17.05 avg = 16.99\n resnet18 min = 12.05 max = 14.90 avg = 13.47\n resnet18_int8 min = 15.10 max = 15.42 avg = 15.27\n alexnet min = 13.85 max = 15.73 avg = 14.50\n vgg16 min = 56.85 max = 57.88 avg = 57.32\n vgg16_int8 min = 70.12 max = 72.99 avg = 71.53\n resnet50 min = 29.45 max = 29.78 avg = 29.64\n resnet50_int8 min = 24.99 max = 25.31 avg = 25.16\n squeezenet_ssd min = 17.51 max = 22.63 avg = 19.25\n squeezenet_ssd_int8 min = 16.81 max = 17.26 avg = 16.98\n mobilenet_ssd min = 15.96 max = 16.52 avg = 16.11\n mobilenet_ssd_int8 min = 13.70 max = 14.26 avg = 13.95\n mobilenet_yolo min = 50.48 max = 52.88 avg = 51.76\n mobilenetv2_yolov3 min = 22.63 max = 22.99 avg = 22.85\n yolov4-tiny min = 29.01 max = 38.20 avg = 32.50\n nanodet_m min = 12.58 max = 15.53 avg = 13.86\n yolo-fastest-1.1 min = 8.57 max = 9.18 avg = 8.86\n yolo-fastestv2 min = 6.85 max = 8.47 avg = 8.05\n vision_transformer min = 548.48 max = 703.29 avg = 614.47\n FastestDet min = 7.71 max = 9.31 avg = 8.15\n \nvenus:/data/local/tmp $ ./benchncnn 8 8 2 0 1\n./benchncnn 8 8 2 0 1\n[0 Adreno (TM) 660] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 660] bugsbn1=1 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Adreno (TM) 660] fp16-p/s/u/a=1/1/0/1 int8-p/s/u/a=1/0/0/1\n[0 Adreno (TM) 660] subgroup=64 basic/vote/ballot/shuffle=1/1/1/1\n[0 Adreno (TM) 660] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 8\nnum_threads = 8\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 10.63 max = 12.41 avg = 11.80\n squeezenet_int8 min = 6.93 max = 8.82 avg = 7.86\n mobilenet min = 12.79 max = 14.12 avg = 13.48\n mobilenet_int8 min = 9.18 max = 9.70 avg = 9.44\n mobilenet_v2 min = 14.73 max = 15.62 avg = 15.13\n mobilenet_v3 min = 14.68 max = 16.72 avg = 15.70\n shufflenet min = 11.28 max = 12.75 avg = 12.17\n shufflenet_v2 min = 11.44 max = 14.27 avg = 12.07\n mnasnet min = 14.54 max = 15.94 avg = 15.35\n proxylessnasnet min = 16.33 max = 17.31 avg = 16.71\n efficientnet_b0 min = 22.64 max = 25.42 avg = 24.35\n efficientnetv2_b0 min = 41.16 max = 52.08 avg = 45.61\n regnety_400m min = 17.56 max = 18.08 avg = 17.85\n blazeface min = 2.87 max = 3.89 avg = 3.34\n googlenet min = 31.64 max = 33.38 avg = 32.14\n googlenet_int8 min = 18.29 max = 19.15 avg = 18.73\n resnet18 min = 23.47 max = 24.60 avg = 23.85\n resnet18_int8 min = 11.89 max = 17.17 avg = 14.54\n alexnet min = 25.62 max = 26.23 avg = 25.98\n vgg16 min = 41.81 max = 42.69 avg = 42.12\n vgg16_int8 min = 79.43 max = 123.88 avg = 93.17\n resnet50 min = 41.28 max = 43.27 avg = 41.79\n resnet50_int8 min = 25.55 max = 26.34 avg = 25.97\n squeezenet_ssd min = 30.10 max = 33.64 avg = 31.39\n squeezenet_ssd_int8 min = 18.12 max = 18.58 avg = 18.30\n mobilenet_ssd min = 28.29 max = 28.90 avg = 28.66\n mobilenet_ssd_int8 min = 13.90 max = 14.31 avg = 14.02\n mobilenet_yolo min = 43.88 max = 45.43 avg = 44.58\n mobilenetv2_yolov3 min = 16.49 max = 37.05 avg = 19.32\n yolov4-tiny min = 22.70 max = 50.58 avg = 34.92\n nanodet_m min = 19.31 max = 19.88 avg = 19.57\n yolo-fastest-1.1 min = 11.17 max = 11.33 avg = 11.26\n yolo-fastestv2 min = 9.72 max = 10.04 avg = 9.85\n vision_transformer min = 744.98 max = 758.15 avg = 751.62\n FastestDet min = 11.95 max = 13.12 avg = 12.46\n```\n\n### Qualcomm Snapdragon X Elite (X1E78100), Oryon 3.4GHz x 12 + Adreno X1-85\n\nTest on Oryon CPU\n\n```\nloop_count = 10\nnum_threads = 12\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 5.13 max = 5.19 avg = 5.16\n squeezenet_int8 min = 4.31 max = 4.81 avg = 4.67\n mobilenet min = 3.73 max = 3.85 avg = 3.78\n mobilenet_int8 min = 2.51 max = 3.11 avg = 2.64\n mobilenet_v2 min = 3.55 max = 3.70 avg = 3.60\n mobilenet_v3 min = 3.28 max = 3.88 avg = 3.40\n shufflenet min = 3.77 max = 5.07 avg = 4.02\n shufflenet_v2 min = 3.24 max = 3.34 avg = 3.29\n mnasnet min = 3.49 max = 4.09 avg = 3.58\n proxylessnasnet min = 4.30 max = 4.93 avg = 4.41\n efficientnet_b0 min = 4.97 max = 17.26 avg = 6.28\n efficientnetv2_b0 min = 6.85 max = 10.19 avg = 7.39\n regnety_400m min = 11.26 max = 11.36 avg = 11.31\n blazeface min = 1.43 max = 1.48 avg = 1.44\n googlenet min = 9.84 max = 9.96 avg = 9.89\n googlenet_int8 min = 8.04 max = 8.33 avg = 8.13\n resnet18 min = 6.63 max = 9.34 avg = 6.94\n resnet18_int8 min = 5.47 max = 6.24 avg = 5.59\n alexnet min = 7.52 max = 7.61 avg = 7.54\n vgg16 min = 29.66 max = 32.27 avg = 30.07\n vgg16_int8 min = 32.97 max = 34.43 avg = 33.32\n resnet50 min = 16.54 max = 16.68 avg = 16.63\n resnet50_int8 min = 11.12 max = 13.84 avg = 11.42\n squeezenet_ssd min = 9.20 max = 9.77 avg = 9.39\n squeezenet_ssd_int8 min = 8.50 max = 9.17 avg = 8.73\n mobilenet_ssd min = 8.28 max = 8.67 avg = 8.36\n mobilenet_ssd_int8 min = 5.59 max = 6.25 avg = 5.74\n mobilenet_yolo min = 21.42 max = 22.77 avg = 21.65\n mobilenetv2_yolov3 min = 14.03 max = 14.34 avg = 14.13\n yolov4-tiny min = 23.60 max = 23.84 avg = 23.70\n nanodet_m min = 6.64 max = 7.40 avg = 6.77\n yolo-fastest-1.1 min = 4.14 max = 7.15 avg = 4.53\n yolo-fastestv2 min = 3.63 max = 3.70 avg = 3.66\n vision_transformer min = 384.74 max = 415.74 avg = 391.28\n FastestDet min = 4.29 max = 4.94 avg = 4.40\n```\n\nTest on X1-85 GPU\n\n```\n[0 Adreno X1-85] queueC=0[1] queueT=0[1]\n[0 Adreno X1-85] fp16-p/s/u/a=1/1/0/1 int8-p/s/u/a=1/0/0/1 bf16-p/s=1/0\n[0 Adreno X1-85] subgroup=128(64~128) ops=1/1/1/1/1/1/1/1/1/1\n[0 Adreno X1-85] fp16-cm=0 int8-cm=0 bf16-cm=0 fp8-cm=0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 3.23 max = 3.99 avg = 3.63\n mobilenet min = 3.33 max = 5.86 avg = 5.20\n mobilenet_v2 min = 4.06 max = 4.77 avg = 4.52\n mobilenet_v3 min = 4.61 max = 8.12 avg = 6.60\n shufflenet min = 3.16 max = 7.45 avg = 4.65\n shufflenet_v2 min = 3.90 max = 6.00 avg = 5.02\n mnasnet min = 4.44 max = 5.12 avg = 4.81\n proxylessnasnet min = 4.91 max = 7.02 avg = 6.15\n efficientnet_b0 min = 6.61 max = 7.25 avg = 7.04\n efficientnetv2_b0 min = 21.48 max = 56.52 avg = 39.03\n regnety_400m min = 7.33 max = 7.60 avg = 7.44\n blazeface min = 2.83 max = 4.59 avg = 4.30\n googlenet min = 11.00 max = 12.98 avg = 12.60\n resnet18 min = 12.11 max = 14.59 avg = 13.27\n alexnet min = 11.64 max = 12.18 avg = 11.96\n vgg16 min = 40.06 max = 45.62 avg = 42.88\n resnet50 min = 18.99 max = 21.93 avg = 20.88\n squeezenet_ssd min = 10.95 max = 14.73 avg = 13.03\n mobilenet_ssd min = 7.92 max = 9.75 avg = 9.46\n mobilenet_yolo min = 9.02 max = 12.54 avg = 11.38\n mobilenetv2_yolov3 min = 12.70 max = 14.70 avg = 13.95\n yolov4-tiny min = 25.88 max = 30.26 avg = 28.12\n nanodet_m min = 9.38 max = 33.46 avg = 20.29\n yolo-fastest-1.1 min = 6.08 max = 6.75 avg = 6.43\n yolo-fastestv2 min = 4.50 max = 6.47 avg = 6.04\n vision_transformer min = 184.89 max = 191.78 avg = 189.07\n FastestDet min = 6.01 max = 7.83 avg = 6.43\n```\n\n### Raspberry Pi 3 Model B+ Broadcom BCM2837B0, Cortex-A53 (ARMv8) (1.4GHz x 4)\n```\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 4 4 0 -1 1\nloop_count = 4\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 84.74 max = 85.60 avg = 85.22\n squeezenet_int8 min = 74.48 max = 74.80 avg = 74.68\n mobilenet min = 107.84 max = 110.13 avg = 108.66\n mobilenet_int8 min = 66.91 max = 67.12 avg = 67.03\n mobilenet_v2 min = 110.64 max = 112.73 avg = 111.68\n mobilenet_v3 min = 85.78 max = 86.74 avg = 86.44\n shufflenet min = 58.38 max = 60.32 avg = 59.33\n shufflenet_v2 min = 46.76 max = 47.53 avg = 47.19\n mnasnet min = 95.53 max = 95.88 avg = 95.78\n proxylessnasnet min = 102.24 max = 105.58 avg = 103.38\n efficientnet_b0 min = 134.87 max = 136.98 avg = 135.86\n efficientnetv2_b0 min = 146.62 max = 148.06 avg = 147.13\n regnety_400m min = 118.60 max = 119.51 avg = 119.03\n blazeface min = 15.42 max = 15.61 avg = 15.52\n googlenet min = 223.78 max = 224.85 avg = 224.22\n googlenet_int8 min = 188.23 max = 190.15 avg = 189.21\n resnet18 min = 270.86 max = 272.66 avg = 271.93\n resnet18_int8 min = 159.57 max = 160.39 avg = 160.07\n alexnet min = 157.79 max = 160.77 avg = 159.09\n resnet50 min = 583.57 max = 591.41 avg = 587.42\n resnet50_int8 min = 383.96 max = 401.37 avg = 391.87\n squeezenet_ssd min = 247.90 max = 249.77 avg = 248.98\n squeezenet_ssd_int8 min = 191.65 max = 192.81 avg = 192.17\n mobilenet_ssd min = 240.11 max = 241.02 avg = 240.62\n mobilenet_ssd_int8 min = 136.30 max = 137.26 avg = 136.73\n mobilenet_yolo min = 523.59 max = 539.91 avg = 529.98\n mobilenetv2_yolov3 min = 356.44 max = 366.85 avg = 362.06\n yolov4-tiny min = 410.25 max = 422.18 avg = 417.17\n nanodet_m min = 114.98 max = 115.83 avg = 115.40\n yolo-fastest-1.1 min = 79.85 max = 80.83 avg = 80.28\n yolo-fastestv2 min = 62.36 max = 62.91 avg = 62.60\n FastestDet min = 67.11 max = 68.51 avg = 67.98\n\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 125.34 max = 125.81 avg = 125.58\n squeezenet_int8 min = 135.56 max = 136.34 avg = 135.98\n mobilenet min = 204.62 max = 207.06 avg = 205.65\n mobilenet_int8 min = 181.34 max = 182.46 avg = 181.91\n mobilenet_v2 min = 158.69 max = 158.94 avg = 158.80\n mobilenet_v3 min = 127.13 max = 127.31 avg = 127.23\n shufflenet min = 84.64 max = 85.29 avg = 84.89\n shufflenet_v2 min = 74.28 max = 74.64 avg = 74.44\n mnasnet min = 148.12 max = 148.65 avg = 148.42\n proxylessnasnet min = 199.56 max = 201.99 avg = 200.42\n efficientnet_b0 min = 240.94 max = 241.75 avg = 241.27\n efficientnetv2_b0 min = 270.71 max = 270.90 avg = 270.83\n regnety_400m min = 186.89 max = 187.08 avg = 187.01\n blazeface min = 22.75 max = 23.24 avg = 22.95\n googlenet min = 450.64 max = 450.96 avg = 450.79\n googlenet_int8 min = 424.66 max = 426.83 avg = 425.78\n resnet18 min = 379.21 max = 380.01 avg = 379.57\n resnet18_int8 min = 312.23 max = 313.21 avg = 312.68\n alexnet min = 270.13 max = 270.88 avg = 270.55\n resnet50 min = 977.51 max = 981.89 avg = 979.75\n resnet50_int8 min = 890.77 max = 896.89 avg = 893.83\n squeezenet_ssd min = 331.52 max = 333.47 avg = 332.46\n squeezenet_ssd_int8 min = 317.71 max = 319.64 avg = 318.62\n mobilenet_ssd min = 425.42 max = 426.52 avg = 425.93\n mobilenet_ssd_int8 min = 370.17 max = 370.90 avg = 370.66\n mobilenet_yolo min = 930.40 max = 932.24 avg = 931.46\n mobilenetv2_yolov3 min = 534.79 max = 543.56 avg = 539.20\n yolov4-tiny min = 675.33 max = 676.83 avg = 676.14\n nanodet_m min = 178.13 max = 178.98 avg = 178.64\n yolo-fastest-1.1 min = 100.83 max = 101.96 avg = 101.49\n yolo-fastestv2 min = 79.73 max = 79.94 avg = 79.84\n FastestDet min = 89.09 max = 90.07 avg = 89.78\n```\n\n### Raspberry Pi 4 Model B Broadcom BCM2711B0, Cortex-A72 (ARMv8) (1.8GHz x 4)\n```\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 10 4 0 -1 1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 46.28 max = 46.91 avg = 46.65\n squeezenet_int8 min = 42.18 max = 44.98 avg = 42.59\n mobilenet min = 60.74 max = 61.79 avg = 61.17\n mobilenet_int8 min = 34.19 max = 34.55 avg = 34.37\n mobilenet_v2 min = 61.63 max = 62.02 avg = 61.88\n mobilenet_v3 min = 47.08 max = 48.40 avg = 47.53\n shufflenet min = 32.91 max = 33.30 avg = 33.09\n shufflenet_v2 min = 24.37 max = 24.73 avg = 24.56\n mnasnet min = 51.80 max = 52.14 avg = 51.98\n proxylessnasnet min = 53.02 max = 53.58 avg = 53.32\n efficientnet_b0 min = 73.92 max = 74.44 avg = 74.19\n efficientnetv2_b0 min = 79.10 max = 79.60 avg = 79.34\n regnety_400m min = 65.27 max = 66.12 avg = 65.70\n blazeface min = 8.62 max = 8.75 avg = 8.69\n googlenet min = 113.74 max = 115.14 avg = 114.35\n googlenet_int8 min = 100.87 max = 101.71 avg = 101.25\n resnet18 min = 122.27 max = 125.39 avg = 123.12\n resnet18_int8 min = 82.19 max = 94.12 avg = 83.92\n alexnet min = 75.75 max = 78.08 avg = 76.40\n vgg16 min = 541.66 max = 552.56 avg = 547.09\n vgg16_int8 min = 391.44 max = 395.73 avg = 394.23\n resnet50 min = 261.90 max = 263.91 avg = 262.83\n resnet50_int8 min = 195.60 max = 198.08 avg = 196.65\n squeezenet_ssd min = 127.01 max = 129.85 avg = 127.61\n squeezenet_ssd_int8 min = 104.98 max = 107.67 avg = 105.47\n mobilenet_ssd min = 120.43 max = 123.28 avg = 121.46\n mobilenet_ssd_int8 min = 70.70 max = 72.85 avg = 71.14\n mobilenet_yolo min = 270.89 max = 273.42 avg = 272.33\n mobilenetv2_yolov3 min = 183.85 max = 185.73 avg = 184.88\n yolov4-tiny min = 205.95 max = 209.90 avg = 207.22\n nanodet_m min = 68.08 max = 68.69 avg = 68.38\n yolo-fastest-1.1 min = 47.97 max = 48.20 avg = 48.06\n yolo-fastestv2 min = 37.17 max = 37.69 avg = 37.47\n vision_transformer min = 1872.31 max = 1964.95 avg = 1909.21\n FastestDet min = 38.39 max = 39.17 avg = 38.69\n\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 10 1 0 -1 1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 73.35 max = 75.10 avg = 73.96\n squeezenet_int8 min = 69.17 max = 69.66 avg = 69.42\n mobilenet min = 123.76 max = 125.35 avg = 124.32\n mobilenet_int8 min = 84.66 max = 85.24 avg = 84.82\n mobilenet_v2 min = 92.98 max = 94.05 avg = 93.48\n mobilenet_v3 min = 72.48 max = 73.14 avg = 72.81\n shufflenet min = 47.17 max = 47.83 avg = 47.51\n shufflenet_v2 min = 41.62 max = 42.60 avg = 42.12\n mnasnet min = 83.60 max = 84.35 avg = 83.98\n proxylessnasnet min = 98.48 max = 99.33 avg = 98.78\n efficientnet_b0 min = 129.45 max = 130.02 avg = 129.73\n efficientnetv2_b0 min = 155.06 max = 156.70 avg = 155.76\n regnety_400m min = 105.39 max = 106.03 avg = 105.70\n blazeface min = 12.54 max = 12.84 avg = 12.65\n googlenet min = 235.38 max = 236.34 avg = 235.94\n googlenet_int8 min = 209.63 max = 210.39 avg = 210.00\n resnet18 min = 190.80 max = 191.43 avg = 191.10\n resnet18_int8 min = 157.92 max = 158.97 avg = 158.50\n alexnet min = 139.34 max = 139.44 avg = 139.40\n vgg16 min = 1066.58 max = 1079.30 avg = 1071.85\n vgg16_int8 min = 866.15 max = 873.75 avg = 869.84\n resnet50 min = 533.15 max = 535.12 avg = 534.11\n resnet50_int8 min = 423.72 max = 424.24 avg = 423.96\n squeezenet_ssd min = 178.90 max = 179.53 avg = 179.30\n squeezenet_ssd_int8 min = 157.05 max = 159.06 avg = 157.89\n mobilenet_ssd min = 250.71 max = 251.26 avg = 251.00\n mobilenet_ssd_int8 min = 170.21 max = 170.96 avg = 170.56\n mobilenet_yolo min = 557.48 max = 560.08 avg = 558.80\n mobilenetv2_yolov3 min = 301.60 max = 307.98 avg = 306.52\n yolov4-tiny min = 370.55 max = 375.69 avg = 372.99\n nanodet_m min = 103.05 max = 103.74 avg = 103.45\n yolo-fastest-1.1 min = 56.58 max = 57.44 avg = 57.01\n yolo-fastestv2 min = 46.69 max = 47.34 avg = 47.03\n vision_transformer min = 6605.19 max = 6606.66 avg = 6605.73\n FastestDet min = 52.11 max = 52.97 avg = 52.61\n```\n### Raspberry Pi 5 Broadcom BCM2712, Cortex-A76 (ARMv8) (2.4GHz x 4)\n```\npi@raspberrypi:~/ncnn/benchmark $ ./benchncnn 10 4 0 -1 -1 >> text.out\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 6.74 max = 8.16 avg = 7.38\n squeezenet_int8 min = 6.97 max = 7.67 avg = 7.21\n mobilenet min = 9.00 max = 72.98 avg = 33.88\n mobilenet_int8 min = 8.68 max = 8.80 avg = 8.74\n mobilenet_v2 min = 10.46 max = 10.63 avg = 10.52\n mobilenet_v3 min = 7.30 max = 7.44 avg = 7.35\n shufflenet min = 4.14 max = 4.18 avg = 4.16\n shufflenet_v2 min = 3.37 max = 3.41 avg = 3.39\n mnasnet min = 6.83 max = 8.55 avg = 7.10\n proxylessnasnet min = 7.85 max = 7.97 avg = 7.88\n efficientnet_b0 min = 12.28 max = 12.37 avg = 12.33\n efficientnetv2_b0 min = 13.54 max = 13.84 avg = 13.69\n regnety_400m min = 10.93 max = 11.07 avg = 10.99\n blazeface min = 1.45 max = 1.48 avg = 1.47\n googlenet min = 25.13 max = 25.47 avg = 25.35\n googlenet_int8 min = 24.00 max = 24.23 avg = 24.12\n resnet18 min = 19.84 max = 20.19 avg = 19.96\n resnet18_int8 min = 16.68 max = 16.83 avg = 16.74\n alexnet min = 21.21 max = 21.54 avg = 21.36\n vgg16 min = 127.75 max = 134.00 avg = 129.24\n vgg16_int8 min = 106.39 max = 110.66 avg = 107.01\n resnet50 min = 45.94 max = 46.54 avg = 46.21\n resnet50_int8 min = 40.16 max = 42.58 avg = 40.75\n squeezenet_ssd min = 30.10 max = 30.95 avg = 30.37\n squeezenet_ssd_int8 min = 27.71 max = 29.03 avg = 28.15\n mobilenet_ssd min = 24.16 max = 24.89 avg = 24.52\n mobilenet_ssd_int8 min = 21.79 max = 22.37 avg = 22.05\n mobilenet_yolo min = 58.06 max = 58.45 avg = 58.19\n mobilenetv2_yolov3 min = 37.49 max = 37.94 avg = 37.68\n yolov4-tiny min = 44.45 max = 60.58 avg = 46.29\n nanodet_m min = 11.01 max = 11.28 avg = 11.18\n yolo-fastest-1.1 min = 5.53 max = 5.97 avg = 5.62\n yolo-fastestv2 min = 4.76 max = 4.84 avg = 4.80\n vision_transformer min = 600.65 max = 622.47 avg = 611.65\n FastestDet min = 4.83 max = 6.94 avg = 5.34\n\n\npi@raspberrypi:~/ncnn/benchmark $ ./benchncnn 10 1 0 -1 -1 \nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 11.77 max = 12.18 avg = 11.87\n squeezenet_int8 min = 11.67 max = 11.98 avg = 11.82\n mobilenet min = 20.24 max = 20.59 avg = 20.30\n mobilenet_int8 min = 14.38 max = 14.51 avg = 14.44\n mobilenet_v2 min = 16.21 max = 16.49 avg = 16.38\n mobilenet_v3 min = 11.64 max = 12.12 avg = 11.80\n shufflenet min = 7.17 max = 7.24 avg = 7.20\n shufflenet_v2 min = 7.07 max = 7.21 avg = 7.14\n mnasnet min = 12.93 max = 13.03 avg = 12.99\n proxylessnasnet min = 15.72 max = 15.80 avg = 15.74\n efficientnet_b0 min = 24.12 max = 24.53 avg = 24.20\n efficientnetv2_b0 min = 27.59 max = 28.04 avg = 27.75\n regnety_400m min = 16.41 max = 16.66 avg = 16.49\n blazeface min = 2.98 max = 3.04 avg = 3.02\n googlenet min = 48.62 max = 48.87 avg = 48.71\n googlenet_int8 min = 49.07 max = 49.26 avg = 49.15\n resnet18 min = 29.54 max = 30.17 avg = 29.68\n resnet18_int8 min = 36.30 max = 36.55 avg = 36.42\n alexnet min = 35.24 max = 35.86 avg = 35.62\n vgg16 min = 188.84 max = 190.87 avg = 189.63\n vgg16_int8 min = 272.27 max = 274.15 avg = 273.10\n resnet50 min = 89.04 max = 89.87 avg = 89.43\n resnet50_int8 min = 80.00 max = 80.50 avg = 80.16\n squeezenet_ssd min = 38.02 max = 38.69 avg = 38.29\n squeezenet_ssd_int8 min = 40.58 max = 41.17 avg = 40.94\n mobilenet_ssd min = 45.42 max = 47.08 avg = 45.90\n mobilenet_ssd_int8 min = 36.05 max = 37.02 avg = 36.35\n mobilenet_yolo min = 104.82 max = 106.56 avg = 105.69\n mobilenetv2_yolov3 min = 60.11 max = 60.29 avg = 60.19\n yolov4-tiny min = 67.61 max = 69.05 avg = 68.02\n nanodet_m min = 19.63 max = 19.81 avg = 19.69\n yolo-fastest-1.1 min = 8.10 max = 8.14 avg = 8.12\n yolo-fastestv2 min = 7.21 max = 7.26 avg = 7.24\n vision_transformer min = 1249.08 max = 1253.32 avg = 1250.30\n FastestDet min = 7.33 max = 7.44 avg = 7.38\n```\n### Raspberry Pi 5 Broadcom BCM2712, VideoCore VII Graphics (Vulkan 1.2)\n```\nfan@raspberrypi:~/ncnn/benchmark $ ../build/benchmark/benchncnn 10 $(nproc) 0 0\n[0 V3D 7.1.7] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 V3D 7.1.7] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 V3D 7.1.7] fp16-p/s/a=1/1/0 int8-p/s/a=1/1/0\n[0 V3D 7.1.7] subgroup=16 basic/vote/ballot/shuffle=1/0/0/0\n[0 V3D 7.1.7] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] subgroup=4 basic/vote/ballot/shuffle=1/1/1/1\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 120.75 max = 121.31 avg = 120.94\n squeezenet_int8 min = 9.57 max = 24.49 avg = 11.23\n mobilenet min = 160.32 max = 160.75 avg = 160.53\n mobilenet_int8 min = 11.29 max = 11.47 avg = 11.37\n mobilenet_v2 min = 121.05 max = 121.93 avg = 121.46\n mobilenet_v3 min = 117.90 max = 119.20 avg = 118.48\n shufflenet min = 70.82 max = 71.55 avg = 71.04\n shufflenet_v2 min = 97.74 max = 98.58 avg = 98.00\n mnasnet min = 118.21 max = 118.76 avg = 118.44\n proxylessnasnet min = 124.28 max = 124.92 avg = 124.52\n efficientnet_b0 min = 187.48 max = 188.38 avg = 187.93\n efficientnetv2_b0 min = 270.11 max = 280.80 avg = 272.26\n regnety_400m min = 142.14 max = 143.25 avg = 142.66\n blazeface min = 31.97 max = 32.41 avg = 32.17\n googlenet min = 346.30 max = 347.47 avg = 346.81\n googlenet_int8 min = 30.77 max = 32.26 avg = 31.52\n resnet18 min = 346.96 max = 347.50 avg = 347.26\n resnet18_int8 min = 19.95 max = 20.95 avg = 20.48\n alexnet min = 181.57 max = 182.03 avg = 181.75\n vgg16 min = 1776.00 max = 1776.66 avg = 1776.40\n vgg16_int8 min = 134.10 max = 141.76 avg = 136.32\n resnet50 min = 841.90 max = 842.50 avg = 842.16\n resnet50_int8 min = 54.29 max = 55.22 avg = 54.54\n squeezenet_ssd min = 461.71 max = 468.09 avg = 466.97\n squeezenet_ssd_int8 min = 38.05 max = 39.00 avg = 38.58\n mobilenet_ssd min = 379.50 max = 381.66 avg = 380.14\n mobilenet_ssd_int8 min = 29.91 max = 30.77 avg = 30.13\n mobilenet_yolo min = 753.61 max = 755.06 avg = 753.97\n mobilenetv2_yolov3 min = 382.18 max = 389.90 avg = 386.97\n yolov4-tiny min = 673.87 max = 674.71 avg = 674.07\n nanodet_m min = 206.55 max = 210.48 avg = 209.69\n yolo-fastest-1.1 min = 109.98 max = 111.18 avg = 110.45\n yolo-fastestv2 min = 86.07 max = 87.16 avg = 86.51\n vision_transformer min = 20594.51 max = 20601.53 avg = 20596.59\n FastestDet min = 90.25 max = 91.00 avg = 90.64\n```\n\n### Raspberry Pi 5 Broadcom BCM2712 Overclock to 2.9Ghz, VideoCore VII Graphics Overclock to 1.1Ghz (Vulkan 1.2)\n```\npi@raspberrypi:~/ncnn/build/benchmark $ sudo echo \"arm_freq=2900\" >> /boot/firmware/config.txt\npi@raspberrypi:~/ncnn/build/benchmark $ sudo echo \"gpu_freq=1100\" >> /boot/firmware/config.txt\npi@raspberrypi:~/ncnn/build/benchmark $ sudo reboot\n\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 10 4 0 0\n[0 V3D 7.1.7] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 V3D 7.1.7] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 V3D 7.1.7] fp16-p/s/u/a=1/1/1/0 int8-p/s/u/a=1/1/1/0\n[0 V3D 7.1.7] subgroup=16 basic/vote/ballot/shuffle=1/0/0/0\n[0 V3D 7.1.7] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] subgroup=4 basic/vote/ballot/shuffle=1/1/1/1\n[1 llvmpipe (LLVM 15.0.6, 128 bits)] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 106.98 max = 107.05 avg = 107.02\n squeezenet_int8 min = 8.51 max = 8.83 avg = 8.65\n mobilenet min = 147.66 max = 147.71 avg = 147.68\n mobilenet_int8 min = 10.21 max = 10.54 avg = 10.37\n mobilenet_v2 min = 110.11 max = 110.23 avg = 110.18\n mobilenet_v3 min = 101.84 max = 102.03 avg = 101.92\n shufflenet min = 59.77 max = 59.84 avg = 59.80\n shufflenet_v2 min = 81.46 max = 81.60 avg = 81.51\n mnasnet min = 105.88 max = 105.98 avg = 105.94\n proxylessnasnet min = 108.82 max = 108.89 avg = 108.86\n efficientnet_b0 min = 168.79 max = 168.93 avg = 168.87\n efficientnetv2_b0 min = 232.52 max = 232.80 avg = 232.65\n regnety_400m min = 130.33 max = 130.49 avg = 130.36\n blazeface min = 22.23 max = 22.49 avg = 22.39\n googlenet min = 299.25 max = 299.37 avg = 299.31\n googlenet_int8 min = 29.21 max = 29.97 avg = 29.58\n resnet18 min = 304.47 max = 304.64 avg = 304.58\n resnet18_int8 min = 19.31 max = 20.77 avg = 20.24\n alexnet min = 203.68 max = 203.79 avg = 203.76\n vgg16 min = 1571.91 max = 1572.22 avg = 1572.06\n vgg16_int8 min = 128.46 max = 130.89 avg = 129.96\n resnet50 min = 754.16 max = 754.33 avg = 754.26\n resnet50_int8 min = 52.65 max = 53.48 avg = 53.09\n squeezenet_ssd min = 398.22 max = 398.36 avg = 398.28\n squeezenet_ssd_int8 min = 34.26 max = 34.67 avg = 34.51\n mobilenet_ssd min = 344.81 max = 344.99 avg = 344.89\n mobilenet_ssd_int8 min = 27.59 max = 28.01 avg = 27.77\n mobilenet_yolo min = 712.53 max = 712.63 avg = 712.59\n mobilenetv2_yolov3 min = 362.81 max = 363.11 avg = 362.90\n yolov4-tiny min = 589.30 max = 589.51 avg = 589.39\n nanodet_m min = 178.83 max = 178.97 avg = 178.88\n yolo-fastest-1.1 min = 92.36 max = 92.58 avg = 92.45\n yolo-fastestv2 min = 70.68 max = 70.84 avg = 70.74\n vision_transformer min = 18615.94 max = 18648.17 avg = 18633.77\n FastestDet min = 74.59 max = 74.68 avg = 74.63\n\npi@raspberrypi:~/ncnn/build/benchmark $ ./benchncnn 10 4 0 -1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 7.61 max = 7.76 avg = 7.70\n squeezenet_int8 min = 7.97 max = 8.68 avg = 8.23\n mobilenet min = 9.65 max = 9.91 avg = 9.80\n mobilenet_int8 min = 10.60 max = 36.93 avg = 13.29\n mobilenet_v2 min = 12.25 max = 12.64 avg = 12.40\n mobilenet_v3 min = 8.14 max = 8.26 avg = 8.20\n shufflenet min = 3.72 max = 3.82 avg = 3.77\n shufflenet_v2 min = 2.99 max = 3.10 avg = 3.05\n mnasnet min = 7.27 max = 7.46 avg = 7.37\n proxylessnasnet min = 8.39 max = 8.55 avg = 8.48\n efficientnet_b0 min = 13.15 max = 13.59 avg = 13.39\n efficientnetv2_b0 min = 14.79 max = 15.30 avg = 14.91\n regnety_400m min = 9.49 max = 9.71 avg = 9.57\n blazeface min = 1.41 max = 1.46 avg = 1.43\n googlenet min = 28.60 max = 28.87 avg = 28.73\n googlenet_int8 min = 27.09 max = 27.77 avg = 27.47\n resnet18 min = 21.47 max = 21.88 avg = 21.65\n resnet18_int8 min = 20.07 max = 20.30 avg = 20.24\n alexnet min = 22.75 max = 23.47 avg = 23.05\n vgg16 min = 154.32 max = 158.51 avg = 157.40\n vgg16_int8 min = 127.78 max = 162.60 avg = 133.21\n resnet50 min = 49.36 max = 49.86 avg = 49.63\n resnet50_int8 min = 46.44 max = 46.89 avg = 46.74\n squeezenet_ssd min = 37.31 max = 74.95 avg = 41.30\n squeezenet_ssd_int8 min = 32.62 max = 33.63 avg = 33.09\n mobilenet_ssd min = 27.40 max = 27.99 avg = 27.68\n mobilenet_ssd_int8 min = 26.70 max = 27.71 avg = 27.23\n mobilenet_yolo min = 60.25 max = 61.10 avg = 60.67\n mobilenetv2_yolov3 min = 43.51 max = 44.29 avg = 43.87\n yolov4-tiny min = 51.63 max = 52.64 avg = 52.24\n nanodet_m min = 11.89 max = 12.06 avg = 11.97\n yolo-fastest-1.1 min = 5.63 max = 5.78 avg = 5.69\n yolo-fastestv2 min = 5.34 max = 5.48 avg = 5.40\n vision_transformer min = 481.78 max = 506.72 avg = 493.05\n FastestDet min = 4.91 max = 5.14 avg = 5.01\n```\n### Raspberry Pi Zero 2 W Broadcom BCM2710A1, Cortex-A53 (ARMv8) (1.0GHz x 4)\n\n```\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 119.52 max = 120.29 avg = 119.93\n squeezenet_int8 min = 96.32 max = 96.96 avg = 96.55\n mobilenet min = 162.60 max = 165.49 avg = 163.19\n mobilenet_int8 min = 90.78 max = 91.39 avg = 91.03\n mobilenet_v2 min = 145.71 max = 148.83 avg = 147.39\n mobilenet_v3 min = 113.89 max = 151.95 avg = 119.04\n shufflenet min = 72.72 max = 73.27 avg = 72.96\n shufflenet_v2 min = 63.64 max = 64.50 avg = 64.13\n mnasnet min = 126.07 max = 126.93 avg = 126.53\n proxylessnasnet min = 139.90 max = 140.84 avg = 140.35\n efficientnet_b0 min = 201.88 max = 202.55 avg = 202.14\n efficientnetv2_b0 min = 227.22 max = 228.84 avg = 228.09\n regnety_400m min = 156.49 max = 157.47 avg = 156.96\n blazeface min = 22.79 max = 23.28 avg = 23.10\n googlenet min = 323.74 max = 324.90 avg = 324.45\n googlenet_int8 min = 250.86 max = 252.82 avg = 251.63\n resnet18 min = 351.37 max = 355.67 avg = 353.45\n resnet18_int8 min = 194.83 max = 196.68 avg = 195.51\n alexnet min = 271.18 max = 273.53 avg = 272.18\n resnet50 min = 777.44 max = 797.47 avg = 782.63\n resnet50_int8 min = 496.78 max = 498.86 avg = 497.57\n squeezenet_ssd min = 376.10 max = 382.41 avg = 379.13\n squeezenet_ssd_int8 min = 255.99 max = 257.57 avg = 256.78\n mobilenet_ssd min = 338.64 max = 339.93 avg = 339.50\n mobilenet_ssd_int8 min = 190.24 max = 190.68 avg = 190.48\n mobilenet_yolo min = 746.83 max = 748.14 avg = 747.53\n mobilenetv2_yolov3 min = 487.99 max = 491.18 avg = 489.37\n yolov4-tiny min = 644.73 max = 652.24 avg = 646.64\n nanodet_m min = 165.27 max = 167.12 avg = 166.27\n yolo-fastest-1.1 min = 98.74 max = 100.02 avg = 99.17\n yolo-fastestv2 min = 80.52 max = 81.86 avg = 81.29\n\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 240.53 max = 241.07 avg = 240.77\n squeezenet_int8 min = 212.63 max = 213.23 avg = 212.94\n mobilenet min = 393.79 max = 394.04 avg = 393.94\n mobilenet_int8 min = 286.58 max = 286.95 avg = 286.75\n mobilenet_v2 min = 273.97 max = 274.51 avg = 274.23\n mobilenet_v3 min = 233.77 max = 234.59 avg = 234.20\n shufflenet min = 133.05 max = 133.36 avg = 133.23\n shufflenet_v2 min = 128.86 max = 129.47 avg = 129.18\n mnasnet min = 265.70 max = 266.17 avg = 265.93\n proxylessnasnet min = 329.78 max = 330.54 avg = 330.13\n efficientnet_b0 min = 518.42 max = 519.38 avg = 519.00\n efficientnetv2_b0 min = 594.37 max = 595.17 avg = 594.74\n regnety_400m min = 329.53 max = 330.44 avg = 329.87\n blazeface min = 42.24 max = 45.56 avg = 43.96\n googlenet min = 780.05 max = 780.63 avg = 780.39\n googlenet_int8 min = 663.83 max = 664.43 avg = 664.15\n resnet18 min = 653.62 max = 657.59 avg = 654.69\n resnet18_int8 min = 479.03 max = 479.72 avg = 479.40\n alexnet min = 687.99 max = 690.34 avg = 689.15\n resnet50 min = 1800.97 max = 1806.11 avg = 1802.79\n resnet50_int8 min = 1311.68 max = 1314.56 avg = 1313.15\n squeezenet_ssd min = 563.63 max = 565.57 avg = 564.44\n squeezenet_ssd_int8 min = 481.24 max = 483.97 avg = 482.20\n mobilenet_ssd min = 799.21 max = 829.10 avg = 803.56\n mobilenet_ssd_int8 min = 568.11 max = 568.88 avg = 568.42\n mobilenet_yolo min = 1815.60 max = 1816.44 avg = 1815.93\n mobilenetv2_yolov3 min = 951.34 max = 952.15 avg = 951.72\n yolov4-tiny min = 1258.21 max = 1259.49 avg = 1258.66\n nanodet_m min = 301.04 max = 304.09 avg = 301.70\n yolo-fastest-1.1 min = 155.04 max = 155.98 avg = 155.53\n yolo-fastestv2 min = 126.77 max = 127.40 avg = 127.05\n```\n\n### Banana Pi M2 Zero 2 AllWinner H2+, Cortex-A7 (ARMv7-A) (1.2GHz x 4)\n\n```\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 230.97 max = 232.18 avg = 231.49\n squeezenet_int8 min = 171.12 max = 172.87 avg = 171.68\n mobilenet min = 327.65 max = 340.92 avg = 329.88\n mobilenet_int8 min = 166.58 max = 169.55 avg = 167.47\n mobilenet_v2 min = 276.81 max = 278.67 avg = 277.55\n mobilenet_v3 min = 220.74 max = 225.14 avg = 222.08\n shufflenet min = 147.97 max = 157.68 avg = 149.40\n shufflenet_v2 min = 146.56 max = 154.90 avg = 148.25\n mnasnet min = 243.06 max = 244.47 avg = 243.80\n proxylessnasnet min = 260.38 max = 261.47 avg = 260.66\n efficientnet_b0 min = 368.98 max = 371.03 avg = 369.96\n efficientnetv2_b0 min = 433.96 max = 459.25 avg = 437.52\n regnety_400m min = 307.53 max = 312.29 avg = 308.68\n blazeface min = 46.54 max = 47.35 avg = 46.98\n googlenet min = 647.86 max = 669.20 avg = 651.19\n googlenet_int8 min = 439.90 max = 442.35 avg = 441.38\n resnet18 min = 642.53 max = 856.58 avg = 698.28\n resnet18_int8 min = 352.10 max = 354.51 avg = 353.44\n alexnet min = 593.16 max = 624.20 avg = 598.66\n resnet50 min = 1556.12 max = 1782.22 avg = 1606.86\n resnet50_int8 min = 911.63 max = 999.42 avg = 924.37\n squeezenet_ssd min = 653.85 max = 658.07 avg = 655.19\n squeezenet_ssd_int8 min = 456.26 max = 467.76 avg = 459.87\n mobilenet_ssd min = 671.93 max = 682.64 avg = 674.88\n mobilenet_ssd_int8 min = 347.18 max = 349.07 avg = 347.81\n mobilenet_yolo min = 1471.16 max = 1492.65 avg = 1479.30\n mobilenetv2_yolov3 min = 895.90 max = 906.60 avg = 899.74\n yolov4-tiny min = 1178.53 max = 1205.79 avg = 1183.98\n nanodet_m min = 358.89 max = 366.07 avg = 362.20\n yolo-fastest-1.1 min = 189.93 max = 192.18 avg = 190.91\n yolo-fastestv2 min = 158.60 max = 161.33 avg = 159.43\n\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 602.97 max = 604.97 avg = 603.46\n squeezenet_int8 min = 431.18 max = 432.42 avg = 431.77\n mobilenet min = 971.52 max = 986.64 avg = 974.04\n mobilenet_int8 min = 556.74 max = 556.98 avg = 556.84\n mobilenet_v2 min = 682.85 max = 684.17 avg = 683.34\n mobilenet_v3 min = 585.10 max = 585.76 avg = 585.57\n shufflenet min = 340.64 max = 342.63 avg = 341.26\n shufflenet_v2 min = 322.41 max = 324.13 avg = 323.35\n mnasnet min = 644.30 max = 645.93 avg = 644.71\n proxylessnasnet min = 732.50 max = 733.30 avg = 732.96\n efficientnet_b0 min = 1084.70 max = 1094.98 avg = 1086.52\n efficientnetv2_b0 min = 1282.27 max = 1283.67 avg = 1282.60\n regnety_400m min = 764.60 max = 768.54 avg = 765.30\n blazeface min = 100.48 max = 106.28 avg = 103.33\n googlenet min = 1878.69 max = 1883.96 avg = 1880.76\n googlenet_int8 min = 1274.31 max = 1296.02 avg = 1279.59\n resnet18 min = 1837.91 max = 1843.95 avg = 1839.17\n resnet18_int8 min = 1011.98 max = 1014.43 avg = 1013.01\n alexnet min = 1997.59 max = 2001.81 avg = 1999.42\n resnet50 min = 4844.31 max = 4857.05 avg = 4847.80\n resnet50_int8 min = 2792.59 max = 2810.08 avg = 2797.30\n squeezenet_ssd min = 1438.96 max = 1443.31 avg = 1441.09\n squeezenet_ssd_int8 min = 1046.76 max = 1053.00 avg = 1049.22\n mobilenet_ssd min = 2018.66 max = 2023.70 avg = 2019.67\n mobilenet_ssd_int8 min = 1129.16 max = 1130.62 avg = 1129.82\n mobilenet_yolo min = 4724.90 max = 4728.57 avg = 4726.41\n mobilenetv2_yolov3 min = 2410.67 max = 2427.95 avg = 2413.89\n yolov4-tiny min = 3177.27 max = 3185.52 avg = 3179.71\n nanodet_m min = 761.38 max = 768.79 avg = 766.53\n yolo-fastest-1.1 min = 391.82 max = 393.32 avg = 392.39\n yolo-fastestv2 min = 316.93 max = 319.86 avg = 318.33\n```\n\n### Radxa Orion O6 (Big Cortex‑A720 2.6Ghz x4 + Medium Cortex‑A720 x 4 + Little Cortex‑A520 x 4 + Arm Immortals G720 MC10 GPU 1.1Ghz)\n\n```\nradxa@orion-o6:~/ncnn/build/benchmark$ ./benchncnn 4 1 2 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 8.52 max = 8.53 avg = 8.53\n squeezenet_int8 min = 6.49 max = 6.50 avg = 6.50\n mobilenet min = 15.56 max = 15.61 avg = 15.58\n mobilenet_int8 min = 8.68 max = 8.70 avg = 8.69\n mobilenet_v2 min = 9.67 max = 9.68 avg = 9.67\n mobilenet_v3 min = 8.05 max = 8.07 avg = 8.06\n shufflenet min = 5.30 max = 5.32 avg = 5.31\n shufflenet_v2 min = 5.55 max = 5.57 avg = 5.56\n mnasnet min = 9.23 max = 9.26 avg = 9.25\n proxylessnasnet min = 11.58 max = 11.58 avg = 11.58\n efficientnet_b0 min = 18.67 max = 18.68 avg = 18.67\n efficientnetv2_b0 min = 21.55 max = 21.59 avg = 21.57\n regnety_400m min = 13.02 max = 13.07 avg = 13.05\n blazeface min = 2.04 max = 2.06 avg = 2.05\n googlenet min = 35.36 max = 35.49 avg = 35.40\n googlenet_int8 min = 27.86 max = 27.97 avg = 27.91\n resnet18 min = 21.68 max = 21.74 avg = 21.70\n resnet18_int8 min = 19.07 max = 19.12 avg = 19.09\n alexnet min = 23.94 max = 24.06 avg = 24.02\n vgg16 min = 123.48 max = 124.36 avg = 123.87\n vgg16_int8 min = 139.53 max = 139.72 avg = 139.64\n resnet50 min = 68.07 max = 68.09 avg = 68.08\n resnet50_int8 min = 39.99 max = 40.07 avg = 40.03\n squeezenet_ssd min = 20.35 max = 20.43 avg = 20.38\n squeezenet_ssd_int8 min = 18.62 max = 18.69 avg = 18.67\n mobilenet_ssd min = 31.40 max = 31.56 avg = 31.48\n mobilenet_ssd_int8 min = 17.44 max = 17.54 avg = 17.49\n mobilenet_yolo min = 70.84 max = 70.94 avg = 70.88\n mobilenetv2_yolov3 min = 35.24 max = 35.30 avg = 35.28\n yolov4-tiny min = 42.96 max = 43.02 avg = 42.99\n nanodet_m min = 13.05 max = 13.11 avg = 13.08\n yolo-fastest-1.1 min = 5.21 max = 5.22 avg = 5.22\n yolo-fastestv2 min = 4.48 max = 4.50 avg = 4.49\n vision_transformer min = 1001.70 max = 1002.06 avg = 1001.90\n FastestDet min = 4.65 max = 4.67 avg = 4.66\nradxa@orion-o6:~/ncnn/build/benchmark$ ./benchncnn 4 12 2 -1 1\nloop_count = 4\nnum_threads = 12\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 38.01 max = 40.45 avg = 39.00\n squeezenet_int8 min = 45.53 max = 45.73 avg = 45.60\n mobilenet min = 33.35 max = 37.73 avg = 35.96\n mobilenet_int8 min = 33.87 max = 34.05 avg = 33.93\n mobilenet_v2 min = 57.97 max = 61.42 avg = 59.74\n mobilenet_v3 min = 65.47 max = 65.76 avg = 65.65\n shufflenet min = 110.95 max = 111.29 avg = 111.12\n shufflenet_v2 min = 63.97 max = 64.20 avg = 64.08\n mnasnet min = 56.06 max = 56.44 avg = 56.23\n proxylessnasnet min = 63.84 max = 64.36 avg = 64.10\n efficientnet_b0 min = 94.52 max = 94.79 avg = 94.65\n efficientnetv2_b0 min = 154.39 max = 158.08 avg = 156.57\n regnety_400m min = 454.18 max = 457.25 avg = 455.08\n blazeface min = 44.79 max = 45.03 avg = 44.92\n googlenet min = 91.22 max = 93.72 avg = 92.01\n googlenet_int8 min = 115.45 max = 118.36 avg = 116.69\n resnet18 min = 42.81 max = 50.61 avg = 45.62\n resnet18_int8 min = 45.26 max = 47.70 avg = 46.52\n alexnet min = 25.74 max = 28.83 avg = 26.66\n vgg16 min = 61.15 max = 64.72 avg = 63.09\n vgg16_int8 min = 67.75 max = 73.18 avg = 69.38\n resnet50 min = 90.29 max = 100.58 avg = 96.62\n resnet50_int8 min = 92.35 max = 97.42 avg = 94.64\n squeezenet_ssd min = 105.26 max = 111.83 avg = 107.89\n squeezenet_ssd_int8 min = 117.49 max = 121.57 avg = 118.91\n mobilenet_ssd min = 89.79 max = 95.18 avg = 92.15\n mobilenet_ssd_int8 min = 97.02 max = 103.84 avg = 99.86\n mobilenet_yolo min = 603.04 max = 606.87 avg = 605.03\n mobilenetv2_yolov3 min = 75.32 max = 80.43 avg = 76.83\n yolov4-tiny min = 51.46 max = 60.43 avg = 56.32\n nanodet_m min = 104.05 max = 109.94 avg = 107.06\n yolo-fastest-1.1 min = 90.31 max = 90.50 avg = 90.41\n yolo-fastestv2 min = 94.72 max = 96.62 avg = 95.52\n vision_transformer min = 323.38 max = 333.42 avg = 329.50\n FastestDet min = 80.86 max = 83.37 avg = 81.84\nradxa@orion-o6:~/ncnn/build/benchmark$ ./benchncnn 4 1 2 0 1\n[0 Mali-G720-Immortalis] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G720-Immortalis] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Mali-G720-Immortalis] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[0 Mali-G720-Immortalis] subgroup=16 basic/vote/ballot/shuffle=1/1/1/1\n[0 Mali-G720-Immortalis] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 16.33 max = 16.59 avg = 16.45\n squeezenet_int8 min = 6.36 max = 10.08 avg = 7.32\n mobilenet min = 3.45 max = 27.79 avg = 14.90\n mobilenet_int8 min = 8.71 max = 8.76 avg = 8.74\n mobilenet_v2 min = 4.31 max = 4.43 avg = 4.40\n mobilenet_v3 min = 19.81 max = 19.86 avg = 19.83\n shufflenet min = 14.76 max = 14.83 avg = 14.79\n shufflenet_v2 min = 15.24 max = 15.33 avg = 15.28\n mnasnet min = 3.71 max = 10.64 avg = 5.55\n proxylessnasnet min = 4.82 max = 4.95 avg = 4.90\n efficientnet_b0 min = 6.58 max = 6.62 avg = 6.60\n efficientnetv2_b0 min = 56.26 max = 57.46 avg = 56.82\n regnety_400m min = 5.30 max = 30.08 avg = 17.72\n blazeface min = 4.36 max = 4.52 avg = 4.46\n googlenet min = 9.03 max = 9.07 avg = 9.05\n googlenet_int8 min = 27.90 max = 27.94 avg = 27.92\n resnet18 min = 6.47 max = 28.26 avg = 11.93\n resnet18_int8 min = 19.79 max = 19.83 avg = 19.81\n alexnet min = 7.76 max = 7.81 avg = 7.77\n vgg16 min = 27.58 max = 27.90 avg = 27.77\n vgg16_int8 min = 143.28 max = 144.19 avg = 143.68\n resnet50 min = 14.06 max = 14.22 avg = 14.15\n resnet50_int8 min = 41.37 max = 41.48 avg = 41.43\n squeezenet_ssd min = 11.11 max = 60.31 avg = 47.93\n squeezenet_ssd_int8 min = 19.29 max = 19.39 avg = 19.35\n mobilenet_ssd min = 8.78 max = 8.88 avg = 8.82\n mobilenet_ssd_int8 min = 17.60 max = 17.66 avg = 17.62\n mobilenet_yolo min = 13.64 max = 13.91 avg = 13.76\n mobilenetv2_yolov3 min = 11.97 max = 15.79 avg = 14.01\n yolov4-tiny min = 26.72 max = 32.41 avg = 28.27\n nanodet_m min = 9.84 max = 13.42 avg = 10.76\n yolo-fastest-1.1 min = 15.38 max = 15.62 avg = 15.56\n yolo-fastestv2 min = 13.56 max = 13.67 avg = 13.61\n vision_transformer min = 831.86 max = 835.66 avg = 833.83\n FastestDet min = 13.85 max = 13.92 avg = 13.88\n```\n\n### Radxa Zero 3W, Cortex-A55 (ARMv82) (1.416 GHz x 4)\n```\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 34.51 max = 106.19 avg = 79.43\n squeezenet_int8 min = 31.48 max = 49.87 avg = 34.65\n mobilenet min = 42.23 max = 45.36 avg = 42.89\n mobilenet_int8 min = 35.97 max = 53.84 avg = 38.77\n mobilenet_v2 min = 39.61 max = 40.35 avg = 40.00\n mobilenet_v3 min = 31.19 max = 31.85 avg = 31.50\n shufflenet min = 24.75 max = 27.74 avg = 25.55\n shufflenet_v2 min = 22.00 max = 22.70 avg = 22.31\n mnasnet min = 34.95 max = 53.55 avg = 37.39\n proxylessnasnet min = 39.96 max = 44.32 avg = 40.81\n efficientnet_b0 min = 49.76 max = 67.77 avg = 52.61\n efficientnetv2_b0 min = 64.00 max = 85.78 avg = 67.06\n regnety_400m min = 55.23 max = 73.22 avg = 57.87\n blazeface min = 7.80 max = 10.39 avg = 8.27\n googlenet min = 98.24 max = 118.27 avg = 101.78\n googlenet_int8 min = 98.81 max = 115.66 avg = 101.52\n resnet18 min = 75.33 max = 88.59 avg = 78.19\n resnet18_int8 min = 76.31 max = 95.17 avg = 79.03\n alexnet min = 65.07 max = 73.80 avg = 67.18\n vgg16 min = 423.20 max = 455.15 avg = 436.32\n vgg16_int8 min = 591.82 max = 620.22 avg = 607.55\n resnet50 min = 185.53 max = 207.10 avg = 193.03\n resnet50_int8 min = 176.84 max = 194.73 avg = 181.81\n squeezenet_ssd min = 96.64 max = 118.46 avg = 100.86\n squeezenet_ssd_int8 min = 96.61 max = 123.48 avg = 104.64\n mobilenet_ssd min = 95.38 max = 110.52 avg = 98.61\n mobilenet_ssd_int8 min = 76.21 max = 95.41 avg = 79.10\n mobilenet_yolo min = 210.73 max = 235.47 avg = 221.72\n mobilenetv2_yolov3 min = 134.59 max = 154.33 avg = 139.54\n yolov4-tiny min = 167.79 max = 191.60 avg = 171.25\n nanodet_m min = 63.22 max = 80.73 avg = 66.25\n yolo-fastest-1.1 min = 32.87 max = 88.05 avg = 47.36\n yolo-fastestv2 min = 26.03 max = 27.01 avg = 26.54\n vision_transformer min = 3682.51 max = 3882.79 avg = 3809.42\n FastestDet min = 30.69 max = 50.65 avg = 33.65\n```\n\n### Avaota Aim T527, Allwinner T527 (Cortex-A55 2.2GHz x 4 + Cortex-A55 1.8GHz x 4)\n\n```\n./benchncnn 4 4 2 -1 1\nloop_count = 4\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 14.15 max = 14.21 avg = 14.17\n squeezenet_int8 min = 21.05 max = 21.12 avg = 21.09\n mobilenet min = 19.22 max = 19.30 avg = 19.25\n mobilenet_int8 min = 18.65 max = 19.52 avg = 19.07\n mobilenet_v2 min = 20.23 max = 21.01 avg = 20.63\n mobilenet_v3 min = 15.34 max = 15.48 avg = 15.41\n shufflenet min = 10.30 max = 10.37 avg = 10.33\n shufflenet_v2 min = 9.18 max = 9.34 avg = 9.23\n mnasnet min = 15.58 max = 15.62 avg = 15.60\n proxylessnasnet min = 19.64 max = 19.73 avg = 19.67\n efficientnet_b0 min = 25.62 max = 25.81 avg = 25.69\n efficientnetv2_b0 min = 36.95 max = 37.46 avg = 37.17\n regnety_400m min = 23.75 max = 24.13 avg = 23.90\n blazeface min = 3.37 max = 3.42 avg = 3.40\n googlenet min = 57.36 max = 58.32 avg = 57.88\n googlenet_int8 min = 60.80 max = 62.30 avg = 61.50\n resnet18 min = 39.99 max = 40.34 avg = 40.17\n resnet18_int8 min = 54.18 max = 56.08 avg = 55.16\n alexnet min = 41.87 max = 42.21 avg = 42.08\n vgg16 min = 260.14 max = 260.94 avg = 260.51\n vgg16_int8 min = 347.42 max = 348.90 avg = 348.30\n resnet50 min = 90.91 max = 91.26 avg = 91.07\n resnet50_int8 min = 121.94 max = 122.56 avg = 122.28\n squeezenet_ssd min = 57.11 max = 57.57 avg = 57.37\n squeezenet_ssd_int8 min = 74.70 max = 75.18 avg = 74.91\n mobilenet_ssd min = 49.60 max = 49.96 avg = 49.71\n mobilenet_ssd_int8 min = 49.45 max = 49.93 avg = 49.63\n mobilenet_yolo min = 114.98 max = 115.37 avg = 115.18\n mobilenetv2_yolov3 min = 75.74 max = 75.97 avg = 75.87\n yolov4-tiny min = 99.09 max = 99.43 avg = 99.25\n nanodet_m min = 29.40 max = 29.77 avg = 29.60\n yolo-fastest-1.1 min = 13.78 max = 13.85 avg = 13.82\n yolo-fastestv2 min = 12.91 max = 13.10 avg = 12.98\n vision_transformer min = 1641.78 max = 1648.71 avg = 1646.65\n FastestDet min = 12.24 max = 12.61 avg = 12.42\n\n```\n\n\n### Khadas VIM3, Amlogic A311D (Cortex-A73 2.2GHz x 4 + Cortex-A53 1.8GHz x 2)\n\n```\nvim3:/data/local/tmp # ./benchncnn 8 4 2 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 30.98 max = 31.26 avg = 31.09\n squeezenet_int8 min = 24.70 max = 24.84 avg = 24.78\n mobilenet min = 42.57 max = 43.37 avg = 42.96\n mobilenet_int8 min = 22.33 max = 22.52 avg = 22.44\n mobilenet_v2 min = 39.36 max = 39.77 avg = 39.56\n mobilenet_v3 min = 30.13 max = 30.45 avg = 30.28\n shufflenet min = 21.62 max = 21.94 avg = 21.80\n shufflenet_v2 min = 18.83 max = 19.24 avg = 19.05\n mnasnet min = 33.54 max = 34.08 avg = 33.80\n proxylessnasnet min = 35.81 max = 36.05 avg = 35.95\n efficientnet_b0 min = 53.82 max = 54.44 avg = 54.21\n efficientnetv2_b0 min = 62.20 max = 62.60 avg = 62.43\n regnety_400m min = 48.82 max = 49.27 avg = 49.05\n blazeface min = 6.34 max = 6.51 avg = 6.43\n googlenet min = 81.96 max = 82.53 avg = 82.23\n googlenet_int8 min = 64.42 max = 65.00 avg = 64.77\n resnet18 min = 77.00 max = 77.83 avg = 77.46\n resnet18_int8 min = 48.91 max = 49.14 avg = 49.05\n alexnet min = 60.43 max = 60.93 avg = 60.69\n vgg16 min = 414.89 max = 423.00 avg = 418.75\n vgg16_int8 min = 245.58 max = 246.37 avg = 245.94\n resnet50 min = 185.53 max = 187.35 avg = 186.18\n resnet50_int8 min = 123.36 max = 124.75 avg = 124.17\n squeezenet_ssd min = 85.87 max = 86.42 avg = 86.23\n squeezenet_ssd_int8 min = 64.90 max = 65.24 avg = 65.08\n mobilenet_ssd min = 88.32 max = 90.02 avg = 89.10\n mobilenet_ssd_int8 min = 46.85 max = 47.18 avg = 46.98\n mobilenet_yolo min = 192.33 max = 195.38 avg = 194.10\n mobilenetv2_yolov3 min = 127.33 max = 128.58 avg = 127.96\n yolov4-tiny min = 150.44 max = 152.02 avg = 151.20\n nanodet_m min = 54.22 max = 54.61 avg = 54.37\n yolo-fastest-1.1 min = 28.13 max = 28.76 avg = 28.40\n yolo-fastestv2 min = 22.10 max = 22.26 avg = 22.19\n\nvim3:/data/local/tmp # ./benchncnn 4 1 2 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 68.25 max = 68.85 avg = 68.67\n squeezenet_int8 min = 51.92 max = 52.08 avg = 52.01\n mobilenet min = 112.69 max = 113.72 avg = 113.33\n mobilenet_int8 min = 66.43 max = 66.89 avg = 66.68\n mobilenet_v2 min = 81.36 max = 81.77 avg = 81.62\n mobilenet_v3 min = 62.33 max = 63.39 avg = 62.94\n shufflenet min = 37.84 max = 38.03 avg = 37.93\n shufflenet_v2 min = 37.33 max = 38.08 avg = 37.68\n mnasnet min = 73.83 max = 74.32 avg = 74.03\n proxylessnasnet min = 85.19 max = 86.43 avg = 85.84\n efficientnet_b0 min = 138.68 max = 139.69 avg = 139.19\n efficientnetv2_b0 min = 167.53 max = 167.99 avg = 167.75\n regnety_400m min = 94.78 max = 95.81 avg = 95.21\n blazeface min = 11.22 max = 11.43 avg = 11.28\n googlenet min = 229.35 max = 230.91 avg = 229.89\n googlenet_int8 min = 173.04 max = 173.48 avg = 173.24\n resnet18 min = 191.54 max = 193.78 avg = 192.49\n resnet18_int8 min = 132.97 max = 133.51 avg = 133.25\n alexnet min = 140.31 max = 141.95 avg = 141.18\n vgg16 min = 1093.71 max = 1100.95 avg = 1097.64\n vgg16_int8 min = 734.44 max = 736.16 avg = 735.05\n resnet50 min = 530.38 max = 533.93 avg = 531.87\n resnet50_int8 min = 332.88 max = 334.22 avg = 333.71\n squeezenet_ssd min = 159.08 max = 160.98 avg = 160.16\n squeezenet_ssd_int8 min = 126.97 max = 127.96 avg = 127.43\n mobilenet_ssd min = 238.92 max = 241.14 avg = 239.70\n mobilenet_ssd_int8 min = 135.57 max = 136.02 avg = 135.78\n mobilenet_yolo min = 539.59 max = 543.88 avg = 541.90\n mobilenetv2_yolov3 min = 281.32 max = 285.05 avg = 283.24\n yolov4-tiny min = 381.99 max = 384.93 avg = 383.53\n nanodet_m min = 98.32 max = 98.85 avg = 98.60\n yolo-fastest-1.1 min = 44.59 max = 44.95 avg = 44.80\n yolo-fastestv2 min = 36.88 max = 37.11 avg = 36.98\n\nvim3:/data/local/tmp $ ./benchncnn 8 6 2 0 1 \n[0 Mali-G52] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G52] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=1\n[0 Mali-G52] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/0/0/0\n[0 Mali-G52] subgroup=8 basic/vote/ballot/shuffle=1/0/0/0\n[0 Mali-G52] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 8\nnum_threads = 6\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 21.29 max = 21.81 avg = 21.56\n squeezenet_int8 min = 37.59 max = 37.85 avg = 37.70\n mobilenet min = 32.08 max = 32.61 avg = 32.42\n mobilenet_int8 min = 40.12 max = 40.46 avg = 40.28\n mobilenet_v2 min = 24.55 max = 24.67 avg = 24.62\n mobilenet_v3 min = 25.35 max = 25.60 avg = 25.47\n shufflenet min = 18.78 max = 89.48 avg = 35.41\n shufflenet_v2 min = 21.15 max = 21.33 avg = 21.22\n mnasnet min = 25.08 max = 25.31 avg = 25.21\n proxylessnasnet min = 26.97 max = 27.18 avg = 27.05\n efficientnet_b0 min = 40.70 max = 40.91 avg = 40.81\n efficientnetv2_b0 min = 189.26 max = 192.84 avg = 191.33\n regnety_400m min = 30.88 max = 31.17 avg = 31.03\n blazeface min = 24.34 max = 24.52 avg = 24.45\n googlenet min = 67.14 max = 67.43 avg = 67.30\n googlenet_int8 min = 98.06 max = 98.57 avg = 98.35\n resnet18 min = 61.13 max = 61.63 avg = 61.44\n resnet18_int8 min = 72.63 max = 73.48 avg = 73.01\n alexnet min = 68.88 max = 70.34 avg = 69.71\n vgg16 min = 347.48 max = 348.48 avg = 347.94\n vgg16_int8 min = 342.50 max = 357.78 avg = 353.13\n resnet50 min = 158.90 max = 160.10 avg = 159.76\n resnet50_int8 min = 211.35 max = 212.68 avg = 212.11\n squeezenet_ssd min = 81.61 max = 82.17 avg = 81.91\n squeezenet_ssd_int8 min = 85.52 max = 85.98 avg = 85.79\n mobilenet_ssd min = 73.38 max = 74.41 avg = 74.02\n mobilenet_ssd_int8 min = 85.13 max = 91.47 avg = 86.13\n mobilenet_yolo min = 154.47 max = 155.23 avg = 154.74\n mobilenetv2_yolov3 min = 100.75 max = 101.96 avg = 101.27\n yolov4-tiny min = 140.52 max = 161.68 avg = 153.85\n nanodet_m min = 85.27 max = 110.53 avg = 94.81\n yolo-fastest-1.1 min = 23.56 max = 42.04 avg = 33.10\n yolo-fastestv2 min = 19.54 max = 21.66 avg = 21.01\n vision_transformer min = 6395.34 max = 6418.70 avg = 6410.43\n FastestDet min = 21.53 max = 23.21 avg = 22.98\n```\n\n### Rockchip RK3588 (Cortex-A76 2.4GHz x 4 + Cortex-A55 1.8GHz x 4)\n\n```\nrk3588_s:/data/local/tmp # ./benchncnn 8 4 2 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 7.57 max = 7.68 avg = 7.60\n squeezenet_int8 min = 8.43 max = 8.52 avg = 8.46\n mobilenet min = 11.01 max = 11.08 avg = 11.05\n mobilenet_int8 min = 8.89 max = 8.96 avg = 8.91\n mobilenet_v2 min = 8.73 max = 8.78 avg = 8.76\n mobilenet_v3 min = 7.90 max = 7.95 avg = 7.92\n shufflenet min = 7.95 max = 8.02 avg = 7.99\n shufflenet_v2 min = 6.09 max = 6.13 avg = 6.11\n mnasnet min = 8.30 max = 8.35 avg = 8.33\n proxylessnasnet min = 9.67 max = 9.72 avg = 9.69\n efficientnet_b0 min = 17.51 max = 17.60 avg = 17.56\n efficientnetv2_b0 min = 28.10 max = 28.17 avg = 28.14\n regnety_400m min = 16.33 max = 16.39 avg = 16.35\n blazeface min = 2.81 max = 2.89 avg = 2.83\n googlenet min = 33.33 max = 33.41 avg = 33.37\n googlenet_int8 min = 33.62 max = 33.87 avg = 33.77\n resnet18 min = 18.83 max = 18.90 avg = 18.86\n resnet18_int8 min = 33.92 max = 34.10 avg = 34.00\n alexnet min = 29.07 max = 29.11 avg = 29.09\n vgg16 min = 106.86 max = 107.40 avg = 107.06\n vgg16_int8 min = 283.66 max = 284.16 avg = 283.94\n resnet50 min = 53.70 max = 54.21 avg = 53.83\n resnet50_int8 min = 66.11 max = 66.24 avg = 66.15\n squeezenet_ssd min = 34.88 max = 35.04 avg = 34.99\n squeezenet_ssd_int8 min = 43.25 max = 43.62 avg = 43.37\n mobilenet_ssd min = 31.32 max = 31.42 avg = 31.37\n mobilenet_ssd_int8 min = 26.11 max = 26.18 avg = 26.13\n mobilenet_yolo min = 58.89 max = 59.02 avg = 58.95\n mobilenetv2_yolov3 min = 37.53 max = 37.64 avg = 37.58\n yolov4-tiny min = 52.95 max = 53.31 avg = 53.03\n nanodet_m min = 16.06 max = 16.14 avg = 16.10\n yolo-fastest-1.1 min = 8.42 max = 8.47 avg = 8.45\n yolo-fastestv2 min = 7.81 max = 7.88 avg = 7.84\n\nrk3588_s:/data/local/tmp # ./benchncnn 8 1 2 -1 1\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 25.04 max = 25.14 avg = 25.07\n squeezenet_int8 min = 26.29 max = 26.38 avg = 26.33\n mobilenet min = 41.17 max = 41.23 avg = 41.19\n mobilenet_int8 min = 32.51 max = 32.57 avg = 32.54\n mobilenet_v2 min = 27.27 max = 27.31 avg = 27.29\n mobilenet_v3 min = 22.49 max = 22.54 avg = 22.51\n shufflenet min = 18.15 max = 18.22 avg = 18.18\n shufflenet_v2 min = 15.82 max = 15.86 avg = 15.85\n mnasnet min = 26.45 max = 26.50 avg = 26.47\n proxylessnasnet min = 31.60 max = 31.66 avg = 31.62\n efficientnet_b0 min = 55.53 max = 55.68 avg = 55.62\n efficientnetv2_b0 min = 96.84 max = 96.92 avg = 96.89\n regnety_400m min = 33.66 max = 33.70 avg = 33.68\n blazeface min = 8.80 max = 8.84 avg = 8.83\n googlenet min = 116.89 max = 117.06 avg = 116.97\n googlenet_int8 min = 107.92 max = 108.03 avg = 107.98\n resnet18 min = 60.97 max = 61.18 avg = 61.05\n resnet18_int8 min = 118.95 max = 119.04 avg = 119.00\n alexnet min = 93.49 max = 93.59 avg = 93.55\n vgg16 min = 333.81 max = 334.52 avg = 334.07\n vgg16_int8 min = 947.19 max = 947.55 avg = 947.35\n resnet50 min = 186.95 max = 187.42 avg = 187.15\n resnet50_int8 min = 225.72 max = 225.86 avg = 225.75\n squeezenet_ssd min = 93.29 max = 93.66 avg = 93.47\n squeezenet_ssd_int8 min = 120.22 max = 120.95 avg = 120.49\n mobilenet_ssd min = 105.84 max = 105.90 avg = 105.87\n mobilenet_ssd_int8 min = 85.95 max = 86.04 avg = 86.01\n mobilenet_yolo min = 194.22 max = 194.64 avg = 194.41\n mobilenetv2_yolov3 min = 103.63 max = 103.72 avg = 103.69\n yolov4-tiny min = 136.59 max = 137.14 avg = 136.91\n nanodet_m min = 41.40 max = 41.49 avg = 41.43\n yolo-fastest-1.1 min = 18.73 max = 18.80 avg = 18.77\n yolo-fastestv2 min = 18.25 max = 18.31 avg = 18.28\n\nrk3588_s:/data/local/tmp # ./benchncnn 8 4 1 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 25.54 max = 25.99 avg = 25.71\n squeezenet_int8 min = 30.88 max = 31.16 avg = 31.01\n mobilenet min = 36.24 max = 62.95 avg = 39.89\n mobilenet_int8 min = 31.90 max = 32.37 avg = 32.06\n mobilenet_v2 min = 27.49 max = 27.82 avg = 27.64\n mobilenet_v3 min = 26.30 max = 26.69 avg = 26.45\n shufflenet min = 25.49 max = 25.72 avg = 25.60\n shufflenet_v2 min = 21.59 max = 22.67 avg = 21.78\n mnasnet min = 27.92 max = 28.10 avg = 28.00\n proxylessnasnet min = 34.18 max = 34.42 avg = 34.28\n efficientnet_b0 min = 57.37 max = 57.60 avg = 57.45\n efficientnetv2_b0 min = 83.50 max = 84.03 avg = 83.66\n regnety_400m min = 50.83 max = 51.27 avg = 50.98\n blazeface min = 14.07 max = 14.29 avg = 14.17\n googlenet min = 100.60 max = 101.00 avg = 100.87\n googlenet_int8 min = 106.58 max = 107.14 avg = 106.71\n resnet18 min = 58.60 max = 59.62 avg = 59.00\n resnet18_int8 min = 84.90 max = 85.15 avg = 84.99\n alexnet min = 86.06 max = 86.58 avg = 86.22\n vgg16 min = 308.42 max = 309.18 avg = 308.81\n vgg16_int8 min = 543.61 max = 545.09 avg = 544.40\n resnet50 min = 163.45 max = 164.44 avg = 163.92\n resnet50_int8 min = 179.51 max = 180.16 avg = 179.83\n squeezenet_ssd min = 96.32 max = 97.24 avg = 96.71\n squeezenet_ssd_int8 min = 116.48 max = 117.65 avg = 116.85\n mobilenet_ssd min = 92.12 max = 93.09 avg = 92.55\n mobilenet_ssd_int8 min = 81.78 max = 82.42 avg = 81.95\n mobilenet_yolo min = 174.95 max = 175.40 avg = 175.15\n mobilenetv2_yolov3 min = 110.63 max = 111.05 avg = 110.81\n yolov4-tiny min = 163.37 max = 164.24 avg = 163.63\n nanodet_m min = 52.96 max = 53.59 avg = 53.12\n yolo-fastest-1.1 min = 28.98 max = 29.33 avg = 29.20\n yolo-fastestv2 min = 23.52 max = 24.16 avg = 23.76\n\nrk3588_s:/data/local/tmp # ./benchncnn 8 1 1 -1 1\nloop_count = 8\nnum_threads = 1\npowersave = 1\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 83.46 max = 83.63 avg = 83.53\n squeezenet_int8 min = 101.39 max = 102.29 avg = 101.77\n mobilenet min = 131.78 max = 132.25 avg = 131.87\n mobilenet_int8 min = 111.66 max = 112.60 avg = 111.94\n mobilenet_v2 min = 92.92 max = 227.19 avg = 132.44\n mobilenet_v3 min = 78.38 max = 78.64 avg = 78.49\n shufflenet min = 62.98 max = 63.17 avg = 63.09\n shufflenet_v2 min = 56.85 max = 57.23 avg = 57.00\n mnasnet min = 87.53 max = 87.71 avg = 87.60\n proxylessnasnet min = 113.25 max = 114.10 avg = 113.58\n efficientnet_b0 min = 180.95 max = 181.16 avg = 181.07\n efficientnetv2_b0 min = 285.34 max = 285.62 avg = 285.51\n regnety_400m min = 109.24 max = 109.36 avg = 109.31\n blazeface min = 41.12 max = 41.53 avg = 41.23\n googlenet min = 358.94 max = 359.55 avg = 359.24\n googlenet_int8 min = 371.32 max = 371.84 avg = 371.51\n resnet18 min = 209.97 max = 210.42 avg = 210.22\n resnet18_int8 min = 302.93 max = 303.51 avg = 303.26\n alexnet min = 318.95 max = 321.70 avg = 319.40\n vgg16 min = 1126.11 max = 1127.83 avg = 1126.98\n vgg16_int8 min = 2026.90 max = 2034.04 avg = 2029.35\n resnet50 min = 602.90 max = 603.70 avg = 603.30\n resnet50_int8 min = 647.33 max = 649.41 avg = 648.65\n squeezenet_ssd min = 280.60 max = 281.50 avg = 281.02\n squeezenet_ssd_int8 min = 359.41 max = 362.07 avg = 360.66\n mobilenet_ssd min = 319.11 max = 319.29 avg = 319.21\n mobilenet_ssd_int8 min = 272.16 max = 273.36 avg = 272.83\n mobilenet_yolo min = 607.07 max = 607.38 avg = 607.21\n mobilenetv2_yolov3 min = 326.66 max = 326.95 avg = 326.80\n yolov4-tiny min = 449.56 max = 450.45 avg = 450.04\n nanodet_m min = 142.09 max = 142.54 avg = 142.32\n yolo-fastest-1.1 min = 63.74 max = 63.80 avg = 63.78\n yolo-fastestv2 min = 57.56 max = 58.17 avg = 57.97\n\nrk3588_s:/data/local/tmp # ./benchncnn 8 1 2 0 0\n[0 Mali-G610] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G610] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Mali-G610] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Mali-G610] subgroup=16 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 7.09 max = 7.20 avg = 7.13\n mobilenet min = 9.16 max = 9.32 avg = 9.22\n mobilenet_v2 min = 10.18 max = 10.32 avg = 10.25\n mobilenet_v3 min = 8.01 max = 8.09 avg = 8.04\n shufflenet min = 5.88 max = 5.93 avg = 5.89\n shufflenet_v2 min = 6.30 max = 6.33 avg = 6.31\n mnasnet min = 7.91 max = 8.00 avg = 7.95\n proxylessnasnet min = 11.20 max = 11.42 avg = 11.30\n regnety_400m min = 11.65 max = 11.84 avg = 11.74\n blazeface min = 2.50 max = 2.59 avg = 2.53\n googlenet min = 17.69 max = 17.78 avg = 17.74\n resnet18 min = 16.04 max = 16.39 avg = 16.25\n alexnet min = 15.47 max = 15.66 avg = 15.56\n vgg16 min = 64.74 max = 65.42 avg = 65.04\n resnet50 min = 37.83 max = 38.31 avg = 38.12\n squeezenet_ssd min = 23.14 max = 23.44 avg = 23.26\n mobilenet_ssd min = 22.48 max = 23.01 avg = 22.74\n mobilenet_yolo min = 40.08 max = 40.72 avg = 40.32\n mobilenetv2_yolov3 min = 31.88 max = 32.57 avg = 32.12\n yolov4-tiny min = 49.64 max = 50.73 avg = 50.13\n nanodet_m min = 10.60 max = 10.70 avg = 10.64\n yolo-fastest-1.1 min = 7.63 max = 7.66 avg = 7.64\n yolo-fastestv2 min = 6.99 max = 7.02 avg = 7.00\n```\n\n### Station-M3/ROC-RK3588S-PC, Rockchip RK3588S (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz + Mali-G610) StationOS (Android)\n\n```\nroc_rk3588s_pc:/data/local/tmp # ./benchncnn 10 1 0 0 0\n./benchncnn 10 1 0 0 0\n[0 Mali-G610] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G610] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Mali-G610] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Mali-G610] subgroup=16 basic/vote/ballot/shuffle=1/1/1/1\n[0 Mali-G610] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 7.83 max = 14.17 avg = 9.76\n squeezenet_int8 min = 13.41 max = 13.52 avg = 13.45\n mobilenet min = 8.73 max = 9.68 avg = 9.07\n mobilenet_int8 min = 17.70 max = 17.89 avg = 17.80\n mobilenet_v2 min = 10.73 max = 21.20 avg = 18.93\n mobilenet_v3 min = 9.00 max = 13.36 avg = 10.64\n shufflenet min = 7.79 max = 7.93 avg = 7.85\n shufflenet_v2 min = 8.01 max = 8.06 avg = 8.03\n mnasnet min = 7.43 max = 8.71 avg = 8.28\n proxylessnasnet min = 10.56 max = 12.07 avg = 11.70\n efficientnet_b0 min = 2.15 max = 2.19 avg = 2.17\n efficientnetv2_b0 min = 0.56 max = 0.62 avg = 0.57\n regnety_400m min = 1.65 max = 1.69 avg = 1.67\n blazeface min = 0.76 max = 0.79 avg = 0.78\n googlenet min = 1.53 max = 1.60 avg = 1.56\n googlenet_int8 min = 60.85 max = 61.01 avg = 60.93\n resnet18 min = 0.63 max = 0.82 avg = 0.65\n resnet18_int8 min = 64.60 max = 65.13 avg = 64.78\n alexnet min = 0.35 max = 0.40 avg = 0.37\n vgg16 min = 0.54 max = 0.60 avg = 0.56\n vgg16_int8 min = 445.21 max = 562.09 avg = 537.10\n resnet50 min = 0.95 max = 0.97 avg = 0.96\n resnet50_int8 min = 113.02 max = 113.38 avg = 113.17\n squeezenet_ssd min = 1.94 max = 2.00 avg = 1.96\n squeezenet_ssd_int8 min = 52.09 max = 56.93 avg = 56.35\n mobilenet_ssd min = 1.19 max = 1.26 avg = 1.21\n mobilenet_ssd_int8 min = 44.33 max = 44.87 avg = 44.66\n mobilenet_yolo min = 1.05 max = 1.24 avg = 1.13\n mobilenetv2_yolov3 min = 1.18 max = 1.25 avg = 1.21\n yolov4-tiny min = 0.78 max = 0.80 avg = 0.78\n nanodet_m min = 3.43 max = 3.80 avg = 3.57\n yolo-fastest-1.1 min = 1.43 max = 1.50 avg = 1.47\n yolo-fastestv2 min = 2.03 max = 2.10 avg = 2.05\n vision_transformer min = 0.32 max = 0.36 avg = 0.35\n FastestDet min = 1.90 max = 1.95 avg = 1.93\n\nroc_rk3588s_pc:/data/local/tmp # ./benchncnn 10 1 0 -1 0\n./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 13.36 max = 13.50 avg = 13.40\n squeezenet_int8 min = 16.22 max = 16.34 avg = 16.30\n mobilenet min = 22.41 max = 22.49 avg = 22.44\n mobilenet_int8 min = 17.76 max = 17.94 avg = 17.84\n mobilenet_v2 min = 17.60 max = 17.80 avg = 17.70\n mobilenet_v3 min = 13.55 max = 13.70 avg = 13.61\n shufflenet min = 7.91 max = 7.95 avg = 7.93\n shufflenet_v2 min = 8.36 max = 8.40 avg = 8.38\n mnasnet min = 14.50 max = 14.60 avg = 14.56\n proxylessnasnet min = 16.99 max = 17.12 avg = 17.06\n efficientnet_b0 min = 26.55 max = 26.78 avg = 26.62\n efficientnetv2_b0 min = 46.96 max = 47.44 avg = 47.30\n regnety_400m min = 18.53 max = 18.63 avg = 18.58\n blazeface min = 2.98 max = 3.02 avg = 3.00\n googlenet min = 62.69 max = 63.14 avg = 62.90\n googlenet_int8 min = 60.86 max = 61.54 avg = 61.05\n resnet18 min = 30.34 max = 31.39 avg = 31.22\n resnet18_int8 min = 57.42 max = 57.67 avg = 57.56\n alexnet min = 40.81 max = 40.87 avg = 40.84\n vgg16 min = 192.71 max = 195.20 avg = 194.26\n vgg16_int8 min = 450.95 max = 534.38 avg = 482.27\n resnet50 min = 105.11 max = 105.64 avg = 105.30\n resnet50_int8 min = 105.94 max = 132.01 avg = 116.48\n squeezenet_ssd min = 51.36 max = 51.59 avg = 51.51\n squeezenet_ssd_int8 min = 69.01 max = 69.83 avg = 69.37\n mobilenet_ssd min = 53.19 max = 55.24 avg = 53.50\n mobilenet_ssd_int8 min = 44.49 max = 44.98 avg = 44.74\n mobilenet_yolo min = 112.65 max = 113.28 avg = 112.94\n mobilenetv2_yolov3 min = 63.38 max = 63.83 avg = 63.55\n yolov4-tiny min = 77.57 max = 78.20 avg = 77.90\n nanodet_m min = 25.21 max = 25.81 avg = 25.58\n yolo-fastest-1.1 min = 8.76 max = 8.84 avg = 8.80\n yolo-fastestv2 min = 8.46 max = 8.53 avg = 8.50\n vision_transformer min = 1499.53 max = 1501.32 avg = 1500.50\n FastestDet min = 7.04 max = 7.08 avg = 7.06\n```\n\n### Station P2, Rockchip RK3568 (Cortex-A55 2.0GHz x 4)\n\n```\n./benchncnn 4 4 0 -1 1\nloop_count = 4\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 26.02 max = 27.15 avg = 26.74\n squeezenet_int8 min = 44.69 max = 45.70 avg = 45.24\n mobilenet min = 32.63 max = 33.49 avg = 33.10\n mobilenet_int8 min = 44.23 max = 45.86 avg = 44.99\n mobilenet_v2 min = 31.59 max = 32.02 avg = 31.86\n mobilenet_v3 min = 25.71 max = 26.44 avg = 26.10\n shufflenet min = 22.12 max = 23.17 avg = 22.52\n shufflenet_v2 min = 17.84 max = 18.21 avg = 17.96\n mnasnet min = 28.26 max = 28.70 avg = 28.45\n proxylessnasnet min = 31.96 max = 32.25 avg = 32.13\n efficientnet_b0 min = 53.17 max = 54.48 avg = 53.60\n efficientnetv2_b0 min = 70.08 max = 70.69 avg = 70.30\n regnety_400m min = 40.80 max = 41.79 avg = 41.10\n blazeface min = 10.79 max = 11.57 avg = 11.11\n googlenet min = 83.66 max = 92.22 avg = 86.23\n googlenet_int8 min = 116.44 max = 118.34 avg = 117.08\n resnet18 min = 61.38 max = 62.52 avg = 61.94\n resnet18_int8 min = 95.58 max = 96.93 avg = 96.28\n alexnet min = 69.90 max = 70.59 avg = 70.19\n vgg16 min = 334.24 max = 343.89 avg = 337.24\n vgg16_int8 min = 464.88 max = 474.71 avg = 468.29\n resnet50 min = 141.65 max = 146.23 avg = 143.78\n resnet50_int8 min = 230.36 max = 254.75 avg = 241.24\n squeezenet_ssd min = 98.38 max = 104.60 avg = 100.50\n squeezenet_ssd_int8 min = 134.73 max = 137.88 avg = 136.12\n mobilenet_ssd min = 77.48 max = 79.92 avg = 78.64\n mobilenet_ssd_int8 min = 101.44 max = 102.61 avg = 102.06\n mobilenet_yolo min = 149.12 max = 150.14 avg = 149.76\n mobilenetv2_yolov3 min = 103.71 max = 107.81 avg = 105.69\n yolov4-tiny min = 145.75 max = 149.35 avg = 147.09\n nanodet_m min = 52.91 max = 54.06 avg = 53.53\n\n./benchncnn 4 2 0 -1 1\nloop_count = 4\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 33.78 max = 34.38 avg = 34.16\n squeezenet_int8 min = 61.66 max = 62.11 avg = 61.85\n mobilenet min = 46.53 max = 46.74 avg = 46.62\n mobilenet_int8 min = 71.06 max = 71.76 avg = 71.38\n mobilenet_v2 min = 39.05 max = 39.38 avg = 39.19\n mobilenet_v3 min = 32.20 max = 32.47 avg = 32.29\n shufflenet min = 27.13 max = 27.40 avg = 27.27\n shufflenet_v2 min = 23.38 max = 23.92 avg = 23.62\n mnasnet min = 35.51 max = 35.73 avg = 35.62\n proxylessnasnet min = 42.98 max = 43.16 avg = 43.06\n efficientnet_b0 min = 75.34 max = 75.79 avg = 75.61\n efficientnetv2_b0 min = 107.34 max = 107.83 avg = 107.60\n regnety_400m min = 47.91 max = 48.20 avg = 48.02\n blazeface min = 16.38 max = 16.63 avg = 16.49\n googlenet min = 124.27 max = 125.24 avg = 124.65\n googlenet_int8 min = 177.78 max = 178.39 avg = 178.06\n resnet18 min = 82.02 max = 82.70 avg = 82.38\n resnet18_int8 min = 148.06 max = 149.03 avg = 148.39\n alexnet min = 105.20 max = 105.91 avg = 105.54\n vgg16 min = 459.65 max = 464.94 avg = 462.02\n vgg16_int8 min = 737.54 max = 750.64 avg = 742.90\n resnet50 min = 204.44 max = 205.20 avg = 204.84\n resnet50_int8 min = 364.47 max = 366.04 avg = 365.53\n squeezenet_ssd min = 124.42 max = 128.01 avg = 125.80\n squeezenet_ssd_int8 min = 179.29 max = 183.83 avg = 181.43\n mobilenet_ssd min = 113.85 max = 115.50 avg = 114.41\n mobilenet_ssd_int8 min = 161.35 max = 162.38 avg = 161.71\n mobilenet_yolo min = 214.95 max = 216.62 avg = 215.72\n mobilenetv2_yolov3 min = 134.23 max = 136.26 avg = 135.07\n yolov4-tiny min = 194.72 max = 195.49 avg = 195.18\n nanodet_m min = 67.67 max = 68.09 avg = 67.90\n\n./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 54.31 max = 55.65 avg = 55.00\n squeezenet_int8 min = 103.96 max = 106.28 avg = 104.92\n mobilenet min = 79.02 max = 79.46 avg = 79.25\n mobilenet_int8 min = 130.06 max = 130.61 avg = 130.36\n mobilenet_v2 min = 60.15 max = 60.66 avg = 60.31\n mobilenet_v3 min = 49.40 max = 49.57 avg = 49.49\n shufflenet min = 39.39 max = 39.78 avg = 39.60\n shufflenet_v2 min = 35.48 max = 35.70 avg = 35.62\n mnasnet min = 55.38 max = 56.10 avg = 55.71\n proxylessnasnet min = 70.29 max = 70.48 avg = 70.35\n efficientnet_b0 min = 128.56 max = 129.96 avg = 129.26\n efficientnetv2_b0 min = 181.00 max = 181.56 avg = 181.24\n regnety_400m min = 67.15 max = 69.62 avg = 67.95\n blazeface min = 26.07 max = 26.58 avg = 26.33\n googlenet min = 219.19 max = 221.32 avg = 220.01\n googlenet_int8 min = 317.62 max = 319.40 avg = 318.37\n resnet18 min = 135.33 max = 136.94 avg = 135.88\n resnet18_int8 min = 264.69 max = 265.51 avg = 265.16\n alexnet min = 190.54 max = 193.50 avg = 191.88\n vgg16 min = 790.99 max = 809.24 avg = 795.85\n vgg16_int8 min = 1354.48 max = 1358.89 avg = 1357.40\n resnet50 min = 358.08 max = 362.96 avg = 360.29\n resnet50_int8 min = 667.92 max = 670.40 avg = 668.78\n squeezenet_ssd min = 193.15 max = 194.02 avg = 193.49\n squeezenet_ssd_int8 min = 291.42 max = 294.70 avg = 293.16\n mobilenet_ssd min = 189.54 max = 190.28 avg = 189.97\n mobilenet_ssd_int8 min = 289.94 max = 290.40 avg = 290.28\n mobilenet_yolo min = 370.37 max = 384.69 avg = 375.11\n mobilenetv2_yolov3 min = 210.93 max = 211.70 avg = 211.40\n yolov4-tiny min = 309.11 max = 310.74 avg = 309.89\n nanodet_m min = 100.42 max = 112.25 avg = 103.66\n```\n\n### Rock3A, Rockchip RK3568 (Cortex-A55 2.0GHz x 4) ubuntu 20.04\n\n```\nrock@rock3a:~/ncnn/build/benchmark$ ./benchncnn 8 4 0 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 29.52 max = 30.30 avg = 29.76\n squeezenet_int8 min = 35.40 max = 36.19 avg = 35.88\n mobilenet min = 34.47 max = 35.44 avg = 34.84\n mobilenet_int8 min = 34.19 max = 34.53 avg = 34.40\n mobilenet_v2 min = 35.75 max = 36.09 avg = 35.88\n mobilenet_v3 min = 28.12 max = 28.82 avg = 28.49\n shufflenet min = 23.62 max = 24.08 avg = 23.84\n shufflenet_v2 min = 19.37 max = 19.64 avg = 19.52\n mnasnet min = 30.84 max = 31.45 avg = 31.02\n proxylessnasnet min = 35.73 max = 36.07 avg = 35.90\n efficientnet_b0 min = 48.16 max = 49.29 avg = 48.64\n efficientnetv2_b0 min = 66.62 max = 67.11 avg = 66.85\n regnety_400m min = 41.11 max = 41.64 avg = 41.34\n blazeface min = 12.38 max = 12.64 avg = 12.56\n googlenet min = 86.73 max = 87.79 avg = 87.11\n googlenet_int8 min = 101.42 max = 103.87 avg = 102.55\n resnet18 min = 64.85 max = 65.84 avg = 65.23\n resnet18_int8 min = 93.55 max = 94.54 avg = 94.03\n alexnet min = 70.89 max = 73.58 avg = 71.57\n vgg16 min = 356.13 max = 358.52 avg = 357.15\n vgg16_int8 min = 521.92 max = 524.13 avg = 523.11\n resnet50 min = 147.65 max = 150.33 avg = 148.52\n resnet50_int8 min = 191.94 max = 192.73 avg = 192.30\n squeezenet_ssd min = 104.32 max = 105.75 avg = 105.00\n squeezenet_ssd_int8 min = 125.97 max = 127.53 avg = 126.70\n mobilenet_ssd min = 82.29 max = 82.65 avg = 82.47\n mobilenet_ssd_int8 min = 79.26 max = 80.93 avg = 79.72\n mobilenet_yolo min = 165.51 max = 165.86 avg = 165.72\n mobilenetv2_yolov3 min = 116.11 max = 116.83 avg = 116.43\n yolov4-tiny min = 152.09 max = 153.39 avg = 152.60\n nanodet_m min = 53.63 max = 54.14 avg = 53.92\n\nrock@rock3a:~/ncnn/build/benchmark$ ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 62.47 max = 63.04 avg = 62.84\n squeezenet_int8 min = 67.23 max = 68.48 avg = 67.93\n mobilenet min = 85.27 max = 85.69 avg = 85.49\n mobilenet_int8 min = 75.00 max = 75.48 avg = 75.26\n mobilenet_v2 min = 68.41 max = 69.09 avg = 68.76\n mobilenet_v3 min = 54.19 max = 54.52 avg = 54.34\n shufflenet min = 45.90 max = 46.30 avg = 46.09\n shufflenet_v2 min = 39.64 max = 40.07 avg = 39.91\n mnasnet min = 62.16 max = 62.41 avg = 62.30\n proxylessnasnet min = 80.79 max = 81.41 avg = 81.12\n efficientnet_b0 min = 113.47 max = 113.68 avg = 113.57\n efficientnetv2_b0 min = 167.30 max = 167.58 avg = 167.44\n regnety_400m min = 72.12 max = 72.24 avg = 72.17\n blazeface min = 31.89 max = 32.04 avg = 31.95\n googlenet min = 224.27 max = 224.86 avg = 224.55\n googlenet_int8 min = 240.02 max = 240.93 avg = 240.45\n resnet18 min = 150.25 max = 150.69 avg = 150.47\n resnet18_int8 min = 226.70 max = 228.19 avg = 227.56\n alexnet min = 197.44 max = 199.16 avg = 198.17\n vgg16 min = 859.80 max = 860.79 avg = 860.35\n vgg16_int8 min = 1409.66 max = 1411.92 avg = 1411.07\n resnet50 min = 381.04 max = 382.73 avg = 381.86\n resnet50_int8 min = 441.78 max = 445.00 avg = 443.29\n squeezenet_ssd min = 208.14 max = 208.67 avg = 208.41\n squeezenet_ssd_int8 min = 248.82 max = 250.80 avg = 249.89\n mobilenet_ssd min = 200.95 max = 201.21 avg = 201.06\n mobilenet_ssd_int8 min = 173.81 max = 174.54 avg = 174.28\n mobilenet_yolo min = 394.65 max = 395.00 avg = 394.78\n mobilenetv2_yolov3 min = 231.80 max = 232.27 avg = 232.08\n yolov4-tiny min = 321.31 max = 322.43 avg = 321.79\n nanodet_m min = 103.81 max = 104.61 avg = 104.25\n```\n\n### Station-M2/ROC-RK3566-PC, Rockchip RK3566 (Cortex-A55 1.8GHz x 4 + Mali-G52) StationOS (Android)\n\n```\nrk3566_roc_pc:/data/local/tmp # ./benchncnn 10 1 0 0 0\n./benchncnn 10 1 0 0 0\n[0 Mali-G52] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G52] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=1\n[0 Mali-G52] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Mali-G52] subgroup=8 basic/vote/ballot/shuffle=1/1/1/1\n[0 Mali-G52] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 43.67 max = 44.15 avg = 43.82\n squeezenet_int8 min = 62.72 max = 63.99 avg = 63.49\n mobilenet min = 74.32 max = 74.82 avg = 74.58\n mobilenet_int8 min = 64.42 max = 65.43 avg = 64.89\n mobilenet_v2 min = 52.96 max = 53.23 avg = 53.09\n mobilenet_v3 min = 51.55 max = 53.12 avg = 51.96\n shufflenet min = 40.73 max = 41.28 avg = 40.98\n shufflenet_v2 min = 41.56 max = 43.62 avg = 42.22\n mnasnet min = 54.37 max = 54.63 avg = 54.52\n proxylessnasnet min = 57.91 max = 59.38 avg = 58.36\n efficientnet_b0 min = 38.40 max = 40.29 avg = 39.06\n efficientnetv2_b0 min = 36.91 max = 38.45 avg = 37.72\n regnety_400m min = 69.07 max = 69.98 avg = 69.40\n blazeface min = 12.26 max = 13.08 avg = 12.57\n googlenet min = 147.08 max = 147.80 avg = 147.48\n googlenet_int8 min = 221.94 max = 225.99 avg = 223.12\n resnet18 min = 137.90 max = 138.50 avg = 138.19\n resnet18_int8 min = 187.84 max = 190.88 avg = 188.81\n alexnet min = 167.56 max = 168.92 avg = 168.17\n vgg16 min = 713.42 max = 715.20 avg = 714.51\n vgg16_int8 min = 1279.97 max = 1302.95 avg = 1294.59\n resnet50 min = 369.74 max = 375.95 avg = 372.60\n resnet50_int8 min = 391.86 max = 397.49 avg = 395.17\n squeezenet_ssd min = 155.18 max = 156.09 avg = 155.62\n squeezenet_ssd_int8 min = 218.83 max = 222.64 avg = 221.11\n mobilenet_ssd min = 161.62 max = 163.22 avg = 162.27\n mobilenet_ssd_int8 min = 147.33 max = 149.16 avg = 148.23\n mobilenet_yolo min = 344.09 max = 349.15 avg = 346.73\n mobilenetv2_yolov3 min = 168.72 max = 169.64 avg = 169.22\n yolov4-tiny min = 239.44 max = 241.11 avg = 240.00\n nanodet_m min = 88.06 max = 89.89 avg = 88.87\n yolo-fastest-1.1 min = 36.05 max = 37.86 avg = 36.47\n yolo-fastestv2 min = 34.80 max = 36.58 avg = 35.37\n vision_transformer min = 356.42 max = 359.37 avg = 358.03\n FastestDet min = 38.03 max = 38.52 avg = 38.24\n\nrk3566_roc_pc:/data/local/tmp # ./benchncnn 10 1 0 -1 0\n./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 47.01 max = 48.12 avg = 47.62\n squeezenet_int8 min = 63.30 max = 64.10 avg = 63.74\n mobilenet min = 70.24 max = 71.52 avg = 70.63\n mobilenet_int8 min = 63.90 max = 65.25 avg = 64.41\n mobilenet_v2 min = 55.75 max = 56.26 avg = 56.02\n mobilenet_v3 min = 45.56 max = 46.47 avg = 46.17\n shufflenet min = 34.16 max = 35.16 avg = 34.64\n shufflenet_v2 min = 32.58 max = 33.86 avg = 33.25\n mnasnet min = 52.43 max = 53.15 avg = 52.80\n proxylessnasnet min = 65.55 max = 67.04 avg = 66.36\n efficientnet_b0 min = 82.52 max = 82.97 avg = 82.64\n efficientnetv2_b0 min = 148.90 max = 150.47 avg = 149.64\n regnety_400m min = 63.33 max = 64.29 avg = 63.70\n blazeface min = 11.55 max = 12.35 avg = 11.77\n googlenet min = 205.85 max = 208.74 avg = 207.17\n googlenet_int8 min = 222.72 max = 225.84 avg = 223.98\n resnet18 min = 134.19 max = 136.81 avg = 135.39\n resnet18_int8 min = 187.26 max = 189.45 avg = 188.36\n alexnet min = 143.01 max = 144.97 avg = 143.42\n vgg16 min = 829.44 max = 839.46 avg = 835.37\n vgg16_int8 min = 1299.25 max = 1306.89 avg = 1301.71\n resnet50 min = 326.54 max = 330.21 avg = 328.27\n resnet50_int8 min = 391.67 max = 395.59 avg = 393.27\n squeezenet_ssd min = 166.12 max = 168.33 avg = 167.08\n squeezenet_ssd_int8 min = 221.82 max = 223.85 avg = 222.69\n mobilenet_ssd min = 163.17 max = 166.55 avg = 164.11\n mobilenet_ssd_int8 min = 146.16 max = 148.20 avg = 147.41\n mobilenet_yolo min = 335.15 max = 338.32 avg = 336.66\n mobilenetv2_yolov3 min = 193.18 max = 195.51 avg = 194.33\n yolov4-tiny min = 288.82 max = 292.16 avg = 290.36\n nanodet_m min = 98.31 max = 100.30 avg = 99.20\n yolo-fastest-1.1 min = 37.73 max = 38.97 avg = 38.40\n yolo-fastestv2 min = 36.21 max = 37.90 avg = 37.13\n vision_transformer min = 7385.59 max = 7410.59 avg = 7402.20\n FastestDet min = 34.55 max = 35.42 avg = 35.06\n```\n\n### Rockchip RK3399 (Cortex-A72 1.8GHz x 2 + Cortex-A53 1.5GHz x 4)\n\n```\nnanopc-t4:/data/local/tmp # ./benchncnn 8 2 2 -1 1\nloop_count = 8\nnum_threads = 2\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 43.73 max = 44.30 avg = 43.97\n squeezenet_int8 min = 37.92 max = 38.39 avg = 38.09\n mobilenet min = 64.28 max = 66.66 avg = 65.14\n mobilenet_int8 min = 43.17 max = 43.73 avg = 43.38\n mobilenet_v2 min = 51.30 max = 52.18 avg = 51.75\n mobilenet_v3 min = 41.51 max = 43.25 avg = 42.10\n shufflenet min = 27.43 max = 28.27 avg = 27.75\n shufflenet_v2 min = 24.96 max = 25.79 avg = 25.55\n mnasnet min = 45.44 max = 46.95 avg = 46.16\n proxylessnasnet min = 51.98 max = 53.52 avg = 52.48\n efficientnet_b0 min = 83.79 max = 84.68 avg = 84.27\n efficientnetv2_b0 min = 97.89 max = 99.27 avg = 98.55\n regnety_400m min = 65.15 max = 65.89 avg = 65.41\n blazeface min = 8.74 max = 8.89 avg = 8.80\n googlenet min = 131.46 max = 140.16 avg = 133.24\n googlenet_int8 min = 115.72 max = 118.34 avg = 116.60\n resnet18 min = 111.77 max = 113.18 avg = 112.37\n resnet18_int8 min = 84.27 max = 84.90 avg = 84.49\n alexnet min = 105.74 max = 109.87 avg = 107.15\n vgg16 min = 619.88 max = 634.59 avg = 629.15\n vgg16_int8 min = 447.14 max = 451.09 avg = 448.53\n resnet50 min = 291.51 max = 296.55 avg = 293.08\n resnet50_int8 min = 224.09 max = 227.03 avg = 225.02\n squeezenet_ssd min = 109.72 max = 112.09 avg = 110.78\n squeezenet_ssd_int8 min = 93.41 max = 94.83 avg = 93.97\n mobilenet_ssd min = 131.30 max = 132.82 avg = 131.94\n mobilenet_ssd_int8 min = 87.52 max = 88.89 avg = 88.35\n mobilenet_yolo min = 288.02 max = 289.84 avg = 288.61\n mobilenetv2_yolov3 min = 168.45 max = 170.94 avg = 169.79\n yolov4-tiny min = 217.45 max = 226.39 avg = 219.76\n nanodet_m min = 65.74 max = 66.84 avg = 66.49\n yolo-fastest-1.1 min = 32.91 max = 33.74 avg = 33.37\n yolo-fastestv2 min = 28.90 max = 37.31 avg = 30.27\n\nnanopc-t4:/data/local/tmp # ./benchncnn 8 1 2 -1 1\nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 71.35 max = 73.02 avg = 71.83\n squeezenet_int8 min = 60.39 max = 60.96 avg = 60.69\n mobilenet min = 111.12 max = 113.02 avg = 111.99\n mobilenet_int8 min = 80.14 max = 81.59 avg = 81.00\n mobilenet_v2 min = 78.18 max = 80.89 avg = 79.18\n mobilenet_v3 min = 63.49 max = 64.26 avg = 63.90\n shufflenet min = 38.90 max = 40.28 avg = 39.26\n shufflenet_v2 min = 37.72 max = 38.45 avg = 38.02\n mnasnet min = 72.34 max = 73.59 avg = 72.87\n proxylessnasnet min = 87.33 max = 89.70 avg = 88.45\n efficientnet_b0 min = 145.14 max = 146.77 avg = 145.93\n efficientnetv2_b0 min = 169.33 max = 171.16 avg = 170.16\n regnety_400m min = 99.08 max = 99.80 avg = 99.47\n blazeface min = 12.28 max = 12.69 avg = 12.48\n googlenet min = 228.18 max = 229.36 avg = 228.64\n googlenet_int8 min = 201.62 max = 203.71 avg = 202.25\n resnet18 min = 175.71 max = 180.53 avg = 176.85\n resnet18_int8 min = 151.42 max = 152.45 avg = 151.83\n alexnet min = 160.81 max = 186.24 avg = 165.30\n vgg16 min = 1044.34 max = 1080.88 avg = 1062.34\n vgg16_int8 min = 844.53 max = 851.71 avg = 848.65\n resnet50 min = 503.25 max = 505.20 avg = 504.18\n resnet50_int8 min = 397.71 max = 400.19 avg = 398.63\n squeezenet_ssd min = 162.98 max = 165.97 avg = 164.34\n squeezenet_ssd_int8 min = 145.93 max = 148.59 avg = 146.94\n mobilenet_ssd min = 226.54 max = 229.80 avg = 227.80\n mobilenet_ssd_int8 min = 159.97 max = 163.18 avg = 161.06\n mobilenet_yolo min = 512.90 max = 517.47 avg = 515.06\n mobilenetv2_yolov3 min = 274.88 max = 280.24 avg = 276.36\n yolov4-tiny min = 351.97 max = 358.70 avg = 355.60\n nanodet_m min = 95.32 max = 97.83 avg = 96.28\n yolo-fastest-1.1 min = 43.47 max = 46.52 avg = 44.55\n yolo-fastestv2 min = 37.22 max = 37.63 avg = 37.45\n\nnanopc-t4:/data/local/tmp # ./benchncnn 8 4 1 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 48.11 max = 48.51 avg = 48.24\n squeezenet_int8 min = 43.19 max = 44.17 avg = 43.40\n mobilenet min = 65.47 max = 66.40 avg = 65.68\n mobilenet_int8 min = 49.15 max = 51.65 avg = 49.76\n mobilenet_v2 min = 53.60 max = 54.19 avg = 53.87\n mobilenet_v3 min = 52.83 max = 92.92 avg = 66.25\n shufflenet min = 35.71 max = 36.03 avg = 35.83\n shufflenet_v2 min = 31.88 max = 32.38 avg = 32.16\n mnasnet min = 51.59 max = 54.01 avg = 52.30\n proxylessnasnet min = 60.11 max = 60.40 avg = 60.24\n efficientnet_b0 min = 98.22 max = 99.40 avg = 98.56\n efficientnetv2_b0 min = 114.19 max = 123.90 avg = 115.89\n regnety_400m min = 85.89 max = 86.20 avg = 86.03\n blazeface min = 11.23 max = 11.37 avg = 11.31\n googlenet min = 142.25 max = 160.88 avg = 145.26\n googlenet_int8 min = 125.45 max = 128.50 avg = 125.96\n resnet18 min = 116.68 max = 118.26 avg = 117.00\n resnet18_int8 min = 88.43 max = 90.95 avg = 89.08\n alexnet min = 150.91 max = 160.01 avg = 152.51\n vgg16 min = 674.91 max = 684.83 avg = 679.08\n vgg16_int8 min = 417.60 max = 422.52 avg = 419.60\n resnet50 min = 297.23 max = 299.37 avg = 298.03\n resnet50_int8 min = 243.99 max = 251.39 avg = 245.99\n squeezenet_ssd min = 127.92 max = 128.53 avg = 128.17\n squeezenet_ssd_int8 min = 112.54 max = 114.63 avg = 113.19\n mobilenet_ssd min = 136.43 max = 140.14 avg = 137.33\n mobilenet_ssd_int8 min = 102.14 max = 105.00 avg = 102.77\n mobilenet_yolo min = 291.45 max = 294.04 avg = 292.63\n mobilenetv2_yolov3 min = 183.13 max = 187.00 avg = 184.05\n yolov4-tiny min = 257.46 max = 268.76 avg = 260.49\n nanodet_m min = 83.16 max = 91.03 avg = 84.77\n yolo-fastest-1.1 min = 43.53 max = 43.87 avg = 43.74\n yolo-fastestv2 min = 35.04 max = 35.54 avg = 35.17\n\nnanopc-t4:/data/local/tmp # ./benchncnn 8 1 1 -1 1\nloop_count = 8\nnum_threads = 1\npowersave = 1\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 129.63 max = 130.58 avg = 129.85\n squeezenet_int8 min = 124.10 max = 126.34 avg = 124.81\n mobilenet min = 207.92 max = 208.72 avg = 208.41\n mobilenet_int8 min = 175.55 max = 176.11 avg = 175.84\n mobilenet_v2 min = 143.02 max = 143.56 avg = 143.25\n mobilenet_v3 min = 133.11 max = 134.05 avg = 133.33\n shufflenet min = 77.97 max = 78.54 avg = 78.19\n shufflenet_v2 min = 75.59 max = 76.05 avg = 75.82\n mnasnet min = 139.86 max = 141.77 avg = 140.19\n proxylessnasnet min = 178.57 max = 179.57 avg = 179.03\n efficientnet_b0 min = 316.10 max = 317.82 avg = 316.86\n efficientnetv2_b0 min = 359.26 max = 362.03 avg = 360.31\n regnety_400m min = 182.64 max = 183.03 avg = 182.82\n blazeface min = 25.81 max = 26.53 avg = 26.20\n googlenet min = 448.45 max = 450.80 avg = 449.35\n googlenet_int8 min = 406.07 max = 410.65 avg = 408.04\n resnet18 min = 351.64 max = 362.12 avg = 354.19\n resnet18_int8 min = 298.10 max = 300.45 avg = 299.26\n alexnet min = 586.92 max = 588.73 avg = 587.80\n vgg16 min = 2170.12 max = 2202.80 avg = 2183.32\n vgg16_int8 min = 1533.65 max = 1542.01 avg = 1537.33\n resnet50 min = 975.40 max = 977.79 avg = 976.61\n resnet50_int8 min = 851.59 max = 855.22 avg = 853.75\n squeezenet_ssd min = 306.35 max = 307.54 avg = 306.96\n squeezenet_ssd_int8 min = 291.32 max = 292.87 avg = 292.18\n mobilenet_ssd min = 423.70 max = 424.63 avg = 424.11\n mobilenet_ssd_int8 min = 358.62 max = 359.42 avg = 359.04\n mobilenet_yolo min = 928.06 max = 929.25 avg = 928.55\n mobilenetv2_yolov3 min = 496.96 max = 499.29 avg = 497.73\n yolov4-tiny min = 712.80 max = 714.15 avg = 713.55\n nanodet_m min = 179.42 max = 180.60 avg = 179.75\n yolo-fastest-1.1 min = 88.06 max = 88.85 avg = 88.35\n yolo-fastestv2 min = 68.68 max = 69.83 avg = 69.08\n\nnanopc-t4:/data/local/tmp # ./benchncnn 4 1 2 0 0\n[0 Mali-T860] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-T860] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=1\n[0 Mali-T860] fp16-p/s/a=1/0/1 int8-p/s/a=1/0/0\n[0 Mali-T860] subgroup=0 basic=0 vote=0 ballot=0 shuffle=0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 24.57 max = 24.71 avg = 24.64\n mobilenet min = 35.86 max = 36.14 avg = 36.04\n mobilenet_v2 min = 30.18 max = 30.19 avg = 30.19\n mobilenet_v3 min = 30.88 max = 31.12 avg = 31.01\n shufflenet min = 33.90 max = 33.98 avg = 33.93\n shufflenet_v2 min = 29.10 max = 29.14 avg = 29.12\n mnasnet min = 30.49 max = 30.59 avg = 30.53\n proxylessnasnet min = 33.56 max = 33.61 avg = 33.59\n efficientnet_b0 min = 51.15 max = 51.54 avg = 51.38\n efficientnetv2_b0 min = 86.26 max = 87.36 avg = 86.91\n regnety_400m min = 38.44 max = 38.54 avg = 38.49\n blazeface min = 9.66 max = 9.74 avg = 9.70\n googlenet min = 80.62 max = 80.96 avg = 80.81\n resnet18 min = 74.07 max = 74.36 avg = 74.23\n alexnet min = 76.84 max = 77.26 avg = 77.08\n vgg16 min = 300.71 max = 300.89 avg = 300.80\n resnet50 min = 175.96 max = 176.72 avg = 176.23\n squeezenet_ssd min = 71.20 max = 71.38 avg = 71.32\n mobilenet_ssd min = 76.99 max = 77.47 avg = 77.19\n mobilenet_yolo min = 160.41 max = 160.84 avg = 160.62\n mobilenetv2_yolov3 min = 91.31 max = 91.37 avg = 91.35\n yolov4-tiny min = 130.78 max = 131.54 avg = 131.16\n nanodet_m min = 55.90 max = 56.03 avg = 55.96\n yolo-fastest-1.1 min = 25.50 max = 25.66 avg = 25.59\n yolo-fastestv2 min = 24.94 max = 25.07 avg = 25.01\n```\n\n### MYIR RemiPi,Renesas RZG2L(Cortex-A55 1.5GHz x 2)\n\n```\nroot@myir-remi-1g:~/ncnn# time ./benchncnn 10 4 0 -1 1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 85.38 max = 87.72 avg = 86.78\n squeezenet_int8 min = 84.23 max = 86.46 avg = 85.59\n mobilenet min = 121.01 max = 122.55 avg = 121.76\n mobilenet_int8 min = 95.64 max = 97.27 avg = 96.25\n mobilenet_v2 min = 101.35 max = 102.24 avg = 101.72\n mobilenet_v3 min = 84.09 max = 86.66 avg = 84.86\n shufflenet min = 63.32 max = 65.16 avg = 64.53\n shufflenet_v2 min = 60.33 max = 62.35 avg = 61.04\n mnasnet min = 95.51 max = 96.70 avg = 95.95\n proxylessnasnet min = 124.46 max = 125.82 avg = 125.14\n efficientnet_b0 min = 144.94 max = 146.46 avg = 145.56\n efficientnetv2_b0 min = 182.87 max = 185.63 avg = 184.56\n regnety_400m min = 105.31 max = 106.42 avg = 105.72\n blazeface min = 21.34 max = 21.90 avg = 21.50\n googlenet min = 313.01 max = 318.42 avg = 314.25\n googlenet_int8 min = 301.87 max = 304.93 avg = 303.66\n resnet18 min = 248.02 max = 253.93 avg = 250.12\n resnet18_int8 min = 244.65 max = 246.62 avg = 245.66\n alexnet min = 204.00 max = 206.39 avg = 205.21\n resnet50 min = 583.13 max = 584.82 avg = 584.11\n resnet50_int8 min = 517.42 max = 520.97 avg = 519.07\n squeezenet_ssd min = 266.63 max = 273.34 avg = 268.60\n squeezenet_ssd_int8 min = 255.42 max = 260.98 avg = 257.15\n mobilenet_ssd min = 267.16 max = 270.41 avg = 268.20\n mobilenet_ssd_int8 min = 205.03 max = 206.43 avg = 205.53\n mobilenet_yolo min = 571.08 max = 576.15 avg = 574.18\n mobilenetv2_yolov3 min = 342.52 max = 344.84 avg = 343.38\n yolov4-tiny min = 499.74 max = 503.13 avg = 501.45\n nanodet_m min = 161.87 max = 163.90 avg = 162.93\n yolo-fastest-1.1 min = 72.84 max = 74.81 avg = 73.35\n yolo-fastestv2 min = 68.24 max = 70.49 avg = 68.74\n vision_transformer min = 12464.09 max = 12491.57 avg = 12475.63\n FastestDet min = 67.92 max = 69.90 avg = 68.94\n```\n\n### OrangePi Zero 2, Allwinner H616 (Cortex-A53 1.5GHz x 4)\n\n```\norangepi@zero2:~/ncnn/benchmark$ ./benchncnn 10 4 0 -1 1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 76.25 max = 90.20 avg = 78.99\n squeezenet_int8 min = 59.92 max = 60.44 avg = 60.10\n mobilenet min = 106.91 max = 132.22 avg = 109.99\n mobilenet_int8 min = 57.96 max = 59.06 avg = 58.19\n mobilenet_v2 min = 97.93 max = 124.48 avg = 100.91\n mobilenet_v3 min = 82.27 max = 83.93 avg = 83.00\n shufflenet min = 55.27 max = 82.06 avg = 58.40\n shufflenet_v2 min = 44.94 max = 71.99 avg = 48.10\n mnasnet min = 90.66 max = 91.41 avg = 90.92\n proxylessnasnet min = 91.55 max = 118.74 avg = 94.71\n efficientnet_b0 min = 127.95 max = 155.13 avg = 131.25\n efficientnetv2_b0 min = 145.96 max = 173.67 avg = 149.36\n regnety_400m min = 102.83 max = 103.52 avg = 103.08\n blazeface min = 14.46 max = 14.95 avg = 14.77\n googlenet min = 217.71 max = 244.16 avg = 221.38\n googlenet_int8 min = 163.04 max = 187.69 avg = 166.20\n resnet18 min = 251.45 max = 277.52 avg = 255.00\n resnet18_int8 min = 136.54 max = 161.95 avg = 141.60\n alexnet min = 212.07 max = 233.27 avg = 215.34\n vgg16 min = 1206.92 max = 1981.79 avg = 1673.28\n vgg16_int8 min = 622.93 max = 702.12 avg = 661.83\n resnet50 min = 555.84 max = 643.69 avg = 576.17\n resnet50_int8 min = 348.11 max = 374.25 avg = 354.17\n squeezenet_ssd min = 224.68 max = 251.32 avg = 230.59\n squeezenet_ssd_int8 min = 154.87 max = 182.66 avg = 159.08\n mobilenet_ssd min = 238.49 max = 426.65 avg = 263.18\n mobilenet_ssd_int8 min = 118.36 max = 138.39 avg = 120.78\n mobilenet_yolo min = 500.28 max = 615.83 avg = 553.59\n mobilenetv2_yolov3 min = 340.27 max = 369.13 avg = 347.17\n yolov4-tiny min = 365.04 max = 408.48 avg = 383.93\n nanodet_m min = 112.88 max = 141.85 avg = 116.13\n yolo-fastest-1.1 min = 72.05 max = 73.46 avg = 72.68\n yolo-fastestv2 min = 54.94 max = 55.35 avg = 55.15\n vision_transformer min = 6842.19 max = 9125.07 avg = 7343.64\n FastestDet min = 59.09 max = 59.87 avg = 59.35\n```\n\n### OrangePi4 LTS, Rockchip RK3399 (Cortex-A72 1.8GHz x 2 + Cortex-A53 1.5GHz x 4)\nTest Ubuntu 22.04 Gnome Desktop\n```\norangepi@orangepi4-lts:~/ncnn/benchmark$ ./benchncnn 10 6 0 -1 0\nloop_count = 10\nnum_threads = 6\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 40.89 max = 50.29 avg = 45.15\n squeezenet_int8 min = 40.36 max = 48.57 avg = 43.56\n mobilenet min = 55.81 max = 67.35 avg = 59.81\n mobilenet_int8 min = 39.96 max = 45.10 avg = 42.09\n mobilenet_v2 min = 53.29 max = 64.12 avg = 57.40\n mobilenet_v3 min = 38.94 max = 51.11 avg = 43.06\n shufflenet min = 27.32 max = 38.53 avg = 31.85\n shufflenet_v2 min = 24.38 max = 31.17 avg = 28.32\n mnasnet min = 47.02 max = 50.68 avg = 48.86\n proxylessnasnet min = 52.31 max = 61.31 avg = 56.66\n efficientnet_b0 min = 68.14 max = 76.07 avg = 72.62\n efficientnetv2_b0 min = 77.23 max = 96.07 avg = 84.83\n regnety_400m min = 60.81 max = 81.72 avg = 72.37\n blazeface min = 7.24 max = 8.19 avg = 7.68\n googlenet min = 122.99 max = 132.67 avg = 128.90\n googlenet_int8 min = 108.45 max = 121.17 avg = 115.37\n resnet18 min = 100.67 max = 115.30 avg = 107.65\n resnet18_int8 min = 80.17 max = 87.56 avg = 84.01\n alexnet min = 71.00 max = 83.09 avg = 76.21\n vgg16 min = 557.67 max = 606.30 avg = 581.12\n vgg16_int8 min = 369.93 max = 393.20 avg = 384.86\n resnet50 min = 254.25 max = 272.90 avg = 265.18\n resnet50_int8 min = 220.70 max = 231.50 avg = 225.03\n squeezenet_ssd min = 118.91 max = 131.52 avg = 123.91\n squeezenet_ssd_int8 min = 98.25 max = 116.42 avg = 110.13\n mobilenet_ssd min = 126.62 max = 134.13 avg = 129.56\n mobilenet_ssd_int8 min = 83.83 max = 91.61 avg = 86.75\n mobilenet_yolo min = 281.19 max = 299.79 avg = 290.05\n mobilenetv2_yolov3 min = 180.37 max = 194.10 avg = 185.61\n yolov4-tiny min = 215.28 max = 227.29 avg = 221.61\n nanodet_m min = 64.63 max = 75.86 avg = 70.46\n yolo-fastest-1.1 min = 39.54 max = 48.30 avg = 44.76\n yolo-fastestv2 min = 29.91 max = 53.15 avg = 37.32\n vision_transformer min = 2520.25 max = 2595.28 avg = 2557.05\n FastestDet min = 32.45 max = 47.38 avg = 40.55\n\norangepi@orangepi4-lts:~/ncnn/benchmark$ ./benchncnn 10 4 1 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 48.90 max = 56.65 avg = 53.09\n squeezenet_int8 min = 48.09 max = 54.69 avg = 51.26\n mobilenet min = 66.06 max = 79.73 avg = 73.96\n mobilenet_int8 min = 51.33 max = 58.30 avg = 54.71\n mobilenet_v2 min = 61.06 max = 88.93 avg = 71.48\n mobilenet_v3 min = 50.41 max = 65.40 avg = 56.51\n shufflenet min = 38.11 max = 63.95 avg = 44.03\n shufflenet_v2 min = 33.27 max = 36.43 avg = 34.89\n mnasnet min = 60.02 max = 72.71 avg = 64.57\n proxylessnasnet min = 66.61 max = 73.25 avg = 70.65\n efficientnet_b0 min = 87.27 max = 94.97 avg = 91.00\n efficientnetv2_b0 min = 99.89 max = 112.09 avg = 106.13\n regnety_400m min = 84.65 max = 92.78 avg = 89.51\n blazeface min = 9.73 max = 11.45 avg = 10.85\n googlenet min = 154.74 max = 164.25 avg = 159.33\n googlenet_int8 min = 140.29 max = 148.08 avg = 144.18\n resnet18 min = 131.51 max = 244.02 avg = 150.56\n resnet18_int8 min = 102.11 max = 114.40 avg = 108.32\n alexnet min = 81.13 max = 92.35 avg = 86.86\n vgg16 min = 649.91 max = 668.62 avg = 660.25\n vgg16_int8 min = 513.75 max = 523.77 avg = 518.17\n resnet50 min = 330.89 max = 378.23 avg = 344.07\n resnet50_int8 min = 280.38 max = 286.93 avg = 284.43\n squeezenet_ssd min = 134.35 max = 146.97 avg = 141.17\n squeezenet_ssd_int8 min = 126.31 max = 137.29 avg = 130.73\n mobilenet_ssd min = 146.83 max = 161.70 avg = 155.08\n mobilenet_ssd_int8 min = 105.74 max = 117.05 avg = 111.62\n mobilenet_yolo min = 339.30 max = 352.16 avg = 345.22\n mobilenetv2_yolov3 min = 223.12 max = 234.18 avg = 229.81\n yolov4-tiny min = 267.30 max = 272.95 avg = 270.47\n nanodet_m min = 78.72 max = 86.18 avg = 81.81\n yolo-fastest-1.1 min = 47.96 max = 55.08 avg = 51.81\n yolo-fastestv2 min = 38.01 max = 44.32 avg = 42.29\n vision_transformer min = 3499.34 max = 3526.15 avg = 3514.43\n FastestDet min = 40.14 max = 44.37 avg = 42.30\n\norangepi@orangepi4-lts:~/ncnn/benchmark$ ./benchncnn 10 2 2 -1 0\nloop_count = 10\nnum_threads = 2\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 45.65 max = 46.72 avg = 46.15\n squeezenet_int8 min = 42.60 max = 43.01 avg = 42.76\n mobilenet min = 69.35 max = 70.59 avg = 69.92\n mobilenet_int8 min = 46.08 max = 46.35 avg = 46.20\n mobilenet_v2 min = 57.47 max = 58.90 avg = 58.08\n mobilenet_v3 min = 44.72 max = 45.47 avg = 45.05\n shufflenet min = 31.74 max = 32.16 avg = 31.97\n shufflenet_v2 min = 26.74 max = 26.98 avg = 26.86\n mnasnet min = 50.47 max = 51.20 avg = 50.82\n proxylessnasnet min = 57.31 max = 58.24 avg = 57.68\n efficientnet_b0 min = 79.61 max = 80.79 avg = 80.02\n efficientnetv2_b0 min = 92.67 max = 93.37 avg = 93.08\n regnety_400m min = 67.08 max = 68.07 avg = 67.59\n blazeface min = 8.56 max = 8.81 avg = 8.70\n googlenet min = 136.82 max = 138.26 avg = 137.44\n googlenet_int8 min = 121.96 max = 122.64 avg = 122.36\n resnet18 min = 118.04 max = 119.24 avg = 118.49\n resnet18_int8 min = 89.55 max = 92.11 avg = 90.38\n alexnet min = 80.75 max = 82.34 avg = 81.24\n vgg16 min = 602.11 max = 628.12 avg = 612.26\n vgg16_int8 min = 481.31 max = 484.49 avg = 482.84\n resnet50 min = 307.31 max = 310.10 avg = 308.88\n resnet50_int8 min = 240.45 max = 243.43 avg = 241.76\n squeezenet_ssd min = 119.65 max = 122.93 avg = 121.34\n squeezenet_ssd_int8 min = 102.71 max = 103.45 avg = 103.20\n mobilenet_ssd min = 142.16 max = 143.58 avg = 142.54\n mobilenet_ssd_int8 min = 93.20 max = 93.81 avg = 93.41\n mobilenet_yolo min = 315.42 max = 318.06 avg = 317.00\n mobilenetv2_yolov3 min = 190.59 max = 191.74 avg = 190.96\n yolov4-tiny min = 228.77 max = 230.49 avg = 229.78\n nanodet_m min = 66.82 max = 67.23 avg = 67.02\n yolo-fastest-1.1 min = 38.20 max = 40.89 avg = 38.85\n yolo-fastestv2 min = 32.53 max = 33.48 avg = 33.03\n vision_transformer min = 3372.17 max = 3516.54 avg = 3461.89\n FastestDet min = 32.92 max = 35.55 avg = 33.62\n\n```\n\n### OrangePicm4, Rockchip Rk3566 (Cortex-A55 1.8GHz x 4)\n```\norangepi@orangepicm4:~/code/ncnn-test$ ./benchncnn 10 4 0 -1 1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 23.91 max = 91.49 avg = 31.03\n squeezenet_int8 min = 24.44 max = 25.39 avg = 24.75\n mobilenet min = 30.67 max = 31.75 avg = 30.98\n mobilenet_int8 min = 27.87 max = 28.48 avg = 28.05\n mobilenet_v2 min = 31.82 max = 32.56 avg = 32.07\n mobilenet_v3 min = 24.63 max = 24.91 avg = 24.81\n shufflenet min = 19.77 max = 20.19 avg = 20.01\n shufflenet_v2 min = 16.67 max = 40.81 avg = 28.79\n mnasnet min = 27.48 max = 28.36 avg = 27.75\n proxylessnasnet min = 33.04 max = 37.30 avg = 33.70\n efficientnet_b0 min = 39.21 max = 175.34 avg = 53.26\n efficientnetv2_b0 min = 48.94 max = 78.68 avg = 52.44\n regnety_400m min = 39.81 max = 40.15 avg = 39.96\n blazeface min = 6.22 max = 6.36 avg = 6.30\n googlenet min = 75.48 max = 120.58 avg = 82.05\n googlenet_int8 min = 74.42 max = 78.70 avg = 75.29\n resnet18 min = 58.21 max = 99.04 avg = 66.07\n resnet18_int8 min = 54.18 max = 79.91 avg = 57.31\n alexnet min = 49.18 max = 161.71 avg = 63.03\n vgg16 min = 323.82 max = 452.63 avg = 360.92\n vgg16_int8 min = 379.18 max = 527.82 avg = 432.99\n resnet50 min = 135.84 max = 200.71 avg = 142.54\n resnet50_int8 min = 126.06 max = 169.65 avg = 136.29\n squeezenet_ssd min = 77.62 max = 137.89 avg = 86.87\n squeezenet_ssd_int8 min = 74.17 max = 76.22 avg = 74.91\n mobilenet_ssd min = 68.60 max = 132.81 avg = 75.30\n mobilenet_ssd_int8 min = 58.01 max = 59.24 avg = 58.81\n mobilenet_yolo min = 151.61 max = 247.03 avg = 168.31\n mobilenetv2_yolov3 min = 106.00 max = 163.45 avg = 111.92\n yolov4-tiny min = 132.99 max = 193.53 avg = 139.88\n nanodet_m min = 51.43 max = 87.10 avg = 58.17\n yolo-fastest-1.1 min = 26.10 max = 66.68 avg = 30.33\n yolo-fastestv2 min = 21.87 max = 69.79 avg = 35.55\n vision_transformer min = 2301.36 max = 2513.89 avg = 2426.14\n FastestDet min = 21.33 max = 21.59 avg = 21.47\norangepi@orangepicm4:~/code/ncnn-test$ ./benchncnn 10 1 0 -1 1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 47.26 max = 48.21 avg = 47.68\n squeezenet_int8 min = 50.80 max = 54.79 avg = 51.64\n mobilenet min = 68.18 max = 71.72 avg = 68.78\n mobilenet_int8 min = 58.34 max = 58.73 avg = 58.56\n mobilenet_v2 min = 56.56 max = 57.38 avg = 57.04\n mobilenet_v3 min = 45.52 max = 53.46 avg = 47.98\n shufflenet min = 34.88 max = 75.06 avg = 46.15\n shufflenet_v2 min = 33.43 max = 49.65 avg = 36.86\n mnasnet min = 53.87 max = 54.08 avg = 53.98\n proxylessnasnet min = 70.99 max = 71.40 avg = 71.14\n efficientnet_b0 min = 83.79 max = 89.78 avg = 84.96\n efficientnetv2_b0 min = 103.89 max = 117.47 avg = 105.81\n regnety_400m min = 63.68 max = 81.25 avg = 66.66\n blazeface min = 12.18 max = 39.24 avg = 21.79\n googlenet min = 179.41 max = 202.18 avg = 185.39\n googlenet_int8 min = 187.88 max = 198.49 avg = 191.01\n resnet18 min = 132.67 max = 148.94 avg = 136.09\n resnet18_int8 min = 150.37 max = 158.14 avg = 153.17\n alexnet min = 115.00 max = 120.17 avg = 116.26\n vgg16 min = 809.99 max = 851.07 avg = 827.73\n vgg16_int8 min = 1149.74 max = 1161.37 avg = 1154.22\n resnet50 min = 327.19 max = 350.42 avg = 332.12\n resnet50_int8 min = 325.08 max = 332.46 avg = 327.17\n squeezenet_ssd min = 150.33 max = 163.00 avg = 153.12\n squeezenet_ssd_int8 min = 152.21 max = 157.94 avg = 155.36\n mobilenet_ssd min = 149.30 max = 150.23 avg = 149.72\n mobilenet_ssd_int8 min = 121.93 max = 127.07 avg = 123.03\n mobilenet_yolo min = 330.91 max = 345.64 avg = 336.21\n mobilenetv2_yolov3 min = 193.25 max = 214.92 avg = 198.82\n yolov4-tiny min = 284.38 max = 332.54 avg = 293.43\n nanodet_m min = 90.69 max = 100.74 avg = 92.56\n yolo-fastest-1.1 min = 38.93 max = 51.96 avg = 42.11\n yolo-fastestv2 min = 35.74 max = 48.11 avg = 38.63\n vision_transformer min = 7280.18 max = 7301.27 avg = 7292.38\n FastestDet min = 36.54 max = 42.31 avg = 38.41\n```\n\n### OrangePi5, Rockchip RK3588s (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz)\n```\norangepi@orangepi5:~/ncnn-master/benchmark$ ./benchncnn 10 8 0 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.22 max = 6.69 avg = 6.37\n squeezenet_int8 min = 7.93 max = 8.32 avg = 8.07\n mobilenet min = 9.08 max = 14.02 avg = 9.81\n mobilenet_int8 min = 7.89 max = 9.02 avg = 8.47\n mobilenet_v2 min = 7.77 max = 8.09 avg = 7.92\n mobilenet_v3 min = 6.87 max = 8.19 avg = 7.46\n shufflenet min = 5.98 max = 10.21 avg = 7.23\n shufflenet_v2 min = 4.82 max = 5.04 avg = 4.93\n mnasnet min = 6.15 max = 6.36 avg = 6.24\n proxylessnasnet min = 9.50 max = 10.50 avg = 9.93\n efficientnet_b0 min = 11.46 max = 11.79 avg = 11.60\n efficientnetv2_b0 min = 18.61 max = 19.48 avg = 18.88\n regnety_400m min = 10.54 max = 12.44 avg = 10.86\n blazeface min = 1.96 max = 5.35 avg = 2.58\n googlenet min = 26.62 max = 32.59 avg = 29.96\n googlenet_int8 min = 28.27 max = 32.80 avg = 30.01\n resnet18 min = 15.52 max = 18.29 avg = 16.37\n resnet18_int8 min = 23.33 max = 26.89 avg = 24.99\n alexnet min = 19.92 max = 22.75 avg = 21.06\n vgg16 min = 101.18 max = 122.44 avg = 107.45\n vgg16_int8 min = 164.69 max = 227.98 avg = 189.73\n resnet50 min = 42.96 max = 59.26 avg = 50.83\n resnet50_int8 min = 54.46 max = 66.72 avg = 61.37\n squeezenet_ssd min = 24.39 max = 31.19 avg = 27.69\n squeezenet_ssd_int8 min = 27.15 max = 41.55 avg = 33.68\n mobilenet_ssd min = 22.26 max = 26.89 avg = 23.95\n mobilenet_ssd_int8 min = 21.18 max = 24.21 avg = 23.05\n mobilenet_yolo min = 52.65 max = 65.53 avg = 58.47\n mobilenetv2_yolov3 min = 31.34 max = 45.15 avg = 34.63\n yolov4-tiny min = 40.55 max = 49.32 avg = 43.85\n nanodet_m min = 16.08 max = 19.51 avg = 17.58\n yolo-fastest-1.1 min = 6.48 max = 7.33 avg = 6.98\n yolo-fastestv2 min = 4.96 max = 11.66 avg = 7.30\n vision_transformer min = 678.22 max = 815.73 avg = 729.16\n FastestDet min = 4.95 max = 10.65 avg = 6.88\n\n\norangepi@orangepi5:~/ncnn-master/benchmark$ ./benchncnn 10 4 1 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 10.91 max = 11.14 avg = 11.03\n squeezenet_int8 min = 14.26 max = 14.55 avg = 14.30\n mobilenet min = 15.92 max = 16.26 avg = 16.11\n mobilenet_int8 min = 14.71 max = 15.22 avg = 14.91\n mobilenet_v2 min = 12.28 max = 12.49 avg = 12.37\n mobilenet_v3 min = 11.31 max = 11.72 avg = 11.46\n shufflenet min = 10.10 max = 10.33 avg = 10.24\n shufflenet_v2 min = 9.38 max = 9.70 avg = 9.55\n mnasnet min = 12.28 max = 12.80 avg = 12.44\n proxylessnasnet min = 16.54 max = 16.66 avg = 16.60\n efficientnet_b0 min = 19.56 max = 20.66 avg = 19.86\n efficientnetv2_b0 min = 34.06 max = 34.65 avg = 34.41\n regnety_400m min = 23.97 max = 24.69 avg = 24.20\n blazeface min = 3.39 max = 3.56 avg = 3.48\n googlenet min = 46.96 max = 47.90 avg = 47.56\n googlenet_int8 min = 49.56 max = 50.23 avg = 49.79\n resnet18 min = 28.44 max = 29.54 avg = 28.77\n resnet18_int8 min = 41.32 max = 42.44 avg = 41.67\n alexnet min = 31.83 max = 32.77 avg = 32.32\n vgg16 min = 170.32 max = 178.30 avg = 173.22\n vgg16_int8 min = 282.55 max = 299.32 avg = 287.78\n resnet50 min = 78.00 max = 81.57 avg = 78.79\n resnet50_int8 min = 89.12 max = 92.31 avg = 90.92\n squeezenet_ssd min = 38.07 max = 39.07 avg = 38.59\n squeezenet_ssd_int8 min = 50.98 max = 52.56 avg = 51.68\n mobilenet_ssd min = 38.79 max = 39.67 avg = 39.34\n mobilenet_ssd_int8 min = 33.53 max = 35.26 avg = 34.66\n mobilenet_yolo min = 90.50 max = 92.32 avg = 90.99\n mobilenetv2_yolov3 min = 51.38 max = 51.93 avg = 51.56\n yolov4-tiny min = 75.65 max = 76.80 avg = 76.17\n nanodet_m min = 21.33 max = 21.68 avg = 21.50\n yolo-fastest-1.1 min = 11.18 max = 12.06 avg = 11.36\n yolo-fastestv2 min = 9.87 max = 10.33 avg = 10.15\n vision_transformer min = 1475.77 max = 1477.97 avg = 1476.77\n FastestDet min = 9.39 max = 9.73 avg = 9.53\n\n\norangepi@orangepi5:~/ncnn-master/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.59 max = 3.70 avg = 3.66\n squeezenet_int8 min = 4.32 max = 4.42 avg = 4.36\n mobilenet min = 5.50 max = 5.55 avg = 5.53\n mobilenet_int8 min = 4.52 max = 4.60 avg = 4.56\n mobilenet_v2 min = 4.50 max = 4.60 avg = 4.54\n mobilenet_v3 min = 4.09 max = 4.28 avg = 4.15\n shufflenet min = 3.49 max = 3.58 avg = 3.51\n shufflenet_v2 min = 2.91 max = 3.07 avg = 2.97\n mnasnet min = 4.18 max = 4.25 avg = 4.21\n proxylessnasnet min = 4.94 max = 5.00 avg = 4.97\n efficientnet_b0 min = 7.50 max = 7.54 avg = 7.52\n efficientnetv2_b0 min = 11.32 max = 11.41 avg = 11.37\n regnety_400m min = 7.92 max = 8.01 avg = 7.95\n blazeface min = 1.21 max = 1.31 avg = 1.24\n googlenet min = 15.03 max = 15.17 avg = 15.10\n googlenet_int8 min = 15.48 max = 15.61 avg = 15.55\n resnet18 min = 9.91 max = 9.97 avg = 9.93\n resnet18_int8 min = 15.80 max = 16.00 avg = 15.89\n alexnet min = 12.35 max = 12.64 avg = 12.48\n vgg16 min = 61.92 max = 65.62 avg = 62.93\n vgg16_int8 min = 129.94 max = 131.65 avg = 130.65\n resnet50 min = 27.41 max = 27.62 avg = 27.52\n resnet50_int8 min = 33.01 max = 33.23 avg = 33.08\n squeezenet_ssd min = 13.92 max = 14.27 avg = 14.02\n squeezenet_ssd_int8 min = 18.04 max = 18.40 avg = 18.15\n mobilenet_ssd min = 13.69 max = 13.80 avg = 13.74\n mobilenet_ssd_int8 min = 10.95 max = 11.10 avg = 11.02\n mobilenet_yolo min = 32.06 max = 32.30 avg = 32.17\n mobilenetv2_yolov3 min = 19.27 max = 20.68 avg = 19.97\n yolov4-tiny min = 25.41 max = 29.51 avg = 27.76\n nanodet_m min = 6.68 max = 6.73 avg = 6.70\n yolo-fastest-1.1 min = 3.77 max = 4.02 avg = 3.83\n yolo-fastestv2 min = 3.41 max = 3.65 avg = 3.48\n vision_transformer min = 548.32 max = 654.71 avg = 579.48\n FastestDet min = 3.38 max = 3.46 avg = 3.42\n\n```\n\n### OrangePi5 Plus, Rockchip RK3588 (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz)\n```\norangepi@orangepi5plus:~/ncnn$ ./benchncnn 8 4 2 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 5.55 max = 5.67 avg = 5.61\n squeezenet_int8 min = 5.39 max = 5.76 avg = 5.60\n mobilenet min = 7.43 max = 7.50 avg = 7.47\n mobilenet_int8 min = 6.91 max = 7.00 avg = 6.96\n mobilenet_v2 min = 8.24 max = 8.47 avg = 8.33\n mobilenet_v3 min = 6.63 max = 7.32 avg = 6.84\n shufflenet min = 4.10 max = 4.23 avg = 4.14\n shufflenet_v2 min = 3.51 max = 3.61 avg = 3.56\n mnasnet min = 5.76 max = 7.79 avg = 6.53\n proxylessnasnet min = 6.66 max = 7.19 avg = 6.79\n efficientnet_b0 min = 10.32 max = 10.73 avg = 10.40\n efficientnetv2_b0 min = 11.48 max = 11.78 avg = 11.61\n regnety_400m min = 9.73 max = 9.85 avg = 9.79\n blazeface min = 1.39 max = 1.62 avg = 1.46\n googlenet min = 21.48 max = 23.08 avg = 22.79\n googlenet_int8 min = 20.82 max = 21.78 avg = 21.01\n resnet18 min = 9.37 max = 10.05 avg = 9.50\n resnet18_int8 min = 14.88 max = 19.64 avg = 15.90\n alexnet min = 24.74 max = 24.93 avg = 24.81\n vgg16 min = 58.75 max = 62.44 avg = 59.52\n vgg16_int8 min = 73.68 max = 75.89 avg = 74.14\n resnet50 min = 44.88 max = 45.10 avg = 44.98\n resnet50_int8 min = 35.54 max = 36.02 avg = 35.71\n squeezenet_ssd min = 12.07 max = 26.66 avg = 19.03\n squeezenet_ssd_int8 min = 21.95 max = 25.51 avg = 23.21\n mobilenet_ssd min = 12.62 max = 12.73 avg = 12.67\n mobilenet_ssd_int8 min = 17.21 max = 17.68 avg = 17.44\n mobilenet_yolo min = 32.82 max = 32.98 avg = 32.91\n mobilenetv2_yolov3 min = 18.67 max = 20.52 avg = 19.57\n yolov4-tiny min = 38.82 max = 40.84 avg = 39.82\n nanodet_m min = 9.05 max = 9.22 avg = 9.13\n yolo-fastest-1.1 min = 4.67 max = 5.04 avg = 4.74\n yolo-fastestv2 min = 4.27 max = 4.32 avg = 4.29\n vision_transformer min = 429.32 max = 431.02 avg = 430.20\n FastestDet min = 4.28 max = 4.72 avg = 4.36\n\n```\n\n### RDK X3 Module (Cortex-A53 1.5GHz x 4) aarch64\n```\nroot@ubuntu:/home/sunrise/ncnn-master/benchmark# ../build-aarch64-linux-gnu/benchmark/benchncnn 10 4 0 -1 1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 49.83 max = 50.57 avg = 50.08\n squeezenet_int8 min = 48.43 max = 49.18 avg = 48.67\n mobilenet min = 68.37 max = 69.09 avg = 68.63\n mobilenet_int8 min = 58.19 max = 58.72 avg = 58.37\n mobilenet_v2 min = 58.76 max = 60.62 avg = 59.20\n mobilenet_v3 min = 49.75 max = 50.60 avg = 50.06\n shufflenet min = 37.17 max = 37.96 avg = 37.50\n shufflenet_v2 min = 32.08 max = 32.42 avg = 32.22\n mnasnet min = 55.51 max = 57.02 avg = 55.90\n proxylessnasnet min = 68.15 max = 69.53 avg = 68.78\n efficientnet_b0 min = 88.64 max = 90.16 avg = 89.43\n efficientnetv2_b0 min = 102.45 max = 103.42 avg = 102.92\n regnety_400m min = 88.22 max = 89.09 avg = 88.62\n blazeface min = 9.78 max = 10.15 avg = 9.93\n googlenet min = 152.20 max = 153.92 avg = 153.28\n googlenet_int8 min = 141.80 max = 143.30 avg = 142.48\n resnet18 min = 116.70 max = 117.59 avg = 117.03\n resnet18_int8 min = 104.42 max = 105.85 avg = 104.94\n alexnet min = 82.55 max = 83.23 avg = 82.82\n vgg16 min = 590.22 max = 598.18 avg = 594.35\n vgg16_int8 min = 504.56 max = 507.21 avg = 505.73\n resnet50 min = 307.36 max = 308.68 avg = 308.03\n resnet50_int8 min = 281.35 max = 283.87 avg = 282.30\n squeezenet_ssd min = 124.93 max = 126.51 avg = 125.51\n squeezenet_ssd_int8 min = 118.07 max = 118.89 avg = 118.29\n mobilenet_ssd min = 142.27 max = 142.57 avg = 142.44\n mobilenet_ssd_int8 min = 116.51 max = 117.60 avg = 117.04\n mobilenet_yolo min = 314.64 max = 317.09 avg = 315.93\n mobilenetv2_yolov3 min = 204.55 max = 205.30 avg = 204.93\n yolov4-tiny min = 246.69 max = 249.64 avg = 247.95\n nanodet_m min = 77.73 max = 78.30 avg = 77.99\n yolo-fastest-1.1 min = 46.29 max = 47.52 avg = 46.93\n yolo-fastestv2 min = 36.55 max = 36.95 avg = 36.73\n vision_transformer min = 3372.85 max = 3409.14 avg = 3377.75\n FastestDet min = 38.23 max = 38.77 avg = 38.49\n\n```\n\n### NanoPi R2S, Rockchip RK3328 (Cortex-A53 1.3GHz x 4) Armbian focal (21.05.1) aarch64\n```\nroot@nanopi-r2s:~/ncnn/build/benchmark# ./benchncnn 8 4 0\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 62.20 max = 62.81 avg = 62.49\n squeezenet_int8 min = 57.92 max = 71.46 avg = 59.76\n mobilenet min = 82.88 max = 89.36 avg = 84.52\n mobilenet_int8 min = 57.16 max = 96.22 avg = 62.29\n mobilenet_v2 min = 73.68 max = 75.92 avg = 74.17\n mobilenet_v3 min = 59.57 max = 60.14 avg = 59.84\n shufflenet min = 52.34 max = 52.70 avg = 52.53\n shufflenet_v2 min = 45.51 max = 45.92 avg = 45.73\n mnasnet min = 67.75 max = 83.15 avg = 69.82\n proxylessnasnet min = 81.70 max = 83.66 avg = 82.31\n efficientnet_b0 min = 121.10 max = 123.22 avg = 121.55\n efficientnetv2_b0 min = 138.93 max = 192.15 avg = 154.94\n regnety_400m min = 99.62 max = 116.29 avg = 101.97\n blazeface min = 18.80 max = 19.15 avg = 19.01\n googlenet min = 176.36 max = 202.84 avg = 181.86\n googlenet_int8 min = 155.50 max = 190.50 avg = 161.20\n resnet18 min = 165.79 max = 201.57 avg = 172.56\n resnet18_int8 min = 122.24 max = 160.53 avg = 134.24\n alexnet min = 227.07 max = 238.09 avg = 232.19\n vgg16_int8 min = 522.14 max = 551.75 avg = 531.68\n resnet50 min = 378.30 max = 440.21 avg = 388.56\n resnet50_int8 min = 315.76 max = 373.97 avg = 329.88\n squeezenet_ssd min = 175.37 max = 200.86 avg = 179.01\n squeezenet_ssd_int8 min = 134.71 max = 147.57 avg = 136.57\n mobilenet_ssd min = 174.43 max = 212.11 avg = 180.61\n mobilenet_ssd_int8 min = 119.41 max = 153.75 avg = 124.21\n mobilenet_yolo min = 366.27 max = 422.67 avg = 383.65\n mobilenetv2_yolov3 min = 238.56 max = 281.97 avg = 247.56\n yolov4-tiny min = 311.45 max = 333.32 avg = 316.79\n nanodet_m min = 114.15 max = 122.39 avg = 115.44\n\nroot@nanopi-r2s:~/ncnn/build/benchmark# ./benchncnn 8 2 0\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 89.02 max = 90.52 avg = 89.35\n squeezenet_int8 min = 81.19 max = 81.90 avg = 81.42\n mobilenet min = 131.47 max = 134.39 avg = 132.34\n mobilenet_int8 min = 102.20 max = 103.03 avg = 102.66\n mobilenet_v2 min = 102.40 max = 108.12 avg = 103.91\n mobilenet_v3 min = 89.17 max = 90.10 avg = 89.53\n shufflenet min = 65.74 max = 68.86 avg = 66.50\n shufflenet_v2 min = 62.83 max = 64.41 avg = 63.25\n mnasnet min = 98.01 max = 98.24 avg = 98.14\n proxylessnasnet min = 121.10 max = 123.55 avg = 121.80\n efficientnet_b0 min = 187.79 max = 188.41 avg = 188.08\n efficientnetv2_b0 min = 211.96 max = 213.99 avg = 212.74\n regnety_400m min = 124.98 max = 125.49 avg = 125.28\n blazeface min = 24.91 max = 25.14 avg = 25.00\n googlenet min = 278.47 max = 283.24 avg = 280.79\n googlenet_int8 min = 243.81 max = 247.82 avg = 245.30\n resnet18 min = 257.46 max = 259.29 avg = 258.29\n resnet18_int8 min = 187.18 max = 188.74 avg = 187.70\n alexnet min = 384.52 max = 387.07 avg = 385.84\n vgg16_int8 min = 897.26 max = 901.68 avg = 899.19\n resnet50 min = 618.85 max = 623.92 avg = 620.85\n resnet50_int8 min = 512.33 max = 514.93 avg = 513.64\n squeezenet_ssd min = 211.21 max = 218.71 avg = 213.02\n squeezenet_ssd_int8 min = 193.32 max = 193.97 avg = 193.70\n mobilenet_ssd min = 271.11 max = 275.58 avg = 272.06\n mobilenet_ssd_int8 min = 208.80 max = 209.59 avg = 209.05\n mobilenet_yolo min = 570.55 max = 575.98 avg = 572.73\n mobilenetv2_yolov3 min = 329.04 max = 353.84 avg = 340.42\n yolov4-tiny min = 435.16 max = 463.68 avg = 457.69\n nanodet_m min = 155.70 max = 159.13 avg = 156.50\n```\n\n### EAIDK 310, Rockchip RK3228H (Cortex-A53 1.3GHz x 4) fedora-28 aarch64\n```\n[openailab@MiWiFi-R1D-srv benchmark]$ ./benchncnn 8 4 0 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 68.97 max = 71.42 avg = 69.65\n squeezenet_int8 min = 58.47 max = 59.58 avg = 58.77\n mobilenet min = 90.87 max = 100.18 avg = 92.48\n mobilenet_int8 min = 59.46 max = 63.02 avg = 60.01\n mobilenet_v2 min = 82.92 max = 112.01 avg = 88.10\n mobilenet_v3 min = 66.65 max = 69.57 avg = 67.27\n shufflenet min = 48.22 max = 48.49 avg = 48.34\n shufflenet_v2 min = 48.52 max = 52.88 avg = 49.17\n mnasnet min = 75.63 max = 79.83 avg = 76.43\n proxylessnasnet min = 84.73 max = 86.69 avg = 85.16\n efficientnet_b0 min = 125.69 max = 129.00 avg = 126.38\n efficientnetv2_b0 min = 144.44 max = 149.01 avg = 145.33\n regnety_400m min = 99.69 max = 101.23 avg = 100.38\n blazeface min = 15.84 max = 16.24 avg = 16.03\n googlenet min = 194.64 max = 199.29 avg = 196.07\n googlenet_int8 min = 158.54 max = 165.64 avg = 160.25\n resnet18 min = 200.65 max = 221.60 avg = 204.30\n resnet18_int8 min = 122.69 max = 126.57 avg = 123.54\n alexnet min = 175.54 max = 200.91 avg = 181.38\n resnet50 min = 428.75 max = 466.51 avg = 439.67\n resnet50_int8 min = 324.95 max = 347.47 avg = 329.74\n squeezenet_ssd min = 199.86 max = 207.51 avg = 201.99\n squeezenet_ssd_int8 min = 150.35 max = 176.92 avg = 154.60\n mobilenet_ssd min = 186.50 max = 189.92 avg = 188.09\n mobilenet_ssd_int8 min = 123.55 max = 127.17 avg = 124.63\n mobilenet_yolo min = 393.83 max = 414.09 avg = 398.57\n mobilenetv2_yolov3 min = 263.49 max = 273.11 avg = 266.11\n yolov4-tiny min = 342.33 max = 363.69 avg = 346.34\n nanodet_m min = 119.66 max = 127.29 avg = 121.26\n yolo-fastest-1.1 min = 61.87 max = 90.26 avg = 65.77\n yolo-fastestv2 min = 48.48 max = 50.82 avg = 48.93\n\n[openailab@MiWiFi-R1D-srv benchmark]$ ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 152.15 max = 152.67 avg = 152.43\n squeezenet_int8 min = 143.22 max = 144.24 avg = 143.61\n mobilenet min = 237.77 max = 239.69 avg = 238.47\n mobilenet_int8 min = 199.91 max = 201.35 avg = 200.50\n mobilenet_v2 min = 169.67 max = 170.18 avg = 169.93\n mobilenet_v3 min = 150.06 max = 151.17 avg = 150.78\n shufflenet min = 91.78 max = 92.38 avg = 92.06\n shufflenet_v2 min = 100.86 max = 101.75 avg = 101.50\n mnasnet min = 165.10 max = 166.74 avg = 166.24\n proxylessnasnet min = 218.42 max = 220.55 avg = 219.12\n efficientnet_b0 min = 348.00 max = 349.03 avg = 348.49\n efficientnetv2_b0 min = 404.06 max = 406.16 avg = 405.00\n regnety_400m min = 209.48 max = 211.36 avg = 210.44\n blazeface min = 31.31 max = 32.61 avg = 32.00\n googlenet min = 510.38 max = 512.43 avg = 511.25\n googlenet_int8 min = 454.38 max = 456.19 avg = 455.02\n resnet18 min = 407.78 max = 409.45 avg = 408.34\n resnet18_int8 min = 357.01 max = 360.72 avg = 358.74\n alexnet min = 504.12 max = 506.74 avg = 505.08\n resnet50 min = 1115.42 max = 1121.91 avg = 1118.67\n resnet50_int8 min = 973.38 max = 976.26 avg = 975.21\n squeezenet_ssd min = 361.52 max = 363.69 avg = 362.38\n squeezenet_ssd_int8 min = 333.81 max = 337.16 avg = 335.24\n mobilenet_ssd min = 477.43 max = 478.36 avg = 477.82\n mobilenet_ssd_int8 min = 409.33 max = 409.67 avg = 409.52\n mobilenet_yolo min = 1048.79 max = 1057.72 avg = 1053.80\n mobilenetv2_yolov3 min = 567.04 max = 571.44 avg = 569.04\n yolov4-tiny min = 788.40 max = 790.74 avg = 789.12\n nanodet_m min = 253.68 max = 254.59 avg = 254.16\n yolo-fastest-1.1 min = 102.44 max = 103.11 avg = 102.67\n yolo-fastestv2 min = 82.19 max = 82.43 avg = 82.35\n```\n\n### NVIDIA Jetson Orin Nano\n```\norin@nano:~/ncnn/benchmark$ ./benchncnn 8 6 0 0 1\n[0 NVIDIA Tegra Orin (nvgpu)] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA Tegra Orin (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra Orin (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra Orin (nvgpu)] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA Tegra Orin (nvgpu)] fp16-matrix-16_8_8/16_8_16/16_16_16=1/1/1\nloop_count = 8\nnum_threads = 6\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 5.31 max = 5.95 avg = 5.44\n squeezenet_int8 min = 5.13 max = 6.24 avg = 5.57\n mobilenet min = 2.98 max = 5.52 avg = 3.66\n mobilenet_int8 min = 5.97 max = 7.76 avg = 6.98\n mobilenet_v2 min = 6.73 max = 6.98 avg = 6.91\n mobilenet_v3 min = 8.58 max = 8.77 avg = 8.71\n shufflenet min = 7.33 max = 7.43 avg = 7.39\n shufflenet_v2 min = 7.59 max = 8.46 avg = 8.27\n mnasnet min = 4.78 max = 6.81 avg = 5.41\n proxylessnasnet min = 7.39 max = 7.65 avg = 7.52\n efficientnet_b0 min = 10.81 max = 15.28 avg = 12.27\n efficientnetv2_b0 min = 46.58 max = 48.56 avg = 47.70\n regnety_400m min = 9.86 max = 10.46 avg = 10.04\n blazeface min = 3.98 max = 4.66 avg = 4.31\n googlenet min = 10.01 max = 14.44 avg = 11.48\n googlenet_int8 min = 18.07 max = 19.55 avg = 18.65\n resnet18 min = 6.52 max = 9.73 avg = 8.26\n resnet18_int8 min = 13.28 max = 20.58 avg = 14.96\n alexnet min = 8.71 max = 9.05 avg = 8.84\n vgg16 min = 19.28 max = 19.49 avg = 19.35\n vgg16_int8 min = 98.14 max = 100.92 avg = 99.76\n resnet50 min = 9.25 max = 9.37 avg = 9.31\n resnet50_int8 min = 31.16 max = 34.44 avg = 32.59\n squeezenet_ssd min = 13.60 max = 18.96 avg = 16.68\n squeezenet_ssd_int8 min = 17.81 max = 19.83 avg = 18.75\n mobilenet_ssd min = 11.88 max = 13.86 avg = 13.27\n mobilenet_ssd_int8 min = 14.05 max = 21.16 avg = 15.64\n mobilenet_yolo min = 14.18 max = 14.41 avg = 14.26\n mobilenetv2_yolov3 min = 16.65 max = 18.78 avg = 18.06\n yolov4-tiny min = 25.60 max = 26.56 avg = 25.92\n nanodet_m min = 15.71 max = 19.89 avg = 19.03\n yolo-fastest-1.1 min = 8.72 max = 9.18 avg = 8.96\n yolo-fastestv2 min = 7.97 max = 8.10 avg = 8.04\n vision_transformer min = 821.34 max = 825.91 avg = 823.26\n FastestDet min = 7.72 max = 8.15 avg = 7.81\norin@nano:~/ncnn/benchmark$ ./benchncnn 8 1 0 0 1\n[0 NVIDIA Tegra Orin (nvgpu)] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA Tegra Orin (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra Orin (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra Orin (nvgpu)] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA Tegra Orin (nvgpu)] fp16-matrix-16_8_8/16_8_16/16_16_16=1/1/1\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 5.05 max = 5.23 avg = 5.09\n squeezenet_int8 min = 15.93 max = 16.09 avg = 16.00\n mobilenet min = 2.97 max = 5.49 avg = 3.84\n mobilenet_int8 min = 23.27 max = 23.38 avg = 23.33\n mobilenet_v2 min = 3.61 max = 4.01 avg = 3.83\n mobilenet_v3 min = 6.12 max = 8.36 avg = 6.67\n shufflenet min = 4.07 max = 7.25 avg = 6.22\n shufflenet_v2 min = 8.49 max = 8.82 avg = 8.67\n mnasnet min = 3.70 max = 8.23 avg = 5.37\n proxylessnasnet min = 6.36 max = 9.16 avg = 7.52\n efficientnet_b0 min = 10.55 max = 10.81 avg = 10.65\n efficientnetv2_b0 min = 28.22 max = 28.62 avg = 28.54\n regnety_400m min = 7.22 max = 10.04 avg = 8.50\n blazeface min = 3.70 max = 3.86 avg = 3.76\n googlenet min = 7.18 max = 9.76 avg = 8.21\n googlenet_int8 min = 63.19 max = 63.54 avg = 63.32\n resnet18 min = 4.67 max = 4.73 avg = 4.70\n resnet18_int8 min = 50.51 max = 50.81 avg = 50.65\n alexnet min = 8.56 max = 10.64 avg = 9.02\n vgg16 min = 19.24 max = 19.50 avg = 19.31\n vgg16_int8 min = 411.02 max = 412.40 avg = 411.60\n resnet50 min = 9.14 max = 9.52 avg = 9.41\n resnet50_int8 min = 112.04 max = 112.43 avg = 112.25\n squeezenet_ssd min = 13.23 max = 13.79 avg = 13.52\n squeezenet_ssd_int8 min = 46.52 max = 46.98 avg = 46.77\n mobilenet_ssd min = 8.89 max = 12.51 avg = 9.95\n mobilenet_ssd_int8 min = 47.66 max = 48.73 avg = 48.13\n mobilenet_yolo min = 9.68 max = 9.75 avg = 9.70\n mobilenetv2_yolov3 min = 15.84 max = 17.54 avg = 16.83\n yolov4-tiny min = 23.32 max = 25.49 avg = 24.56\n nanodet_m min = 13.59 max = 19.53 avg = 15.85\n yolo-fastest-1.1 min = 7.68 max = 11.32 avg = 8.20\n yolo-fastestv2 min = 7.75 max = 7.84 avg = 7.78\n vision_transformer min = 822.27 max = 829.73 avg = 825.74\n FastestDet min = 7.51 max = 8.05 avg = 7.68\n \norin@nano:~/ncnn/benchmark$ ./benchncnn 8 6 0 -1 1\nloop_count = 8\nnum_threads = 6\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 5.07 max = 6.99 avg = 5.69\n squeezenet_int8 min = 5.08 max = 5.79 avg = 5.42\n mobilenet min = 6.96 max = 8.20 avg = 7.45\n mobilenet_int8 min = 5.91 max = 7.33 avg = 6.37\n mobilenet_v2 min = 5.86 max = 7.55 avg = 6.51\n mobilenet_v3 min = 5.60 max = 7.22 avg = 6.14\n shufflenet min = 5.20 max = 5.79 avg = 5.44\n shufflenet_v2 min = 4.56 max = 5.90 avg = 4.86\n mnasnet min = 5.43 max = 6.44 avg = 5.83\n proxylessnasnet min = 5.92 max = 8.70 avg = 6.83\n efficientnet_b0 min = 10.09 max = 11.57 avg = 10.65\n efficientnetv2_b0 min = 12.79 max = 15.96 avg = 14.12\n regnety_400m min = 14.04 max = 21.23 avg = 15.88\n blazeface min = 1.76 max = 1.90 avg = 1.81\n googlenet min = 19.45 max = 25.43 avg = 21.21\n googlenet_int8 min = 17.67 max = 18.59 avg = 18.20\n resnet18 min = 12.26 max = 19.47 avg = 15.13\n resnet18_int8 min = 13.02 max = 14.78 avg = 13.86\n alexnet min = 12.27 max = 19.18 avg = 15.02\n vgg16 min = 59.43 max = 89.43 avg = 65.11\n vgg16_int8 min = 97.71 max = 141.28 avg = 108.00\n resnet50 min = 38.69 max = 40.67 avg = 39.26\n resnet50_int8 min = 28.67 max = 31.63 avg = 29.93\n squeezenet_ssd min = 14.52 max = 26.92 avg = 17.89\n squeezenet_ssd_int8 min = 16.61 max = 19.27 avg = 17.82\n mobilenet_ssd min = 16.61 max = 22.65 avg = 17.89\n mobilenet_ssd_int8 min = 13.22 max = 14.83 avg = 14.04\n mobilenet_yolo min = 40.10 max = 44.28 avg = 41.48\n mobilenetv2_yolov3 min = 21.48 max = 22.83 avg = 22.01\n yolov4-tiny min = 33.30 max = 37.31 avg = 34.59\n nanodet_m min = 10.80 max = 12.62 avg = 11.54\n yolo-fastest-1.1 min = 5.51 max = 6.03 avg = 5.75\n yolo-fastestv2 min = 4.98 max = 6.35 avg = 5.44\n vision_transformer min = 610.40 max = 681.89 avg = 628.84\n FastestDet min = 4.82 max = 6.19 avg = 5.32\norin@nano:~/ncnn/benchmark$ ./benchncnn 8 1 0 -1 1\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 15.94 max = 16.23 avg = 16.04\n squeezenet_int8 min = 15.91 max = 16.09 avg = 15.98\n mobilenet min = 28.77 max = 28.91 avg = 28.83\n mobilenet_int8 min = 23.29 max = 23.63 avg = 23.46\n mobilenet_v2 min = 19.32 max = 19.43 avg = 19.37\n mobilenet_v3 min = 16.57 max = 16.65 avg = 16.61\n shufflenet min = 10.39 max = 10.48 avg = 10.44\n shufflenet_v2 min = 10.61 max = 10.69 avg = 10.65\n mnasnet min = 18.61 max = 18.69 avg = 18.65\n proxylessnasnet min = 21.97 max = 22.17 avg = 22.05\n efficientnet_b0 min = 36.73 max = 36.89 avg = 36.83\n efficientnetv2_b0 min = 41.72 max = 41.97 avg = 41.83\n regnety_400m min = 25.71 max = 26.03 avg = 25.85\n blazeface min = 3.59 max = 3.63 avg = 3.60\n googlenet min = 66.85 max = 67.38 avg = 67.12\n googlenet_int8 min = 63.65 max = 63.85 avg = 63.74\n resnet18 min = 48.49 max = 49.21 avg = 48.83\n resnet18_int8 min = 50.82 max = 51.16 avg = 50.92\n alexnet min = 57.67 max = 58.24 avg = 58.03\n vgg16 min = 280.03 max = 281.34 avg = 280.77\n vgg16_int8 min = 413.51 max = 414.67 avg = 414.08\n resnet50 min = 138.19 max = 138.94 avg = 138.48\n resnet50_int8 min = 112.53 max = 112.86 avg = 112.68\n squeezenet_ssd min = 46.26 max = 46.46 avg = 46.37\n squeezenet_ssd_int8 min = 47.56 max = 48.33 avg = 47.85\n mobilenet_ssd min = 60.51 max = 60.81 avg = 60.68\n mobilenet_ssd_int8 min = 47.47 max = 47.76 avg = 47.58\n mobilenet_yolo min = 136.20 max = 136.54 avg = 136.37\n mobilenetv2_yolov3 min = 69.80 max = 70.04 avg = 69.93\n yolov4-tiny min = 87.71 max = 88.63 avg = 88.12\n nanodet_m min = 25.73 max = 26.06 avg = 25.85\n yolo-fastest-1.1 min = 10.25 max = 10.35 avg = 10.29\n yolo-fastestv2 min = 9.25 max = 9.38 avg = 9.33\n vision_transformer min = 2282.07 max = 2690.34 avg = 2481.94\n FastestDet min = 9.80 max = 9.88 avg = 9.84\n```\n\n### NVIDIA Jetson Nano\n```\n[0 NVIDIA Tegra X1 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]\n[0 NVIDIA Tegra X1 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra X1 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra X1 (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 12.15 max = 26.48 avg = 18.11\n squeezenet_int8 min = 27.60 max = 42.50 avg = 29.89\n mobilenet min = 16.07 max = 16.10 avg = 16.09\n mobilenet_int8 min = 30.65 max = 32.15 avg = 31.07\n mobilenet_v2 min = 12.87 max = 13.15 avg = 12.99\n mobilenet_v3 min = 13.32 max = 16.65 avg = 14.57\n shufflenet min = 14.21 max = 14.34 avg = 14.29\n shufflenet_v2 min = 13.03 max = 21.97 avg = 19.02\n mnasnet min = 13.33 max = 13.64 avg = 13.49\n proxylessnasnet min = 14.65 max = 14.91 avg = 14.76\n efficientnet_b0 min = 21.26 max = 21.41 avg = 21.35\n efficientnetv2_b0 min = 54.66 max = 60.81 avg = 57.16\n regnety_400m min = 17.91 max = 18.08 avg = 18.01\n blazeface min = 6.87 max = 7.03 avg = 6.94\n googlenet min = 43.30 max = 43.54 avg = 43.43\n googlenet_int8 min = 80.07 max = 84.28 avg = 81.10\n resnet18 min = 43.89 max = 44.06 avg = 43.98\n resnet18_int8 min = 60.70 max = 63.43 avg = 61.60\n alexnet min = 74.21 max = 75.20 avg = 74.45\n vgg16 min = 310.39 max = 310.65 avg = 310.52\n vgg16_int8 min = 293.15 max = 297.28 avg = 294.93\n resnet50 min = 93.03 max = 93.22 avg = 93.12\n resnet50_int8 min = 158.54 max = 161.25 avg = 159.56\n squeezenet_ssd min = 55.88 max = 57.43 avg = 56.46\n squeezenet_ssd_int8 min = 72.42 max = 73.25 avg = 72.73\n mobilenet_ssd min = 35.38 max = 37.57 avg = 36.63\n mobilenet_ssd_int8 min = 62.92 max = 64.97 avg = 63.63\n mobilenet_yolo min = 76.56 max = 80.44 avg = 78.05\n mobilenetv2_yolov3 min = 46.35 max = 48.14 avg = 47.26\n yolov4-tiny min = 95.38 max = 97.55 avg = 96.45\n nanodet_m min = 22.82 max = 26.01 avg = 24.48\n yolo-fastest-1.1 min = 20.23 max = 25.51 avg = 21.52\n yolo-fastestv2 min = 20.67 max = 20.82 avg = 20.75\n\n[0 NVIDIA Tegra X1 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]\n[0 NVIDIA Tegra X1 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra X1 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra X1 (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 12.00 max = 15.41 avg = 13.55\n squeezenet_int8 min = 78.76 max = 79.14 avg = 78.91\n mobilenet min = 16.03 max = 16.25 avg = 16.15\n mobilenet_int8 min = 107.58 max = 107.68 avg = 107.61\n mobilenet_v2 min = 12.84 max = 13.13 avg = 12.99\n mobilenet_v3 min = 13.29 max = 16.64 avg = 14.38\n shufflenet min = 14.23 max = 14.54 avg = 14.34\n shufflenet_v2 min = 12.94 max = 13.21 avg = 13.02\n mnasnet min = 13.42 max = 13.66 avg = 13.53\n proxylessnasnet min = 14.64 max = 14.94 avg = 14.76\n efficientnet_b0 min = 21.28 max = 21.51 avg = 21.36\n efficientnetv2_b0 min = 74.32 max = 78.50 avg = 77.79\n regnety_400m min = 17.94 max = 18.26 avg = 18.07\n blazeface min = 6.83 max = 6.94 avg = 6.89\n googlenet min = 43.45 max = 43.63 avg = 43.52\n googlenet_int8 min = 255.68 max = 256.33 avg = 255.92\n resnet18 min = 43.96 max = 44.06 avg = 44.01\n resnet18_int8 min = 192.01 max = 192.64 avg = 192.33\n alexnet min = 74.04 max = 74.23 avg = 74.14\n vgg16 min = 310.32 max = 310.64 avg = 310.44\n vgg16_int8 min = 1003.05 max = 1004.27 avg = 1003.66\n resnet50 min = 93.05 max = 93.34 avg = 93.21\n resnet50_int8 min = 516.27 max = 517.12 avg = 516.69\n squeezenet_ssd min = 56.67 max = 56.86 avg = 56.73\n squeezenet_ssd_int8 min = 182.96 max = 184.26 avg = 183.71\n mobilenet_ssd min = 35.61 max = 35.70 avg = 35.65\n mobilenet_ssd_int8 min = 217.02 max = 217.50 avg = 217.23\n mobilenet_yolo min = 78.10 max = 78.36 avg = 78.20\n mobilenetv2_yolov3 min = 49.86 max = 57.83 avg = 53.18\n yolov4-tiny min = 96.76 max = 96.86 avg = 96.82\n nanodet_m min = 25.26 max = 25.36 avg = 25.31\n yolo-fastest-1.1 min = 21.55 max = 24.22 avg = 23.78\n yolo-fastestv2 min = 20.80 max = 21.01 avg = 20.90\n\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 30.03 max = 31.41 avg = 30.59\n squeezenet_int8 min = 27.32 max = 27.76 avg = 27.50\n mobilenet min = 41.74 max = 42.57 avg = 42.05\n mobilenet_int8 min = 30.48 max = 31.57 avg = 30.85\n mobilenet_v2 min = 33.49 max = 34.18 avg = 33.83\n mobilenet_v3 min = 30.59 max = 30.96 avg = 30.79\n shufflenet min = 21.07 max = 31.68 avg = 22.53\n shufflenet_v2 min = 19.55 max = 20.01 avg = 19.71\n mnasnet min = 31.70 max = 32.26 avg = 31.93\n proxylessnasnet min = 36.90 max = 38.55 avg = 37.27\n efficientnet_b0 min = 68.42 max = 77.60 avg = 70.60\n efficientnetv2_b0 min = 73.72 max = 81.05 avg = 75.31\n regnety_400m min = 56.67 max = 66.82 avg = 58.24\n blazeface min = 6.55 max = 6.96 avg = 6.74\n googlenet min = 92.74 max = 94.22 avg = 93.12\n googlenet_int8 min = 80.86 max = 87.28 avg = 82.41\n resnet18 min = 83.10 max = 84.30 avg = 83.44\n resnet18_int8 min = 59.40 max = 65.86 avg = 60.70\n alexnet min = 89.21 max = 92.45 avg = 89.98\n vgg16 min = 445.72 max = 451.09 avg = 447.39\n vgg16_int8 min = 292.81 max = 295.55 avg = 294.34\n resnet50 min = 203.42 max = 204.45 avg = 204.08\n resnet50_int8 min = 157.87 max = 160.30 avg = 158.67\n squeezenet_ssd min = 85.60 max = 87.24 avg = 86.18\n squeezenet_ssd_int8 min = 73.10 max = 85.64 avg = 74.94\n mobilenet_ssd min = 86.75 max = 96.51 avg = 88.49\n mobilenet_ssd_int8 min = 63.40 max = 71.57 avg = 64.97\n mobilenet_yolo min = 193.84 max = 195.24 avg = 194.62\n mobilenetv2_yolov3 min = 115.80 max = 117.27 avg = 116.27\n yolov4-tiny min = 156.30 max = 158.26 avg = 156.81\n nanodet_m min = 46.64 max = 47.97 avg = 47.12\n yolo-fastest-1.1 min = 25.78 max = 27.86 avg = 26.29\n yolo-fastestv2 min = 20.54 max = 30.73 avg = 22.18\n\nloop_count = 8\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 85.91 max = 86.86 avg = 86.14\n squeezenet_int8 min = 77.57 max = 78.10 avg = 77.69\n mobilenet min = 137.43 max = 138.03 avg = 137.63\n mobilenet_int8 min = 108.06 max = 108.21 avg = 108.13\n mobilenet_v2 min = 93.81 max = 94.70 avg = 93.99\n mobilenet_v3 min = 81.77 max = 82.49 avg = 81.99\n shufflenet min = 47.84 max = 48.46 avg = 48.17\n shufflenet_v2 min = 47.93 max = 48.23 avg = 48.09\n mnasnet min = 91.73 max = 92.55 avg = 91.98\n proxylessnasnet min = 115.41 max = 115.75 avg = 115.56\n efficientnet_b0 min = 225.64 max = 226.21 avg = 225.94\n efficientnetv2_b0 min = 239.71 max = 240.20 avg = 239.89\n regnety_400m min = 118.46 max = 118.84 avg = 118.61\n blazeface min = 15.58 max = 17.14 avg = 16.21\n googlenet min = 286.85 max = 287.51 avg = 287.11\n googlenet_int8 min = 256.44 max = 256.74 avg = 256.53\n resnet18 min = 221.27 max = 221.93 avg = 221.60\n resnet18_int8 min = 189.95 max = 191.34 avg = 190.74\n alexnet min = 284.30 max = 285.40 avg = 284.87\n vgg16 min = 1241.51 max = 1244.53 avg = 1242.90\n vgg16_int8 min = 1003.92 max = 1004.47 avg = 1004.29\n resnet50 min = 624.43 max = 625.34 avg = 624.84\n resnet50_int8 min = 516.64 max = 517.26 avg = 516.99\n squeezenet_ssd min = 190.21 max = 191.35 avg = 190.71\n squeezenet_ssd_int8 min = 182.97 max = 184.19 avg = 183.38\n mobilenet_ssd min = 275.60 max = 276.17 avg = 275.90\n mobilenet_ssd_int8 min = 216.67 max = 217.58 avg = 216.94\n mobilenet_yolo min = 616.16 max = 617.45 avg = 616.71\n mobilenetv2_yolov3 min = 324.88 max = 325.73 avg = 325.19\n yolov4-tiny min = 421.01 max = 423.52 avg = 422.14\n nanodet_m min = 117.39 max = 117.75 avg = 117.54\n yolo-fastest-1.1 min = 54.55 max = 55.61 avg = 54.87\n yolo-fastestv2 min = 44.40 max = 44.78 avg = 44.57\n```\n\n### NVIDIA Jetson TX2 NX(NV-Denver2 2.0Ghz x 2 + Cortex-A57 2.0Ghz x 4 + 256-core NVIDIA Pascal iGPU)\n```\nfan@ubuntu:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 $(nproc) 0 0\n[0 NVIDIA Tegra X2 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]\n[0 NVIDIA Tegra X2 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra X2 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra X2 (nvgpu)] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA Tegra X2 (nvgpu)] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 6\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 4.84 max = 6.12 avg = 5.33\n squeezenet_int8 min = 23.14 max = 148.62 avg = 52.65\n mobilenet min = 7.23 max = 7.57 avg = 7.40\n mobilenet_int8 min = 19.69 max = 101.50 avg = 44.15\n mobilenet_v2 min = 6.65 max = 6.86 avg = 6.76\n mobilenet_v3 min = 7.22 max = 8.34 avg = 8.01\n shufflenet min = 6.14 max = 6.73 avg = 6.51\n shufflenet_v2 min = 5.33 max = 5.43 avg = 5.39\n mnasnet min = 6.98 max = 7.47 avg = 7.16\n proxylessnasnet min = 6.90 max = 7.52 avg = 7.09\n efficientnet_b0 min = 11.42 max = 11.89 avg = 11.67\n efficientnetv2_b0 min = 26.48 max = 51.57 avg = 36.25\n regnety_400m min = 8.94 max = 9.45 avg = 9.13\n blazeface min = 2.08 max = 3.21 avg = 2.42\n googlenet min = 15.33 max = 15.78 avg = 15.53\n googlenet_int8 min = 64.02 max = 158.22 avg = 79.32\n resnet18 min = 12.25 max = 13.28 avg = 12.78\n resnet18_int8 min = 41.89 max = 156.59 avg = 57.07\n alexnet min = 20.15 max = 20.51 avg = 20.32\n vgg16 min = 62.45 max = 64.63 avg = 63.06\n vgg16_int8 min = 198.24 max = 271.71 avg = 217.63\n resnet50 min = 30.05 max = 31.11 avg = 30.39\n resnet50_int8 min = 129.03 max = 205.33 avg = 154.72\n squeezenet_ssd min = 18.48 max = 22.90 avg = 20.26\n squeezenet_ssd_int8 min = 48.18 max = 71.20 avg = 60.89\n mobilenet_ssd min = 15.56 max = 15.76 avg = 15.67\n mobilenet_ssd_int8 min = 55.10 max = 114.34 avg = 67.41\n mobilenet_yolo min = 28.75 max = 32.54 avg = 30.30\n mobilenetv2_yolov3 min = 26.15 max = 32.36 avg = 29.57\n yolov4-tiny min = 23.08 max = 37.19 avg = 25.43\n nanodet_m min = 15.81 max = 19.99 avg = 18.10\n yolo-fastest-1.1 min = 7.35 max = 11.26 avg = 8.69\n yolo-fastestv2 min = 6.16 max = 6.61 avg = 6.31\n vision_transformer min = 1301.45 max = 1356.58 avg = 1321.51\n FastestDet min = 5.64 max = 6.60 avg = 5.90\nfan@ubuntu:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 1 0 0\n[0 NVIDIA Tegra X2 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]\n[0 NVIDIA Tegra X2 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra X2 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra X2 (nvgpu)] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA Tegra X2 (nvgpu)] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 5.10 max = 6.33 avg = 5.51\n squeezenet_int8 min = 56.36 max = 59.23 avg = 57.79\n mobilenet min = 6.61 max = 9.93 avg = 7.27\n mobilenet_int8 min = 95.73 max = 107.69 avg = 99.35\n mobilenet_v2 min = 6.66 max = 9.87 avg = 7.22\n mobilenet_v3 min = 7.20 max = 8.77 avg = 7.61\n shufflenet min = 5.87 max = 6.13 avg = 5.97\n shufflenet_v2 min = 5.63 max = 8.24 avg = 6.10\n mnasnet min = 6.55 max = 9.05 avg = 7.10\n proxylessnasnet min = 7.29 max = 7.86 avg = 7.50\n efficientnet_b0 min = 11.22 max = 12.13 avg = 11.49\n efficientnetv2_b0 min = 20.21 max = 24.55 avg = 21.42\n regnety_400m min = 8.94 max = 10.77 avg = 9.37\n blazeface min = 2.30 max = 2.45 avg = 2.35\n googlenet min = 15.48 max = 17.88 avg = 16.32\n googlenet_int8 min = 197.08 max = 205.18 avg = 200.93\n resnet18 min = 12.69 max = 13.38 avg = 13.01\n resnet18_int8 min = 147.42 max = 154.63 avg = 149.94\n alexnet min = 20.49 max = 20.83 avg = 20.62\n vgg16 min = 62.43 max = 63.41 avg = 62.81\n vgg16_int8 min = 802.28 max = 810.33 avg = 805.66\n resnet50 min = 29.96 max = 30.56 avg = 30.26\n resnet50_int8 min = 488.38 max = 494.67 avg = 491.09\n squeezenet_ssd min = 18.35 max = 18.84 avg = 18.59\n squeezenet_ssd_int8 min = 121.27 max = 124.52 avg = 122.21\n mobilenet_ssd min = 15.13 max = 15.60 avg = 15.30\n mobilenet_ssd_int8 min = 206.22 max = 225.98 avg = 222.55\n mobilenet_yolo min = 30.12 max = 31.28 avg = 30.41\n mobilenetv2_yolov3 min = 26.65 max = 27.08 avg = 26.87\n yolov4-tiny min = 22.91 max = 23.32 avg = 23.04\n nanodet_m min = 11.57 max = 11.99 avg = 11.75\n yolo-fastest-1.1 min = 7.06 max = 7.49 avg = 7.25\n yolo-fastestv2 min = 6.17 max = 6.65 avg = 6.34\n vision_transformer min = 1185.13 max = 1193.94 avg = 1189.50\n FastestDet min = 5.78 max = 6.87 avg = 6.11\nfan@ubuntu:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 $(nproc) 0 -1\nloop_count = 10\nnum_threads = 6\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 19.92 max = 22.96 avg = 21.43\n squeezenet_int8 min = 20.33 max = 25.17 avg = 22.63\n mobilenet min = 27.25 max = 80.19 avg = 36.64\n mobilenet_int8 min = 21.22 max = 31.14 avg = 27.05\n mobilenet_v2 min = 21.95 max = 25.77 avg = 24.10\n mobilenet_v3 min = 20.10 max = 34.13 avg = 25.30\n shufflenet min = 14.96 max = 108.36 avg = 28.88\n shufflenet_v2 min = 13.25 max = 29.33 avg = 16.43\n mnasnet min = 19.41 max = 111.63 avg = 30.57\n proxylessnasnet min = 22.58 max = 27.29 avg = 24.43\n efficientnet_b0 min = 32.95 max = 35.53 avg = 34.46\n efficientnetv2_b0 min = 36.91 max = 52.12 avg = 41.72\n regnety_400m min = 43.87 max = 152.33 avg = 56.15\n blazeface min = 4.51 max = 16.71 avg = 6.79\n googlenet min = 59.37 max = 93.96 avg = 70.88\n googlenet_int8 min = 57.95 max = 124.06 avg = 71.47\n resnet18 min = 51.99 max = 134.81 avg = 68.50\n resnet18_int8 min = 40.54 max = 130.18 avg = 54.10\n alexnet min = 41.42 max = 67.03 avg = 52.66\n vgg16 min = 253.75 max = 295.39 avg = 265.01\n vgg16_int8 min = 183.96 max = 334.83 avg = 206.81\n resnet50 min = 305.79 max = 330.68 avg = 316.55\n resnet50_int8 min = 120.10 max = 133.19 avg = 125.92\n squeezenet_ssd min = 51.06 max = 125.69 avg = 67.34\n squeezenet_ssd_int8 min = 44.56 max = 156.68 avg = 61.47\n mobilenet_ssd min = 52.27 max = 123.50 avg = 64.86\n mobilenet_ssd_int8 min = 48.18 max = 183.44 avg = 63.25\n mobilenet_yolo min = 120.27 max = 160.73 avg = 130.75\n mobilenetv2_yolov3 min = 74.39 max = 167.08 avg = 86.50\n yolov4-tiny min = 108.39 max = 123.62 avg = 112.81\n nanodet_m min = 32.38 max = 91.62 avg = 42.01\n yolo-fastest-1.1 min = 17.97 max = 157.78 avg = 34.93\n yolo-fastestv2 min = 16.12 max = 19.55 avg = 18.03\n vision_transformer min = 2317.30 max = 2437.95 avg = 2375.98\n FastestDet min = 15.52 max = 127.95 avg = 27.40\nfan@ubuntu:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 1 0 -1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 48.72 max = 50.66 avg = 49.98\n squeezenet_int8 min = 56.50 max = 61.58 avg = 58.64\n mobilenet min = 88.10 max = 89.76 avg = 88.92\n mobilenet_int8 min = 95.08 max = 96.92 avg = 95.82\n mobilenet_v2 min = 58.72 max = 61.48 avg = 59.54\n mobilenet_v3 min = 48.58 max = 49.95 avg = 49.24\n shufflenet min = 30.42 max = 32.03 avg = 31.17\n shufflenet_v2 min = 28.27 max = 29.37 avg = 28.65\n mnasnet min = 56.85 max = 58.22 avg = 57.37\n proxylessnasnet min = 68.67 max = 71.23 avg = 69.64\n efficientnet_b0 min = 89.27 max = 92.67 avg = 90.33\n efficientnetv2_b0 min = 107.72 max = 109.86 avg = 108.53\n regnety_400m min = 85.19 max = 91.74 avg = 86.95\n blazeface min = 8.60 max = 8.80 avg = 8.71\n googlenet min = 161.58 max = 166.70 avg = 163.60\n googlenet_int8 min = 183.79 max = 189.43 avg = 186.17\n resnet18 min = 123.43 max = 126.29 avg = 124.86\n resnet18_int8 min = 140.80 max = 144.92 avg = 142.60\n alexnet min = 93.16 max = 100.47 avg = 96.44\n vgg16 min = 664.14 max = 671.67 avg = 667.90\n vgg16_int8 min = 799.67 max = 813.66 avg = 803.50\n resnet50 min = 384.10 max = 388.46 avg = 386.49\n resnet50_int8 min = 448.11 max = 473.27 avg = 465.12\n squeezenet_ssd min = 106.58 max = 109.62 avg = 107.39\n squeezenet_ssd_int8 min = 118.39 max = 122.62 avg = 120.43\n mobilenet_ssd min = 178.89 max = 183.37 avg = 180.47\n mobilenet_ssd_int8 min = 201.46 max = 207.18 avg = 203.00\n mobilenet_yolo min = 407.54 max = 411.12 avg = 409.33\n mobilenetv2_yolov3 min = 211.83 max = 214.46 avg = 213.20\n yolov4-tiny min = 249.11 max = 254.22 avg = 251.38\n nanodet_m min = 69.41 max = 71.26 avg = 70.28\n yolo-fastest-1.1 min = 30.99 max = 33.29 avg = 32.03\n yolo-fastestv2 min = 27.70 max = 28.90 avg = 27.93\n vision_transformer min = 3203.45 max = 3402.10 avg = 3286.58\n FastestDet min = 29.05 max = 32.57 avg = 30.53\n```\n\n### Rockchip RK3288-CG.W (Cortex-A17 1.8GHz x 4)\n```\nWW_Tinker_Board:/data/local/tmp # ./benchncnn 8 4 0 -1 1\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 56.61 max = 56.80 avg = 56.69\n squeezenet_int8 min = 40.63 max = 41.05 avg = 40.89\n mobilenet min = 83.91 max = 84.59 avg = 84.23\n mobilenet_int8 min = 36.15 max = 36.44 avg = 36.25\n mobilenet_v2 min = 71.12 max = 71.73 avg = 71.54\n mobilenet_v3 min = 56.08 max = 56.56 avg = 56.28\n shufflenet min = 37.39 max = 37.75 avg = 37.55\n shufflenet_v2 min = 35.19 max = 35.52 avg = 35.34\n mnasnet min = 62.08 max = 62.36 avg = 62.24\n proxylessnasnet min = 66.98 max = 67.38 avg = 67.16\n efficientnet_b0 min = 109.95 max = 110.71 avg = 110.15\n efficientnetv2_b0 min = 122.56 max = 123.31 avg = 122.94\n regnety_400m min = 88.84 max = 89.19 avg = 88.99\n blazeface min = 11.79 max = 11.92 avg = 11.85\n googlenet min = 162.56 max = 165.39 avg = 163.19\n googlenet_int8 min = 110.35 max = 110.91 avg = 110.60\n resnet18 min = 172.39 max = 173.99 avg = 173.24\n resnet18_int8 min = 84.00 max = 84.40 avg = 84.19\n alexnet min = 156.71 max = 158.23 avg = 157.59\n vgg16 min = 956.95 max = 964.32 avg = 960.60\n vgg16_int8 min = 388.10 max = 389.52 avg = 388.68\n resnet50 min = 403.05 max = 404.80 avg = 404.01\n resnet50_int8 min = 205.12 max = 207.42 avg = 206.19\n squeezenet_ssd min = 163.61 max = 165.79 avg = 164.93\n squeezenet_ssd_int8 min = 125.88 max = 126.35 avg = 126.12\n mobilenet_ssd min = 175.97 max = 176.86 avg = 176.39\n mobilenet_ssd_int8 min = 76.90 max = 77.74 avg = 77.35\n mobilenet_yolo min = 385.59 max = 387.19 avg = 386.60\n mobilenetv2_yolov3 min = 234.88 max = 236.22 avg = 235.66\n yolov4-tiny min = 307.44 max = 310.64 avg = 308.54\n nanodet_m min = 92.54 max = 93.15 avg = 92.82\n yolo-fastest-1.1 min = 46.69 max = 47.02 avg = 46.83\n yolo-fastestv2 min = 38.37 max = 38.68 avg = 38.54\n\nWW_Tinker_Board:/data/local/tmp # ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 138.27 max = 138.57 avg = 138.41\n squeezenet_int8 min = 85.97 max = 86.23 avg = 86.05\n mobilenet min = 234.90 max = 235.08 avg = 235.00\n mobilenet_int8 min = 99.92 max = 100.45 avg = 100.12\n mobilenet_v2 min = 157.76 max = 157.99 avg = 157.86\n mobilenet_v3 min = 130.05 max = 130.23 avg = 130.17\n shufflenet min = 74.48 max = 74.62 avg = 74.55\n shufflenet_v2 min = 74.05 max = 74.25 avg = 74.13\n mnasnet min = 150.74 max = 151.03 avg = 150.87\n proxylessnasnet min = 171.09 max = 171.23 avg = 171.16\n efficientnet_b0 min = 306.85 max = 307.02 avg = 306.97\n efficientnetv2_b0 min = 347.40 max = 347.87 avg = 347.64\n regnety_400m min = 190.26 max = 190.33 avg = 190.29\n blazeface min = 25.25 max = 25.68 avg = 25.47\n googlenet min = 432.09 max = 432.48 avg = 432.32\n googlenet_int8 min = 275.55 max = 276.07 avg = 275.88\n resnet18 min = 355.11 max = 358.56 avg = 356.90\n resnet18_int8 min = 205.80 max = 206.68 avg = 206.26\n alexnet min = 330.09 max = 330.29 avg = 330.15\n vgg16 min = 2122.95 max = 2124.45 avg = 2123.68\n vgg16_int8 min = 1048.53 max = 1049.29 avg = 1048.86\n resnet50 min = 1047.27 max = 1048.33 avg = 1047.63\n resnet50_int8 min = 517.75 max = 519.28 avg = 518.81\n squeezenet_ssd min = 304.69 max = 305.75 avg = 305.16\n squeezenet_ssd_int8 min = 219.16 max = 219.94 avg = 219.45\n mobilenet_ssd min = 483.73 max = 484.12 avg = 484.01\n mobilenet_ssd_int8 min = 208.89 max = 209.19 avg = 209.09\n mobilenet_yolo min = 1092.75 max = 1093.70 avg = 1093.13\n mobilenetv2_yolov3 min = 560.66 max = 560.92 avg = 560.77\n yolov4-tiny min = 704.69 max = 705.38 avg = 705.12\n nanodet_m min = 187.13 max = 187.57 avg = 187.39\n yolo-fastest-1.1 min = 83.05 max = 83.11 avg = 83.08\n yolo-fastestv2 min = 72.19 max = 72.23 avg = 72.21\n\nWW_Tinker_Board:/data/local/tmp # ./benchncnn 4 1 0 0 0\n[0 Mali-T760] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-T760] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=1\n[0 Mali-T760] fp16-p/s/a=1/0/1 int8-p/s/a=1/0/0\n[0 Mali-T760] subgroup=0 basic=0 vote=0 ballot=0 shuffle=0\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 41.78 max = 41.82 avg = 41.79\n mobilenet min = 62.67 max = 62.80 avg = 62.74\n mobilenet_v2 min = 51.08 max = 51.26 avg = 51.17\n mobilenet_v3 min = 51.43 max = 51.70 avg = 51.51\n shufflenet min = 56.83 max = 56.94 avg = 56.87\n shufflenet_v2 min = 48.46 max = 48.63 avg = 48.53\n mnasnet min = 52.31 max = 52.63 avg = 52.42\n proxylessnasnet min = 57.33 max = 57.46 avg = 57.41\n efficientnet_b0 min = 87.52 max = 87.80 avg = 87.62\n efficientnetv2_b0 min = 123.83 max = 124.67 avg = 124.34\n regnety_400m min = 65.52 max = 65.81 avg = 65.64\n blazeface min = 14.56 max = 14.73 avg = 14.62\n googlenet min = 138.52 max = 139.39 avg = 138.89\n resnet18 min = 124.45 max = 124.81 avg = 124.58\n alexnet min = 130.46 max = 130.68 avg = 130.54\n```\n### HiSilicon Hi3519V101 (Cortex-A17 1.2GHz x 1)\n```\nroot@Hi3519:/ncnn-benchmark # taskset 2 ./benchncnn 8 1 0\nloop_count = 8\nnum_threads = 1\npowersave = 0\n squeezenet min = 272.97 max = 275.84 avg = 274.85\n squeezenet-int8 min = 200.87 max = 202.47 avg = 201.74\n mobilenet min = 480.90 max = 482.16 avg = 481.64\n mobilenet_v2 min = 350.01 max = 352.39 avg = 350.81\n shufflenet min = 152.40 max = 153.17 avg = 152.80\n googlenet min = 1096.65 max = 1101.35 avg = 1099.21\n resnet18 min = 983.92 max = 987.00 avg = 985.25\n alexnet min = 1140.30 max = 1141.55 avg = 1140.92\n squeezenet-ssd min = 574.62 max = 580.12 avg = 577.23\n mobilenet-ssd min = 960.26 max = 969.13 avg = 965.93\n mobilenet-yolo min = 1867.78 max = 1880.08 avg = 1873.89\n```\n\n### iPhone 5S (Apple A7 1.3GHz x 2)\n```\niPhone:~ root# ./benchncnn 8 2 0 -1\n[0 Apple A7 GPU] queueC=0[8] queueT=0[8] memU=1 memDL=1 memHV=1\n[0 Apple A7 GPU] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = -1\n squeezenet min = 49.21 max = 50.40 avg = 49.74\n squeezenet_int8 min = 54.73 max = 57.39 avg = 56.70\n mobilenet min = 79.03 max = 80.00 avg = 79.44\n mobilenet_int8 min = 109.95 max = 112.69 avg = 111.38\n mobilenet_v2 min = 57.34 max = 57.88 avg = 57.47\n mobilenet_v3 min = 52.66 max = 53.73 avg = 53.12\n shufflenet min = 32.78 max = 36.12 avg = 35.12\n shufflenet_v2 min = 31.25 max = 32.10 avg = 31.61\n mnasnet min = 54.58 max = 56.12 avg = 55.44\n proxylessnasnet min = 69.52 max = 72.42 avg = 70.40\n googlenet min = 192.82 max = 194.20 avg = 193.35\n googlenet_int8 min = 235.43 max = 244.71 avg = 239.64\n resnet18 min = 164.33 max = 167.27 avg = 165.51\n resnet18_int8 min = 176.16 max = 179.73 avg = 178.60\n alexnet min = 224.50 max = 228.21 avg = 226.51\n vgg16 min = 4262.28 max = 4400.29 avg = 4300.34\n vgg16_int8 min = 2835.84 max = 2955.22 avg = 2890.26\n resnet50 min = 542.66 max = 1344.49 avg = 737.05\n resnet50_int8 min = 426.08 max = 435.34 avg = 431.87\n squeezenet_ssd min = 129.03 max = 131.44 avg = 129.99\n squeezenet_ssd_int8 min = 155.52 max = 161.42 avg = 158.51\n mobilenet_ssd min = 168.18 max = 170.17 avg = 169.42\n mobilenet_ssd_int8 min = 205.78 max = 212.07 avg = 209.66\n mobilenet_yolo min = 347.32 max = 363.15 avg = 355.72\n mobilenetv2_yolov3 min = 193.11 max = 196.64 avg = 194.31\n\niPhone:~ root# ./benchncnn 4 1 0 -1\n[0 Apple A7 GPU] queueC=0[8] queueT=0[8] memU=1 memDL=1 memHV=1\n[0 Apple A7 GPU] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\n squeezenet min = 86.36 max = 86.81 avg = 86.57\n squeezenet_int8 min = 99.62 max = 100.07 avg = 99.83\n mobilenet min = 143.11 max = 146.50 avg = 145.38\n mobilenet_int8 min = 202.25 max = 203.32 avg = 203.02\n mobilenet_v2 min = 97.56 max = 98.55 avg = 98.09\n mobilenet_v3 min = 87.45 max = 87.68 avg = 87.52\n shufflenet min = 54.01 max = 54.13 avg = 54.08\n shufflenet_v2 min = 48.11 max = 48.65 avg = 48.36\n mnasnet min = 95.02 max = 95.77 avg = 95.25\n proxylessnasnet min = 123.91 max = 124.61 avg = 124.18\n googlenet min = 344.23 max = 348.95 avg = 345.97\n googlenet_int8 min = 420.30 max = 420.99 avg = 420.65\n resnet18 min = 300.44 max = 301.36 avg = 300.99\n resnet18_int8 min = 308.60 max = 310.52 avg = 309.70\n alexnet min = 423.92 max = 429.84 avg = 427.24\n vgg16 min = 4787.59 max = 5015.23 avg = 4900.43\n vgg16_int8 min = 3560.59 max = 3722.75 avg = 3639.88\n resnet50 min = 797.88 max = 1294.57 avg = 985.63\n resnet50_int8 min = 751.15 max = 760.25 avg = 757.89\n squeezenet_ssd min = 193.75 max = 196.13 avg = 195.29\n squeezenet_ssd_int8 min = 243.78 max = 245.19 avg = 244.74\n mobilenet_ssd min = 299.69 max = 307.22 avg = 305.12\n mobilenet_ssd_int8 min = 385.91 max = 389.82 avg = 388.48\n mobilenet_yolo min = 657.00 max = 659.31 avg = 658.08\n mobilenetv2_yolov3 min = 335.59 max = 342.22 avg = 339.37\n\niPhone:~ root# ./benchncnn 4 1 0 0\n[0 Apple A7 GPU] queueC=0[8] queueT=0[8] memU=1 memDL=1 memHV=1\n[0 Apple A7 GPU] fp16p=1 fp16s=0 fp16a=0 int8s=0 int8a=0\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = 0\n squeezenet min = 260.18 max = 262.55 avg = 261.09\n mobilenet min = 288.73 max = 291.83 avg = 289.67\n mobilenet_v2 min = 265.72 max = 267.05 avg = 266.14\n mobilenet_v3 min = 255.86 max = 257.35 avg = 256.43\n shufflenet min = 236.66 max = 239.49 avg = 237.98\n shufflenet_v2 min = 244.92 max = 247.75 avg = 246.22\n mnasnet min = 254.75 max = 256.48 avg = 255.85\n proxylessnasnet min = 281.42 max = 282.62 avg = 282.11\n googlenet min = 745.36 max = 764.91 avg = 754.16\n resnet18 min = 721.26 max = 741.98 avg = 734.78\n alexnet min = 521.43 max = 530.95 avg = 527.01\n resnet50 min = 1494.86 max = 1505.79 avg = 1501.49\n squeezenet_ssd min = 1096.45 max = 1102.84 avg = 1098.55\n mobilenet_ssd min = 639.50 max = 641.81 avg = 640.83\n mobilenet_yolo min = 1445.16 max = 1450.94 avg = 1447.42\n mobilenetv2_yolov3 min = 1047.24 max = 1060.97 avg = 1052.86\n```\n\n### Freescale i.MX7 Dual (Cortex A7 1.0GHz x 2)\n```\nimx7d_pico:/data/local/tmp $ ./benchncnn 8 2 0 -1 1\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 220.10 max = 226.46 avg = 222.89\n squeezenet_int8 min = 159.26 max = 165.25 avg = 161.71\n mobilenet min = 366.92 max = 373.78 avg = 371.55\n mobilenet_int8 min = 223.14 max = 229.66 avg = 225.66\n mobilenet_v2 min = 252.32 max = 259.41 avg = 255.54\n mobilenet_v3 min = 214.05 max = 222.24 avg = 217.53\n shufflenet min = 137.02 max = 144.79 avg = 138.85\n shufflenet_v2 min = 134.89 max = 140.75 avg = 137.18\n mnasnet min = 250.64 max = 256.75 avg = 253.33\n proxylessnasnet min = 285.35 max = 291.43 avg = 288.37\n efficientnet_b0 min = 430.47 max = 436.63 avg = 434.75\n regnety_400m min = 317.69 max = 325.77 avg = 321.24\n blazeface min = 42.93 max = 43.30 avg = 43.14\n googlenet min = 721.84 max = 728.40 avg = 724.23\n googlenet_int8 min = 504.07 max = 511.06 avg = 507.39\n resnet18 min = 645.61 max = 653.08 avg = 648.51\n resnet18_int8 min = 370.84 max = 514.38 avg = 392.80\n alexnet min = 783.64 max = 794.83 avg = 786.95\n squeezenet_ssd min = 508.71 max = 513.70 avg = 511.29\n squeezenet_ssd_int8 min = 402.85 max = 409.32 avg = 406.45\n mobilenet_ssd min = 763.70 max = 771.52 avg = 767.61\n mobilenet_ssd_int8 min = 457.99 max = 460.85 avg = 459.76\n mobilenet_yolo min = 1730.90 max = 1746.52 avg = 1741.26\n mobilenetv2_yolov3 min = 884.00 max = 892.97 avg = 889.38\n yolov4-tiny min = 1181.20 max = 1218.20 avg = 1202.28\n nanodet_m min = 331.53 max = 339.89 avg = 334.62\n\nimx7d_pico:/data/local/tmp $ ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 408.39 max = 410.27 avg = 408.95\n squeezenet_int8 min = 290.25 max = 290.95 avg = 290.61\n mobilenet min = 707.10 max = 711.64 avg = 708.47\n mobilenet_int8 min = 434.95 max = 436.16 avg = 435.66\n mobilenet_v2 min = 466.52 max = 467.41 avg = 466.96\n mobilenet_v3 min = 407.03 max = 408.29 avg = 407.56\n shufflenet min = 240.65 max = 241.07 avg = 240.85\n shufflenet_v2 min = 229.27 max = 235.66 avg = 231.51\n mnasnet min = 471.21 max = 471.48 avg = 471.35\n proxylessnasnet min = 544.74 max = 547.62 avg = 546.20\n efficientnet_b0 min = 824.09 max = 824.44 avg = 824.20\n regnety_400m min = 570.20 max = 571.73 avg = 570.82\n blazeface min = 76.46 max = 77.05 avg = 76.81\n googlenet min = 1368.82 max = 1369.99 avg = 1369.33\n googlenet_int8 min = 945.51 max = 946.61 avg = 945.91\n resnet18 min = 1237.79 max = 1257.12 avg = 1246.80\n resnet18_int8 min = 705.09 max = 706.72 avg = 705.63\n alexnet min = 1516.35 max = 1522.82 avg = 1519.52\n squeezenet_ssd min = 906.97 max = 908.48 avg = 907.68\n squeezenet_ssd_int8 min = 727.15 max = 728.16 avg = 727.77\n mobilenet_ssd min = 1475.19 max = 1478.52 avg = 1476.81\n mobilenet_ssd_int8 min = 883.88 max = 890.68 avg = 885.90\n mobilenet_yolo min = 3408.43 max = 3418.63 avg = 3412.52\n mobilenetv2_yolov3 min = 1685.18 max = 1695.89 avg = 1689.23\n yolov4-tiny min = 2168.24 max = 2183.24 avg = 2175.93\n nanodet_m min = 561.56 max = 562.05 avg = 561.72\n```\n\n### Z7-Lite 7020 XC7Z020CLG400-2 (Cortex-A9 766MHz x 2)\n```\nroot@petalinux_hdmi:~# LD_LIBRARY_PATH=. ./benchncnn 8 2 0 -1 1\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 389.18 max = 390.13 avg = 389.60\n squeezenet_int8 min = 254.33 max = 255.24 avg = 254.85\n mobilenet min = 623.71 max = 625.01 avg = 624.46\n mobilenet_int8 min = 240.40 max = 241.03 avg = 240.87\n mobilenet_v2 min = 450.00 max = 450.89 avg = 450.40\n mobilenet_v3 min = 362.99 max = 363.66 avg = 363.28\n shufflenet min = 212.20 max = 213.28 avg = 212.84\n shufflenet_v2 min = 210.26 max = 212.64 avg = 211.53\n mnasnet min = 408.67 max = 409.64 avg = 409.17\n proxylessnasnet min = 449.86 max = 450.94 avg = 450.45\n efficientnet_b0 min = 737.40 max = 739.58 avg = 738.32\n efficientnetv2_b0 min = 848.58 max = 849.74 avg = 849.24\n regnety_400m min = 501.32 max = 503.02 avg = 501.87\n blazeface min = 70.89 max = 72.22 avg = 71.61\n squeezenet_ssd min = 978.55 max = 979.86 avg = 979.22\n squeezenet_ssd_int8 min = 691.90 max = 694.18 avg = 692.73\n mobilenet_ssd min = 1353.12 max = 1354.13 avg = 1353.53\n mobilenet_ssd_int8 min = 496.26 max = 497.29 avg = 496.61\n nanodet_m min = 542.04 max = 546.29 avg = 544.73\n yolo-fastest-1.1 min = 282.75 max = 286.11 avg = 284.24\n yolo-fastestv2 min = 230.91 max = 232.74 avg = 231.56\n\nroot@petalinux_hdmi:~# LD_LIBRARY_PATH=. ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 637.19 max = 639.33 avg = 637.82\n squeezenet_int8 min = 390.31 max = 391.63 avg = 390.94\n mobilenet min = 1085.54 max = 1085.96 avg = 1085.71\n mobilenet_int8 min = 437.28 max = 437.65 avg = 437.44\n mobilenet_v2 min = 716.03 max = 716.75 avg = 716.35\n mobilenet_v3 min = 587.83 max = 588.55 avg = 588.21\n shufflenet min = 331.28 max = 331.97 avg = 331.63\n shufflenet_v2 min = 331.03 max = 333.19 avg = 331.76\n mnasnet min = 682.68 max = 683.11 avg = 682.82\n proxylessnasnet min = 763.89 max = 764.80 avg = 764.35\n efficientnet_b0 min = 1288.61 max = 1289.10 avg = 1288.81\n efficientnetv2_b0 min = 1499.12 max = 1500.11 avg = 1499.65\n regnety_400m min = 852.03 max = 853.16 avg = 852.68\n blazeface min = 109.40 max = 111.51 avg = 110.41\n squeezenet_ssd min = 1493.25 max = 1497.00 avg = 1494.87\n squeezenet_ssd_int8 min = 1016.77 max = 1019.31 avg = 1017.99\n mobilenet_ssd min = 2379.20 max = 2379.83 avg = 2379.64\n mobilenet_ssd_int8 min = 881.70 max = 881.89 avg = 881.83\n nanodet_m min = 831.13 max = 832.58 avg = 831.87\n yolo-fastest-1.1 min = 466.80 max = 469.90 avg = 468.79\n yolo-fastestv2 min = 352.07 max = 355.20 avg = 353.36\n```\n\n### Loongson 2K1000 (GS264 1.0GHz x 2)\n```\nroot@ls2k:~/ncnn/build/benchmark# ./benchncnn 10 2 0 -1 1\nloop_count = 10\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 184.33 max = 184.94 avg = 184.65\n squeezenet_int8 min = 201.42 max = 201.99 avg = 201.72\n mobilenet min = 277.17 max = 278.04 avg = 277.66\n mobilenet_int8 min = 234.61 max = 235.17 avg = 234.81\n mobilenet_v2 min = 223.10 max = 274.92 avg = 228.71\n mobilenet_v3 min = 185.79 max = 201.76 avg = 187.60\n shufflenet min = 129.78 max = 131.09 avg = 130.28\n shufflenet_v2 min = 115.86 max = 116.77 avg = 116.42\n mnasnet min = 213.92 max = 214.72 avg = 214.26\n proxylessnasnet min = 240.05 max = 242.02 avg = 240.86\n efficientnet_b0 min = 347.52 max = 348.53 avg = 348.13\n efficientnetv2_b0 min = 382.78 max = 479.58 avg = 398.18\n regnety_400m min = 270.00 max = 312.84 avg = 274.66\n blazeface min = 37.60 max = 38.02 avg = 37.79\n googlenet min = 659.55 max = 693.17 avg = 666.17\n googlenet_int8 min = 678.26 max = 718.39 avg = 682.79\n resnet18 min = 499.75 max = 766.88 avg = 532.49\n resnet18_int8 min = 500.38 max = 533.97 avg = 504.56\n alexnet min = 508.49 max = 542.94 avg = 516.13\n vgg16 min = 2654.06 max = 3082.44 avg = 2762.51\n vgg16_int8 min = 2628.96 max = 2665.35 avg = 2647.12\n resnet50 min = 1256.97 max = 1417.45 avg = 1283.04\n resnet50_int8 min = 1232.55 max = 1276.94 avg = 1244.59\n squeezenet_ssd min = 538.83 max = 588.03 avg = 553.44\n squeezenet_ssd_int8 min = 501.67 max = 532.61 avg = 505.72\n mobilenet_ssd min = 571.14 max = 600.93 avg = 578.22\n mobilenet_ssd_int8 min = 478.67 max = 515.39 avg = 483.06\n mobilenet_yolo min = 1644.48 max = 1729.17 avg = 1669.18\n mobilenetv2_yolov3 min = 752.22 max = 792.40 avg = 760.10\n yolov4-tiny min = 994.48 max = 1096.10 avg = 1016.49\n nanodet_m min = 299.12 max = 343.99 avg = 303.98\n yolo-fastest-1.1 min = 141.56 max = 142.93 avg = 142.04\n yolo-fastestv2 min = 125.66 max = 168.88 avg = 130.28\n\nroot@ls2k:~/ncnn/build/benchmark# ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 295.48 max = 296.42 avg = 295.98\n squeezenet_int8 min = 334.05 max = 336.31 avg = 335.35\n mobilenet min = 476.33 max = 479.00 avg = 477.41\n mobilenet_int8 min = 446.03 max = 448.21 avg = 446.73\n mobilenet_v2 min = 343.26 max = 343.97 avg = 343.69\n mobilenet_v3 min = 296.84 max = 297.31 avg = 297.11\n shufflenet min = 202.31 max = 203.96 avg = 202.79\n shufflenet_v2 min = 181.69 max = 182.42 avg = 182.08\n mnasnet min = 353.73 max = 354.12 avg = 353.99\n proxylessnasnet min = 404.49 max = 405.00 avg = 404.75\n efficientnet_b0 min = 592.54 max = 593.81 avg = 593.14\n efficientnetv2_b0 min = 649.91 max = 651.49 avg = 650.54\n regnety_400m min = 425.96 max = 426.33 avg = 426.12\n blazeface min = 59.74 max = 60.19 avg = 59.90\n googlenet min = 1120.13 max = 1217.54 avg = 1146.27\n googlenet_int8 min = 1205.17 max = 1213.43 avg = 1208.13\n resnet18 min = 803.07 max = 997.37 avg = 856.09\n resnet18_int8 min = 911.74 max = 916.16 avg = 913.31\n alexnet min = 883.47 max = 903.08 avg = 889.06\n vgg16 min = 4425.52 max = 4587.36 avg = 4467.61\n vgg16_int8 min = 4896.90 max = 4993.15 avg = 4924.44\n resnet50 min = 2163.22 max = 2169.90 avg = 2167.49\n resnet50_int8 min = 2202.87 max = 2218.00 avg = 2210.51\n squeezenet_ssd min = 831.06 max = 926.94 avg = 856.24\n squeezenet_ssd_int8 min = 800.52 max = 803.28 avg = 801.72\n mobilenet_ssd min = 979.74 max = 980.82 avg = 980.22\n mobilenet_ssd_int8 min = 893.79 max = 895.41 avg = 894.51\n mobilenet_yolo min = 2578.17 max = 2586.30 avg = 2582.55\n mobilenetv2_yolov3 min = 1190.77 max = 1207.67 avg = 1196.06\n yolov4-tiny min = 1558.29 max = 1570.18 avg = 1561.52\n nanodet_m min = 442.90 max = 444.27 avg = 443.72\n yolo-fastest-1.1 min = 203.60 max = 208.43 avg = 205.20\n yolo-fastestv2 min = 184.61 max = 185.05 avg = 184.75\n```\n\n### Loongson 2K1000LA (LA264 1.0GHz * 2)\n```\nroot@ls2kla:~/ncnn/build/benchmark# ./benchncnn 10 2 0 -1 1\nloop_count = 10\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 151.11 max = 162.36 avg = 153.30\n squeezenet_int8 min = 195.32 max = 198.63 avg = 196.12\n mobilenet min = 279.27 max = 283.42 avg = 280.40\n mobilenet_int8 min = 264.78 max = 268.41 avg = 265.76\n mobilenet_v2 min = 204.39 max = 207.69 avg = 205.77\n mobilenet_v3 min = 171.32 max = 187.07 avg = 173.15\n shufflenet min = 147.43 max = 150.72 avg = 147.89\n shufflenet_v2 min = 169.42 max = 172.58 avg = 170.35\n mnasnet min = 204.87 max = 208.01 avg = 205.63\n proxylessnasnet min = 226.79 max = 237.74 avg = 229.02\n efficientnet_b0 min = 302.30 max = 310.91 avg = 303.87\n efficientnetv2_b0 min = 327.65 max = 361.15 avg = 334.45\n regnety_400m min = 264.08 max = 278.49 avg = 266.35\n blazeface min = 31.80 max = 39.18 avg = 32.88\n googlenet min = 562.95 max = 578.42 avg = 566.28\n googlenet_int8 min = 598.16 max = 613.56 avg = 601.68\n resnet18 min = 466.73 max = 472.08 avg = 469.58\n resnet18_int8 min = 489.69 max = 493.74 avg = 491.63\n alexnet min = 381.35 max = 388.12 avg = 384.78\n vgg16 min = 2321.29 max = 2345.89 avg = 2330.29\n vgg16_int8 min = 2562.86 max = 2568.06 avg = 2565.68\n resnet50 min = 1219.09 max = 1225.67 avg = 1221.36\n resnet50_int8 min = 1263.44 max = 1266.74 avg = 1265.09\n squeezenet_ssd min = 433.23 max = 441.06 avg = 437.07\n squeezenet_ssd_int8 min = 438.69 max = 443.17 avg = 440.81\n mobilenet_ssd min = 587.37 max = 598.57 avg = 589.99\n mobilenet_ssd_int8 min = 539.62 max = 552.57 avg = 542.87\n mobilenet_yolo min = 1485.30 max = 1491.17 avg = 1487.81\n mobilenetv2_yolov3 min = 711.57 max = 722.91 avg = 715.07\n yolov4-tiny min = 954.76 max = 961.66 avg = 957.28\n nanodet_m min = 364.22 max = 369.32 avg = 365.94\n yolo-fastest-1.1 min = 154.81 max = 160.45 avg = 156.23\n yolo-fastestv2 min = 157.39 max = 168.82 avg = 159.51\n vision_transformer min = 18926.46 max = 18980.43 avg = 18951.29\n FastestDet min = 168.81 max = 176.77 avg = 170.26\n\nroot@ls2kla:~/ncnn/build/benchmark# ./benchncnn 4 1 0 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 272.76 max = 280.89 avg = 275.29\n squeezenet_int8 min = 352.02 max = 353.25 avg = 352.40\n mobilenet min = 519.09 max = 519.68 avg = 519.34\n mobilenet_int8 min = 509.85 max = 510.23 avg = 510.04\n mobilenet_v2 min = 352.06 max = 352.74 avg = 352.37\n mobilenet_v3 min = 295.13 max = 295.70 avg = 295.39\n shufflenet min = 241.58 max = 241.94 avg = 241.73\n shufflenet_v2 min = 282.88 max = 283.39 avg = 283.18\n mnasnet min = 357.74 max = 358.21 avg = 357.98\n proxylessnasnet min = 403.26 max = 411.69 avg = 406.02\n efficientnet_b0 min = 546.11 max = 546.88 avg = 546.53\n efficientnetv2_b0 min = 596.83 max = 597.05 avg = 596.93\n regnety_400m min = 441.94 max = 442.02 avg = 441.98\n blazeface min = 54.08 max = 54.59 avg = 54.38\n googlenet min = 1042.19 max = 1048.03 avg = 1044.40\n googlenet_int8 min = 1118.22 max = 1121.18 avg = 1119.79\n resnet18 min = 838.79 max = 839.81 avg = 839.43\n resnet18_int8 min = 939.62 max = 940.72 avg = 940.23\n alexnet min = 729.36 max = 740.65 avg = 734.19\n vgg16 min = 4326.68 max = 4335.10 avg = 4330.97\n vgg16_int8 min = 4896.71 max = 4909.63 avg = 4905.14\n resnet50 min = 2277.36 max = 2280.34 avg = 2279.14\n resnet50_int8 min = 2399.07 max = 2402.21 avg = 2400.78\n squeezenet_ssd min = 751.49 max = 753.79 avg = 752.20\n squeezenet_ssd_int8 min = 771.01 max = 774.08 avg = 771.91\n mobilenet_ssd min = 1063.41 max = 1065.65 avg = 1064.16\n mobilenet_ssd_int8 min = 1031.59 max = 1033.03 avg = 1032.09\n mobilenet_yolo min = 2585.33 max = 2586.65 avg = 2586.11\n mobilenetv2_yolov3 min = 1246.35 max = 1248.43 avg = 1247.32\n yolov4-tiny min = 1639.13 max = 1642.47 avg = 1640.87\n nanodet_m min = 606.40 max = 607.14 avg = 606.86\n yolo-fastest-1.1 min = 242.15 max = 244.64 avg = 243.43\n yolo-fastestv2 min = 246.92 max = 247.84 avg = 247.27\n vision_transformer min = 36607.51 max = 36870.44 avg = 36724.88\n FastestDet min = 266.96 max = 268.86 avg = 267.94\n```\n\n### Loongson 2K2000 (LA364 1.5GHz * 2 with lsx)\n```\nloongson@loongson-pc:~/ncnn/build/benchmark$ ./benchncnn 4 2 0 -1 1\nloop_count = 4\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 58.54 max = 61.57 avg = 60.37\n squeezenet_int8 min = 66.79 max = 72.05 avg = 70.49\n mobilenet min = 110.46 max = 112.72 avg = 111.84\n mobilenet_int8 min = 117.83 max = 126.51 avg = 123.42\n mobilenet_v2 min = 65.19 max = 70.78 avg = 67.73\n mobilenet_v3 min = 51.30 max = 56.61 avg = 54.52\n shufflenet min = 32.78 max = 35.11 avg = 33.99\n shufflenet_v2 min = 31.58 max = 32.59 avg = 32.15\n mnasnet min = 64.18 max = 78.53 avg = 68.72\n proxylessnasnet min = 73.49 max = 85.30 avg = 77.35\n efficientnet_b0 min = 101.83 max = 106.26 avg = 104.91\n efficientnetv2_b0 min = 126.55 max = 131.95 avg = 127.91\n regnety_400m min = 88.19 max = 92.58 avg = 89.60\n blazeface min = 8.57 max = 8.68 avg = 8.63\n googlenet min = 207.97 max = 214.47 avg = 211.07\n googlenet_int8 min = 237.92 max = 241.06 avg = 239.76\n resnet18 min = 153.42 max = 161.54 avg = 158.21\n resnet18_int8 min = 177.77 max = 183.83 avg = 181.90\n alexnet min = 145.71 max = 149.41 avg = 147.97\n vgg16 min = 937.03 max = 961.65 avg = 945.20\n vgg16_int8 min = 850.20 max = 869.47 avg = 859.99\n resnet50 min = 497.95 max = 524.29 avg = 511.85\n resnet50_int8 min = 541.22 max = 549.09 avg = 544.30\n squeezenet_ssd min = 155.11 max = 163.01 avg = 159.72\n squeezenet_ssd_int8 min = 136.11 max = 138.38 avg = 137.36\n mobilenet_ssd min = 226.97 max = 231.33 avg = 229.20\n mobilenet_ssd_int8 min = 248.61 max = 253.10 avg = 250.83\n mobilenet_yolo min = 613.25 max = 626.75 avg = 619.83\n mobilenetv2_yolov3 min = 249.50 max = 258.17 avg = 255.75\n yolov4-tiny min = 312.41 max = 349.24 avg = 328.38\n nanodet_m min = 81.50 max = 84.20 avg = 83.14\n yolo-fastest-1.1 min = 30.46 max = 30.91 avg = 30.71\n yolo-fastestv2 min = 26.78 max = 28.80 avg = 28.10\n vision_transformer min = 4483.37 max = 4519.06 avg = 4507.04\n FastestDet min = 31.15 max = 32.37 avg = 32.06\n```\n\n### Loongson 3A3000 (GS464E 1.45GHz * 4)\n```\nroot@3A3K:~/Desktop/ncnn-20221128/build/benchmark$ ./benchncnn 5 4 2 -1 0\nloop_count = 5\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 88.82 max = 116.74 avg = 94.92\n squeezenet_int8 min = 140.62 max = 162.48 avg = 146.32\n mobilenet min = 144.80 max = 244.58 avg = 172.14\n mobilenet_int8 min = 265.21 max = 293.89 avg = 281.80\n mobilenet_v2 min = 109.80 max = 156.74 avg = 120.48\n mobilenet_v3 min = 90.18 max = 93.25 avg = 91.50\n shufflenet min = 56.64 max = 216.12 avg = 100.68\n shufflenet_v2 min = 45.70 max = 142.00 avg = 65.20\n mnasnet min = 106.99 max = 229.11 avg = 134.22\n proxylessnasnet min = 123.68 max = 261.01 avg = 155.97\n efficientnet_b0 min = 160.98 max = 191.14 avg = 171.55\n efficientnetv2_b0 min = 162.75 max = 187.67 avg = 176.19\n regnety_400m min = 135.06 max = 174.12 avg = 151.30\n blazeface min = 15.26 max = 43.81 avg = 23.91\n googlenet min = 327.16 max = 386.02 avg = 350.25\n googlenet_int8 min = 500.45 max = 637.39 avg = 540.62\n resnet18 min = 254.45 max = 421.56 avg = 304.48\n resnet18_int8 min = 385.14 max = 559.01 avg = 439.74\n alexnet min = 179.19 max = 220.91 avg = 190.63\n vgg16 min = 1563.99 max = 1645.01 avg = 1619.63\n vgg16_int8 min = 1436.00 max = 1530.45 avg = 1473.00\n resnet50 min = 702.35 max = 833.23 avg = 764.14\n resnet50_int8 min = 1099.40 max = 1208.84 avg = 1154.51\n squeezenet_ssd min = 191.40 max = 270.10 avg = 218.75\n squeezenet_ssd_int8 min = 304.51 max = 387.51 avg = 344.98\n mobilenet_ssd min = 315.77 max = 417.37 avg = 344.40\n mobilenet_ssd_int8 min = 554.28 max = 656.07 avg = 580.72\n mobilenet_yolo min = 806.48 max = 851.22 avg = 825.50\n mobilenetv2_yolov3 min = 382.38 max = 503.38 avg = 421.03\n yolov4-tiny min = 502.87 max = 620.30 avg = 550.08\n nanodet_m min = 126.00 max = 314.03 avg = 184.93\n yolo-fastest-1.1 min = 64.68 max = 189.47 avg = 110.89\n yolo-fastestv2 min = 69.03 max = 116.31 avg = 82.36\n vision_transformer min = 14737.56 max = 15012.35 avg = 14890.56\n FastestDet min = 84.30 max = 139.87 avg = 102.23\n```\n\n### Loongson 3A4000 (GS464V 1.8GHz * 4 with MSA128)\n```\nroot@3A4K:~/Desktop/ncnn-20221128/build/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 17.04 max = 39.86 avg = 20.39\n squeezenet_int8 min = 21.77 max = 25.93 avg = 23.02\n mobilenet min = 26.34 max = 97.11 avg = 38.24\n mobilenet_int8 min = 32.93 max = 33.31 avg = 33.07\n mobilenet_v2 min = 19.40 max = 19.91 avg = 19.63\n mobilenet_v3 min = 16.48 max = 45.31 avg = 19.68\n shufflenet min = 12.23 max = 116.79 avg = 22.86\n shufflenet_v2 min = 11.14 max = 11.59 avg = 11.37\n mnasnet min = 18.33 max = 51.66 avg = 24.52\n proxylessnasnet min = 22.03 max = 22.46 avg = 22.19\n efficientnet_b0 min = 34.94 max = 129.52 avg = 45.76\n efficientnetv2_b0 min = 38.58 max = 67.86 avg = 41.84\n regnety_400m min = 35.53 max = 38.59 avg = 36.14\n blazeface min = 4.08 max = 4.34 avg = 4.17\n googlenet min = 72.60 max = 100.31 avg = 76.25\n googlenet_int8 min = 82.09 max = 107.09 avg = 86.78\n resnet18 min = 53.99 max = 100.21 avg = 63.52\n resnet18_int8 min = 57.20 max = 77.00 avg = 60.47\n alexnet min = 61.95 max = 80.86 avg = 65.01\n vgg16 min = 329.58 max = 438.99 avg = 360.40\n vgg16_int8 min = 293.27 max = 366.16 avg = 311.23\n resnet50 min = 138.06 max = 260.50 avg = 169.27\n resnet50_int8 min = 154.06 max = 244.31 avg = 173.37\n squeezenet_ssd min = 60.44 max = 97.92 avg = 65.41\n squeezenet_ssd_int8 min = 55.34 max = 136.72 avg = 68.15\n mobilenet_ssd min = 57.97 max = 139.16 avg = 69.27\n mobilenet_ssd_int8 min = 66.66 max = 89.91 avg = 71.00\n mobilenet_yolo min = 169.38 max = 711.10 avg = 242.62\n mobilenetv2_yolov3 min = 75.61 max = 97.83 avg = 80.23\n yolov4-tiny min = 110.52 max = 143.67 avg = 118.53\n nanodet_m min = 24.04 max = 92.81 avg = 32.45\n yolo-fastest-1.1 min = 10.97 max = 32.77 avg = 15.05\n yolo-fastestv2 min = 11.54 max = 12.09 avg = 11.84\n vision_transformer min = 4193.41 max = 4274.03 avg = 4213.64\n FastestDet min = 12.54 max = 13.01 avg = 12.78\n```\n\n\n\n### Loongson 3A4000 (GS464V 1.8GHz * 4 with MSA128)\n\nTest on UOS V20 E1050\n\n```\nuos@uos-PC:~/ncnn/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 25.28 max = 38.19 avg = 27.81\n squeezenet_int8 min = 21.61 max = 22.13 avg = 21.85\n mobilenet min = 44.77 max = 69.54 avg = 55.37\n mobilenet_int8 min = 32.96 max = 44.00 avg = 36.08\n mobilenet_v2 min = 29.21 max = 52.70 avg = 35.47\n mobilenet_v3 min = 24.62 max = 27.32 avg = 25.18\n shufflenet min = 18.90 max = 49.70 avg = 22.95\n shufflenet_v2 min = 15.87 max = 22.38 avg = 17.67\n mnasnet min = 29.08 max = 69.37 avg = 35.53\n proxylessnasnet min = 33.30 max = 94.15 avg = 42.81\n efficientnet_b0 min = 49.34 max = 61.22 avg = 52.01\n efficientnetv2_b0 min = 57.89 max = 72.55 avg = 60.72\n regnety_400m min = 50.65 max = 74.16 avg = 57.56\n blazeface min = 4.97 max = 5.33 avg = 5.11\n googlenet min = 101.45 max = 119.73 avg = 106.85\n googlenet_int8 min = 83.94 max = 99.75 avg = 87.36\n resnet18 min = 81.65 max = 99.76 avg = 85.96\n resnet18_int8 min = 58.60 max = 75.88 avg = 60.62\n alexnet min = 77.05 max = 208.05 avg = 120.39\n vgg16 min = 427.51 max = 676.57 avg = 531.53\n vgg16_int8 min = 326.59 max = 487.96 avg = 417.74\n resnet50 min = 221.51 max = 580.11 avg = 305.64\n resnet50_int8 min = 158.00 max = 190.71 avg = 167.50\n squeezenet_ssd min = 98.87 max = 135.55 avg = 115.54\n squeezenet_ssd_int8 min = 66.33 max = 361.40 avg = 148.19\n mobilenet_ssd min = 94.12 max = 340.16 avg = 184.85\n mobilenet_ssd_int8 min = 88.26 max = 150.47 avg = 112.35\n mobilenet_yolo min = 252.07 max = 510.61 avg = 327.21\n mobilenetv2_yolov3 min = 115.31 max = 183.63 avg = 147.28\n yolov4-tiny min = 153.92 max = 259.18 avg = 196.70\n nanodet_m min = 34.95 max = 66.15 avg = 46.41\n yolo-fastest-1.1 min = 15.34 max = 15.94 avg = 15.62\n yolo-fastestv2 min = 15.53 max = 16.06 avg = 15.80\n vision_transformer min = 4200.48 max = 5853.43 avg = 4555.42\n FastestDet min = 16.73 max = 18.72 avg = 17.08\n\n\nuos@uos-PC:~/ncnn/benchmark$ ./benchncnn 10 4 1 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 25.93 max = 47.61 avg = 28.45\n squeezenet_int8 min = 21.84 max = 27.09 avg = 22.84\n mobilenet min = 44.61 max = 83.44 avg = 52.52\n mobilenet_int8 min = 32.91 max = 45.99 avg = 34.52\n mobilenet_v2 min = 29.44 max = 37.14 avg = 30.43\n mobilenet_v3 min = 24.54 max = 42.68 avg = 27.25\n shufflenet min = 17.16 max = 42.10 avg = 20.08\n shufflenet_v2 min = 15.99 max = 16.43 avg = 16.29\n mnasnet min = 29.14 max = 43.37 avg = 30.79\n proxylessnasnet min = 33.15 max = 34.12 avg = 33.52\n efficientnet_b0 min = 49.35 max = 87.75 avg = 54.03\n efficientnetv2_b0 min = 57.69 max = 84.67 avg = 64.12\n regnety_400m min = 50.55 max = 75.35 avg = 55.31\n blazeface min = 5.01 max = 5.16 avg = 5.05\n googlenet min = 101.51 max = 116.33 avg = 105.38\n googlenet_int8 min = 84.34 max = 102.58 avg = 89.89\n resnet18 min = 80.58 max = 94.47 avg = 86.27\n resnet18_int8 min = 59.00 max = 76.66 avg = 62.15\n alexnet min = 91.72 max = 117.98 avg = 102.20\n vgg16 min = 435.57 max = 453.90 avg = 441.39\n vgg16_int8 min = 308.39 max = 332.69 avg = 321.09\n resnet50 min = 219.93 max = 249.30 avg = 231.93\n resnet50_int8 min = 156.78 max = 179.34 avg = 163.43\n squeezenet_ssd min = 109.48 max = 153.84 avg = 123.75\n squeezenet_ssd_int8 min = 74.33 max = 117.03 avg = 93.81\n mobilenet_ssd min = 94.91 max = 161.38 avg = 127.78\n mobilenet_ssd_int8 min = 82.35 max = 112.79 avg = 91.86\n mobilenet_yolo min = 252.05 max = 285.16 avg = 266.33\n mobilenetv2_yolov3 min = 113.98 max = 173.83 avg = 139.60\n yolov4-tiny min = 150.06 max = 210.96 avg = 164.94\n nanodet_m min = 34.62 max = 67.81 avg = 48.43\n yolo-fastest-1.1 min = 15.78 max = 16.09 avg = 15.93\n yolo-fastestv2 min = 15.54 max = 32.82 avg = 17.62\n vision_transformer min = 4202.89 max = 5573.15 avg = 4426.38\n FastestDet min = 16.39 max = 17.06 avg = 16.75\n\n\n\n\nuos@uos-PC:~/ncnn/benchmark$ ./benchncnn 10 4 0 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 25.98 max = 36.75 avg = 28.86\n squeezenet_int8 min = 22.04 max = 30.86 avg = 23.28\n mobilenet min = 44.82 max = 60.73 avg = 46.72\n mobilenet_int8 min = 33.00 max = 48.45 avg = 34.70\n mobilenet_v2 min = 29.53 max = 56.78 avg = 33.98\n mobilenet_v3 min = 24.69 max = 45.60 avg = 28.13\n shufflenet min = 17.25 max = 24.72 avg = 18.18\n shufflenet_v2 min = 16.00 max = 31.27 avg = 17.62\n mnasnet min = 28.95 max = 44.73 avg = 32.58\n proxylessnasnet min = 32.99 max = 45.42 avg = 34.66\n efficientnet_b0 min = 49.71 max = 53.47 avg = 50.25\n efficientnetv2_b0 min = 57.51 max = 78.56 avg = 61.47\n regnety_400m min = 50.18 max = 71.85 avg = 54.77\n blazeface min = 4.98 max = 9.36 avg = 5.48\n googlenet min = 101.25 max = 121.71 avg = 105.71\n googlenet_int8 min = 82.97 max = 111.81 avg = 89.49\n resnet18 min = 75.66 max = 87.19 avg = 78.72\n resnet18_int8 min = 58.92 max = 108.67 avg = 76.70\n alexnet min = 79.12 max = 144.22 avg = 101.91\n vgg16 min = 430.14 max = 460.46 avg = 444.56\n vgg16_int8 min = 308.08 max = 350.15 avg = 324.86\n resnet50 min = 219.60 max = 258.59 avg = 237.46\n resnet50_int8 min = 156.54 max = 180.28 avg = 163.11\n squeezenet_ssd min = 77.71 max = 137.36 avg = 119.68\n squeezenet_ssd_int8 min = 78.88 max = 113.64 avg = 95.83\n mobilenet_ssd min = 94.82 max = 156.99 avg = 119.67\n mobilenet_ssd_int8 min = 77.17 max = 98.29 avg = 86.90\n mobilenet_yolo min = 252.29 max = 295.62 avg = 265.58\n mobilenetv2_yolov3 min = 114.28 max = 159.82 avg = 140.03\n yolov4-tiny min = 150.99 max = 203.07 avg = 165.18\n nanodet_m min = 34.48 max = 71.56 avg = 49.84\n yolo-fastest-1.1 min = 15.36 max = 30.00 avg = 17.11\n yolo-fastestv2 min = 15.42 max = 26.96 avg = 16.78\n vision_transformer min = 4187.60 max = 4319.84 avg = 4220.05\n FastestDet min = 16.30 max = 24.88 avg = 17.49\n\n```\n\n\n\n\n### Loongson 3A5000 (LA464 2.5GHz * 4)\n```\nroot@3A5K:~/Desktop/ncnn-20230223/build/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 11.97 max = 19.38 avg = 13.61\n squeezenet_int8 min = 14.96 max = 15.36 avg = 15.12\n mobilenet min = 20.14 max = 27.50 avg = 21.12\n mobilenet_int8 min = 25.28 max = 35.06 avg = 27.37\n mobilenet_v2 min = 12.82 max = 13.20 avg = 12.98\n mobilenet_v3 min = 11.39 max = 25.03 avg = 12.86\n shufflenet min = 7.35 max = 7.50 avg = 7.40\n shufflenet_v2 min = 7.12 max = 7.23 avg = 7.18\n mnasnet min = 12.85 max = 21.69 avg = 13.83\n proxylessnasnet min = 15.35 max = 15.79 avg = 15.43\n efficientnet_b0 min = 24.20 max = 24.46 avg = 24.30\n efficientnetv2_b0 min = 26.80 max = 42.43 avg = 29.25\n regnety_400m min = 22.85 max = 38.30 avg = 24.51\n blazeface min = 2.57 max = 2.67 avg = 2.60\n googlenet min = 49.09 max = 85.91 avg = 67.57\n googlenet_int8 min = 64.89 max = 95.28 avg = 76.41\n resnet18 min = 42.43 max = 62.39 avg = 52.38\n resnet18_int8 min = 47.96 max = 68.69 avg = 56.75\n alexnet min = 46.01 max = 59.26 avg = 49.20\n vgg16 min = 246.82 max = 261.80 avg = 252.81\n vgg16_int8 min = 247.13 max = 256.81 avg = 252.37\n resnet50 min = 102.17 max = 138.16 avg = 117.65\n resnet50_int8 min = 115.09 max = 151.30 avg = 129.13\n squeezenet_ssd min = 43.62 max = 70.64 avg = 53.89\n squeezenet_ssd_int8 min = 38.66 max = 60.12 avg = 47.66\n mobilenet_ssd min = 42.67 max = 68.78 avg = 53.95\n mobilenet_ssd_int8 min = 56.29 max = 68.31 avg = 59.86\n mobilenet_yolo min = 129.04 max = 188.26 avg = 149.64\n mobilenetv2_yolov3 min = 61.80 max = 71.41 avg = 66.43\n yolov4-tiny min = 88.64 max = 108.17 avg = 95.48\n nanodet_m min = 16.24 max = 16.57 avg = 16.34\n yolo-fastest-1.1 min = 6.98 max = 7.16 avg = 7.05\n yolo-fastestv2 min = 6.95 max = 7.29 avg = 7.08\n vision_transformer min = 2910.63 max = 3109.29 avg = 2949.04\n FastestDet min = 7.66 max = 7.90 avg = 7.80\n```\n\n### Loongson 3A6000 (LA664 2.5GHz * 4+4)\n\n```\n~/ncnn/build/benchmark$ ./benchncnn 10 8 2 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 7.12 max = 7.20 avg = 7.16\n squeezenet_int8 min = 8.93 max = 9.20 avg = 8.98\n mobilenet min = 11.81 max = 11.88 avg = 11.84\n mobilenet_int8 min = 14.25 max = 14.33 avg = 14.28\n mobilenet_v2 min = 8.06 max = 8.16 avg = 8.08\n mobilenet_v3 min = 6.84 max = 6.90 avg = 6.87\n shufflenet min = 5.38 max = 5.44 avg = 5.39\n shufflenet_v2 min = 5.20 max = 5.22 avg = 5.20\n mnasnet min = 8.06 max = 8.10 avg = 8.07\n proxylessnasnet min = 8.94 max = 9.09 avg = 8.99\n efficientnet_b0 min = 13.43 max = 13.65 avg = 13.48\n efficientnetv2_b0 min = 16.06 max = 16.18 avg = 16.11\n regnety_400m min = 18.11 max = 18.18 avg = 18.14\n blazeface min = 1.59 max = 1.61 avg = 1.60\n googlenet min = 26.08 max = 26.24 avg = 26.17\n googlenet_int8 min = 31.25 max = 31.42 avg = 31.34\n resnet18 min = 19.65 max = 19.73 avg = 19.69\n resnet18_int8 min = 25.55 max = 25.66 avg = 25.60\n alexnet min = 19.56 max = 19.81 avg = 19.67\n vgg16 min = 115.32 max = 116.38 avg = 115.99\n vgg16_int8 min = 135.94 max = 136.73 avg = 136.34\n resnet50 min = 56.46 max = 56.96 avg = 56.81\n resnet50_int8 min = 66.13 max = 66.40 avg = 66.27\n squeezenet_ssd min = 22.84 max = 22.99 avg = 22.89\n squeezenet_ssd_int8 min = 22.34 max = 22.76 avg = 22.54\n mobilenet_ssd min = 24.67 max = 24.75 avg = 24.71\n mobilenet_ssd_int8 min = 29.32 max = 29.37 avg = 29.34\n mobilenet_yolo min = 82.82 max = 84.02 avg = 83.40\n mobilenetv2_yolov3 min = 30.31 max = 30.45 avg = 30.38\n yolov4-tiny min = 42.49 max = 42.74 avg = 42.62\n nanodet_m min = 11.00 max = 11.08 avg = 11.02\n yolo-fastest-1.1 min = 5.28 max = 5.40 avg = 5.31\n yolo-fastestv2 min = 5.09 max = 5.10 avg = 5.10\n vision_transformer min = 869.40 max = 898.18 avg = 874.07\n FastestDet min = 5.28 max = 5.37 avg = 5.31\n```\n\n### Phytium FT-2000/4 (FTC663 armv8 2.2GHz x 4)\nTest on Kylin OS V10\n```\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 40.92 max = 43.43 avg = 41.34\n squeezenet_int8 min = 35.48 max = 36.07 avg = 35.75\n mobilenet min = 72.23 max = 72.53 avg = 72.39\n mobilenet_int8 min = 48.10 max = 48.59 avg = 48.31\n mobilenet_v2 min = 47.94 max = 48.45 avg = 48.13\n mobilenet_v3 min = 37.95 max = 39.59 avg = 38.41\n shufflenet min = 21.51 max = 21.84 avg = 21.64\n shufflenet_v2 min = 21.10 max = 21.45 avg = 21.26\n mnasnet min = 44.53 max = 45.15 avg = 44.74\n proxylessnasnet min = 53.02 max = 53.62 avg = 53.21\n efficientnet_b0 min = 79.81 max = 80.51 avg = 80.15\n efficientnetv2_b0 min = 92.55 max = 103.10 avg = 97.53\n regnety_400m min = 58.52 max = 70.04 avg = 64.20\n blazeface min = 6.06 max = 9.85 avg = 6.88\n googlenet min = 146.49 max = 162.69 avg = 152.98\n googlenet_int8 min = 127.38 max = 132.11 avg = 128.51\n resnet18 min = 107.79 max = 108.83 avg = 108.37\n resnet18_int8 min = 97.28 max = 99.03 avg = 97.73\n alexnet min = 89.95 max = 91.63 avg = 90.28\n vgg16 min = 642.27 max = 647.16 avg = 644.09\n vgg16_int8 min = 567.03 max = 574.11 avg = 568.74\n resnet50 min = 329.12 max = 331.79 avg = 330.10\n resnet50_int8 min = 252.48 max = 253.65 avg = 252.93\n squeezenet_ssd min = 96.46 max = 96.95 avg = 96.69\n squeezenet_ssd_int8 min = 92.35 max = 93.24 avg = 92.72\n mobilenet_ssd min = 149.14 max = 150.56 avg = 149.40\n mobilenet_ssd_int8 min = 97.56 max = 98.03 avg = 97.82\n mobilenet_yolo min = 339.71 max = 340.60 avg = 339.89\n mobilenetv2_yolov3 min = 174.53 max = 175.80 avg = 175.01\n yolov4-tiny min = 213.72 max = 214.94 avg = 214.08\n nanodet_m min = 49.95 max = 50.47 avg = 50.19\n yolo-fastest-1.1 min = 23.80 max = 24.42 avg = 23.91\n yolo-fastestv2 min = 19.78 max = 19.95 avg = 19.84\n vision_transformer min = 3927.51 max = 4025.76 avg = 3947.06\n FastestDet min = 21.78 max = 22.17 avg = 21.88\n\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 4 1 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 70.80 max = 76.55 avg = 72.49\n squeezenet_int8 min = 110.36 max = 133.06 avg = 114.23\n mobilenet min = 77.97 max = 85.73 avg = 79.98\n mobilenet_int8 min = 80.05 max = 84.09 avg = 81.76\n mobilenet_v2 min = 101.07 max = 192.92 avg = 139.32\n mobilenet_v3 min = 108.60 max = 129.37 avg = 113.80\n shufflenet min = 160.96 max = 188.96 avg = 168.62\n shufflenet_v2 min = 96.20 max = 190.31 avg = 119.77\n mnasnet min = 97.34 max = 104.00 avg = 99.85\n proxylessnasnet min = 112.58 max = 276.49 avg = 145.74\n efficientnet_b0 min = 171.01 max = 238.15 avg = 195.53\n efficientnetv2_b0 min = 235.31 max = 299.00 avg = 254.12\n regnety_400m min = 1059.87 max = 1173.49 avg = 1084.13\n blazeface min = 58.69 max = 64.83 avg = 60.83\n googlenet min = 190.47 max = 257.76 avg = 207.71\n googlenet_int8 min = 285.67 max = 327.20 avg = 300.87\n resnet18 min = 111.87 max = 118.36 avg = 114.48\n resnet18_int8 min = 143.08 max = 147.98 avg = 144.93\n alexnet min = 72.83 max = 76.52 avg = 74.01\n vgg16 min = 390.35 max = 406.58 avg = 397.19\n vgg16_int8 min = 358.54 max = 369.89 avg = 364.31\n resnet50 min = 275.57 max = 300.14 avg = 283.21\n resnet50_int8 min = 315.18 max = 371.22 avg = 328.43\n squeezenet_ssd min = 170.14 max = 200.18 avg = 175.23\n squeezenet_ssd_int8 min = 259.01 max = 271.23 avg = 263.35\n mobilenet_ssd min = 166.85 max = 170.64 avg = 168.74\n mobilenet_ssd_int8 min = 191.71 max = 195.91 avg = 193.44\n mobilenet_yolo min = 960.70 max = 1080.81 avg = 983.68\n mobilenetv2_yolov3 min = 187.72 max = 207.92 avg = 192.60\n yolov4-tiny min = 172.72 max = 177.62 avg = 174.63\n nanodet_m min = 128.79 max = 137.31 avg = 131.04\n yolo-fastest-1.1 min = 132.39 max = 148.06 avg = 137.90\n yolo-fastestv2 min = 130.97 max = 137.73 avg = 133.53\n vision_transformer min = 2229.10 max = 2392.59 avg = 2304.21\n FastestDet min = 119.98 max = 126.26 avg = 122.40\n\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 70.93 max = 75.55 avg = 72.93\n squeezenet_int8 min = 109.65 max = 153.48 avg = 124.20\n mobilenet min = 78.02 max = 85.80 avg = 81.97\n mobilenet_int8 min = 80.34 max = 89.31 avg = 83.20\n mobilenet_v2 min = 99.51 max = 110.36 avg = 102.54\n mobilenet_v3 min = 109.04 max = 116.28 avg = 111.75\n shufflenet min = 160.04 max = 166.21 avg = 163.59\n shufflenet_v2 min = 88.90 max = 91.82 avg = 90.24\n mnasnet min = 97.02 max = 103.09 avg = 98.70\n proxylessnasnet min = 111.21 max = 117.47 avg = 113.97\n efficientnet_b0 min = 167.99 max = 175.35 avg = 171.26\n efficientnetv2_b0 min = 228.59 max = 245.97 avg = 232.79\n regnety_400m min = 1049.34 max = 1085.18 avg = 1064.68\n blazeface min = 59.35 max = 64.91 avg = 60.35\n googlenet min = 187.87 max = 195.29 avg = 190.56\n googlenet_int8 min = 283.22 max = 301.69 avg = 287.66\n resnet18 min = 111.48 max = 116.76 avg = 112.88\n resnet18_int8 min = 142.41 max = 148.79 avg = 145.14\n alexnet min = 72.59 max = 75.37 avg = 73.62\n vgg16 min = 389.61 max = 452.95 avg = 424.36\n vgg16_int8 min = 365.57 max = 465.13 avg = 422.84\n resnet50 min = 283.07 max = 411.14 avg = 332.88\n resnet50_int8 min = 323.21 max = 381.13 avg = 340.59\n squeezenet_ssd min = 178.21 max = 252.82 avg = 211.62\n squeezenet_ssd_int8 min = 263.82 max = 372.38 avg = 284.38\n mobilenet_ssd min = 166.29 max = 281.36 avg = 195.16\n mobilenet_ssd_int8 min = 194.00 max = 220.95 avg = 204.07\n mobilenet_yolo min = 964.99 max = 1027.13 avg = 989.45\n mobilenetv2_yolov3 min = 218.58 max = 512.86 avg = 265.12\n yolov4-tiny min = 172.20 max = 177.27 avg = 174.14\n nanodet_m min = 128.78 max = 222.66 avg = 150.88\n yolo-fastest-1.1 min = 132.52 max = 196.41 avg = 149.03\n yolo-fastestv2 min = 131.39 max = 138.72 avg = 134.96\n vision_transformer min = 2243.31 max = 2659.56 avg = 2395.76\n FastestDet min = 119.44 max = 126.07 avg = 122.27\n\n```\n\n\n\n### Phytium FT-2000+/64 (FTC662 armv8 2.4GHz x 8)\n```\n[root@bogon benchmark]# ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 57.60 max = 59.78 avg = 58.51\n squeezenet_int8 min = 47.05 max = 47.89 avg = 47.40\n mobilenet min = 91.08 max = 95.16 avg = 91.89\n mobilenet_int8 min = 60.27 max = 61.17 avg = 60.74\n mobilenet_v2 min = 63.38 max = 68.12 avg = 66.96\n mobilenet_v3 min = 53.34 max = 54.71 avg = 54.01\n shufflenet min = 37.87 max = 41.78 avg = 39.37\n shufflenet_v2 min = 35.89 max = 37.30 avg = 36.40\n mnasnet min = 59.57 max = 63.23 avg = 60.25\n proxylessnasnet min = 71.24 max = 71.93 avg = 71.51\n efficientnet_b0 min = 134.34 max = 141.14 avg = 137.74\n efficientnetv2_b0 min = 143.82 max = 145.63 avg = 144.36\n regnety_400m min = 76.96 max = 77.66 avg = 77.27\n blazeface min = 11.57 max = 11.90 avg = 11.70\n googlenet min = 188.10 max = 191.27 avg = 189.02\n googlenet_int8 min = 167.54 max = 169.63 avg = 168.38\n resnet18 min = 144.76 max = 163.39 avg = 154.95\n resnet18_int8 min = 124.14 max = 129.84 avg = 127.83\n alexnet min = 198.22 max = 208.86 avg = 205.35\n vgg16 min = 848.10 max = 891.00 avg = 859.94\n vgg16_int8 min = 686.54 max = 742.77 avg = 704.74\n resnet50 min = 413.45 max = 428.84 avg = 417.81\n resnet50_int8 min = 306.32 max = 324.27 avg = 316.47\n squeezenet_ssd min = 147.62 max = 149.58 avg = 148.48\n squeezenet_ssd_int8 min = 116.18 max = 134.86 avg = 126.93\n mobilenet_ssd min = 188.49 max = 191.97 avg = 189.48\n mobilenet_ssd_int8 min = 120.28 max = 121.36 avg = 120.83\n mobilenet_yolo min = 421.79 max = 425.68 avg = 423.51\n mobilenetv2_yolov3 min = 222.86 max = 225.58 avg = 224.01\n yolov4-tiny min = 303.77 max = 310.70 avg = 307.45\n nanodet_m min = 80.87 max = 82.11 avg = 81.35\n\n[root@bogon benchmark]# ./benchncnn 10 8 0 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 14.53 max = 14.92 avg = 14.68\n squeezenet_int8 min = 11.67 max = 11.89 avg = 11.82\n mobilenet min = 17.60 max = 20.05 avg = 18.34\n mobilenet_int8 min = 9.94 max = 10.22 avg = 10.08\n mobilenet_v2 min = 18.46 max = 19.18 avg = 18.81\n mobilenet_v3 min = 16.30 max = 16.71 avg = 16.45\n shufflenet min = 14.65 max = 14.93 avg = 14.78\n shufflenet_v2 min = 11.23 max = 11.56 avg = 11.35\n mnasnet min = 15.65 max = 16.08 avg = 15.92\n proxylessnasnet min = 18.78 max = 21.72 avg = 19.68\n efficientnet_b0 min = 29.16 max = 29.62 avg = 29.37\n efficientnetv2_b0 min = 33.28 max = 35.48 avg = 34.23\n regnety_400m min = 44.90 max = 47.36 avg = 46.32\n blazeface min = 4.23 max = 4.43 avg = 4.30\n googlenet min = 42.11 max = 42.98 avg = 42.38\n googlenet_int8 min = 33.24 max = 38.21 avg = 34.10\n resnet18 min = 33.27 max = 34.00 avg = 33.57\n resnet18_int8 min = 23.66 max = 24.78 avg = 24.24\n alexnet min = 35.78 max = 37.68 avg = 36.46\n vgg16 min = 219.60 max = 235.79 avg = 222.11\n vgg16_int8 min = 128.64 max = 135.19 avg = 130.73\n resnet50 min = 84.15 max = 85.48 avg = 84.66\n resnet50_int8 min = 58.87 max = 61.98 avg = 59.85\n squeezenet_ssd min = 47.60 max = 50.24 avg = 48.54\n squeezenet_ssd_int8 min = 36.42 max = 37.89 avg = 36.99\n mobilenet_ssd min = 39.37 max = 42.63 avg = 41.06\n mobilenet_ssd_int8 min = 21.59 max = 22.05 avg = 21.83\n mobilenet_yolo min = 83.16 max = 88.75 avg = 85.29\n mobilenetv2_yolov3 min = 58.13 max = 59.50 avg = 58.62\n yolov4-tiny min = 74.18 max = 76.56 avg = 75.13\n nanodet_m min = 25.16 max = 31.45 avg = 26.71\n\nroot@FT2K:~/Desktop/ncnn-20221128/build/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 14.19 max = 21.46 avg = 15.16\n squeezenet_int8 min = 11.63 max = 12.08 avg = 11.91\n mobilenet min = 20.52 max = 37.00 avg = 23.66\n mobilenet_int8 min = 13.38 max = 25.95 avg = 15.01\n mobilenet_v2 min = 15.80 max = 16.59 avg = 16.12\n mobilenet_v3 min = 13.38 max = 17.62 avg = 14.21\n shufflenet min = 10.62 max = 11.10 avg = 10.85\n shufflenet_v2 min = 9.09 max = 12.30 avg = 9.66\n mnasnet min = 14.85 max = 15.67 avg = 15.14\n proxylessnasnet min = 16.83 max = 17.10 avg = 16.98\n efficientnet_b0 min = 24.59 max = 26.40 avg = 25.06\n efficientnetv2_b0 min = 30.25 max = 34.46 avg = 31.42\n regnety_400m min = 32.37 max = 41.10 avg = 35.17\n blazeface min = 3.00 max = 3.56 avg = 3.18\n googlenet min = 49.52 max = 64.98 avg = 56.29\n googlenet_int8 min = 38.65 max = 52.51 avg = 43.90\n resnet18 min = 42.81 max = 53.94 avg = 45.38\n resnet18_int8 min = 32.53 max = 53.62 avg = 37.26\n alexnet min = 33.92 max = 47.88 avg = 37.12\n vgg16 min = 214.19 max = 228.96 avg = 220.16\n vgg16_int8 min = 164.22 max = 224.51 avg = 180.15\n resnet50 min = 106.90 max = 189.61 avg = 133.34\n resnet50_int8 min = 79.62 max = 94.41 avg = 83.56\n squeezenet_ssd min = 48.00 max = 49.11 avg = 48.43\n squeezenet_ssd_int8 min = 33.59 max = 47.60 avg = 37.57\n mobilenet_ssd min = 43.97 max = 58.84 avg = 49.64\n mobilenet_ssd_int8 min = 27.94 max = 32.89 avg = 29.56\n mobilenet_yolo min = 107.29 max = 118.80 avg = 114.24\n mobilenetv2_yolov3 min = 63.44 max = 106.75 avg = 70.69\n yolov4-tiny min = 89.93 max = 155.39 avg = 101.90\n nanodet_m min = 20.34 max = 28.67 avg = 21.44\n yolo-fastest-1.1 min = 11.74 max = 12.24 avg = 11.96\n yolo-fastestv2 min = 9.81 max = 9.98 avg = 9.91\n vision_transformer min = 1617.60 max = 1634.13 avg = 1625.87\n FastestDet min = 10.19 max = 10.55 avg = 10.36\n```\n### HUAWEI KunPeng 920 2251K (x8 cores)\ntest on UOS 1050\n```\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 12.11 max = 12.40 avg = 12.25\n squeezenet_int8 min = 14.24 max = 14.50 avg = 14.36\n mobilenet min = 20.52 max = 21.11 avg = 20.63\n mobilenet_int8 min = 18.29 max = 18.63 avg = 18.45\n mobilenet_v2 min = 13.73 max = 13.90 avg = 13.79\n mobilenet_v3 min = 11.37 max = 11.49 avg = 11.41\n shufflenet min = 7.90 max = 7.96 avg = 7.92\n shufflenet_v2 min = 8.09 max = 8.13 avg = 8.11\n mnasnet min = 13.26 max = 13.44 avg = 13.30\n proxylessnasnet min = 16.19 max = 16.39 avg = 16.26\n efficientnet_b0 min = 34.92 max = 35.22 avg = 35.04\n efficientnetv2_b0 min = 43.82 max = 44.39 avg = 43.94\n regnety_400m min = 17.55 max = 18.02 avg = 17.65\n blazeface min = 3.05 max = 3.08 avg = 3.07\n googlenet min = 58.65 max = 59.26 avg = 58.89\n googlenet_int8 min = 60.55 max = 63.00 avg = 61.96\n resnet18 min = 34.27 max = 35.43 avg = 34.84\n resnet18_int8 min = 60.79 max = 62.15 avg = 61.47\n alexnet min = 42.01 max = 44.43 avg = 43.36\n vgg16 min = 174.46 max = 177.33 avg = 175.57\n vgg16_int8 min = 453.93 max = 457.03 avg = 454.79\n resnet50 min = 95.36 max = 96.27 avg = 95.55\n resnet50_int8 min = 119.77 max = 121.26 avg = 120.46\n squeezenet_ssd min = 39.05 max = 39.69 avg = 39.20\n squeezenet_ssd_int8 min = 55.06 max = 56.23 avg = 55.72\n mobilenet_ssd min = 45.20 max = 45.96 avg = 45.49\n mobilenet_ssd_int8 min = 39.40 max = 40.13 avg = 39.76\n mobilenet_yolo min = 98.86 max = 99.85 avg = 99.34\n mobilenetv2_yolov3 min = 51.17 max = 52.89 avg = 51.89\n yolov4-tiny min = 66.43 max = 67.23 avg = 66.70\n nanodet_m min = 20.59 max = 20.79 avg = 20.71\n yolo-fastest-1.1 min = 7.90 max = 7.99 avg = 7.93\n yolo-fastestv2 min = 7.45 max = 7.49 avg = 7.47\n vision_transformer min = 1586.33 max = 1595.34 avg = 1589.76\n FastestDet min = 7.45 max = 7.52 avg = 7.47\n\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 8 0 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.93 max = 3.10 avg = 3.00\n squeezenet_int8 min = 3.47 max = 3.56 avg = 3.52\n mobilenet min = 3.89 max = 4.04 avg = 3.94\n mobilenet_int8 min = 3.29 max = 3.39 avg = 3.33\n mobilenet_v2 min = 3.95 max = 4.08 avg = 3.98\n mobilenet_v3 min = 3.45 max = 3.59 avg = 3.49\n shufflenet min = 3.42 max = 4.66 avg = 3.62\n shufflenet_v2 min = 2.60 max = 2.94 avg = 2.68\n mnasnet min = 3.46 max = 3.57 avg = 3.52\n proxylessnasnet min = 3.94 max = 12.34 avg = 4.88\n efficientnet_b0 min = 7.31 max = 7.60 avg = 7.38\n efficientnetv2_b0 min = 9.01 max = 9.22 avg = 9.08\n regnety_400m min = 8.56 max = 9.36 avg = 8.70\n blazeface min = 1.36 max = 3.52 avg = 1.60\n googlenet min = 11.80 max = 12.02 avg = 11.93\n googlenet_int8 min = 11.87 max = 23.09 avg = 13.16\n resnet18 min = 7.27 max = 7.64 avg = 7.38\n resnet18_int8 min = 11.02 max = 11.73 avg = 11.20\n alexnet min = 9.05 max = 9.35 avg = 9.17\n vgg16 min = 44.13 max = 50.84 avg = 46.89\n vgg16_int8 min = 75.15 max = 80.73 avg = 77.52\n resnet50 min = 18.72 max = 27.49 avg = 19.96\n resnet50_int8 min = 22.72 max = 36.80 avg = 26.78\n squeezenet_ssd min = 13.96 max = 27.42 avg = 15.62\n squeezenet_ssd_int8 min = 15.01 max = 29.53 avg = 19.51\n mobilenet_ssd min = 9.37 max = 13.34 avg = 10.44\n mobilenet_ssd_int8 min = 8.07 max = 24.28 avg = 9.83\n mobilenet_yolo min = 22.06 max = 24.89 avg = 22.91\n mobilenetv2_yolov3 min = 14.41 max = 15.97 avg = 14.78\n yolov4-tiny min = 20.71 max = 23.96 avg = 21.42\n nanodet_m min = 6.37 max = 6.59 avg = 6.45\n yolo-fastest-1.1 min = 4.27 max = 4.52 avg = 4.34\n yolo-fastestv2 min = 3.53 max = 3.63 avg = 3.58\n vision_transformer min = 435.60 max = 523.43 avg = 479.70\n FastestDet min = 3.54 max = 7.95 avg = 5.24\n\nmobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 4.04 max = 4.22 avg = 4.09\n squeezenet_int8 min = 4.64 max = 4.76 avg = 4.69\n mobilenet min = 6.04 max = 6.06 avg = 6.05\n mobilenet_int8 min = 5.23 max = 5.32 avg = 5.25\n mobilenet_v2 min = 5.00 max = 5.03 avg = 5.01\n mobilenet_v3 min = 4.49 max = 4.69 avg = 4.52\n shufflenet min = 3.90 max = 3.94 avg = 3.91\n shufflenet_v2 min = 3.27 max = 3.48 avg = 3.33\n mnasnet min = 4.80 max = 4.83 avg = 4.82\n proxylessnasnet min = 5.20 max = 5.28 avg = 5.23\n efficientnet_b0 min = 10.53 max = 11.06 avg = 10.68\n efficientnetv2_b0 min = 13.18 max = 13.37 avg = 13.25\n regnety_400m min = 9.20 max = 9.25 avg = 9.22\n blazeface min = 1.43 max = 1.45 avg = 1.44\n googlenet min = 17.63 max = 17.78 avg = 17.71\n googlenet_int8 min = 17.63 max = 18.03 avg = 17.85\n resnet18 min = 10.34 max = 10.59 avg = 10.40\n resnet18_int8 min = 17.93 max = 18.84 avg = 18.25\n alexnet min = 13.28 max = 13.37 avg = 13.31\n vgg16 min = 55.41 max = 56.60 avg = 55.70\n vgg16_int8 min = 123.71 max = 125.34 avg = 124.48\n resnet50 min = 27.82 max = 28.22 avg = 27.95\n resnet50_int8 min = 34.50 max = 34.89 avg = 34.70\n squeezenet_ssd min = 14.67 max = 15.19 avg = 14.85\n squeezenet_ssd_int8 min = 19.76 max = 20.32 avg = 19.87\n mobilenet_ssd min = 13.15 max = 13.38 avg = 13.21\n mobilenet_ssd_int8 min = 11.52 max = 11.70 avg = 11.60\n mobilenet_yolo min = 30.95 max = 31.28 avg = 31.05\n mobilenetv2_yolov3 min = 20.04 max = 20.36 avg = 20.16\n yolov4-tiny min = 25.61 max = 26.73 avg = 25.80\n nanodet_m min = 7.93 max = 7.97 avg = 7.95\n yolo-fastest-1.1 min = 4.52 max = 4.59 avg = 4.53\n yolo-fastestv2 min = 3.74 max = 3.88 avg = 3.77\n vision_transformer min = 546.94 max = 726.81 avg = 698.27\n FastestDet min = 3.59 max = 3.61 avg = 3.60\n```\n\n### HUAWEI KunPeng 920 3211K (x24 cores)\ntest on ubuntu 22.04\n```\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 12.11 max = 12.20 avg = 12.14\n squeezenet_int8 min = 14.34 max = 14.46 avg = 14.41\n mobilenet min = 20.27 max = 20.36 avg = 20.31\n mobilenet_int8 min = 17.45 max = 17.74 avg = 17.58\n mobilenet_v2 min = 13.72 max = 13.87 avg = 13.78\n mobilenet_v3 min = 11.51 max = 11.69 avg = 11.61\n shufflenet min = 8.07 max = 8.36 avg = 8.20\n shufflenet_v2 min = 8.13 max = 8.17 avg = 8.14\n mnasnet min = 13.34 max = 13.45 avg = 13.41\n proxylessnasnet min = 16.22 max = 16.35 avg = 16.29\n efficientnet_b0 min = 34.69 max = 35.14 avg = 34.82\n efficientnetv2_b0 min = 44.54 max = 44.68 avg = 44.61\n regnety_400m min = 18.06 max = 18.15 avg = 18.10\n blazeface min = 3.06 max = 3.22 avg = 3.12\n googlenet min = 56.80 max = 57.60 avg = 57.08\n googlenet_int8 min = 58.64 max = 59.98 avg = 59.42\n resnet18 min = 35.02 max = 35.35 avg = 35.10\n resnet18_int8 min = 61.13 max = 61.68 avg = 61.33\n alexnet min = 42.56 max = 43.05 avg = 42.69\n vgg16 min = 186.32 max = 188.73 avg = 187.20\n vgg16_int8 min = 459.01 max = 461.48 avg = 460.29\n resnet50 min = 97.59 max = 98.32 avg = 97.83\n resnet50_int8 min = 118.67 max = 120.45 avg = 119.78\n squeezenet_ssd min = 39.62 max = 39.95 avg = 39.81\n squeezenet_ssd_int8 min = 56.72 max = 57.63 avg = 57.00\n mobilenet_ssd min = 45.44 max = 45.82 avg = 45.63\n mobilenet_ssd_int8 min = 38.99 max = 40.08 avg = 39.39\n mobilenet_yolo min = 98.71 max = 99.27 avg = 98.94\n mobilenetv2_yolov3 min = 51.50 max = 52.41 avg = 51.87\n yolov4-tiny min = 68.02 max = 68.43 avg = 68.24\n nanodet_m min = 20.49 max = 20.64 avg = 20.59\n yolo-fastest-1.1 min = 8.17 max = 8.45 avg = 8.23\n yolo-fastestv2 min = 7.73 max = 8.06 avg = 7.87\n vision_transformer min = 1620.65 max = 1630.45 avg = 1625.64\n FastestDet min = 7.65 max = 7.77 avg = 7.69\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 2 0 -1 0\nloop_count = 10\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.77 max = 6.85 avg = 6.81\n squeezenet_int8 min = 7.98 max = 8.07 avg = 8.03\n mobilenet min = 10.70 max = 10.78 avg = 10.73\n mobilenet_int8 min = 9.21 max = 9.36 avg = 9.28\n mobilenet_v2 min = 7.91 max = 7.99 avg = 7.94\n mobilenet_v3 min = 6.72 max = 6.92 avg = 6.78\n shufflenet min = 5.34 max = 5.55 avg = 5.38\n shufflenet_v2 min = 5.12 max = 5.15 avg = 5.14\n mnasnet min = 7.74 max = 7.86 avg = 7.80\n proxylessnasnet min = 9.00 max = 9.03 avg = 9.02\n efficientnet_b0 min = 18.51 max = 18.58 avg = 18.54\n efficientnetv2_b0 min = 23.68 max = 23.83 avg = 23.74\n regnety_400m min = 12.65 max = 12.68 avg = 12.66\n blazeface min = 1.99 max = 2.14 avg = 2.03\n googlenet min = 30.83 max = 31.29 avg = 30.91\n googlenet_int8 min = 31.97 max = 33.12 avg = 32.45\n resnet18 min = 18.81 max = 18.87 avg = 18.84\n resnet18_int8 min = 32.80 max = 32.99 avg = 32.90\n alexnet min = 22.88 max = 23.16 avg = 22.94\n vgg16 min = 100.58 max = 101.12 avg = 100.90\n vgg16_int8 min = 235.81 max = 237.97 avg = 236.20\n resnet50 min = 51.12 max = 51.43 avg = 51.28\n resnet50_int8 min = 62.46 max = 63.02 avg = 62.72\n squeezenet_ssd min = 23.26 max = 23.73 avg = 23.38\n squeezenet_ssd_int8 min = 31.91 max = 32.30 avg = 32.13\n mobilenet_ssd min = 24.73 max = 24.95 avg = 24.84\n mobilenet_ssd_int8 min = 20.99 max = 21.52 avg = 21.21\n mobilenet_yolo min = 54.91 max = 55.70 avg = 55.15\n mobilenetv2_yolov3 min = 30.18 max = 30.52 avg = 30.31\n yolov4-tiny min = 40.46 max = 40.61 avg = 40.55\n nanodet_m min = 12.56 max = 12.72 avg = 12.62\n yolo-fastest-1.1 min = 6.00 max = 6.15 avg = 6.04\n yolo-fastestv2 min = 5.32 max = 5.59 avg = 5.43\n vision_transformer min = 894.51 max = 896.28 avg = 895.57\n FastestDet min = 5.33 max = 5.42 avg = 5.36\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 4 0 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 4.18 max = 4.35 avg = 4.22\n squeezenet_int8 min = 4.85 max = 4.98 avg = 4.89\n mobilenet min = 5.80 max = 5.95 avg = 5.89\n mobilenet_int8 min = 4.86 max = 4.94 avg = 4.89\n mobilenet_v2 min = 4.66 max = 4.73 avg = 4.69\n mobilenet_v3 min = 4.46 max = 4.50 avg = 4.48\n shufflenet min = 4.01 max = 4.17 avg = 4.04\n shufflenet_v2 min = 3.39 max = 3.41 avg = 3.39\n mnasnet min = 4.81 max = 4.93 avg = 4.85\n proxylessnasnet min = 5.47 max = 5.54 avg = 5.49\n efficientnet_b0 min = 10.49 max = 10.55 avg = 10.52\n efficientnetv2_b0 min = 13.67 max = 13.77 avg = 13.72\n regnety_400m min = 10.20 max = 10.24 avg = 10.21\n blazeface min = 1.52 max = 1.58 avg = 1.54\n googlenet min = 17.65 max = 17.69 avg = 17.68\n googlenet_int8 min = 18.14 max = 18.27 avg = 18.19\n resnet18 min = 10.52 max = 10.63 avg = 10.57\n resnet18_int8 min = 17.42 max = 17.53 avg = 17.49\n alexnet min = 13.12 max = 13.20 avg = 13.16\n vgg16 min = 55.24 max = 55.45 avg = 55.35\n vgg16_int8 min = 123.46 max = 124.23 avg = 123.75\n resnet50 min = 28.31 max = 28.57 avg = 28.39\n resnet50_int8 min = 34.10 max = 34.39 avg = 34.23\n squeezenet_ssd min = 14.85 max = 14.96 avg = 14.91\n squeezenet_ssd_int8 min = 19.71 max = 19.88 avg = 19.82\n mobilenet_ssd min = 13.49 max = 13.58 avg = 13.52\n mobilenet_ssd_int8 min = 11.60 max = 11.70 avg = 11.66\n mobilenet_yolo min = 31.74 max = 31.96 avg = 31.81\n mobilenetv2_yolov3 min = 17.87 max = 18.03 avg = 17.93\n yolov4-tiny min = 25.63 max = 25.78 avg = 25.72\n nanodet_m min = 8.16 max = 8.22 avg = 8.20\n yolo-fastest-1.1 min = 4.72 max = 4.86 avg = 4.75\n yolo-fastestv2 min = 3.98 max = 4.15 avg = 4.00\n vision_transformer min = 501.18 max = 503.51 avg = 502.12\n FastestDet min = 3.74 max = 3.76 avg = 3.75\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 8 0 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.91 max = 3.10 avg = 2.97\n squeezenet_int8 min = 3.42 max = 3.74 avg = 3.51\n mobilenet min = 3.57 max = 3.70 avg = 3.61\n mobilenet_int8 min = 3.06 max = 3.14 avg = 3.10\n mobilenet_v2 min = 3.73 max = 3.75 avg = 3.75\n mobilenet_v3 min = 3.50 max = 3.66 avg = 3.56\n shufflenet min = 3.63 max = 3.65 avg = 3.64\n shufflenet_v2 min = 2.85 max = 3.02 avg = 2.95\n mnasnet min = 3.60 max = 3.67 avg = 3.62\n proxylessnasnet min = 4.00 max = 4.08 avg = 4.03\n efficientnet_b0 min = 7.31 max = 7.34 avg = 7.33\n efficientnetv2_b0 min = 9.44 max = 9.51 avg = 9.47\n regnety_400m min = 9.76 max = 10.07 avg = 9.90\n blazeface min = 1.56 max = 1.75 avg = 1.61\n googlenet min = 11.22 max = 11.28 avg = 11.25\n googlenet_int8 min = 11.40 max = 12.82 avg = 11.76\n resnet18 min = 6.83 max = 6.96 avg = 6.90\n resnet18_int8 min = 10.28 max = 10.38 avg = 10.33\n alexnet min = 8.75 max = 8.88 avg = 8.80\n vgg16 min = 36.00 max = 36.72 avg = 36.29\n vgg16_int8 min = 67.38 max = 67.72 avg = 67.54\n resnet50 min = 17.63 max = 17.82 avg = 17.68\n resnet50_int8 min = 20.05 max = 20.21 avg = 20.15\n squeezenet_ssd min = 11.18 max = 11.45 avg = 11.26\n squeezenet_ssd_int8 min = 14.09 max = 14.23 avg = 14.18\n mobilenet_ssd min = 8.60 max = 8.69 avg = 8.64\n mobilenet_ssd_int8 min = 7.75 max = 7.87 avg = 7.81\n mobilenet_yolo min = 21.97 max = 22.25 avg = 22.09\n mobilenetv2_yolov3 min = 14.04 max = 14.18 avg = 14.12\n yolov4-tiny min = 19.66 max = 19.93 avg = 19.81\n nanodet_m min = 6.52 max = 6.67 avg = 6.57\n yolo-fastest-1.1 min = 4.61 max = 4.76 avg = 4.66\n yolo-fastestv2 min = 3.78 max = 3.91 avg = 3.82\n vision_transformer min = 323.01 max = 327.38 avg = 323.75\n FastestDet min = 3.50 max = 3.54 avg = 3.51\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 16 0 -1 0\nloop_count = 10\nnum_threads = 16\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.00 max = 3.25 avg = 3.08\n squeezenet_int8 min = 4.13 max = 4.47 avg = 4.21\n mobilenet min = 3.27 max = 3.42 avg = 3.34\n mobilenet_int8 min = 3.49 max = 3.58 avg = 3.56\n mobilenet_v2 min = 3.86 max = 4.10 avg = 3.97\n mobilenet_v3 min = 3.72 max = 3.80 avg = 3.76\n shufflenet min = 4.67 max = 4.78 avg = 4.72\n shufflenet_v2 min = 3.16 max = 3.24 avg = 3.20\n mnasnet min = 3.51 max = 3.65 avg = 3.57\n proxylessnasnet min = 4.08 max = 4.35 avg = 4.15\n efficientnet_b0 min = 7.51 max = 7.80 avg = 7.63\n efficientnetv2_b0 min = 8.92 max = 9.39 avg = 9.05\n regnety_400m min = 14.80 max = 15.05 avg = 14.89\n blazeface min = 2.14 max = 2.28 avg = 2.20\n googlenet min = 9.91 max = 10.00 avg = 9.96\n googlenet_int8 min = 11.51 max = 11.65 avg = 11.60\n resnet18 min = 6.39 max = 6.56 avg = 6.46\n resnet18_int8 min = 9.76 max = 9.91 avg = 9.84\n alexnet min = 6.99 max = 7.10 avg = 7.04\n vgg16 min = 27.52 max = 28.64 avg = 27.88\n vgg16_int8 min = 45.64 max = 45.93 avg = 45.78\n resnet50 min = 13.96 max = 14.17 avg = 14.07\n resnet50_int8 min = 16.82 max = 16.93 avg = 16.89\n squeezenet_ssd min = 11.11 max = 11.54 avg = 11.23\n squeezenet_ssd_int8 min = 13.77 max = 14.00 avg = 13.88\n mobilenet_ssd min = 8.21 max = 8.46 avg = 8.35\n mobilenet_ssd_int8 min = 8.87 max = 9.03 avg = 8.94\n mobilenet_yolo min = 30.77 max = 31.35 avg = 31.08\n mobilenetv2_yolov3 min = 12.11 max = 13.10 avg = 12.43\n yolov4-tiny min = 18.25 max = 18.68 avg = 18.41\n nanodet_m min = 6.55 max = 6.68 avg = 6.59\n yolo-fastest-1.1 min = 6.00 max = 6.22 avg = 6.09\n yolo-fastestv2 min = 4.86 max = 5.01 avg = 4.94\n vision_transformer min = 218.18 max = 220.49 avg = 218.79\n FastestDet min = 5.01 max = 5.14 avg = 5.07\n(base) mobtgzhang@mobtgzhang-PC:~/ncnn/benchmark$ ./benchncnn 10 24 0 -1 0\nloop_count = 10\nnum_threads = 24\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.52 max = 3.96 avg = 3.70\n squeezenet_int8 min = 5.49 max = 5.83 avg = 5.65\n mobilenet min = 3.42 max = 3.83 avg = 3.55\n mobilenet_int8 min = 3.69 max = 45.17 avg = 11.59\n mobilenet_v2 min = 4.63 max = 5.44 avg = 4.84\n mobilenet_v3 min = 4.51 max = 4.89 avg = 4.68\n shufflenet min = 6.21 max = 6.52 avg = 6.36\n shufflenet_v2 min = 3.98 max = 17.54 avg = 5.45\n mnasnet min = 4.28 max = 4.56 avg = 4.39\n proxylessnasnet min = 4.76 max = 5.13 avg = 4.92\n efficientnet_b0 min = 7.45 max = 111.76 avg = 22.59\n efficientnetv2_b0 min = 10.87 max = 33.13 avg = 13.51\n regnety_400m min = 20.97 max = 21.73 avg = 21.46\n blazeface min = 2.56 max = 2.82 avg = 2.67\n googlenet min = 10.54 max = 105.87 avg = 21.85\n googlenet_int8 min = 14.21 max = 77.02 avg = 22.23\n resnet18 min = 7.08 max = 7.51 avg = 7.31\n resnet18_int8 min = 11.25 max = 50.66 avg = 19.14\n alexnet min = 7.13 max = 8.67 avg = 7.44\n vgg16 min = 27.59 max = 35.35 avg = 29.12\n vgg16_int8 min = 44.43 max = 51.76 avg = 46.90\n resnet50 min = 15.16 max = 105.98 avg = 24.91\n resnet50_int8 min = 19.82 max = 20.50 avg = 20.16\n squeezenet_ssd min = 13.03 max = 13.69 avg = 13.40\n squeezenet_ssd_int8 min = 17.62 max = 187.55 avg = 39.92\n mobilenet_ssd min = 8.83 max = 71.97 avg = 15.37\n mobilenet_ssd_int8 min = 10.22 max = 49.61 avg = 15.26\n mobilenet_yolo min = 35.19 max = 46.43 avg = 36.93\n mobilenetv2_yolov3 min = 12.96 max = 15.57 avg = 13.41\n yolov4-tiny min = 19.22 max = 21.43 avg = 19.89\n nanodet_m min = 7.71 max = 8.74 avg = 8.09\n yolo-fastest-1.1 min = 6.71 max = 78.72 avg = 14.16\n yolo-fastestv2 min = 5.72 max = 6.08 avg = 5.88\n vision_transformer min = 192.16 max = 221.86 avg = 202.73\n FastestDet min = 5.13 max = 5.47 avg = 5.30\n```\n\n### HUAWEI Kunpeng 920 7260 (x64 cores)\ntest on Ubuntu 20.04 (gcc 9.4.0)\n```\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 1 0 -1 0\nloop_count = 300\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 11.64 max = 12.11 avg = 11.71\n squeezenet_int8 min = 12.22 max = 13.22 avg = 12.37\n mobilenet min = 20.00 max = 20.79 avg = 20.08\n mobilenet_int8 min = 17.44 max = 19.09 avg = 17.64\n mobilenet_v2 min = 13.29 max = 14.25 avg = 13.39\n mobilenet_v3 min = 11.06 max = 11.84 avg = 11.11\n shufflenet min = 7.56 max = 7.74 avg = 7.59\n shufflenet_v2 min = 7.84 max = 8.37 avg = 7.88\n mnasnet min = 13.07 max = 13.78 avg = 13.14\n proxylessnasnet min = 15.71 max = 16.31 avg = 15.77\n efficientnet_b0 min = 34.79 max = 35.98 avg = 34.92\n efficientnetv2_b0 min = 35.28 max = 36.36 avg = 35.41\n regnety_400m min = 17.06 max = 17.74 avg = 17.16\n blazeface min = 2.99 max = 3.04 avg = 3.01\n googlenet min = 50.76 max = 51.74 avg = 51.00\n googlenet_int8 min = 50.31 max = 52.27 avg = 50.65\n resnet18 min = 34.97 max = 37.17 avg = 35.82\n resnet18_int8 min = 40.47 max = 42.03 avg = 40.78\n alexnet min = 39.19 max = 39.80 avg = 39.32\n vgg16 min = 176.62 max = 181.29 avg = 177.07\n vgg16_int8 min = 352.35 max = 358.38 avg = 355.15\n resnet50 min = 96.76 max = 98.63 avg = 97.09\n resnet50_int8 min = 90.00 max = 92.74 avg = 90.81\n squeezenet_ssd min = 33.23 max = 33.99 avg = 33.39\n squeezenet_ssd_int8 min = 38.50 max = 41.53 avg = 39.28\n mobilenet_ssd min = 42.49 max = 44.78 avg = 42.72\n mobilenet_ssd_int8 min = 37.06 max = 39.97 avg = 37.57\n mobilenet_yolo min = 96.34 max = 98.91 avg = 96.73\n mobilenetv2_yolov3 min = 50.88 max = 52.97 avg = 51.15\n yolov4-tiny min = 65.56 max = 67.13 avg = 65.80\n nanodet_m min = 19.94 max = 20.82 avg = 20.04\n yolo-fastest-1.1 min = 7.66 max = 7.81 avg = 7.71\n yolo-fastestv2 min = 6.82 max = 7.23 avg = 6.87\n vision_transformer min = 1535.03 max = 1552.84 avg = 1543.73\n FastestDet min = 7.17 max = 7.50 avg = 7.21\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 2 0 -1 0\nloop_count = 300\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.35 max = 9.15 avg = 7.33\n squeezenet_int8 min = 8.06 max = 8.60 avg = 8.14\n mobilenet min = 10.30 max = 11.86 avg = 11.48\n mobilenet_int8 min = 8.93 max = 11.87 avg = 10.47\n mobilenet_v2 min = 9.05 max = 11.50 avg = 9.19\n mobilenet_v3 min = 6.32 max = 6.42 avg = 6.36\n shufflenet min = 6.73 max = 8.55 avg = 6.81\n shufflenet_v2 min = 4.94 max = 6.65 avg = 6.32\n mnasnet min = 7.38 max = 10.77 avg = 8.82\n proxylessnasnet min = 8.57 max = 9.72 avg = 8.63\n efficientnet_b0 min = 18.61 max = 22.53 avg = 20.42\n efficientnetv2_b0 min = 18.75 max = 21.93 avg = 20.79\n regnety_400m min = 11.86 max = 15.09 avg = 14.60\n blazeface min = 1.95 max = 3.37 avg = 2.06\n googlenet min = 28.66 max = 32.24 avg = 28.94\n googlenet_int8 min = 27.64 max = 32.15 avg = 30.84\n resnet18 min = 20.33 max = 20.77 avg = 20.47\n resnet18_int8 min = 22.63 max = 23.72 avg = 22.88\n alexnet min = 20.41 max = 29.37 avg = 27.22\n vgg16 min = 101.72 max = 140.33 avg = 103.29\n vgg16_int8 min = 187.56 max = 211.44 avg = 189.92\n resnet50 min = 51.07 max = 59.25 avg = 58.35\n resnet50_int8 min = 46.50 max = 52.55 avg = 48.93\n squeezenet_ssd min = 22.48 max = 28.59 avg = 22.98\n squeezenet_ssd_int8 min = 25.56 max = 26.82 avg = 25.99\n mobilenet_ssd min = 22.81 max = 26.21 avg = 24.88\n mobilenet_ssd_int8 min = 19.31 max = 25.53 avg = 21.74\n mobilenet_yolo min = 59.58 max = 62.04 avg = 59.99\n mobilenetv2_yolov3 min = 33.26 max = 35.74 avg = 33.51\n yolov4-tiny min = 41.14 max = 45.34 avg = 42.46\n nanodet_m min = 12.10 max = 16.69 avg = 15.02\n yolo-fastest-1.1 min = 5.44 max = 7.78 avg = 7.24\n yolo-fastestv2 min = 5.03 max = 8.08 avg = 6.75\n vision_transformer min = 994.46 max = 1090.68 avg = 1045.50\n FastestDet min = 6.76 max = 6.91 avg = 6.83\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 4 0 -1 0\nloop_count = 300\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.79 max = 6.99 avg = 4.55\n squeezenet_int8 min = 5.13 max = 5.68 avg = 5.20\n mobilenet min = 6.25 max = 6.55 avg = 6.30\n mobilenet_int8 min = 5.96 max = 6.10 avg = 6.03\n mobilenet_v2 min = 5.34 max = 7.15 avg = 5.62\n mobilenet_v3 min = 4.05 max = 5.74 avg = 5.01\n shufflenet min = 3.69 max = 5.81 avg = 5.15\n shufflenet_v2 min = 4.31 max = 6.02 avg = 4.56\n mnasnet min = 4.48 max = 6.05 avg = 5.54\n proxylessnasnet min = 5.05 max = 8.08 avg = 6.03\n efficientnet_b0 min = 10.17 max = 12.21 avg = 11.58\n efficientnetv2_b0 min = 10.86 max = 15.78 avg = 12.70\n regnety_400m min = 9.24 max = 14.13 avg = 11.98\n blazeface min = 1.89 max = 1.97 avg = 1.93\n googlenet min = 15.19 max = 20.31 avg = 16.90\n googlenet_int8 min = 17.97 max = 19.40 avg = 18.11\n resnet18 min = 11.18 max = 11.48 avg = 11.29\n resnet18_int8 min = 12.26 max = 12.78 avg = 12.44\n alexnet min = 14.43 max = 16.94 avg = 14.68\n vgg16 min = 62.40 max = 78.42 avg = 64.96\n vgg16_int8 min = 101.52 max = 109.42 avg = 104.46\n resnet50 min = 29.19 max = 39.69 avg = 32.99\n resnet50_int8 min = 26.94 max = 28.82 avg = 27.16\n squeezenet_ssd min = 12.90 max = 16.52 avg = 15.20\n squeezenet_ssd_int8 min = 15.58 max = 18.40 avg = 16.28\n mobilenet_ssd min = 13.68 max = 14.45 avg = 13.87\n mobilenet_ssd_int8 min = 12.20 max = 14.58 avg = 12.84\n mobilenet_yolo min = 34.85 max = 36.54 avg = 35.05\n mobilenetv2_yolov3 min = 18.61 max = 20.93 avg = 19.92\n yolov4-tiny min = 26.09 max = 32.32 avg = 28.03\n nanodet_m min = 7.85 max = 12.48 avg = 11.00\n yolo-fastest-1.1 min = 6.19 max = 6.49 avg = 6.31\n yolo-fastestv2 min = 3.66 max = 6.83 avg = 5.11\n vision_transformer min = 605.95 max = 624.99 avg = 609.79\n FastestDet min = 4.32 max = 5.41 avg = 5.17\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 8 0 -1 0\nloop_count = 300\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.72 max = 3.74 avg = 3.05\n squeezenet_int8 min = 3.80 max = 4.71 avg = 4.03\n mobilenet min = 3.94 max = 5.15 avg = 4.00\n mobilenet_int8 min = 3.73 max = 3.87 avg = 3.80\n mobilenet_v2 min = 4.51 max = 6.57 avg = 4.68\n mobilenet_v3 min = 4.12 max = 4.38 avg = 4.28\n shufflenet min = 4.60 max = 6.27 avg = 4.88\n shufflenet_v2 min = 4.07 max = 4.20 avg = 4.11\n mnasnet min = 4.26 max = 4.51 avg = 4.36\n proxylessnasnet min = 4.71 max = 7.40 avg = 4.80\n efficientnet_b0 min = 8.49 max = 8.74 avg = 8.56\n efficientnetv2_b0 min = 9.34 max = 9.68 avg = 9.41\n regnety_400m min = 8.00 max = 12.85 avg = 10.64\n blazeface min = 1.76 max = 1.84 avg = 1.80\n googlenet min = 10.89 max = 11.33 avg = 10.98\n googlenet_int8 min = 11.66 max = 14.07 avg = 11.83\n resnet18 min = 6.48 max = 6.61 avg = 6.54\n resnet18_int8 min = 7.30 max = 7.79 avg = 7.51\n alexnet min = 8.33 max = 8.95 avg = 8.62\n vgg16 min = 29.94 max = 47.54 avg = 31.95\n vgg16_int8 min = 54.67 max = 60.76 avg = 56.03\n resnet50 min = 16.13 max = 20.79 avg = 20.03\n resnet50_int8 min = 15.64 max = 20.13 avg = 16.11\n squeezenet_ssd min = 11.58 max = 12.02 avg = 11.77\n squeezenet_ssd_int8 min = 11.14 max = 13.72 avg = 12.10\n mobilenet_ssd min = 8.27 max = 10.77 avg = 8.76\n mobilenet_ssd_int8 min = 8.13 max = 9.09 avg = 8.29\n mobilenet_yolo min = 23.90 max = 24.69 avg = 24.17\n mobilenetv2_yolov3 min = 14.83 max = 15.72 avg = 15.19\n yolov4-tiny min = 19.78 max = 23.66 avg = 20.05\n nanodet_m min = 8.92 max = 10.76 avg = 9.09\n yolo-fastest-1.1 min = 5.49 max = 5.77 avg = 5.63\n yolo-fastestv2 min = 5.04 max = 5.21 avg = 5.10\n vision_transformer min = 318.42 max = 379.40 avg = 363.66\n FastestDet min = 4.18 max = 4.54 avg = 4.38\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 16 0 -1 0\nloop_count = 300\nnum_threads = 16\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.70 max = 3.14 avg = 2.81\n squeezenet_int8 min = 3.21 max = 4.22 avg = 3.39\n mobilenet min = 3.13 max = 3.26 avg = 3.20\n mobilenet_int8 min = 3.17 max = 5.05 avg = 3.30\n mobilenet_v2 min = 4.31 max = 6.24 avg = 4.62\n mobilenet_v3 min = 3.57 max = 3.77 avg = 3.68\n shufflenet min = 4.70 max = 6.45 avg = 4.80\n shufflenet_v2 min = 3.73 max = 4.27 avg = 3.87\n mnasnet min = 3.67 max = 3.87 avg = 3.75\n proxylessnasnet min = 4.28 max = 4.81 avg = 4.35\n efficientnet_b0 min = 7.31 max = 7.77 avg = 7.53\n efficientnetv2_b0 min = 9.87 max = 12.33 avg = 10.07\n regnety_400m min = 17.95 max = 18.53 avg = 18.26\n blazeface min = 2.26 max = 2.40 avg = 2.33\n googlenet min = 9.51 max = 9.99 avg = 9.68\n googlenet_int8 min = 10.98 max = 11.36 avg = 11.18\n resnet18 min = 5.59 max = 6.08 avg = 5.71\n resnet18_int8 min = 6.55 max = 7.28 avg = 6.77\n alexnet min = 6.26 max = 6.50 avg = 6.36\n vgg16 min = 23.98 max = 27.37 avg = 24.89\n vgg16_int8 min = 38.07 max = 39.66 avg = 39.02\n resnet50 min = 12.81 max = 14.19 avg = 13.76\n resnet50_int8 min = 12.42 max = 12.84 avg = 12.55\n squeezenet_ssd min = 10.80 max = 11.49 avg = 11.12\n squeezenet_ssd_int8 min = 11.57 max = 12.21 avg = 11.74\n mobilenet_ssd min = 7.46 max = 8.08 avg = 7.84\n mobilenet_ssd_int8 min = 7.47 max = 8.07 avg = 7.63\n mobilenet_yolo min = 21.70 max = 23.43 avg = 21.92\n mobilenetv2_yolov3 min = 12.55 max = 14.56 avg = 12.90\n yolov4-tiny min = 17.68 max = 19.85 avg = 18.18\n nanodet_m min = 8.35 max = 8.70 avg = 8.45\n yolo-fastest-1.1 min = 5.70 max = 7.11 avg = 6.05\n yolo-fastestv2 min = 4.85 max = 5.70 avg = 5.37\n vision_transformer min = 214.36 max = 259.56 avg = 245.47\n FastestDet min = 5.01 max = 5.42 avg = 5.17\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 32 0 -1 0\nloop_count = 300\nnum_threads = 32\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.30 max = 2.94 avg = 2.46\n squeezenet_int8 min = 3.08 max = 4.88 avg = 4.03\n mobilenet min = 2.49 max = 2.76 avg = 2.53\n mobilenet_int8 min = 2.86 max = 3.73 avg = 2.95\n mobilenet_v2 min = 4.51 max = 5.20 avg = 4.74\n mobilenet_v3 min = 5.11 max = 6.91 avg = 6.10\n shufflenet min = 5.57 max = 6.51 avg = 5.78\n shufflenet_v2 min = 4.37 max = 4.66 avg = 4.48\n mnasnet min = 3.72 max = 4.08 avg = 3.90\n proxylessnasnet min = 4.19 max = 6.18 avg = 4.79\n efficientnet_b0 min = 6.80 max = 7.22 avg = 6.89\n efficientnetv2_b0 min = 13.98 max = 17.55 avg = 15.06\n regnety_400m min = 16.10 max = 16.72 avg = 16.26\n blazeface min = 2.12 max = 2.53 avg = 2.17\n googlenet min = 8.63 max = 9.89 avg = 8.77\n googlenet_int8 min = 9.90 max = 11.09 avg = 10.08\n resnet18 min = 6.54 max = 6.99 avg = 6.73\n resnet18_int8 min = 8.34 max = 9.00 avg = 8.67\n alexnet min = 6.64 max = 7.15 avg = 6.93\n vgg16 min = 22.79 max = 23.91 avg = 23.50\n vgg16_int8 min = 32.37 max = 37.51 avg = 33.13\n resnet50 min = 11.19 max = 16.40 avg = 11.47\n resnet50_int8 min = 11.92 max = 12.55 avg = 12.13\n squeezenet_ssd min = 10.75 max = 12.28 avg = 11.12\n squeezenet_ssd_int8 min = 11.31 max = 12.29 avg = 11.57\n mobilenet_ssd min = 10.25 max = 11.26 avg = 10.79\n mobilenet_ssd_int8 min = 11.39 max = 16.99 avg = 11.98\n mobilenet_yolo min = 52.11 max = 60.46 avg = 53.84\n mobilenetv2_yolov3 min = 12.07 max = 12.47 avg = 12.20\n yolov4-tiny min = 17.48 max = 17.79 avg = 17.58\n nanodet_m min = 13.06 max = 14.71 avg = 13.64\n yolo-fastest-1.1 min = 5.70 max = 5.89 avg = 5.79\n yolo-fastestv2 min = 8.89 max = 9.99 avg = 9.21\n vision_transformer min = 158.92 max = 187.40 avg = 168.21\n FastestDet min = 8.70 max = 9.43 avg = 9.00\nroot@8d46e508165f:/home/lkl/ARM_CHAR/ncnn/benchmark# ../build/benchmark/benchncnn 300 64 0 -1 0\nloop_count = 300\nnum_threads = 64\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.85 max = 78.56 avg = 7.81\n squeezenet_int8 min = 8.06 max = 88.91 avg = 9.23\n mobilenet min = 3.02 max = 86.86 avg = 5.89\n mobilenet_int8 min = 3.58 max = 4.55 avg = 3.68\n mobilenet_v2 min = 5.05 max = 150.06 avg = 13.04\n mobilenet_v3 min = 4.85 max = 125.22 avg = 8.34\n shufflenet min = 17.80 max = 220.55 avg = 21.01\n shufflenet_v2 min = 11.23 max = 381.95 avg = 13.71\n mnasnet min = 9.83 max = 128.42 avg = 11.10\n proxylessnasnet min = 10.53 max = 68.52 avg = 12.03\n efficientnet_b0 min = 16.78 max = 968.87 avg = 23.94\n efficientnetv2_b0 min = 26.23 max = 551.18 avg = 31.34\n regnety_400m min = 70.14 max = 407.92 avg = 78.30\n blazeface min = 7.27 max = 191.44 avg = 9.37\n googlenet min = 16.69 max = 820.58 avg = 25.06\n googlenet_int8 min = 20.58 max = 849.09 avg = 29.87\n resnet18 min = 8.67 max = 349.00 avg = 11.33\n resnet18_int8 min = 10.40 max = 128.98 avg = 11.45\n alexnet min = 6.15 max = 196.01 avg = 10.24\n vgg16 min = 21.11 max = 288.66 avg = 29.37\n vgg16_int8 min = 30.72 max = 251.95 avg = 37.68\n resnet50 min = 19.10 max = 114.08 avg = 22.00\n resnet50_int8 min = 18.99 max = 436.89 avg = 24.36\n squeezenet_ssd min = 22.22 max = 510.52 avg = 28.76\n squeezenet_ssd_int8 min = 23.42 max = 614.70 avg = 30.82\n mobilenet_ssd min = 7.62 max = 202.66 avg = 14.59\n mobilenet_ssd_int8 min = 7.89 max = 109.82 avg = 8.80\n mobilenet_yolo min = 31.43 max = 742.10 avg = 45.52\n mobilenetv2_yolov3 min = 18.31 max = 273.05 avg = 20.78\n yolov4-tiny min = 21.03 max = 400.05 avg = 33.64\n nanodet_m min = 19.94 max = 114.18 avg = 21.89\n yolo-fastest-1.1 min = 7.20 max = 174.60 avg = 9.13\n yolo-fastestv2 min = 7.50 max = 170.55 avg = 9.01\n vision_transformer min = 126.90 max = 335.71 avg = 157.38\n FastestDet min = 6.59 max = 19.77 avg = 6.77\n```\n\n### Intel Atom x5-Z8350\n```\nnihui@nihui-ROCK-Pi-X:~/ncnn/build/benchmark$ ./benchncnn 20 4 0 -1 1\nloop_count = 20\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 50.22 max = 50.53 avg = 50.32\n squeezenet_int8 min = 77.92 max = 78.37 avg = 78.07\n mobilenet min = 80.12 max = 81.53 avg = 80.35\n mobilenet_int8 min = 120.54 max = 124.10 avg = 120.84\n mobilenet_v2 min = 56.62 max = 60.12 avg = 58.37\n mobilenet_v3 min = 50.19 max = 50.41 avg = 50.27\n shufflenet min = 37.96 max = 38.28 avg = 38.10\n shufflenet_v2 min = 35.28 max = 35.59 avg = 35.45\n mnasnet min = 54.91 max = 55.10 avg = 55.01\n proxylessnasnet min = 62.25 max = 62.59 avg = 62.40\n efficientnet_b0 min = 101.92 max = 105.73 avg = 102.27\n efficientnetv2_b0 min = 115.48 max = 117.25 avg = 115.89\n regnety_400m min = 79.66 max = 81.70 avg = 79.95\n blazeface min = 10.43 max = 10.60 avg = 10.49\n googlenet min = 170.41 max = 173.44 avg = 170.68\n googlenet_int8 min = 253.06 max = 257.34 avg = 253.57\n resnet18 min = 127.19 max = 130.69 avg = 127.65\n resnet18_int8 min = 200.54 max = 204.25 avg = 200.88\n alexnet min = 104.89 max = 110.89 avg = 105.56\n vgg16 min = 653.78 max = 661.34 avg = 655.44\n vgg16_int8 min = 974.72 max = 1006.48 avg = 978.76\n resnet50 min = 367.63 max = 371.74 avg = 368.27\n resnet50_int8 min = 574.94 max = 584.08 avg = 576.18\n squeezenet_ssd min = 115.35 max = 116.47 avg = 115.62\n squeezenet_ssd_int8 min = 169.95 max = 170.75 avg = 170.26\n mobilenet_ssd min = 167.00 max = 172.02 avg = 168.95\n mobilenet_ssd_int8 min = 244.91 max = 248.30 avg = 245.27\n mobilenet_yolo min = 382.80 max = 393.23 avg = 385.79\n mobilenetv2_yolov3 min = 208.23 max = 211.54 avg = 209.64\n yolov4-tiny min = 251.10 max = 263.77 avg = 256.37\n nanodet_m min = 84.48 max = 84.95 avg = 84.70\n yolo-fastest-1.1 min = 44.11 max = 45.15 avg = 44.26\n yolo-fastestv2 min = 37.95 max = 38.52 avg = 38.34\n\nnihui@nihui-ROCK-Pi-X:~/ncnn/build/benchmark$ ./benchncnn 10 1 0 -1 1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 130.52 max = 131.08 avg = 130.64\n squeezenet_int8 min = 231.03 max = 231.38 avg = 231.19\n mobilenet min = 231.40 max = 231.74 avg = 231.61\n mobilenet_int8 min = 409.74 max = 410.02 avg = 409.85\n mobilenet_v2 min = 150.23 max = 150.72 avg = 150.47\n mobilenet_v3 min = 119.08 max = 119.34 avg = 119.20\n shufflenet min = 72.62 max = 72.81 avg = 72.73\n shufflenet_v2 min = 73.63 max = 73.71 avg = 73.68\n mnasnet min = 140.87 max = 141.09 avg = 140.98\n proxylessnasnet min = 166.39 max = 166.75 avg = 166.54\n efficientnet_b0 min = 280.55 max = 281.30 avg = 280.77\n efficientnetv2_b0 min = 321.05 max = 321.24 avg = 321.16\n regnety_400m min = 183.78 max = 184.64 avg = 183.91\n blazeface min = 18.94 max = 19.08 avg = 19.01\n googlenet min = 453.56 max = 454.71 avg = 454.15\n googlenet_int8 min = 791.40 max = 791.93 avg = 791.61\n resnet18 min = 365.87 max = 366.40 avg = 366.15\n resnet18_int8 min = 652.86 max = 653.39 avg = 653.09\n alexnet min = 289.15 max = 290.25 avg = 289.65\n vgg16 min = 1887.16 max = 1887.73 avg = 1887.41\n vgg16_int8 min = 3211.44 max = 3213.39 avg = 3212.55\n resnet50 min = 1060.37 max = 1061.40 avg = 1060.80\n resnet50_int8 min = 1869.41 max = 1870.59 avg = 1870.17\n squeezenet_ssd min = 277.23 max = 277.83 avg = 277.50\n squeezenet_ssd_int8 min = 455.54 max = 458.06 avg = 456.28\n mobilenet_ssd min = 478.03 max = 478.83 avg = 478.32\n mobilenet_ssd_int8 min = 822.61 max = 822.96 avg = 822.79\n mobilenet_yolo min = 1136.89 max = 1138.51 avg = 1137.74\n mobilenetv2_yolov3 min = 551.81 max = 552.53 avg = 552.14\n yolov4-tiny min = 685.49 max = 686.15 avg = 685.79\n nanodet_m min = 181.21 max = 181.52 avg = 181.32\n yolo-fastest-1.1 min = 82.21 max = 82.68 avg = 82.30\n yolo-fastestv2 min = 67.62 max = 68.36 avg = 68.10\n\nroot@nihui-ROCK-Pi-X:/home/nihui/osd/ncnn/build/benchmark# ./benchncnn 10 1 0 0 0\n[0 Intel(R) HD Graphics (CHV)] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 Intel(R) HD Graphics (CHV)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Intel(R) HD Graphics (CHV)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Intel(R) HD Graphics (CHV)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 29.14 max = 29.76 avg = 29.45\n mobilenet min = 36.19 max = 37.03 avg = 36.52\n mobilenet_v2 min = 30.39 max = 31.62 avg = 30.76\n mobilenet_v3 min = 31.60 max = 32.25 avg = 31.92\n shufflenet min = 22.47 max = 23.19 avg = 22.70\n shufflenet_v2 min = 22.30 max = 24.16 avg = 23.12\n mnasnet min = 29.40 max = 30.23 avg = 29.84\n proxylessnasnet min = 31.00 max = 31.91 avg = 31.41\n efficientnet_b0 min = 58.03 max = 58.74 avg = 58.42\n efficientnetv2_b0 min = 131.17 max = 191.61 avg = 161.37\n regnety_400m min = 40.30 max = 42.27 avg = 41.04\n blazeface min = 15.06 max = 15.96 avg = 15.48\n googlenet min = 85.37 max = 86.49 avg = 85.84\n resnet18 min = 93.87 max = 95.00 avg = 94.53\n alexnet min = 110.96 max = 120.83 avg = 115.14\n vgg16 min = 798.75 max = 812.60 avg = 804.93\n resnet50 min = 213.12 max = 214.81 avg = 213.79\n squeezenet_ssd min = 124.48 max = 125.18 avg = 124.87\n mobilenet_ssd min = 84.04 max = 84.70 avg = 84.49\n mobilenet_yolo min = 186.52 max = 189.61 avg = 188.53\n mobilenetv2_yolov3 min = 102.07 max = 102.97 avg = 102.39\n yolov4-tiny min = 212.49 max = 214.75 avg = 213.77\n nanodet_m min = 42.97 max = 45.58 avg = 44.05\n yolo-fastest-1.1 min = 27.14 max = 32.53 avg = 28.76\n yolo-fastestv2 min = 20.73 max = 25.90 avg = 22.97\n```\n\n### Intel Celeron N5105\n```\nloop_count = 8\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 18.06 max = 18.21 avg = 18.12\n squeezenet_int8 min = 24.55 max = 25.16 avg = 24.69\n mobilenet min = 32.22 max = 32.70 avg = 32.40\n mobilenet_int8 min = 40.52 max = 40.59 avg = 40.54\n mobilenet_v2 min = 22.54 max = 22.71 avg = 22.65\n mobilenet_v3 min = 17.86 max = 19.02 avg = 18.09\n shufflenet min = 11.23 max = 11.30 avg = 11.28\n shufflenet_v2 min = 11.04 max = 11.19 avg = 11.13\n mnasnet min = 19.93 max = 20.09 avg = 20.01\n proxylessnasnet min = 21.91 max = 22.00 avg = 21.95\n efficientnet_b0 min = 33.29 max = 33.66 avg = 33.50\n efficientnetv2_b0 min = 40.16 max = 40.63 avg = 40.34\n regnety_400m min = 27.38 max = 27.59 avg = 27.50\n blazeface min = 3.01 max = 3.11 avg = 3.04\n googlenet min = 64.78 max = 65.16 avg = 65.01\n googlenet_int8 min = 80.11 max = 80.79 avg = 80.46\n resnet18 min = 53.91 max = 54.28 avg = 54.07\n resnet18_int8 min = 63.95 max = 64.20 avg = 64.06\n alexnet min = 51.84 max = 52.17 avg = 52.00\n vgg16 min = 322.01 max = 324.34 avg = 322.72\n vgg16_int8 min = 323.83 max = 324.17 avg = 324.02\n resnet50 min = 152.66 max = 153.33 avg = 153.03\n resnet50_int8 min = 193.40 max = 194.55 avg = 194.03\n squeezenet_ssd min = 44.07 max = 44.51 avg = 44.37\n squeezenet_ssd_int8 min = 51.08 max = 52.26 avg = 51.60\n mobilenet_ssd min = 67.73 max = 68.21 avg = 67.98\n mobilenet_ssd_int8 min = 82.41 max = 82.70 avg = 82.55\n mobilenet_yolo min = 157.38 max = 159.44 avg = 158.23\n mobilenetv2_yolov3 min = 83.35 max = 83.68 avg = 83.55\n yolov4-tiny min = 107.25 max = 107.72 avg = 107.50\n nanodet_m min = 26.93 max = 27.24 avg = 27.09\n yolo-fastest-1.1 min = 12.47 max = 12.71 avg = 12.61\n yolo-fastestv2 min = 10.65 max = 10.95 avg = 10.81\n\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 54.43 max = 54.48 avg = 54.46\n squeezenet_int8 min = 79.32 max = 79.64 avg = 79.43\n mobilenet min = 105.92 max = 106.12 avg = 106.03\n mobilenet_int8 min = 152.24 max = 152.28 avg = 152.26\n mobilenet_v2 min = 62.44 max = 62.83 avg = 62.57\n mobilenet_v3 min = 49.47 max = 49.55 avg = 49.50\n shufflenet min = 27.32 max = 27.37 avg = 27.34\n shufflenet_v2 min = 29.85 max = 30.00 avg = 29.93\n mnasnet min = 59.83 max = 60.09 avg = 59.98\n proxylessnasnet min = 66.66 max = 66.84 avg = 66.76\n efficientnet_b0 min = 104.00 max = 104.19 avg = 104.08\n efficientnetv2_b0 min = 128.05 max = 128.39 avg = 128.21\n regnety_400m min = 77.95 max = 78.03 avg = 78.00\n blazeface min = 6.66 max = 6.77 avg = 6.70\n googlenet min = 195.32 max = 195.75 avg = 195.52\n googlenet_int8 min = 275.81 max = 276.25 avg = 275.98\n resnet18 min = 160.94 max = 161.17 avg = 161.03\n resnet18_int8 min = 223.88 max = 224.12 avg = 224.03\n alexnet min = 120.96 max = 121.16 avg = 121.05\n vgg16 min = 852.50 max = 853.66 avg = 853.04\n vgg16_int8 min = 1081.07 max = 1083.31 avg = 1082.18\n resnet50 min = 497.54 max = 497.85 avg = 497.67\n resnet50_int8 min = 681.79 max = 682.60 avg = 682.29\n squeezenet_ssd min = 101.81 max = 102.49 avg = 102.13\n squeezenet_ssd_int8 min = 147.77 max = 148.52 avg = 148.04\n mobilenet_ssd min = 215.63 max = 216.07 avg = 215.91\n mobilenet_ssd_int8 min = 305.65 max = 305.97 avg = 305.78\n mobilenet_yolo min = 494.99 max = 495.41 avg = 495.16\n mobilenetv2_yolov3 min = 233.51 max = 234.26 avg = 233.84\n yolov4-tiny min = 287.26 max = 287.89 avg = 287.50\n nanodet_m min = 70.48 max = 70.73 avg = 70.61\n yolo-fastest-1.1 min = 27.32 max = 27.36 avg = 27.34\n yolo-fastestv2 min = 23.51 max = 23.85 avg = 23.76\n\n[0 Intel(R) UHD Graphics (JSL)] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 Intel(R) UHD Graphics (JSL)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Intel(R) UHD Graphics (JSL)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Intel(R) UHD Graphics (JSL)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 14.71 max = 15.37 avg = 14.90\n mobilenet min = 15.38 max = 16.34 avg = 16.07\n mobilenet_v2 min = 13.58 max = 14.52 avg = 14.23\n mobilenet_v3 min = 14.95 max = 15.81 avg = 15.20\n shufflenet min = 11.93 max = 12.73 avg = 12.31\n shufflenet_v2 min = 14.47 max = 14.74 avg = 14.60\n mnasnet min = 15.32 max = 17.13 avg = 15.95\n proxylessnasnet min = 15.34 max = 16.25 avg = 15.66\n efficientnet_b0 min = 26.02 max = 26.19 avg = 26.11\n efficientnetv2_b0 min = 75.92 max = 76.18 avg = 76.07\n regnety_400m min = 17.79 max = 18.00 avg = 17.91\n blazeface min = 5.03 max = 5.96 avg = 5.65\n googlenet min = 35.20 max = 35.40 avg = 35.32\n resnet18 min = 35.49 max = 35.61 avg = 35.56\n alexnet min = 40.93 max = 41.25 avg = 41.11\n vgg16 min = 220.66 max = 222.18 avg = 221.42\n resnet50 min = 78.10 max = 78.48 avg = 78.28\n squeezenet_ssd min = 46.90 max = 47.46 avg = 47.26\n mobilenet_ssd min = 33.33 max = 33.54 avg = 33.44\n mobilenet_yolo min = 67.54 max = 67.77 avg = 67.64\n mobilenetv2_yolov3 min = 38.98 max = 39.69 avg = 39.37\n yolov4-tiny min = 68.01 max = 69.74 avg = 68.86\n nanodet_m min = 17.41 max = 18.13 avg = 17.78\n yolo-fastest-1.1 min = 13.91 max = 14.18 avg = 14.03\n yolo-fastestv2 min = 15.94 max = 16.02 avg = 15.97\n```\n\n### nVIDIA RTX2060 of Notebook\n```\nC:\\Users\\ai\\AppData\\Local\\Temp\\benchmark>benchncnn.exe 64 1 0 0 0\n[0 GeForce RTX 2060] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 GeForce RTX 2060] buglssc=0 bugihfa=0\n[0 GeForce RTX 2060] fp16p=1 fp16s=1 fp16a=1 int8s=1 int8a=1\nloop_count = 64\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 2.14 max = 2.93 avg = 2.26\n mobilenet min = 2.08 max = 2.53 avg = 2.22\n mobilenet_v2 min = 2.81 max = 4.03 avg = 3.05\n mobilenet_v3 min = 2.90 max = 3.53 avg = 3.08\n shufflenet min = 1.94 max = 4.27 avg = 2.55\n shufflenet_v2 min = 2.34 max = 2.97 avg = 2.49\n mnasnet min = 2.11 max = 2.86 avg = 2.37\n proxylessnasnet min = 2.27 max = 3.25 avg = 2.49\n googlenet min = 4.34 max = 6.79 avg = 5.25\n resnet18 min = 2.60 max = 4.36 avg = 2.90\n alexnet min = 2.79 max = 4.70 avg = 3.04\n vgg16 min = 11.40 max = 14.32 avg = 12.42\n resnet50 min = 5.26 max = 5.86 avg = 5.51\n squeezenet_ssd min = 5.58 max = 7.94 avg = 6.56\n mobilenet_ssd min = 3.47 max = 5.29 avg = 3.77\n mobilenet_yolo min = 5.49 max = 6.19 avg = 5.70\n mobilenetv2_yolov3 min = 3.69 max = 5.14 avg = 3.91\n```\n\n### nVIDIA RTX A3000 of Notebook (6GB)\n```\ncx@HP-ZBook-Fury-15-6-inch-G8-Mobile-Workstation-PC:~/ncnn/build/benchmark$ ./benchncnn 10 1 0 1\n[0 Intel(R) UHD Graphics (TGL GT1)] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 Intel(R) UHD Graphics (TGL GT1)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Intel(R) UHD Graphics (TGL GT1)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Intel(R) UHD Graphics (TGL GT1)] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 Intel(R) UHD Graphics (TGL GT1)] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\n[1 NVIDIA RTX A3000 Laptop GPU] queueC=2[8] queueG=0[16] queueT=1[2]\n[1 NVIDIA RTX A3000 Laptop GPU] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 NVIDIA RTX A3000 Laptop GPU] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 NVIDIA RTX A3000 Laptop GPU] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[1 NVIDIA RTX A3000 Laptop GPU] fp16-matrix-16_8_8/16_8_16/16_16_16=1/1/1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 1\ncooling_down = 1\n squeezenet min = 1.49 max = 1.94 avg = 1.74\n squeezenet_int8 min = 6.13 max = 6.20 avg = 6.16\n mobilenet min = 4.05 max = 4.82 avg = 4.65\n mobilenet_int8 min = 10.24 max = 10.29 avg = 10.26\n mobilenet_v2 min = 0.98 max = 1.14 avg = 1.03\n mobilenet_v3 min = 1.74 max = 1.82 avg = 1.77\n shufflenet min = 1.43 max = 30.51 avg = 9.51\n shufflenet_v2 min = 3.43 max = 3.89 avg = 3.77\n mnasnet min = 6.50 max = 6.75 avg = 6.62\n proxylessnasnet min = 6.46 max = 7.28 avg = 7.00\n efficientnet_b0 min = 3.14 max = 15.11 avg = 7.29\n efficientnetv2_b0 min = 18.50 max = 20.13 avg = 19.17\n regnety_400m min = 2.16 max = 3.57 avg = 2.70\n blazeface min = 2.52 max = 2.76 avg = 2.65\n googlenet min = 2.67 max = 14.67 avg = 9.85\n googlenet_int8 min = 19.08 max = 19.40 avg = 19.19\n resnet18 min = 5.19 max = 9.44 avg = 8.48\n resnet18_int8 min = 16.57 max = 17.69 avg = 16.96\n alexnet min = 1.98 max = 3.24 avg = 2.23\n vgg16 min = 3.59 max = 12.34 avg = 10.99\n vgg16_int8 min = 110.63 max = 124.31 avg = 118.16\n resnet50 min = 3.01 max = 4.93 avg = 3.77\n resnet50_int8 min = 41.58 max = 44.80 avg = 43.24\n squeezenet_ssd min = 4.08 max = 4.70 avg = 4.32\n squeezenet_ssd_int8 min = 17.32 max = 17.92 avg = 17.46\n mobilenet_ssd min = 2.26 max = 8.23 avg = 5.57\n mobilenet_ssd_int8 min = 20.35 max = 21.89 avg = 20.76\n mobilenet_yolo min = 2.14 max = 16.94 avg = 6.44\n mobilenetv2_yolov3 min = 3.64 max = 5.09 avg = 4.02\n yolov4-tiny min = 10.94 max = 17.46 avg = 13.58\n nanodet_m min = 6.57 max = 13.91 avg = 9.82\n yolo-fastest-1.1 min = 5.40 max = 14.22 avg = 10.78\n yolo-fastestv2 min = 7.49 max = 9.43 avg = 7.99\n vision_transformer min = 76.04 max = 76.96 avg = 76.43\n FastestDet min = 6.31 max = 6.60 avg = 6.43\n```\n\n### nVIDIA RTX2080 of Desktop\n```\nE:\\projects\\framework\\ncnn\\benchmark>benchncnn.exe 4096 1 0 0 0\n[0 GeForce RTX 2080] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 GeForce RTX 2080] buglssc=0 bugihfa=0\n[0 GeForce RTX 2080] fp16p=1 fp16s=1 fp16a=1 int8s=1 int8a=1\nloop_count = 4096\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 1.39 max = 16.70 avg = 1.49\n mobilenet min = 1.32 max = 2.55 avg = 1.42\n mobilenet_v2 min = 1.88 max = 5.02 avg = 2.00\n mobilenet_v3 min = 2.31 max = 3.58 avg = 2.45\n shufflenet min = 1.45 max = 2.65 avg = 1.55\n shufflenet_v2 min = 1.90 max = 3.21 avg = 2.03\n mnasnet min = 1.95 max = 3.17 avg = 2.09\n proxylessnasnet min = 2.02 max = 2.95 avg = 2.16\n googlenet min = 3.81 max = 5.91 avg = 4.05\n resnet18 min = 2.10 max = 3.28 avg = 2.24\n alexnet min = 2.15 max = 3.35 avg = 2.30\n vgg16 min = 7.33 max = 11.12 avg = 7.80\n resnet50 min = 4.21 max = 6.70 avg = 4.49\n squeezenet_ssd min = 4.58 max = 6.86 avg = 4.88\n mobilenet_ssd min = 2.90 max = 4.52 avg = 3.09\n mobilenet_yolo min = 4.15 max = 6.09 avg = 4.40\n mobilenetv2_yolov3 min = 3.04 max = 9.13 avg = 3.28\n```\n\n### NVIDIA Jetson AGX Xavier (Carmel 2.2 GHz x 8 + Volta Tensor Cores 64)\n```\ni@ubuntu:~/projects/ncnn/benchmark$ ./benchncnn 32 1 0 -1 0\nloop_count = 32\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 22.31 max = 23.29 avg = 22.68\n squeezenet_int8 min = 47.64 max = 52.88 avg = 49.72\n mobilenet min = 37.50 max = 38.45 avg = 37.85\n mobilenet_int8 min = 89.14 max = 92.38 avg = 90.95\n mobilenet_v2 min = 24.31 max = 25.53 avg = 24.68\n mobilenet_v3 min = 20.20 max = 21.21 avg = 20.56\n shufflenet min = 14.85 max = 15.64 avg = 15.15\n shufflenet_v2 min = 14.34 max = 16.11 avg = 14.86\n mnasnet min = 23.42 max = 23.86 avg = 23.56\n proxylessnasnet min = 27.44 max = 28.83 avg = 27.83\n efficientnet_b0 min = 34.57 max = 37.84 avg = 35.13\n efficientnetv2_b0 min = 65.16 max = 68.67 avg = 66.76\n regnety_400m min = 33.86 max = 34.49 avg = 34.17\n blazeface min = 11.86 max = 14.15 avg = 12.52\n googlenet min = 83.19 max = 89.84 avg = 85.14\n googlenet_int8 min = 146.74 max = 155.25 avg = 151.14\n resnet18 min = 50.46 max = 57.80 avg = 53.40\n resnet18_int8 min = 108.43 max = 116.14 avg = 110.78\n alexnet min = 56.59 max = 64.93 avg = 59.51\n vgg16 min = 266.78 max = 272.16 avg = 269.14\n vgg16_int8 min = 538.71 max = 551.55 avg = 544.78\n resnet50 min = 169.11 max = 172.26 avg = 170.51\n resnet50_int8 min = 370.55 max = 384.36 avg = 377.75\n squeezenet_ssd min = 58.51 max = 67.88 avg = 62.78\n squeezenet_ssd_int8 min = 95.34 max = 106.49 avg = 97.99\n mobilenet_ssd min = 83.52 max = 86.84 avg = 84.86\n mobilenet_ssd_int8 min = 172.70 max = 181.84 avg = 176.25\n mobilenet_yolo min = 165.26 max = 167.74 avg = 166.51\n mobilenetv2_yolov3 min = 88.11 max = 90.29 avg = 89.19\n yolov4-tiny min = 105.44 max = 109.24 avg = 107.07\n nanodet_m min = 33.60 max = 37.02 avg = 34.39\n yolo-fastest-1.1 min = 13.56 max = 14.22 avg = 13.75\n yolo-fastestv2 min = 13.76 max = 14.59 avg = 14.02\ni@ubuntu:~/projects/ncnn/benchmark$ ./benchncnn 32 2 0 -1 0\nloop_count = 32\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 13.05 max = 13.76 avg = 13.36\n squeezenet_int8 min = 26.08 max = 28.09 avg = 26.69\n mobilenet min = 20.61 max = 21.21 avg = 20.81\n mobilenet_int8 min = 44.72 max = 47.33 avg = 45.76\n mobilenet_v2 min = 14.67 max = 15.23 avg = 14.86\n mobilenet_v3 min = 12.59 max = 15.50 avg = 13.36\n shufflenet min = 12.74 max = 14.14 avg = 13.31\n shufflenet_v2 min = 10.05 max = 10.89 avg = 10.40\n mnasnet min = 14.02 max = 14.75 avg = 14.19\n proxylessnasnet min = 16.05 max = 16.94 avg = 16.31\n efficientnet_b0 min = 20.47 max = 23.05 avg = 20.81\n efficientnetv2_b0 min = 37.51 max = 41.53 avg = 39.19\n regnety_400m min = 25.21 max = 25.73 avg = 25.39\n blazeface min = 7.30 max = 8.44 avg = 7.43\n googlenet min = 42.52 max = 47.38 avg = 44.39\n googlenet_int8 min = 76.38 max = 81.63 avg = 77.93\n resnet18 min = 26.76 max = 28.72 avg = 27.22\n resnet18_int8 min = 55.97 max = 61.57 avg = 57.26\n alexnet min = 29.29 max = 33.20 avg = 31.03\n vgg16 min = 134.32 max = 138.65 avg = 136.05\n vgg16_int8 min = 267.70 max = 281.71 avg = 272.79\n resnet50 min = 87.22 max = 88.75 avg = 87.65\n resnet50_int8 min = 183.80 max = 192.17 avg = 187.25\n squeezenet_ssd min = 35.80 max = 39.00 avg = 37.32\n squeezenet_ssd_int8 min = 53.56 max = 60.43 avg = 55.58\n mobilenet_ssd min = 44.17 max = 48.30 avg = 44.70\n mobilenet_ssd_int8 min = 90.32 max = 94.09 avg = 92.27\n mobilenet_yolo min = 87.50 max = 89.63 avg = 88.33\n mobilenetv2_yolov3 min = 49.76 max = 51.58 avg = 50.44\n yolov4-tiny min = 61.17 max = 64.41 avg = 62.15\n nanodet_m min = 21.43 max = 22.47 avg = 21.82\n yolo-fastest-1.1 min = 10.90 max = 12.63 avg = 11.12\n yolo-fastestv2 min = 10.61 max = 11.11 avg = 10.82\ni@ubuntu:~/projects/ncnn/benchmark$ ./benchncnn 32 4 0 -1 0\nloop_count = 32\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 8.06 max = 8.79 avg = 8.39\n squeezenet_int8 min = 14.96 max = 16.64 avg = 15.37\n mobilenet min = 11.24 max = 11.91 avg = 11.48\n mobilenet_int8 min = 23.63 max = 24.75 avg = 23.81\n mobilenet_v2 min = 9.27 max = 9.97 avg = 9.44\n mobilenet_v3 min = 8.81 max = 10.06 avg = 9.07\n shufflenet min = 11.22 max = 11.53 avg = 11.37\n shufflenet_v2 min = 7.81 max = 8.17 avg = 7.97\n mnasnet min = 9.40 max = 10.49 avg = 10.06\n proxylessnasnet min = 10.53 max = 10.73 avg = 10.62\n efficientnet_b0 min = 13.55 max = 15.14 avg = 13.80\n efficientnetv2_b0 min = 19.83 max = 21.95 avg = 21.09\n regnety_400m min = 21.80 max = 22.91 avg = 22.13\n blazeface min = 5.17 max = 6.27 avg = 5.31\n googlenet min = 22.67 max = 25.35 avg = 23.10\n googlenet_int8 min = 43.19 max = 45.68 avg = 43.72\n resnet18 min = 15.19 max = 16.14 avg = 15.42\n resnet18_int8 min = 31.22 max = 34.76 avg = 31.81\n alexnet min = 15.20 max = 17.65 avg = 15.56\n vgg16 min = 70.76 max = 73.21 avg = 71.70\n vgg16_int8 min = 137.94 max = 143.50 avg = 139.54\n resnet50 min = 47.15 max = 47.91 avg = 47.40\n resnet50_int8 min = 99.80 max = 102.94 avg = 100.29\n squeezenet_ssd min = 22.10 max = 24.11 avg = 22.46\n squeezenet_ssd_int8 min = 33.21 max = 35.98 avg = 33.98\n mobilenet_ssd min = 25.09 max = 26.81 avg = 25.50\n mobilenet_ssd_int8 min = 48.15 max = 50.96 avg = 49.49\n mobilenet_yolo min = 48.63 max = 49.02 avg = 48.84\n mobilenetv2_yolov3 min = 30.93 max = 31.41 avg = 31.13\n yolov4-tiny min = 38.43 max = 41.20 avg = 39.28\n nanodet_m min = 14.95 max = 15.74 avg = 15.35\n yolo-fastest-1.1 min = 8.89 max = 9.18 avg = 9.01\n yolo-fastestv2 min = 8.36 max = 9.28 avg = 8.50\ni@ubuntu:~/projects/ncnn/benchmark$ ./benchncnn 32 8 0 -1 0\nloop_count = 32\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.52 max = 74.10 avg = 12.94\n squeezenet_int8 min = 10.44 max = 18.81 avg = 12.15\n mobilenet min = 7.49 max = 14.63 avg = 8.67\n mobilenet_int8 min = 13.80 max = 15.89 avg = 14.53\n mobilenet_v2 min = 8.15 max = 11.42 avg = 8.78\n mobilenet_v3 min = 7.60 max = 10.92 avg = 8.38\n shufflenet min = 11.51 max = 19.48 avg = 12.97\n shufflenet_v2 min = 7.06 max = 15.58 avg = 9.48\n mnasnet min = 7.77 max = 15.12 avg = 8.68\n proxylessnasnet min = 8.54 max = 42.73 avg = 10.00\n efficientnet_b0 min = 11.11 max = 12.86 avg = 11.89\n efficientnetv2_b0 min = 17.17 max = 29.03 avg = 20.48\n regnety_400m min = 22.41 max = 36.72 avg = 25.49\n blazeface min = 4.93 max = 11.62 avg = 6.13\n googlenet min = 17.02 max = 31.61 avg = 19.92\n googlenet_int8 min = 27.70 max = 35.49 avg = 29.18\n resnet18 min = 9.74 max = 18.78 avg = 11.40\n resnet18_int8 min = 18.52 max = 24.70 avg = 19.32\n alexnet min = 10.70 max = 15.41 avg = 11.39\n vgg16 min = 40.80 max = 54.47 avg = 42.72\n vgg16_int8 min = 74.71 max = 79.66 avg = 76.37\n resnet50 min = 28.21 max = 36.62 avg = 29.41\n resnet50_int8 min = 54.53 max = 76.02 avg = 56.81\n squeezenet_ssd min = 19.01 max = 30.68 avg = 24.89\n squeezenet_ssd_int8 min = 27.61 max = 35.87 avg = 29.22\n mobilenet_ssd min = 17.35 max = 22.87 avg = 18.55\n mobilenet_ssd_int8 min = 29.92 max = 36.35 avg = 31.15\n mobilenet_yolo min = 31.63 max = 55.61 avg = 34.31\n mobilenetv2_yolov3 min = 23.75 max = 35.45 avg = 25.68\n yolov4-tiny min = 29.23 max = 70.12 avg = 31.94\n nanodet_m min = 13.00 max = 21.72 avg = 15.39\n yolo-fastest-1.1 min = 9.72 max = 17.94 avg = 11.45\n yolo-fastestv2 min = 9.16 max = 16.35 avg = 11.08\ni@ubuntu:~/projects/ncnn/benchmark$ ./benchncnn 128 1 0 0 0\n[0 NVIDIA Tegra Xavier (nvgpu)] queueC=2[8] queueG=0[16] queueT=1[1]\n[0 NVIDIA Tegra Xavier (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA Tegra Xavier (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA Tegra Xavier (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 128\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 4.85 max = 19.65 avg = 6.83\n squeezenet_int8 min = 46.38 max = 49.70 avg = 47.22\n mobilenet min = 5.62 max = 6.61 avg = 6.33\n mobilenet_int8 min = 87.42 max = 92.95 avg = 90.52\n mobilenet_v2 min = 5.96 max = 7.53 avg = 6.50\n mobilenet_v3 min = 6.77 max = 7.83 avg = 7.01\n shufflenet min = 10.58 max = 18.46 avg = 13.68\n shufflenet_v2 min = 20.06 max = 21.09 avg = 20.37\n mnasnet min = 6.49 max = 26.49 avg = 8.26\n proxylessnasnet min = 6.75 max = 27.37 avg = 7.88\n efficientnet_b0 min = 12.11 max = 48.35 avg = 14.63\n efficientnetv2_b0 min = 24.61 max = 69.68 avg = 34.33\n regnety_400m min = 9.02 max = 34.40 avg = 10.84\n blazeface min = 7.55 max = 8.10 avg = 7.78\n googlenet min = 12.57 max = 65.14 avg = 18.91\n googlenet_int8 min = 145.74 max = 155.87 avg = 151.06\n resnet18 min = 8.88 max = 30.48 avg = 9.34\n resnet18_int8 min = 109.19 max = 116.78 avg = 111.52\n alexnet min = 9.06 max = 54.53 avg = 19.04\n vgg16 min = 18.12 max = 37.31 avg = 19.65\n vgg16_int8 min = 530.60 max = 551.58 avg = 542.33\n resnet50 min = 11.62 max = 20.64 avg = 12.17\n resnet50_int8 min = 374.83 max = 384.79 avg = 379.50\n squeezenet_ssd min = 14.01 max = 55.88 avg = 23.64\n squeezenet_ssd_int8 min = 89.86 max = 95.80 avg = 92.18\n mobilenet_ssd min = 13.20 max = 13.61 avg = 13.37\n mobilenet_ssd_int8 min = 170.17 max = 181.48 avg = 174.93\n mobilenet_yolo min = 11.78 max = 20.42 avg = 13.34\n mobilenetv2_yolov3 min = 18.08 max = 62.94 avg = 26.70\n yolov4-tiny min = 26.44 max = 34.83 avg = 31.83\n nanodet_m min = 7.93 max = 9.91 avg = 9.01\n yolo-fastest-1.1 min = 6.03 max = 20.85 avg = 8.42\n yolo-fastestv2 min = 9.01 max = 20.60 avg = 12.51\n```\n\n### MacBook Pro (13-inch, M1, 2020)\n```\nMacBook-Pro benchmark % ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 4.80 max = 5.05 avg = 4.86\n squeezenet_int8 min = 4.02 max = 4.13 avg = 4.04\n mobilenet min = 9.09 max = 9.41 avg = 9.22\n mobilenet_int8 min = 4.65 max = 4.76 avg = 4.70\n mobilenet_v2 min = 5.64 max = 5.83 avg = 5.73\n mobilenet_v3 min = 4.64 max = 4.85 avg = 4.76\n shufflenet min = 3.48 max = 3.63 avg = 3.56\n shufflenet_v2 min = 3.69 max = 3.81 avg = 3.73\n mnasnet min = 5.67 max = 5.94 avg = 5.77\n proxylessnasnet min = 7.03 max = 7.28 avg = 7.20\n efficientnet_b0 min = 9.13 max = 9.53 avg = 9.28\n efficientnetv2_b0 min = 17.37 max = 18.47 avg = 17.63\n regnety_400m min = 7.64 max = 8.08 avg = 7.72\n blazeface min = 1.80 max = 1.89 avg = 1.83\n googlenet min = 25.71 max = 25.90 avg = 25.81\n googlenet_int8 min = 16.89 max = 17.10 avg = 16.97\n resnet18 min = 17.16 max = 17.28 avg = 17.20\n resnet18_int8 min = 15.55 max = 15.75 avg = 15.64\n alexnet min = 30.60 max = 31.11 avg = 30.69\n vgg16 min = 73.41 max = 75.37 avg = 73.91\n vgg16_int8 min = 103.81 max = 105.15 avg = 104.19\n resnet50 min = 43.47 max = 44.24 avg = 43.68\n resnet50_int8 min = 30.37 max = 35.25 avg = 31.61\n squeezenet_ssd min = 20.97 max = 21.21 avg = 21.12\n squeezenet_ssd_int8 min = 19.34 max = 19.54 avg = 19.42\n mobilenet_ssd min = 22.18 max = 22.58 avg = 22.28\n mobilenet_ssd_int8 min = 13.27 max = 15.31 avg = 14.05\n mobilenet_yolo min = 40.78 max = 41.04 avg = 40.89\n mobilenetv2_yolov3 min = 20.87 max = 21.92 avg = 21.02\n yolov4-tiny min = 30.73 max = 32.37 avg = 31.29\n nanodet_m min = 8.54 max = 8.86 avg = 8.65\n\n\nMacBook-Pro benchmark % ./benchncnn 10 8 0 0 0\n[0 Apple M1] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 Apple M1] bugsbn1=0 bugbilz=151 bugcopc=0 bugihfa=0\n[0 Apple M1] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Apple M1] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 1.86 max = 2.22 avg = 2.01\n squeezenet_int8 min = 2.38 max = 8.40 avg = 5.13\n mobilenet min = 2.50 max = 2.91 avg = 2.64\n mobilenet_int8 min = 2.29 max = 5.26 avg = 3.54\n mobilenet_v2 min = 2.93 max = 3.12 avg = 2.98\n mobilenet_v3 min = 3.36 max = 3.61 avg = 3.48\n shufflenet min = 1.99 max = 2.54 avg = 2.18\n shufflenet_v2 min = 2.35 max = 2.84 avg = 2.52\n mnasnet min = 2.81 max = 3.33 avg = 2.92\n proxylessnasnet min = 3.21 max = 3.62 avg = 3.36\n efficientnet_b0 min = 4.74 max = 5.73 avg = 5.07\n efficientnetv2_b0 min = 12.04 max = 13.04 avg = 12.61\n regnety_400m min = 3.86 max = 4.04 avg = 3.98\n blazeface min = 0.98 max = 1.11 avg = 1.03\n googlenet min = 4.86 max = 5.38 avg = 5.02\n googlenet_int8 min = 9.43 max = 15.72 avg = 10.44\n resnet18 min = 3.92 max = 4.59 avg = 4.24\n resnet18_int8 min = 6.83 max = 7.57 avg = 7.35\n alexnet min = 7.49 max = 7.87 avg = 7.65\n vgg16 min = 34.10 max = 35.29 avg = 34.60\n vgg16_int8 min = 40.09 max = 44.66 avg = 41.95\n resnet50 min = 7.22 max = 7.83 avg = 7.42\n resnet50_int8 min = 14.52 max = 20.56 avg = 15.78\n squeezenet_ssd min = 8.52 max = 13.79 avg = 9.98\n squeezenet_ssd_int8 min = 12.38 max = 15.44 avg = 13.37\n mobilenet_ssd min = 4.83 max = 6.00 avg = 5.31\n mobilenet_ssd_int8 min = 7.26 max = 13.12 avg = 9.01\n mobilenet_yolo min = 7.22 max = 8.66 avg = 7.99\n mobilenetv2_yolov3 min = 7.46 max = 8.06 avg = 7.80\n yolov4-tiny min = 12.17 max = 13.95 avg = 12.82\n nanodet_m min = 3.54 max = 4.78 avg = 3.86\n```\n\n### MacBook Air (13-inch, M3, 2024)\n```\nMacBook-Air benchmark % ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 3.59 max = 4.20 avg = 3.80\n squeezenet_int8 min = 2.61 max = 2.82 avg = 2.74\n mobilenet min = 6.67 max = 6.92 avg = 6.85\n mobilenet_int8 min = 3.61 max = 3.66 avg = 3.62\n mobilenet_v2 min = 4.08 max = 4.15 avg = 4.10\n mobilenet_v3 min = 3.32 max = 3.44 avg = 3.34\n shufflenet min = 2.08 max = 2.13 avg = 2.10\n shufflenet_v2 min = 2.35 max = 2.44 avg = 2.37\n mnasnet min = 4.14 max = 4.23 avg = 4.18\n proxylessnasnet min = 5.09 max = 5.15 avg = 5.11\n efficientnet_b0 min = 6.67 max = 6.75 avg = 6.70\n efficientnetv2_b0 min = 8.79 max = 8.83 avg = 8.81\n regnety_400m min = 5.68 max = 5.73 avg = 5.69\n blazeface min = 0.75 max = 0.77 avg = 0.76\n googlenet min = 15.94 max = 15.97 avg = 15.96\n googlenet_int8 min = 10.88 max = 10.92 avg = 10.89\n resnet18 min = 12.60 max = 12.63 avg = 12.61\n resnet18_int8 min = 9.88 max = 9.95 avg = 9.90\n alexnet min = 12.72 max = 12.82 avg = 12.77\n vgg16 min = 57.85 max = 61.44 avg = 58.40\n vgg16_int8 min = 78.53 max = 79.85 avg = 78.83\n resnet50 min = 34.79 max = 34.85 avg = 34.81\n resnet50_int8 min = 20.56 max = 20.62 avg = 20.58\n squeezenet_ssd min = 9.64 max = 9.82 avg = 9.69\n squeezenet_ssd_int8 min = 8.21 max = 8.34 avg = 8.25\n mobilenet_ssd min = 14.21 max = 14.34 avg = 14.25\n mobilenet_ssd_int8 min = 7.35 max = 7.41 avg = 7.37\n mobilenet_yolo min = 31.61 max = 31.74 avg = 31.64\n mobilenetv2_yolov3 min = 15.79 max = 15.87 avg = 15.83\n yolov4-tiny min = 22.93 max = 22.99 avg = 22.96\n nanodet_m min = 5.58 max = 5.62 avg = 5.59\n yolo-fastest-1.1 min = 2.00 max = 2.05 avg = 2.01\n yolo-fastestv2 min = 1.75 max = 1.77 avg = 1.76\n vision_transformer min = 1020.57 max = 1046.02 avg = 1028.75\n FastestDet min = 1.88 max = 1.93 avg = 1.89\n\nMacBook-Air benchmark % ./benchncnn 10 8 0 0 0\n[0 Apple M3] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 Apple M3] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Apple M3] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[0 Apple M3] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 Apple M3] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 1.79 max = 2.48 avg = 2.16\n squeezenet_int8 min = 2.78 max = 2.93 avg = 2.80\n mobilenet min = 1.40 max = 1.85 avg = 1.68\n mobilenet_int8 min = 3.60 max = 3.67 avg = 3.61\n mobilenet_v2 min = 1.68 max = 2.28 avg = 1.97\n mobilenet_v3 min = 1.71 max = 2.29 avg = 2.00\n shufflenet min = 1.18 max = 2.49 avg = 1.78\n shufflenet_v2 min = 1.45 max = 2.09 avg = 1.70\n mnasnet min = 1.74 max = 2.25 avg = 2.05\n proxylessnasnet min = 1.75 max = 2.18 avg = 2.02\n efficientnet_b0 min = 2.71 max = 3.19 avg = 2.99\n efficientnetv2_b0 min = 6.77 max = 7.04 avg = 6.88\n regnety_400m min = 1.94 max = 2.40 avg = 2.10\n blazeface min = 1.05 max = 1.43 avg = 1.24\n googlenet min = 3.99 max = 4.42 avg = 4.27\n googlenet_int8 min = 10.83 max = 10.86 avg = 10.85\n resnet18 min = 2.50 max = 2.77 avg = 2.70\n resnet18_int8 min = 9.86 max = 9.91 avg = 9.88\n alexnet min = 2.99 max = 3.28 avg = 3.11\n vgg16 min = 12.41 max = 13.13 avg = 12.54\n vgg16_int8 min = 78.52 max = 78.67 avg = 78.61\n resnet50 min = 5.46 max = 5.52 avg = 5.49\n resnet50_int8 min = 20.57 max = 20.59 avg = 20.58\n squeezenet_ssd min = 3.86 max = 4.53 avg = 4.17\n squeezenet_ssd_int8 min = 8.20 max = 8.35 avg = 8.25\n mobilenet_ssd min = 3.19 max = 3.75 avg = 3.52\n mobilenet_ssd_int8 min = 7.35 max = 7.41 avg = 7.37\n mobilenet_yolo min = 4.77 max = 4.88 avg = 4.81\n mobilenetv2_yolov3 min = 4.28 max = 4.88 avg = 4.62\n yolov4-tiny min = 6.76 max = 7.38 avg = 7.21\n nanodet_m min = 2.92 max = 4.71 avg = 3.46\n yolo-fastest-1.1 min = 1.48 max = 2.04 avg = 1.87\n yolo-fastestv2 min = 1.41 max = 1.97 avg = 1.74\n vision_transformer min = 80.34 max = 80.66 avg = 80.44\n FastestDet min = 1.43 max = 2.04 avg = 1.73\n```\n\n### Ingenic T40XP Xburst2 Core X2 1.4Ghz (without MSA)\n```\nloop_count = 8\nnum_threads = 2\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 921.23 max = 944.03 avg = 930.71\n squeezenet_int8 min = 3280.89 max = 3404.83 avg = 3359.68\n mobilenet min = 1277.61 max = 1298.51 avg = 1284.38\n mobilenet_int8 min = 4342.67 max = 4350.21 avg = 4345.85\n mobilenet_v2 min = 780.92 max = 783.93 avg = 782.79\n mobilenet_v3 min = 650.59 max = 655.08 avg = 652.06\n shufflenet min = 352.75 max = 353.69 avg = 353.24\n shufflenet_v2 min = 362.82 max = 364.08 avg = 363.38\n mnasnet min = 790.45 max = 791.89 avg = 790.99\n proxylessnasnet min = 868.71 max = 870.47 avg = 869.17\n efficientnet_b0 min = 1491.44 max = 1492.36 avg = 1491.95\n efficientnetv2_b0 min = 2135.04 max = 2148.02 avg = 2139.99\n regnety_400m min = 1000.53 max = 1005.29 avg = 1001.81\n blazeface min = 102.72 max = 104.18 avg = 103.51\n googlenet min = 3652.89 max = 3705.40 avg = 3675.43\n googlenet_int8 min = 8067.30 max = 8070.22 avg = 8069.21\n```\n### MacBook Pro (15-inch, 2019) - 2.6GHz six cores Intel Core i7 && Radeon Pro 555X 4GB && Intel UHD Graphics 630 1536MB\n```\n\n➜ benchmark git:(master) ✗ ./benchncnn 10 1 0 -1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 14.68 max = 17.06 avg = 15.55\n squeezenet_int8 min = 51.64 max = 57.85 avg = 54.01\n mobilenet min = 20.74 max = 25.38 avg = 22.77\n mobilenet_int8 min = 66.84 max = 91.01 avg = 75.69\n mobilenet_v2 min = 14.04 max = 20.06 avg = 16.36\n mobilenet_v3 min = 11.89 max = 16.22 avg = 13.58\n shufflenet min = 13.74 max = 17.10 avg = 15.02\n shufflenet_v2 min = 12.73 max = 14.36 avg = 13.53\n mnasnet min = 11.05 max = 17.79 avg = 13.82\n proxylessnasnet min = 12.60 max = 27.38 avg = 17.55\n efficientnet_b0 min = 23.73 max = 26.82 avg = 25.45\n efficientnetv2_b0 min = 27.03 max = 33.89 avg = 30.78\n regnety_400m min = 13.81 max = 21.50 avg = 15.40\n blazeface min = 3.72 max = 4.98 avg = 4.43\n googlenet min = 65.88 max = 76.62 avg = 69.40\n googlenet_int8 min = 192.07 max = 227.85 avg = 203.81\n resnet18 min = 79.45 max = 90.41 avg = 85.32\n resnet18_int8 min = 201.71 max = 222.31 avg = 207.39\n alexnet min = 70.67 max = 80.13 avg = 74.43\n vgg16 min = 233.74 max = 261.62 avg = 250.99\n vgg16_int8 min = 1722.78 max = 1997.14 avg = 1772.71\n resnet50 min = 130.39 max = 135.31 avg = 133.27\n resnet50_int8 min = 439.69 max = 483.78 avg = 461.33\n squeezenet_ssd min = 108.54 max = 122.15 avg = 115.02\n squeezenet_ssd_int8 min = 175.58 max = 185.09 avg = 181.33\n mobilenet_ssd min = 51.89 max = 59.32 avg = 54.30\n mobilenet_ssd_int8 min = 140.15 max = 192.10 avg = 164.47\n mobilenet_yolo min = 117.37 max = 131.89 avg = 126.34\n mobilenetv2_yolov3 min = 57.57 max = 72.29 avg = 64.92\n yolov4-tiny min = 114.45 max = 123.15 avg = 116.91\n nanodet_m min = 25.65 max = 33.27 avg = 28.75\n\n➜ benchmark git:(master) ✗ ./benchncnn 10 1 0 0\n[0 AMD Radeon Pro 555X] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 AMD Radeon Pro 555X] bugsbn1=0 bugbilz=196 bugcopc=0 bugihfa=0\n[0 AMD Radeon Pro 555X] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 AMD Radeon Pro 555X] subgroup=64 basic=0 vote=0 ballot=0 shuffle=0\n[1 Intel(R) UHD Graphics 630] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 Intel(R) UHD Graphics 630] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Intel(R) UHD Graphics 630] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 Intel(R) UHD Graphics 630] subgroup=32 basic=0 vote=0 ballot=0 shuffle=0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 6.66 max = 7.30 avg = 6.91\n squeezenet_int8 min = 49.97 max = 60.92 avg = 53.86\n mobilenet min = 6.99 max = 7.48 avg = 7.17\n mobilenet_int8 min = 70.46 max = 83.20 avg = 79.33\n mobilenet_v2 min = 9.56 max = 10.87 avg = 10.34\n mobilenet_v3 min = 11.48 max = 12.20 avg = 11.94\n shufflenet min = 4.52 max = 5.25 avg = 4.96\n shufflenet_v2 min = 7.29 max = 9.65 avg = 7.99\n mnasnet min = 9.82 max = 11.88 avg = 10.62\n proxylessnasnet min = 7.85 max = 8.41 avg = 8.07\n efficientnet_b0 min = 17.34 max = 17.85 avg = 17.56\n efficientnetv2_b0 min = 21.95 max = 24.10 avg = 23.15\n regnety_400m min = 13.54 max = 14.83 avg = 14.11\n blazeface min = 3.26 max = 6.59 avg = 5.50\n googlenet min = 17.62 max = 19.47 avg = 18.27\n googlenet_int8 min = 198.88 max = 247.97 avg = 223.31\n resnet18 min = 11.10 max = 12.01 avg = 11.59\n resnet18_int8 min = 225.56 max = 259.39 avg = 238.97\n alexnet min = 17.66 max = 19.19 avg = 18.24\n vgg16 min = 53.20 max = 54.88 avg = 53.73\n vgg16_int8 min = 1747.52 max = 2130.08 avg = 1880.42\n resnet50 min = 27.38 max = 28.84 avg = 28.34\n resnet50_int8 min = 461.86 max = 579.83 avg = 528.15\n squeezenet_ssd min = 19.99 max = 20.98 avg = 20.50\n squeezenet_ssd_int8 min = 185.20 max = 209.66 avg = 196.81\n mobilenet_ssd min = 12.81 max = 14.21 avg = 13.48\n mobilenet_ssd_int8 min = 139.29 max = 168.38 avg = 148.20\n mobilenet_yolo min = 19.50 max = 20.51 avg = 19.97\n mobilenetv2_yolov3 min = 15.95 max = 19.28 avg = 16.85\n yolov4-tiny min = 21.43 max = 23.42 avg = 22.28\n nanodet_m min = 7.95 max = 9.23 avg = 8.48\n\n➜ benchmark git:(master) ✗ ./benchncnn 10 1 0 1\n[0 AMD Radeon Pro 555X] queueC=0[1] queueG=0[1] queueT=0[1]\n[0 AMD Radeon Pro 555X] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 AMD Radeon Pro 555X] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 AMD Radeon Pro 555X] subgroup=64 basic=0 vote=0 ballot=0 shuffle=0\n[1 Intel(R) UHD Graphics 630] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 Intel(R) UHD Graphics 630] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Intel(R) UHD Graphics 630] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 Intel(R) UHD Graphics 630] subgroup=32 basic=0 vote=0 ballot=0 shuffle=0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 1\ncooling_down = 1\n squeezenet min = 11.06 max = 13.22 avg = 12.09\n squeezenet_int8 min = 54.87 max = 64.55 avg = 59.84\n mobilenet min = 13.65 max = 16.70 avg = 14.81\n mobilenet_int8 min = 72.36 max = 93.58 avg = 86.40\n mobilenet_v2 min = 11.88 max = 15.90 avg = 13.47\n mobilenet_v3 min = 12.68 max = 16.16 avg = 14.56\n shufflenet min = 13.87 max = 16.68 avg = 14.93\n shufflenet_v2 min = 11.73 max = 13.65 avg = 12.87\n mnasnet min = 12.71 max = 15.56 avg = 14.22\n proxylessnasnet min = 14.03 max = 17.28 avg = 15.37\n efficientnet_b0 min = 17.50 max = 21.46 avg = 19.30\n efficientnetv2_b0 min = 35.47 max = 38.58 avg = 36.89\n regnety_400m min = 16.00 max = 19.45 avg = 17.48\n blazeface min = 6.08 max = 7.18 avg = 6.39\n googlenet min = 23.35 max = 29.68 avg = 25.77\n googlenet_int8 min = 198.49 max = 254.38 avg = 222.77\n resnet18 min = 21.85 max = 28.10 avg = 24.70\n resnet18_int8 min = 211.21 max = 279.55 avg = 222.64\n alexnet min = 24.45 max = 30.47 avg = 26.87\n vgg16 min = 115.20 max = 117.76 avg = 116.48\n vgg16_int8 min = 1715.92 max = 1960.02 avg = 1800.21\n resnet50 min = 45.65 max = 46.25 avg = 46.05\n resnet50_int8 min = 448.13 max = 555.53 avg = 485.47\n squeezenet_ssd min = 28.43 max = 33.26 avg = 29.85\n squeezenet_ssd_int8 min = 180.91 max = 202.51 avg = 190.84\n mobilenet_ssd min = 21.03 max = 26.93 avg = 23.48\n mobilenet_ssd_int8 min = 154.41 max = 184.64 avg = 165.04\n mobilenet_yolo min = 37.04 max = 38.64 avg = 37.52\n mobilenetv2_yolov3 min = 24.98 max = 30.03 avg = 27.70\n yolov4-tiny min = 39.29 max = 50.25 avg = 44.18\n nanodet_m min = 15.97 max = 20.27 avg = 17.93\n```\n\n### Sunway SW421 (sw_64 1.7GHz * 4)\n```\nroot@SW421:~/Desktop/ncnn-20220420/ncnn-20220420/build/benchmark$ ./benchncnn\nloop_count = 4\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 943.61 max = 966.98 avg = 955.24\n squeezenet_int8 min = 654.75 max = 731.28 avg = 674.87\n mobilenet min = 1584.87 max = 1612.88 avg = 1597.47\n mobilenet_int8 min = 1198.21 max = 1204.82 avg = 1201.61\n mobilenet_v2 min = 733.94 max = 754.79 avg = 744.48\n mobilenet_v3 min = 665.26 max = 683.81 avg = 675.18\n shufflenet min = 401.53 max = 435.21 avg = 420.32\n shufflenet_v2 min = 294.65 max = 316.50 avg = 309.08\n mnasnet min = 671.22 max = 808.46 avg = 713.01\n proxylessnasnet min = 686.12 max = 698.13 avg = 692.29\n efficientnet_b0 min = 1151.75 max = 1184.86 avg = 1161.33\n efficientnetv2_b0 min = 1372.05 max = 1395.22 avg = 1379.47\n regnety_400m min = 933.93 max = 949.42 avg = 942.43\n blazeface min = 104.72 max = 136.77 avg = 112.86\n googlenet min = 2574.02 max = 4330.81 avg = 3015.56\n googlenet_int8 min = 2136.42 max = 2183.61 avg = 2166.45\n resnet18 min = 2511.12 max = 2537.42 avg = 2526.08\n resnet18_int8 min = 2003.84 max = 2027.50 avg = 2012.48\n alexnet min = 668.28 max = 686.35 avg = 673.95\n vgg16 min = 24863.92 max = 24967.94 avg = 24907.39\n vgg16_int8 min = 18735.54 max = 18926.83 avg = 18859.32\n resnet50 min = 9896.47 max = 9981.13 avg = 9929.77\n resnet50_int8 min = 6971.01 max = 7085.29 avg = 7017.88\n squeezenet_ssd min = 1798.23 max = 1814.25 avg = 1806.57\n squeezenet_ssd_int8 min = 1586.11 max = 1606.83 avg = 1596.75\n mobilenet_ssd min = 3995.54 max = 4018.27 avg = 4002.78\n mobilenet_ssd_int8 min = 2753.65 max = 2766.06 avg = 2760.04\n mobilenet_yolo min = 10892.22 max = 10978.84 avg = 10921.00\n mobilenetv2_yolov3 min = 3600.80 max = 3607.72 avg = 3603.18\n yolov4-tiny min = 5565.82 max = 5582.22 avg = 5571.78\n nanodet_m min = 1182.97 max = 1220.47 avg = 1199.30\n yolo-fastest-1.1 min = 340.63 max = 360.95 avg = 349.15\n yolo-fastestv2 min = 255.47 max = 281.79 avg = 268.82\n```\n\n### Sunway SW831 (sw_64 2.5GHz * 8)\n```\nroot@SW831:~/Desktop/ncnn_20221128/build/benchmark$ ./benchncnn 5 8 2 -1 0\nloop_count = 5\nnum_threads = 8\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 343.27 max = 420.86 avg = 364.97\n squeezenet_int8 min = 237.91 max = 251.71 avg = 243.84\n mobilenet min = 607.80 max = 696.04 avg = 646.61\n mobilenet_int8 min = 428.37 max = 499.32 avg = 460.21\n mobilenet_v2 min = 291.29 max = 381.93 avg = 311.76\n mobilenet_v3 min = 262.01 max = 287.93 avg = 277.29\n shufflenet min = 144.89 max = 169.10 avg = 150.84\n shufflenet_v2 min = 121.44 max = 139.62 avg = 126.96\n mnasnet min = 265.59 max = 353.84 avg = 288.79\n proxylessnasnet min = 272.08 max = 293.19 avg = 284.61\n efficientnet_b0 min = 445.40 max = 508.36 avg = 467.84\n efficientnetv2_b0 min = 550.57 max = 619.16 avg = 581.85\n regnety_400m min = 374.02 max = 460.64 avg = 394.49\n blazeface min = 39.93 max = 59.19 avg = 44.14\n googlenet min = 941.35 max = 1014.23 avg = 976.37\n googlenet_int8 min = 770.66 max = 827.44 avg = 797.93\n resnet18 min = 815.02 max = 895.13 avg = 843.57\n resnet18_int8 min = 701.10 max = 776.40 avg = 729.49\n alexnet min = 216.74 max = 273.39 avg = 228.99\n vgg16 min = 8645.55 max = 8699.60 avg = 8681.61\n vgg16_int8 min = 6786.91 max = 6930.90 avg = 6854.29\n resnet50 min = 3624.02 max = 3698.91 avg = 3652.31\n resnet50_int8 min = 2537.92 max = 2618.10 avg = 2567.88\n squeezenet_ssd min = 635.25 max = 693.23 avg = 663.56\n squeezenet_ssd_int8 min = 577.37 max = 641.12 avg = 603.34\n mobilenet_ssd min = 1529.35 max = 1711.54 avg = 1582.10\n mobilenet_ssd_int8 min = 982.65 max = 1042.82 avg = 1016.62\n mobilenet_yolo min = 4053.62 max = 4124.84 avg = 4094.38\n mobilenetv2_yolov3 min = 1367.81 max = 1527.79 avg = 1433.04\n yolov4-tiny min = 1943.20 max = 2028.02 avg = 1978.31\n nanodet_m min = 433.66 max = 498.83 avg = 457.77\n yolo-fastest-1.1 min = 140.07 max = 284.35 avg = 192.46\n yolo-fastestv2 min = 123.91 max = 225.70 avg = 152.54\n vision_transformer min = 2470.70 max = 2509.73 avg = 2486.40\n FastestDet min = 145.30 max = 163.43 avg = 154.35\n```\n\n### AXERA AX620A (Cortex-A7 1.0GHz * 4)\n```\n/root/axera # ./benchncnn 4 1 0 -1 0\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 530.57 max = 533.11 avg = 532.22\n squeezenet_int8 min = 359.74 max = 360.02 avg = 359.86\n mobilenet min = 920.12 max = 921.04 avg = 920.52\n mobilenet_int8 min = 532.60 max = 533.08 avg = 532.81\n mobilenet_v2 min = 608.81 max = 609.49 avg = 609.18\n mobilenet_v3 min = 531.43 max = 532.34 avg = 531.90\n shufflenet min = 297.91 max = 300.08 avg = 299.06\n shufflenet_v2 min = 288.44 max = 289.30 avg = 288.79\n mnasnet min = 590.29 max = 590.99 avg = 590.63\n proxylessnasnet min = 678.22 max = 679.22 avg = 678.63\n efficientnet_b0 min = 1041.41 max = 1043.79 avg = 1042.61\n efficientnetv2_b0 min = 1222.41 max = 1223.63 avg = 1222.91\n regnety_400m min = 723.83 max = 725.37 avg = 724.64\n blazeface min = 86.77 max = 87.21 avg = 86.92\n googlenet min = 1740.32 max = 1741.44 avg = 1740.81\n googlenet_int8 min = 1167.95 max = 1169.18 avg = 1168.54\n resnet18 min = 1584.41 max = 1585.36 avg = 1584.97\n resnet18_int8 min = 915.78 max = 918.77 avg = 917.16\n alexnet min = 1811.30 max = 1812.86 avg = 1812.07\n resnet50 min = 4516.48 max = 4523.48 avg = 4519.03\n resnet50_int8 min = 2573.18 max = 2574.29 avg = 2573.69\n squeezenet_ssd min = 1191.79 max = 1193.71 avg = 1193.02\n squeezenet_ssd_int8 min = 862.36 max = 863.69 avg = 862.83\n mobilenet_ssd min = 1950.48 max = 1950.98 avg = 1950.65\n mobilenet_ssd_int8 min = 1081.70 max = 1082.64 avg = 1082.20\n mobilenet_yolo min = 4629.22 max = 4630.23 avg = 4629.69\n mobilenetv2_yolov3 min = 2233.05 max = 2234.14 avg = 2233.42\n yolov4-tiny min = 2942.58 max = 2946.55 avg = 2944.81\n nanodet_m min = 692.19 max = 693.36 avg = 692.79\n yolo-fastest-1.1 min = 333.62 max = 334.43 avg = 334.00\n yolo-fastestv2 min = 256.41 max = 257.32 avg = 256.83\n\n\n/root/axera # ./benchncnn 4 4 0 -1 0\nloop_count = 4\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 150.38 max = 179.83 avg = 157.90\n squeezenet_int8 min = 106.97 max = 107.43 avg = 107.22\n mobilenet min = 248.92 max = 273.98 avg = 255.72\n mobilenet_int8 min = 139.49 max = 139.65 avg = 139.60\n mobilenet_v2 min = 174.67 max = 204.35 avg = 182.30\n mobilenet_v3 min = 152.17 max = 152.54 avg = 152.30\n shufflenet min = 98.74 max = 125.99 avg = 105.74\n shufflenet_v2 min = 103.44 max = 103.88 avg = 103.65\n mnasnet min = 167.63 max = 197.54 avg = 175.28\n proxylessnasnet min = 186.02 max = 186.32 avg = 186.15\n efficientnet_b0 min = 284.35 max = 318.17 avg = 292.90\n efficientnetv2_b0 min = 329.56 max = 359.71 avg = 337.22\n regnety_400m min = 246.91 max = 277.08 avg = 254.71\n blazeface min = 30.95 max = 31.31 avg = 31.16\n googlenet min = 474.87 max = 504.38 avg = 489.43\n googlenet_int8 min = 322.06 max = 331.97 avg = 324.57\n resnet18 min = 440.03 max = 475.28 avg = 456.70\n resnet18_int8 min = 252.01 max = 280.64 avg = 259.22\n alexnet min = 453.16 max = 478.80 avg = 465.88\n resnet50 min = 1214.70 max = 1252.42 avg = 1229.22\n resnet50_int8 min = 684.53 max = 715.65 avg = 706.14\n squeezenet_ssd min = 358.84 max = 393.45 avg = 367.77\n squeezenet_ssd_int8 min = 281.56 max = 312.86 avg = 289.85\n mobilenet_ssd min = 519.11 max = 559.14 avg = 538.41\n mobilenet_ssd_int8 min = 284.58 max = 310.02 avg = 291.02\n mobilenet_yolo min = 1238.87 max = 1284.74 avg = 1260.51\n mobilenetv2_yolov3 min = 624.42 max = 665.81 avg = 642.15\n yolov4-tiny min = 826.46 max = 852.97 avg = 844.88\n nanodet_m min = 246.76 max = 279.09 avg = 255.04\n yolo-fastest-1.1 min = 116.12 max = 116.95 avg = 116.50\n yolo-fastestv2 min = 91.08 max = 102.93 avg = 94.41\n```\n\n### AMD Ryzen 5700g (Zen3 3.8 GHz ~ 4.6 GHz x 8)\ntest in wsl2 with ubuntu 20.04\n```\n$ ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.53 max = 7.05 avg = 6.77\n squeezenet_int8 min = 17.72 max = 17.86 avg = 17.79\n mobilenet min = 11.43 max = 11.98 avg = 11.64\n mobilenet_int8 min = 22.91 max = 24.48 avg = 23.26\n mobilenet_v2 min = 8.28 max = 9.29 avg = 8.66\n mobilenet_v3 min = 6.86 max = 6.98 avg = 6.94\n shufflenet min = 3.75 max = 4.64 avg = 3.91\n shufflenet_v2 min = 5.08 max = 5.80 avg = 5.22\n mnasnet min = 7.54 max = 8.60 avg = 7.81\n proxylessnasnet min = 9.18 max = 10.33 avg = 9.41\n efficientnet_b0 min = 22.57 max = 23.67 avg = 22.93\n efficientnetv2_b0 min = 21.23 max = 22.08 avg = 21.45\n regnety_400m min = 10.56 max = 10.80 avg = 10.63\n blazeface min = 1.08 max = 1.17 avg = 1.11\n googlenet min = 27.91 max = 29.51 avg = 28.28\n googlenet_int8 min = 71.00 max = 86.86 avg = 72.74\n resnet18 min = 20.11 max = 20.56 avg = 20.26\n resnet18_int8 min = 63.80 max = 65.13 avg = 64.19\n alexnet min = 20.64 max = 24.25 avg = 21.65\n vgg16 min = 119.99 max = 125.45 avg = 121.59\n vgg16_int8 min = 268.11 max = 270.41 avg = 269.15\n resnet50 min = 55.42 max = 56.29 avg = 55.70\n resnet50_int8 min = 126.73 max = 132.37 avg = 128.72\n squeezenet_ssd min = 28.41 max = 30.30 avg = 29.20\n squeezenet_ssd_int8 min = 41.12 max = 42.53 avg = 41.52\n mobilenet_ssd min = 24.15 max = 24.91 avg = 24.33\n mobilenet_ssd_int8 min = 46.06 max = 59.19 avg = 49.87\n mobilenet_yolo min = 67.58 max = 73.19 avg = 68.99\n mobilenetv2_yolov3 min = 29.44 max = 30.46 avg = 29.78\n yolov4-tiny min = 41.89 max = 43.47 avg = 42.37\n nanodet_m min = 11.23 max = 11.47 avg = 11.36\n yolo-fastest-1.1 min = 3.86 max = 4.64 avg = 4.04\n yolo-fastestv2 min = 3.43 max = 3.99 avg = 3.56\n vision_transformer min = 1590.86 max = 1593.97 avg = 1591.91\n\n\n$ ./benchncnn 10 16 0 -1 0\nloop_count = 10\nnum_threads = 16\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.94 max = 4.66 avg = 3.31\n squeezenet_int8 min = 3.53 max = 5.26 avg = 3.92\n mobilenet min = 3.96 max = 5.30 avg = 4.21\n mobilenet_int8 min = 4.27 max = 4.56 avg = 4.35\n mobilenet_v2 min = 3.63 max = 4.20 avg = 3.82\n mobilenet_v3 min = 3.25 max = 4.79 avg = 3.58\n shufflenet min = 2.98 max = 3.59 avg = 3.12\n shufflenet_v2 min = 2.62 max = 5.93 avg = 3.04\n mnasnet min = 3.09 max = 3.49 avg = 3.28\n proxylessnasnet min = 3.57 max = 4.18 avg = 3.76\n efficientnet_b0 min = 5.98 max = 6.48 avg = 6.18\n efficientnetv2_b0 min = 6.96 max = 7.48 avg = 7.13\n regnety_400m min = 8.71 max = 11.89 avg = 9.61\n blazeface min = 0.86 max = 0.96 avg = 0.89\n googlenet min = 10.75 max = 11.33 avg = 11.00\n googlenet_int8 min = 12.75 max = 15.47 avg = 13.50\n resnet18 min = 8.92 max = 16.08 avg = 10.08\n resnet18_int8 min = 10.55 max = 10.99 avg = 10.69\n alexnet min = 9.95 max = 10.45 avg = 10.17\n vgg16 min = 52.28 max = 53.69 avg = 52.89\n vgg16_int8 min = 44.90 max = 47.90 avg = 45.61\n resnet50 min = 17.80 max = 21.43 avg = 18.66\n resnet50_int8 min = 21.80 max = 25.42 avg = 22.75\n squeezenet_ssd min = 14.49 max = 16.36 avg = 14.90\n squeezenet_ssd_int8 min = 10.02 max = 10.49 avg = 10.28\n mobilenet_ssd min = 7.20 max = 7.86 avg = 7.51\n mobilenet_ssd_int8 min = 8.51 max = 10.90 avg = 9.09\n mobilenet_yolo min = 35.67 max = 44.84 avg = 37.33\n mobilenetv2_yolov3 min = 12.72 max = 17.16 avg = 13.67\n yolov4-tiny min = 20.81 max = 22.11 avg = 21.33\n nanodet_m min = 5.13 max = 42.12 avg = 9.07\n yolo-fastest-1.1 min = 3.05 max = 4.72 avg = 3.39\n yolo-fastestv2 min = 3.33 max = 3.73 avg = 3.44\n vision_transformer min = 214.91 max = 229.91 avg = 220.82\n```\n\n### Intel Celeron M 420 (Yonah 1.60 GHz x 1)\n\nTested on `Debian GNU/Linux 11 (bullseye) i686` with `cmake -DNCNN_RUNTIME_CPU=OFF -DNCNN_AVX=OFF -DNCNN_AVX2=OFF -DNCNN_AVX512=OFF-DNCNN_BUILD_TESTS=ON ..`.\n\n```\nmouri@Mouri-Laptop-2:~/ncnn/benchmark$ ./../build/benchmark/benchncnn\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 289.23 max = 301.83 avg = 292.90\n squeezenet_int8 min = 442.82 max = 457.21 avg = 446.89\n mobilenet min = 549.62 max = 561.20 avg = 554.78\n mobilenet_int8 min = 823.92 max = 837.70 avg = 830.52\n mobilenet_v2 min = 341.72 max = 353.77 avg = 345.34\n mobilenet_v3 min = 267.68 max = 282.08 avg = 273.10\n shufflenet min = 151.66 max = 153.02 avg = 152.24\n shufflenet_v2 min = 161.54 max = 163.38 avg = 162.13\n mnasnet min = 322.66 max = 336.91 avg = 326.86\n proxylessnasnet min = 356.63 max = 368.79 avg = 360.66\n efficientnet_b0 min = 489.92 max = 505.11 avg = 497.32\n efficientnetv2_b0 min = 618.16 max = 632.02 avg = 622.82\n regnety_400m min = 414.83 max = 428.42 avg = 419.28\n blazeface min = 38.56 max = 40.05 avg = 39.05\n googlenet min = 1022.54 max = 1037.53 avg = 1029.48\n googlenet_int8 min = 1493.35 max = 1495.46 avg = 1494.31\n resnet18 min = 803.32 max = 818.27 avg = 812.49\n resnet18_int8 min = 1188.26 max = 1200.88 avg = 1192.56\n alexnet min = 613.78 max = 623.88 avg = 619.99\n vgg16 min = 4465.44 max = 4478.12 avg = 4474.16\n vgg16_int8 min = 6042.40 max = 6114.37 avg = 6077.07\n resnet50 min = 2517.75 max = 2528.42 avg = 2522.83\n resnet50_int8 min = 3746.28 max = 3771.09 avg = 3756.88\n squeezenet_ssd min = 585.56 max = 636.01 avg = 602.62\n squeezenet_ssd_int8 min = 822.43 max = 968.77 avg = 862.33\n mobilenet_ssd min = 1116.98 max = 1139.17 avg = 1127.65\n mobilenet_ssd_int8 min = 1665.03 max = 1670.55 avg = 1668.37\n mobilenet_yolo min = 2638.61 max = 2666.54 avg = 2652.26\n mobilenetv2_yolov3 min = 1248.56 max = 1255.98 avg = 1251.22\n yolov4-tiny min = 1507.31 max = 1525.56 avg = 1514.66\n nanodet_m min = 386.41 max = 400.63 avg = 391.21\n yolo-fastest-1.1 min = 159.97 max = 164.53 avg = 161.41\n yolo-fastestv2 min = 134.29 max = 135.47 avg = 134.70\n vision_transformer min = 22201.32 max = 22510.75 avg = 22315.09\n FastestDet min = 146.94 max = 148.50 avg = 147.44\n```\n### VisionFive2 , JH7110 (SiFive-U74(RV64GC) 1.5GHz x 4) riscv64 with PowerVR B-Series BXE-4-32\nTest on Debian 11 with g++ 12.2.0 and vulkan 1.3.231\n```\nuser@starfive:~/Downloads/ncnn-master/benchmark$ ./benchncnn 10 4 0 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 149.06 max = 149.33 avg = 149.17\n squeezenet_int8 min = 1318.66 max = 1349.04 avg = 1328.87\n mobilenet min = 255.13 max = 255.71 avg = 255.39\n mobilenet_int8 min = 2025.40 max = 2036.00 avg = 2031.67\n mobilenet_v2 min = 173.92 max = 174.60 avg = 174.31\n mobilenet_v3 min = 166.58 max = 167.30 avg = 167.02\n shufflenet min = 91.36 max = 91.72 avg = 91.57\n shufflenet_v2 min = 83.50 max = 83.95 avg = 83.76\n mnasnet min = 190.42 max = 191.15 avg = 190.66\n proxylessnasnet min = 226.35 max = 226.81 avg = 226.52\n efficientnet_b0 min = 342.74 max = 343.62 avg = 343.15\n efficientnetv2_b0 min = 343.31 max = 344.23 avg = 343.80\n regnety_400m min = 227.04 max = 227.75 avg = 227.43\n blazeface min = 26.18 max = 26.43 avg = 26.28\n googlenet min = 506.76 max = 508.58 avg = 507.84\n googlenet_int8 min = 3827.36 max = 3856.05 avg = 3835.67\n resnet18 min = 401.12 max = 402.27 avg = 401.61\n resnet18_int8 min = 4053.06 max = 4069.98 avg = 4061.63\n alexnet min = 297.81 max = 320.09 avg = 301.39\n vgg16 min = 2338.76 max = 2351.23 avg = 2346.19\n vgg16_int8 min = 36846.41 max = 36929.56 avg = 36886.26\n resnet50 min = 1189.88 max = 1211.10 avg = 1193.34\n resnet50_int8 min = 11819.59 max = 11884.94 avg = 11845.22\n squeezenet_ssd min = 351.71 max = 352.73 avg = 352.30\n squeezenet_ssd_int8 min = 2872.00 max = 2903.35 avg = 2891.01\n mobilenet_ssd min = 530.92 max = 531.73 avg = 531.28\n mobilenet_ssd_int8 min = 4511.56 max = 4553.41 avg = 4523.51\n mobilenet_yolo min = 1357.14 max = 1359.82 avg = 1358.83\n mobilenetv2_yolov3 min = 621.15 max = 622.29 avg = 621.66\n yolov4-tiny min = 803.06 max = 809.19 avg = 805.79\n nanodet_m min = 220.82 max = 221.18 avg = 221.06\n yolo-fastest-1.1 min = 102.59 max = 103.98 avg = 102.93\n yolo-fastestv2 min = 89.61 max = 90.03 avg = 89.76\n vision_transformer min = 15862.96 max = 15897.17 avg = 15878.22\n FastestDet min = 108.69 max = 109.00 avg = 108.84\n\nuser@starfive:~/Downloads/ncnn-master/benchmark$ ./benchncnn 10 4 1 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 148.62 max = 148.95 avg = 148.82\n squeezenet_int8 min = 1324.10 max = 1339.58 avg = 1332.57\n mobilenet min = 255.67 max = 256.20 avg = 255.93\n mobilenet_int8 min = 2024.72 max = 2028.23 avg = 2026.29\n mobilenet_v2 min = 173.76 max = 174.73 avg = 174.31\n mobilenet_v3 min = 166.66 max = 167.28 avg = 166.99\n shufflenet min = 91.18 max = 91.68 avg = 91.46\n shufflenet_v2 min = 83.88 max = 84.84 avg = 84.26\n mnasnet min = 190.23 max = 190.84 avg = 190.45\n proxylessnasnet min = 226.02 max = 226.82 avg = 226.38\n efficientnet_b0 min = 342.95 max = 343.52 avg = 343.25\n efficientnetv2_b0 min = 343.07 max = 343.80 avg = 343.39\n regnety_400m min = 226.96 max = 227.62 avg = 227.24\n blazeface min = 26.08 max = 26.32 avg = 26.18\n googlenet min = 508.30 max = 510.34 avg = 509.27\n googlenet_int8 min = 3825.65 max = 3858.90 avg = 3833.79\n resnet18 min = 400.69 max = 403.18 avg = 401.74\n resnet18_int8 min = 4055.41 max = 4123.79 avg = 4067.55\n alexnet min = 296.35 max = 300.46 avg = 299.11\n vgg16 min = 2337.68 max = 2349.78 avg = 2344.77\n vgg16_int8 min = 36760.47 max = 36985.40 avg = 36918.31\n resnet50 min = 1190.13 max = 1221.98 avg = 1196.77\n resnet50_int8 min = 11816.03 max = 11869.41 avg = 11843.72\n squeezenet_ssd min = 351.24 max = 352.20 avg = 351.89\n squeezenet_ssd_int8 min = 2873.40 max = 2902.55 avg = 2891.58\n mobilenet_ssd min = 530.45 max = 531.85 avg = 530.91\n mobilenet_ssd_int8 min = 4504.87 max = 4564.64 avg = 4528.56\n mobilenet_yolo min = 1357.83 max = 1360.48 avg = 1358.75\n mobilenetv2_yolov3 min = 621.00 max = 621.76 avg = 621.35\n yolov4-tiny min = 803.54 max = 808.00 avg = 806.16\n nanodet_m min = 221.08 max = 222.57 avg = 221.72\n yolo-fastest-1.1 min = 102.79 max = 103.15 avg = 102.95\n yolo-fastestv2 min = 89.56 max = 89.79 avg = 89.70\n vision_transformer min = 15874.12 max = 15907.97 avg = 15883.26\n FastestDet min = 108.22 max = 108.64 avg = 108.36\n\nuser@starfive:~/Downloads/ncnn-master/benchmark$ ./benchncnn 10 1 1 0 0\n[0 PowerVR B-Series BXE-4-32] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 PowerVR B-Series BXE-4-32] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 PowerVR B-Series BXE-4-32] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 PowerVR B-Series BXE-4-32] subgroup=1 basic/vote/ballot/shuffle=1/1/1/1\n[0 PowerVR B-Series BXE-4-32] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 1\npowersave = 1\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 355.26 max = 356.42 avg = 355.75\n squeezenet_int8 min = 5171.49 max = 5187.42 avg = 5178.45\n mobilenet min = 757.04 max = 762.74 avg = 759.77\n mobilenet_int8 min = 7695.03 max = 7715.39 avg = 7705.16\n mobilenet_v2 min = 476.20 max = 477.19 avg = 476.94\n mobilenet_v3 min = 403.12 max = 405.44 avg = 405.09\n shufflenet min = 181.02 max = 182.32 avg = 181.96\n shufflenet_v2 min = 257.29 max = 259.06 avg = 258.57\n mnasnet min = 495.78 max = 497.44 avg = 496.89\n proxylessnasnet min = 562.60 max = 563.02 avg = 562.83\n efficientnet_b0 min = 660.29 max = 664.73 avg = 662.97\n efficientnetv2_b0 min = 856.88 max = 864.96 avg = 861.30\n regnety_400m min = 492.79 max = 495.44 avg = 494.51\n blazeface min = 65.95 max = 68.72 avg = 68.19\n googlenet min = 1132.70 max = 1134.65 avg = 1133.50\n googlenet_int8 min = 14978.60 max = 15000.89 avg = 14988.56\n resnet18 min = 1155.15 max = 1172.06 avg = 1160.64\n resnet18_int8 min = 15776.36 max = 15790.48 avg = 15782.76\n alexnet min = 601.09 max = 606.63 avg = 603.81\n vgg16 min = 5558.47 max = 5613.23 avg = 5586.98\n vgg16_int8 min = 143936.04 max = 144068.45 avg = 143991.58\n resnet50 min = 3425.81 max = 3440.51 avg = 3434.73\n resnet50_int8 min = 44780.92 max = 45144.97 avg = 45038.46\n squeezenet_ssd min = 967.46 max = 978.39 avg = 972.76\n squeezenet_ssd_int8 min = 10842.39 max = 10999.00 avg = 10940.15\n mobilenet_ssd min = 1565.15 max = 1570.11 avg = 1568.87\n mobilenet_ssd_int8 min = 17317.40 max = 17386.46 avg = 17361.80\n mobilenet_yolo min = 3559.36 max = 3570.38 avg = 3568.84\n mobilenetv2_yolov3 min = 1731.98 max = 1739.52 avg = 1735.33\n yolov4-tiny min = 1984.22 max = 2001.65 avg = 1993.20\n nanodet_m min = 603.06 max = 609.65 avg = 607.79\n yolo-fastest-1.1 min = 306.30 max = 312.33 avg = 310.63\n yolo-fastestv2 min = 201.45 max = 207.44 avg = 205.93\n vision_transformer min = 27310.74 max = 27358.54 avg = 27327.23\n FastestDet min = 245.07 max = 248.81 avg = 248.14\n```\n### T-Head TH1520 (C910V, 1.848 GHz x 4 + BXM-4-64 PowerVR)\n\nTested on `Linux anolis-riscv 5.10.112-00579-g8e3db308d5a5 #23 SMP PREEMPT Fri Aug 12 10:17:32 CST 2022 riscv64 riscv64 riscv64 GNU/Linux`\n\n```\n[root@anolis-riscv benchmark]# ./benchncnn\nsyscall error -1\nloop_count = 4\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 187.88 max = 188.82 avg = 188.13\n squeezenet_int8 min = 2388.26 max = 2446.92 avg = 2411.46\n mobilenet min = 321.46 max = 323.34 avg = 322.19\n mobilenet_int8 min = 2318.93 max = 2458.55 avg = 2400.99\n mobilenet_v2 min = 214.01 max = 216.00 avg = 215.35\n mobilenet_v3 min = 247.71 max = 248.18 avg = 247.96\n shufflenet min = 155.58 max = 155.85 avg = 155.67\n shufflenet_v2 min = 99.50 max = 99.75 avg = 99.63\n mnasnet min = 261.46 max = 263.83 avg = 262.53\n proxylessnasnet min = 315.40 max = 316.89 avg = 316.28\n efficientnet_b0 min = 484.97 max = 486.16 avg = 485.55\n efficientnetv2_b0 min = 453.03 max = 453.40 avg = 453.21\n regnety_400m min = 314.09 max = 315.33 avg = 314.77\n blazeface min = 46.14 max = 46.69 avg = 46.39\n googlenet min = 650.99 max = 653.60 avg = 651.69\n googlenet_int8 min = 5435.11 max = 6391.98 avg = 6012.81\n resnet18 min = 505.48 max = 506.70 avg = 506.06\n resnet18_int8 min = 5053.33 max = 6599.94 avg = 6001.86\n alexnet min = 403.68 max = 404.60 avg = 404.23\n vgg16 min = 2731.55 max = 2746.48 avg = 2738.82\n```\n\ntest on `Beaglev-ahead(Linux ahead 5.10.113-ahead #2023.08.02.13.12+2c2096a98 SMP PREEMPT Wed Aug 2 13:13:02 UTC 2 riscv64 GNU/Linux)`\n\n```\ndebian@ahead:~/ncnn/build/benchmark$ sudo ./benchncnn 10 1 0 0 0\n[0 PowerVR B-Series BXM-4-64] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 PowerVR B-Series BXM-4-64] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 PowerVR B-Series BXM-4-64] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 PowerVR B-Series BXM-4-64] subgroup=1 basic/vote/ballot/shuffle=1/1/1/1\n[0 PowerVR B-Series BXM-4-64] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 287.88 max = 296.84 avg = 295.68\n squeezenet_int8 min = 2289.46 max = 2320.97 avg = 2306.60\n mobilenet min = 584.32 max = 588.48 avg = 587.41\n mobilenet_int8 min = 2487.91 max = 2492.12 avg = 2489.64\n mobilenet_v2 min = 380.02 max = 386.67 avg = 385.75\n mobilenet_v3 min = 314.73 max = 328.84 avg = 325.76\n shufflenet min = 146.96 max = 158.29 avg = 156.38\n shufflenet_v2 min = 203.94 max = 211.77 avg = 210.82\n mnasnet min = 395.80 max = 404.95 avg = 403.80\n proxylessnasnet min = 447.74 max = 456.89 avg = 454.87\n efficientnet_b0 min = 532.23 max = 543.05 avg = 538.53\n efficientnetv2_b0 min = 659.43 max = 681.64 avg = 669.13\n regnety_400m min = 393.16 max = 407.27 avg = 403.81\n blazeface min = 50.41 max = 61.83 avg = 56.92\n googlenet min = 890.79 max = 898.09 avg = 896.25\n googlenet_int8 min = 4713.76 max = 5296.61 avg = 5044.39\n resnet18 min = 814.16 max = 824.53 avg = 820.35\n resnet18_int8 min = 4800.73 max = 6015.34 avg = 5765.47\n alexnet min = 453.80 max = 465.51 avg = 462.11\n vgg16 min = 4016.26 max = 4027.30 avg = 4021.94\n vgg16_int8 min = 55069.69 max = 64814.86 avg = 59096.20\n resnet50 min = 2494.42 max = 2502.38 avg = 2500.28\n resnet50_int8 min = 15366.90 max = 17179.36 avg = 16701.20\n squeezenet_ssd min = 724.36 max = 738.28 avg = 730.44\n squeezenet_ssd_int8 min = 4550.62 max = 5235.87 avg = 4684.19\n mobilenet_ssd min = 1207.04 max = 1218.80 avg = 1212.86\n mobilenet_ssd_int8 min = 6019.61 max = 6349.35 avg = 6184.49\n mobilenet_yolo min = 2736.28 max = 2747.06 avg = 2743.21\n mobilenetv2_yolov3 min = 1339.16 max = 1349.46 avg = 1344.81\n yolov4-tiny min = 1457.05 max = 1459.04 avg = 1457.81\n nanodet_m min = 443.40 max = 444.58 avg = 444.00\n yolo-fastest-1.1 min = 240.39 max = 248.05 avg = 247.04\n yolo-fastestv2 min = 162.71 max = 173.30 avg = 169.39\n vision_transformer min = 17148.14 max = 17250.66 avg = 17202.60\n FastestDet min = 199.71 max = 200.38 avg = 199.90\n```\n\n### CVITEK SG2000 (C906, 1 GHz x 1 + 700MHz x 1)\n```\n[root@milkv-duo]~/ncnn# ./benchncnn 4 1 2 -1 0\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 221.53 max = 229.14 avg = 225.53\n squeezenet_int8 min = 8153.49 max = 8163.26 avg = 8160.17\n mobilenet min = 329.60 max = 338.58 avg = 335.00\n mobilenet_int8 min = 12725.12 max = 12733.70 avg = 12728.52\n mobilenet_v2 min = 253.83 max = 260.60 avg = 257.20\n mobilenet_v3 min = 205.51 max = 212.72 avg = 209.26\n shufflenet min = 358.73 max = 367.05 avg = 364.52\n shufflenet_v2 min = 238.44 max = 246.05 avg = 242.09\n mnasnet min = 254.39 max = 258.26 avg = 255.63\n proxylessnasnet min = 294.99 max = 302.80 avg = 300.65\n regnety_400m min = 407.72 max = 409.69 avg = 409.03\n blazeface min = 117.08 max = 124.26 avg = 119.00\n googlenet min = 817.28 max = 824.70 avg = 820.70\n googlenet_int8 min = 18246.97 max = 18276.23 avg = 18261.11\n resnet18 min = 610.81 max = 618.87 avg = 613.91\n resnet18_int8 min = 18772.96 max = 18808.53 avg = 18786.88\n alexnet min = 568.11 max = 577.02 avg = 570.66\n squeezenet_ssd min = 890.76 max = 896.30 avg = 893.57\n squeezenet_ssd_int8 min = 31680.48 max = 31938.09 avg = 31810.68\n mobilenet_ssd min = 746.38 max = 762.07 avg = 752.19\n mobilenet_ssd_int8 min = 41140.62 max = 41540.85 avg = 41356.70\n mobilenet_yolo min = 1744.59 max = 1755.90 avg = 1750.05\n mobilenetv2_yolov3 min = 890.20 max = 897.86 avg = 895.14\n yolov4-tiny min = 1056.03 max = 1059.44 avg = 1058.21\n nanodet_m min = 547.85 max = 554.80 avg = 549.81\n yolo-fastest-1.1 min = 290.89 max = 298.31 avg = 296.24\n yolo-fastestv2 min = 188.59 max = 196.79 avg = 190.96\n FastestDet min = 196.19 max = 205.96 avg = 200.99\n```\n\n### Rockchip RK3588 (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz)\ntest in ROCK5 MODEL B\n\n```\nrock@rock-5b:~/ncnn/build/benchmark$ ./benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 15.22 max = 16.03 avg = 15.70\n squeezenet_int8 min = 16.77 max = 16.96 avg = 16.86\n mobilenet min = 23.07 max = 23.58 avg = 23.36\n mobilenet_int8 min = 18.58 max = 18.90 avg = 18.72\n mobilenet_v2 min = 18.74 max = 19.10 avg = 18.96\n mobilenet_v3 min = 14.40 max = 14.65 avg = 14.50\n shufflenet min = 9.74 max = 9.88 avg = 9.84\n shufflenet_v2 min = 9.44 max = 9.55 avg = 9.50\n mnasnet min = 14.73 max = 15.03 avg = 14.87\n proxylessnasnet min = 18.37 max = 18.59 avg = 18.46\n efficientnet_b0 min = 29.11 max = 30.18 avg = 29.63\n efficientnetv2_b0 min = 46.40 max = 46.95 avg = 46.76\n regnety_400m min = 19.18 max = 19.39 avg = 19.28\n blazeface min = 5.16 max = 5.23 avg = 5.20\n googlenet min = 64.64 max = 65.33 avg = 65.00\n googlenet_int8 min = 61.86 max = 63.41 avg = 62.42\n resnet18 min = 42.00 max = 43.34 avg = 42.48\n resnet18_int8 min = 67.22 max = 67.80 avg = 67.45\n alexnet min = 57.65 max = 58.21 avg = 58.01\n vgg16 min = 192.35 max = 193.36 avg = 192.84\n vgg16_int8 min = 570.86 max = 578.81 avg = 574.50\n resnet50 min = 107.86 max = 109.52 avg = 108.70\n resnet50_int8 min = 134.41 max = 135.86 avg = 135.18\n squeezenet_ssd min = 40.85 max = 41.24 avg = 41.02\n squeezenet_ssd_int8 min = 52.23 max = 53.70 avg = 52.54\n mobilenet_ssd min = 45.11 max = 45.50 avg = 45.32\n mobilenet_ssd_int8 min = 36.53 max = 36.63 avg = 36.59\n mobilenet_yolo min = 95.18 max = 96.79 avg = 95.90\n mobilenetv2_yolov3 min = 65.50 max = 65.88 avg = 65.72\n yolov4-tiny min = 86.13 max = 88.84 avg = 87.29\n nanodet_m min = 22.57 max = 22.87 avg = 22.74\n yolo-fastest-1.1 min = 9.23 max = 9.35 avg = 9.29\n yolo-fastestv2 min = 8.62 max = 8.83 avg = 8.73\n vision_transformer min = 3077.54 max = 3396.13 avg = 3339.58\n FastestDet min = 9.11 max = 9.30 avg = 9.20\n\nrock@rock-5b:~/ncnn/build/benchmark$ ./benchncnn 10 8 0 -1 0\nloop_count = 10\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 10.02 max = 11.01 avg = 10.43\n squeezenet_int8 min = 11.78 max = 13.77 avg = 12.55\n mobilenet min = 12.75 max = 13.58 avg = 13.12\n mobilenet_int8 min = 12.23 max = 14.29 avg = 13.54\n mobilenet_v2 min = 12.76 max = 14.27 avg = 13.40\n mobilenet_v3 min = 9.51 max = 9.81 avg = 9.71\n shufflenet min = 7.06 max = 7.23 avg = 7.13\n shufflenet_v2 min = 6.21 max = 7.32 avg = 6.38\n mnasnet min = 9.32 max = 12.49 avg = 10.75\n proxylessnasnet min = 13.79 max = 15.51 avg = 14.70\n efficientnet_b0 min = 16.59 max = 17.99 avg = 17.08\n efficientnetv2_b0 min = 28.26 max = 32.26 avg = 30.52\n regnety_400m min = 13.43 max = 15.00 avg = 13.72\n blazeface min = 3.87 max = 7.38 avg = 5.65\n googlenet min = 29.18 max = 44.00 avg = 36.31\n googlenet_int8 min = 31.14 max = 37.48 avg = 34.58\n resnet18 min = 21.47 max = 24.40 avg = 22.35\n resnet18_int8 min = 26.68 max = 29.89 avg = 28.45\n alexnet min = 29.35 max = 38.09 avg = 31.65\n vgg16 min = 112.37 max = 122.94 avg = 117.05\n vgg16_int8 min = 161.08 max = 215.29 avg = 176.89\n resnet50 min = 54.54 max = 57.50 avg = 55.71\n resnet50_int8 min = 54.76 max = 65.05 avg = 60.59\n squeezenet_ssd min = 26.21 max = 35.05 avg = 30.76\n squeezenet_ssd_int8 min = 33.34 max = 40.88 avg = 36.19\n mobilenet_ssd min = 26.71 max = 28.85 avg = 27.88\n mobilenet_ssd_int8 min = 22.03 max = 25.31 avg = 24.21\n mobilenet_yolo min = 60.51 max = 74.65 avg = 65.45\n mobilenetv2_yolov3 min = 37.27 max = 44.13 avg = 41.20\n yolov4-tiny min = 49.84 max = 58.12 avg = 53.93\n nanodet_m min = 16.54 max = 22.41 avg = 20.60\n yolo-fastest-1.1 min = 8.49 max = 13.50 avg = 9.91\n yolo-fastestv2 min = 6.28 max = 11.22 avg = 8.00\n vision_transformer min = 968.62 max = 1063.47 avg = 1019.12\n FastestDet min = 6.14 max = 11.92 avg = 7.85\n\nrock@rock-5b:~/ncnn/build/benchmark$ ./benchncnn 10 4 2 -1 0\nloop_count = 10\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.78 max = 7.27 avg = 7.07\n squeezenet_int8 min = 4.58 max = 4.73 avg = 4.63\n mobilenet min = 5.67 max = 5.78 avg = 5.72\n mobilenet_int8 min = 5.01 max = 5.20 avg = 5.15\n mobilenet_v2 min = 5.44 max = 5.76 avg = 5.50\n mobilenet_v3 min = 4.67 max = 5.03 avg = 4.74\n shufflenet min = 4.22 max = 4.30 avg = 4.27\n shufflenet_v2 min = 3.48 max = 3.60 avg = 3.53\n mnasnet min = 4.52 max = 4.83 avg = 4.61\n proxylessnasnet min = 5.44 max = 6.01 avg = 5.56\n efficientnet_b0 min = 8.33 max = 8.52 avg = 8.41\n efficientnetv2_b0 min = 12.95 max = 13.08 avg = 13.02\n regnety_400m min = 8.60 max = 8.73 avg = 8.66\n blazeface min = 1.86 max = 1.95 avg = 1.90\n googlenet min = 16.58 max = 16.85 avg = 16.65\n googlenet_int8 min = 16.99 max = 17.13 avg = 17.06\n resnet18 min = 14.98 max = 15.30 avg = 15.08\n resnet18_int8 min = 20.10 max = 20.22 avg = 20.15\n alexnet min = 19.78 max = 20.21 avg = 19.87\n vgg16 min = 66.35 max = 94.16 avg = 75.24\n vgg16_int8 min = 131.02 max = 131.98 avg = 131.51\n resnet50 min = 28.07 max = 28.78 avg = 28.28\n resnet50_int8 min = 33.56 max = 35.53 avg = 33.84\n squeezenet_ssd min = 16.40 max = 16.80 avg = 16.49\n squeezenet_ssd_int8 min = 18.64 max = 19.00 avg = 18.76\n mobilenet_ssd min = 13.66 max = 13.78 avg = 13.72\n mobilenet_ssd_int8 min = 11.23 max = 11.42 avg = 11.33\n mobilenet_yolo min = 30.76 max = 31.03 avg = 30.86\n mobilenetv2_yolov3 min = 19.28 max = 21.07 avg = 20.30\n yolov4-tiny min = 33.44 max = 37.68 avg = 34.70\n nanodet_m min = 8.28 max = 8.55 avg = 8.38\n yolo-fastest-1.1 min = 4.30 max = 4.40 avg = 4.34\n yolo-fastestv2 min = 4.07 max = 4.18 avg = 4.13\n vision_transformer min = 815.67 max = 819.27 avg = 817.49\n FastestDet min = 4.34 max = 7.47 avg = 5.18\n```\n\n### AWS c5.4xlarge Instance\n\n- OS: Ubuntu 20.04.6 LTS x86_64\n- CPU: Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz\n- Compiler: gcc version 9.4.0 (Ubuntu 9.4.0-1ubuntu1~20.04.2)\n- ncnn tag: 20240102\n\n```\nloop_count = 4\nnum_threads = 8\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 3.31 max = 3.33 avg = 3.32\n squeezenet_int8 min = 3.87 max = 4.34 avg = 4.07\n mobilenet min = 3.12 max = 3.20 avg = 3.17\n mobilenet_int8 min = 3.32 max = 3.45 avg = 3.38\n mobilenet_v2 min = 4.23 max = 4.43 avg = 4.33\n mobilenet_v3 min = 3.82 max = 3.92 avg = 3.87\n shufflenet min = 3.67 max = 3.72 avg = 3.69\n shufflenet_v2 min = 4.08 max = 4.22 avg = 4.15\n mnasnet min = 3.62 max = 3.69 avg = 3.64\n proxylessnasnet min = 4.29 max = 4.59 avg = 4.37\n efficientnet_b0 min = 5.32 max = 5.64 avg = 5.50\n efficientnetv2_b0 min = 6.81 max = 6.88 avg = 6.85\n regnety_400m min = 9.71 max = 9.77 avg = 9.74\n blazeface min = 1.71 max = 2.57 avg = 2.10\n googlenet min = 10.00 max = 10.09 avg = 10.05\n googlenet_int8 min = 8.76 max = 8.79 avg = 8.77\n resnet18 min = 6.55 max = 6.91 avg = 6.70\n resnet18_int8 min = 5.63 max = 5.95 avg = 5.81\n alexnet min = 4.88 max = 4.91 avg = 4.89\n vgg16 min = 36.99 max = 37.04 avg = 37.01\n vgg16_int8 min = 28.13 max = 28.57 avg = 28.31\n resnet50 min = 13.99 max = 14.13 avg = 14.06\n resnet50_int8 min = 12.49 max = 12.56 avg = 12.53\n squeezenet_ssd min = 9.93 max = 10.04 avg = 9.98\n squeezenet_ssd_int8 min = 9.51 max = 9.70 avg = 9.59\n mobilenet_ssd min = 6.60 max = 6.63 avg = 6.61\n mobilenet_ssd_int8 min = 6.95 max = 7.10 avg = 7.02\n mobilenet_yolo min = 18.28 max = 18.44 avg = 18.35\n mobilenetv2_yolov3 min = 13.26 max = 13.39 avg = 13.32\n yolov4-tiny min = 25.14 max = 25.58 avg = 25.37\n nanodet_m min = 7.71 max = 7.77 avg = 7.75\n yolo-fastest-1.1 min = 4.69 max = 4.96 avg = 4.81\n yolo-fastestv2 min = 4.84 max = 5.17 avg = 5.01\n vision_transformer min = 139.34 max = 140.38 avg = 139.96\n FastestDet min = 4.95 max = 5.12 avg = 5.06\n```\n\n### Hyper-V Linux Guest with GPU-PV enabled (Intel Core i7-11800H, NVIDIA GeForce RTX 3070 Laptop GPU)\n\n- Host OS: Microsoft Windows 11 Enterprise (10.0.22621.1635)\n- Guest OS: openSUSE Tumbleweed x86_64 20230507\n- Mesa 3D source tree: https://gitlab.freedesktop.org/mesa/mesa/-/tree/ce6430067613e3e64cabf79918a3d96122b0c4c4\n- Mesa 3D configuration command\n > meson --prefix=\"${PWD}/build/install\" -D gallium-drivers=swrast,d3d12 -D vulkan-drivers=swrast,microsoft-experimental build/\n- ncnn configuration command\n > cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON ..\n\n```\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark> VK_ICD_FILENAMES=/home/mouri/Workspace/mesa/build/install/share/vulkan/icd.d/dzn_icd.x86_64.json ./../build/benchmark/benchncnn 10 1 0 0 0\nWARNING: dzn is not a conformant Vulkan implementation, testing use only.\nWARNING: dzn is not a conformant Vulkan implementation, testing use only.\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] queueC=1[8] queueG=0[4] queueT=2[1]\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] fp16-p/s/a=1/1/1 int8-p/s/a=1/0/0\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] queueC=1[8] queueG=0[4] queueT=2[1]\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] fp16-p/s/a=1/1/1 int8-p/s/a=1/0/0\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] subgroup=16 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 52.30 max = 65.51 avg = 56.65\n squeezenet_int8 min = 14.53 max = 15.55 avg = 14.88\n mobilenet min = 37.42 max = 52.07 avg = 42.48\n mobilenet_int8 min = 19.01 max = 19.82 avg = 19.46\n mobilenet_v2 min = 55.34 max = 73.39 avg = 63.94\n mobilenet_v3 min = 97.02 max = 123.14 avg = 109.90\n shufflenet min = 72.75 max = 100.26 avg = 88.26\n shufflenet_v2 min = 93.34 max = 119.64 avg = 105.76\n mnasnet min = 63.49 max = 74.11 avg = 69.05\n proxylessnasnet min = 65.87 max = 83.87 avg = 76.33\n efficientnet_b0 min = 162.86 max = 210.51 avg = 184.03\n efficientnetv2_b0 min = 200.88 max = 220.40 avg = 210.85\n regnety_400m min = 106.92 max = 134.68 avg = 123.04\n blazeface min = 58.64 max = 66.50 avg = 60.54\n googlenet min = 117.34 max = 145.28 avg = 134.84\n googlenet_int8 min = 62.50 max = 65.07 avg = 63.44\n resnet18 min = 67.30 max = 92.40 avg = 80.23\n resnet18_int8 min = 56.09 max = 58.40 avg = 56.97\n alexnet min = 29.94 max = 47.51 avg = 38.83\n vgg16 min = 59.72 max = 73.08 avg = 65.46\n vgg16_int8 min = 136.35 max = 148.39 avg = 143.96\n resnet50 min = 115.92 max = 152.34 avg = 129.64\n resnet50_int8 min = 93.86 max = 101.51 avg = 97.96\n squeezenet_ssd min = 139.82 max = 149.15 avg = 144.78\n squeezenet_ssd_int8 min = 32.09 max = 35.96 avg = 33.41\n mobilenet_ssd min = 88.14 max = 102.62 avg = 97.79\n mobilenet_ssd_int8 min = 33.93 max = 36.42 avg = 34.41\n mobilenet_yolo min = 52.22 max = 65.25 avg = 58.81\n mobilenetv2_yolov3 min = 75.09 max = 94.12 avg = 85.23\n yolov4-tiny min = 73.27 max = 88.69 avg = 81.44\n nanodet_m min = 110.98 max = 150.70 avg = 127.60\n yolo-fastest-1.1 min = 104.72 max = 135.40 avg = 116.92\n yolo-fastestv2 min = 113.84 max = 142.19 avg = 128.24\n vision_transformer min = 412.19 max = 474.25 avg = 444.15\n FastestDet min = 96.31 max = 131.51 avg = 117.27\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark> VK_ICD_FILENAMES=/home/mouri/Workspace/mesa/build/install/share/vulkan/icd.d/dzn_icd.x86_64.json ./../build/benchmark/benchncnn 10 1 0 1 0\nWARNING: dzn is not a conformant Vulkan implementation, testing use only.\nWARNING: dzn is not a conformant Vulkan implementation, testing use only.\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] queueC=1[8] queueG=0[4] queueT=2[1]\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] fp16-p/s/a=1/1/1 int8-p/s/a=1/0/0\n[0 Microsoft Direct3D12 (NVIDIA GeForce RTX 3070 Laptop GPU)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] queueC=1[8] queueG=0[4] queueT=2[1]\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] fp16-p/s/a=1/1/1 int8-p/s/a=1/0/0\n[1 Microsoft Direct3D12 (Intel(R) UHD Graphics)] subgroup=16 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 1\ncooling_down = 0\n squeezenet min = 36.86 max = 62.04 avg = 44.48\n squeezenet_int8 min = 15.31 max = 16.14 avg = 15.63\n mobilenet min = 30.79 max = 34.67 avg = 32.95\n mobilenet_int8 min = 19.23 max = 19.72 avg = 19.42\n mobilenet_v2 min = 36.56 max = 40.53 avg = 38.20\n mobilenet_v3 min = 52.11 max = 61.72 avg = 56.58\n shufflenet min = 41.50 max = 74.61 avg = 49.24\n shufflenet_v2 min = 44.49 max = 52.30 avg = 49.04\n mnasnet min = 35.66 max = 43.45 avg = 37.98\n proxylessnasnet min = 41.27 max = 47.63 avg = 43.63\n efficientnet_b0 min = 67.66 max = 80.88 avg = 73.64\n efficientnetv2_b0 min = 111.10 max = 156.52 avg = 126.70\n regnety_400m min = 62.66 max = 89.16 avg = 68.99\n blazeface min = 24.86 max = 33.52 avg = 26.91\n googlenet min = 70.55 max = 84.22 avg = 75.19\n googlenet_int8 min = 58.78 max = 64.81 avg = 62.99\n resnet18 min = 44.17 max = 49.37 avg = 46.73\n resnet18_int8 min = 59.99 max = 66.91 avg = 62.35\n alexnet min = 41.54 max = 57.16 avg = 44.30\n vgg16 min = 138.74 max = 165.03 avg = 146.90\n vgg16_int8 min = 135.36 max = 165.89 avg = 142.61\n resnet50 min = 97.46 max = 107.18 avg = 100.89\n resnet50_int8 min = 92.90 max = 100.45 avg = 95.91\n squeezenet_ssd min = 72.27 max = 90.71 avg = 76.09\n squeezenet_ssd_int8 min = 34.66 max = 40.46 avg = 36.58\n mobilenet_ssd min = 59.90 max = 68.74 avg = 62.40\n mobilenet_ssd_int8 min = 37.02 max = 38.59 avg = 37.82\n mobilenet_yolo min = 73.19 max = 80.40 avg = 76.42\n mobilenetv2_yolov3 min = 58.56 max = 66.71 avg = 62.02\n yolov4-tiny min = 63.75 max = 84.29 avg = 69.54\n nanodet_m min = 54.66 max = 67.89 avg = 60.82\n yolo-fastest-1.1 min = 40.89 max = 51.03 avg = 43.15\n yolo-fastestv2 min = 50.43 max = 77.46 avg = 60.66\n vision_transformer min = 1330.82 max = 1388.73 avg = 1354.10\n FastestDet min = 85.75 max = 112.67 avg = 98.62\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark> VK_ICD_FILENAMES=/home/mouri/Workspace/mesa/build/install/share/vulkan/icd.d/dzn_icd.x86_64.json ./../build/benchmark/benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 6.30 max = 10.16 avg = 8.21\n squeezenet_int8 min = 14.53 max = 14.94 avg = 14.67\n mobilenet min = 10.71 max = 11.26 avg = 10.91\n mobilenet_int8 min = 17.66 max = 18.46 avg = 17.91\n mobilenet_v2 min = 7.74 max = 8.05 avg = 7.89\n mobilenet_v3 min = 6.25 max = 6.70 avg = 6.38\n shufflenet min = 3.78 max = 7.87 avg = 5.37\n shufflenet_v2 min = 4.19 max = 7.83 avg = 5.25\n mnasnet min = 7.29 max = 7.61 avg = 7.44\n proxylessnasnet min = 8.10 max = 8.43 avg = 8.24\n efficientnet_b0 min = 11.77 max = 12.66 avg = 12.06\n efficientnetv2_b0 min = 13.80 max = 15.02 avg = 14.11\n regnety_400m min = 10.09 max = 10.26 avg = 10.17\n blazeface min = 1.24 max = 4.02 avg = 2.45\n googlenet min = 24.05 max = 25.78 avg = 24.64\n googlenet_int8 min = 58.75 max = 62.45 avg = 59.54\n resnet18 min = 20.31 max = 21.48 avg = 20.74\n resnet18_int8 min = 53.82 max = 55.27 avg = 54.43\n alexnet min = 17.37 max = 18.69 avg = 17.66\n vgg16 min = 114.49 max = 117.62 avg = 115.96\n vgg16_int8 min = 133.82 max = 144.40 avg = 137.07\n resnet50 min = 54.40 max = 58.74 avg = 55.54\n resnet50_int8 min = 92.95 max = 104.71 avg = 99.18\n squeezenet_ssd min = 17.30 max = 18.65 avg = 17.71\n squeezenet_ssd_int8 min = 32.27 max = 33.88 avg = 32.82\n mobilenet_ssd min = 24.01 max = 25.94 avg = 25.02\n mobilenet_ssd_int8 min = 34.68 max = 36.09 avg = 35.43\n mobilenet_yolo min = 53.32 max = 63.48 avg = 56.58\n mobilenetv2_yolov3 min = 30.06 max = 34.24 avg = 31.46\n yolov4-tiny min = 41.49 max = 43.55 avg = 42.50\n nanodet_m min = 10.24 max = 11.08 avg = 10.43\n yolo-fastest-1.1 min = 3.85 max = 8.34 avg = 5.40\n yolo-fastestv2 min = 4.33 max = 7.61 avg = 6.01\n vision_transformer min = 556.38 max = 599.49 avg = 567.98\n FastestDet min = 4.20 max = 11.37 avg = 6.51\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark>\n```\n\n### Hyper-V Linux Guest with GPU-PV enabled (Intel Core i7-7700K, NVIDIA GeForce GTX 1050 Ti)\n\n- Host OS: Microsoft Windows 10 Enterprise LTSC 2021 (10.0.19044.2846)\n- Guest OS: openSUSE Tumbleweed x86_64 20230507\n- Mesa 3D source tree: https://gitlab.freedesktop.org/mesa/mesa/-/tree/ce6430067613e3e64cabf79918a3d96122b0c4c4\n- Mesa 3D configuration command\n > meson --prefix=\"${PWD}/build/install\" -D gallium-drivers=swrast,d3d12 -D vulkan-drivers=swrast,microsoft-experimental build/\n- ncnn configuration command\n > cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_TESTS=ON ..\n\n```\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark> VK_ICD_FILENAMES=/home/mouri/Workspace/mesa/build/install/share/vulkan/icd.d/dzn_icd.x86_64.json ./../build/benchmark/benchncnn 10 1 0 0 0\nWARNING: dzn is not a conformant Vulkan implementation, testing use only.\n[0 Microsoft Direct3D12 (NVIDIA GeForce GTX 1050 Ti)] queueC=1[8] queueG=0[4] queueT=2[1]\n[0 Microsoft Direct3D12 (NVIDIA GeForce GTX 1050 Ti)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Microsoft Direct3D12 (NVIDIA GeForce GTX 1050 Ti)] fp16-p/s/a=1/0/0 int8-p/s/a=1/0/0\n[0 Microsoft Direct3D12 (NVIDIA GeForce GTX 1050 Ti)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 53.80 max = 64.22 avg = 59.91\n squeezenet_int8 min = 23.21 max = 25.98 avg = 24.44\n mobilenet min = 47.63 max = 55.22 avg = 49.79\n mobilenet_int8 min = 23.27 max = 25.05 avg = 23.77\n mobilenet_v2 min = 58.17 max = 83.14 avg = 68.48\n mobilenet_v3 min = 92.14 max = 114.74 avg = 101.66\n shufflenet min = 75.96 max = 106.54 avg = 89.64\n shufflenet_v2 min = 90.66 max = 114.69 avg = 103.25\n mnasnet min = 58.40 max = 85.74 avg = 67.75\n proxylessnasnet min = 66.73 max = 84.82 avg = 77.73\n efficientnet_b0 min = 134.28 max = 164.39 avg = 155.40\n efficientnetv2_b0 min = 171.97 max = 220.43 avg = 198.26\n regnety_400m min = 124.15 max = 145.61 avg = 135.99\n blazeface min = 53.18 max = 72.10 avg = 60.21\n googlenet min = 119.34 max = 159.93 avg = 134.71\n googlenet_int8 min = 96.71 max = 102.44 avg = 98.57\n resnet18 min = 68.14 max = 89.99 avg = 80.76\n resnet18_int8 min = 88.07 max = 108.62 avg = 91.09\n alexnet min = 44.12 max = 51.57 avg = 48.09\n vgg16 min = 88.49 max = 99.87 avg = 93.42\n vgg16_int8 min = 196.17 max = 211.99 avg = 201.27\n resnet50 min = 115.36 max = 138.65 avg = 125.57\n resnet50_int8 min = 138.15 max = 148.55 avg = 141.08\n squeezenet_ssd min = 138.42 max = 168.49 avg = 155.66\n squeezenet_ssd_int8 min = 46.01 max = 47.83 avg = 46.85\n mobilenet_ssd min = 82.39 max = 134.74 avg = 101.22\n mobilenet_ssd_int8 min = 45.53 max = 46.67 avg = 45.96\n mobilenet_yolo min = 70.39 max = 87.83 avg = 80.01\n mobilenetv2_yolov3 min = 75.71 max = 90.59 avg = 84.04\n yolov4-tiny min = 72.16 max = 87.76 avg = 76.81\n nanodet_m min = 98.27 max = 129.60 avg = 112.34\n yolo-fastest-1.1 min = 101.01 max = 118.45 avg = 106.47\n yolo-fastestv2 min = 109.89 max = 137.23 avg = 123.97\n vision_transformer min = 688.60 max = 750.54 avg = 723.30\n FastestDet min = 104.16 max = 139.23 avg = 123.75\nmouri@MouriVM-openSUSE:~/Workspace/ncnn/benchmark> VK_ICD_FILENAMES=/home/mouri/Workspace/mesa/build/install/share/vulkan/icd.d/dzn_icd.x86_64.json ./../build/benchmark/benchncnn 10 1 0 -1 0\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 8.90 max = 9.48 avg = 9.15\n squeezenet_int8 min = 22.54 max = 24.13 avg = 22.85\n mobilenet min = 14.85 max = 16.15 avg = 15.18\n mobilenet_int8 min = 23.56 max = 23.98 avg = 23.74\n mobilenet_v2 min = 11.03 max = 11.73 avg = 11.22\n mobilenet_v3 min = 8.61 max = 9.29 avg = 8.79\n shufflenet min = 5.26 max = 5.96 avg = 5.42\n shufflenet_v2 min = 5.56 max = 7.06 avg = 5.82\n mnasnet min = 10.46 max = 11.04 avg = 10.68\n proxylessnasnet min = 12.18 max = 12.55 avg = 12.33\n efficientnet_b0 min = 22.46 max = 23.15 avg = 22.86\n efficientnetv2_b0 min = 23.33 max = 23.80 avg = 23.55\n regnety_400m min = 13.03 max = 14.25 avg = 13.28\n blazeface min = 1.49 max = 1.95 avg = 1.61\n googlenet min = 35.26 max = 46.31 avg = 39.63\n googlenet_int8 min = 96.25 max = 98.15 avg = 96.93\n resnet18 min = 29.34 max = 31.00 avg = 29.92\n resnet18_int8 min = 87.84 max = 89.85 avg = 88.73\n alexnet min = 22.91 max = 23.87 avg = 23.18\n vgg16 min = 151.26 max = 174.79 avg = 155.94\n vgg16_int8 min = 193.66 max = 210.63 avg = 199.14\n resnet50 min = 74.89 max = 77.27 avg = 75.91\n resnet50_int8 min = 136.59 max = 162.13 avg = 141.22\n squeezenet_ssd min = 24.48 max = 34.00 avg = 26.19\n squeezenet_ssd_int8 min = 46.31 max = 48.87 avg = 47.09\n mobilenet_ssd min = 31.56 max = 34.45 avg = 32.50\n mobilenet_ssd_int8 min = 45.15 max = 46.53 avg = 45.93\n mobilenet_yolo min = 72.09 max = 78.05 avg = 74.31\n mobilenetv2_yolov3 min = 40.44 max = 41.54 avg = 40.86\n yolov4-tiny min = 56.73 max = 60.59 avg = 57.93\n nanodet_m min = 13.22 max = 19.28 avg = 14.65\n yolo-fastest-1.1 min = 5.47 max = 5.70 avg = 5.58\n yolo-fastestv2 min = 5.68 max = 7.20 avg = 5.88\n vision_transformer min = 600.83 max = 666.35 avg = 617.33\n FastestDet min = 6.05 max = 6.72 avg = 6.23\n```\n\n### AMD Ryzen 9 5950X 16-Core of Desktop[2023-10-12]\n```\nE:\\github\\ncnn\\build-ncnn-vs2019\\benchmark\\Release>benchncnn.exe 100 16 0 -1 0\nloop_count = 100\nnum_threads = 16\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 2.68 max = 3.10 avg = 2.77\n squeezenet_int8 min = 3.57 max = 4.72 avg = 4.04\n mobilenet min = 3.09 max = 5.44 avg = 3.38\n mobilenet_int8 min = 2.36 max = 3.40 avg = 2.74\n mobilenet_v2 min = 4.24 max = 4.81 avg = 4.40\n mobilenet_v3 min = 3.46 max = 3.93 avg = 3.58\n shufflenet min = 3.21 max = 4.54 avg = 4.01\n shufflenet_v2 min = 2.99 max = 4.49 avg = 3.34\n mnasnet min = 3.62 max = 4.31 avg = 3.83\n proxylessnasnet min = 4.06 max = 5.70 avg = 4.23\n efficientnet_b0 min = 5.60 max = 6.55 avg = 5.81\n efficientnetv2_b0 min = 6.83 max = 8.82 avg = 7.12\n regnety_400m min = 8.02 max = 9.75 avg = 8.34\n blazeface min = 1.34 max = 1.77 avg = 1.46\n googlenet min = 11.62 max = 15.95 avg = 12.70\n googlenet_int8 min = 7.43 max = 10.06 avg = 7.92\n resnet18 min = 8.39 max = 10.39 avg = 9.04\n resnet18_int8 min = 6.23 max = 8.64 avg = 6.75\n alexnet min = 7.78 max = 12.51 avg = 8.51\n vgg16 min = 53.85 max = 63.39 avg = 56.36\n vgg16_int8 min = 35.61 max = 46.94 avg = 38.08\n resnet50 min = 18.55 max = 24.46 avg = 19.81\n resnet50_int8 min = 11.95 max = 23.21 avg = 13.51\n squeezenet_ssd min = 10.01 max = 13.16 avg = 10.69\n squeezenet_ssd_int8 min = 9.29 max = 14.02 avg = 10.47\n mobilenet_ssd min = 6.38 max = 10.26 avg = 7.15\n mobilenet_ssd_int8 min = 4.69 max = 6.98 avg = 5.42\n mobilenet_yolo min = 17.63 max = 22.59 avg = 19.45\n mobilenetv2_yolov3 min = 11.79 max = 15.67 avg = 12.76\n yolov4-tiny min = 21.53 max = 25.79 avg = 22.46\n nanodet_m min = 7.16 max = 9.99 avg = 8.01\n yolo-fastest-1.1 min = 3.66 max = 5.00 avg = 4.38\n yolo-fastestv2 min = 3.52 max = 5.20 avg = 4.60\n vision_transformer min = 67.01 max = 93.71 avg = 78.48\n FastestDet min = 4.44 max = 8.62 avg = 4.69\n```\n\n### AMD Radeon RX 6900 XT of Desktop[2023-10-12]\n```\nE:\\github\\ncnn\\build-ncnn-vs2019\\benchmark\\Release>benchncnn.exe 100 16 0 0 0\n[0 AMD Radeon RX 6900 XT] queueC=1[2] queueG=0[1] queueT=2[2]\n[0 AMD Radeon RX 6900 XT] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 AMD Radeon RX 6900 XT] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 AMD Radeon RX 6900 XT] subgroup=64 basic/vote/ballot/shuffle=1/1/1/1\n[0 AMD Radeon RX 6900 XT] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 100\nnum_threads = 16\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 2.19 max = 2.70 avg = 2.47\n squeezenet_int8 min = 3.94 max = 4.51 avg = 4.18\n mobilenet min = 2.03 max = 2.63 avg = 2.28\n mobilenet_int8 min = 2.56 max = 3.34 avg = 2.69\n mobilenet_v2 min = 2.29 max = 2.98 avg = 2.62\n mobilenet_v3 min = 2.31 max = 3.10 avg = 2.75\n shufflenet min = 1.89 max = 2.61 avg = 2.30\n shufflenet_v2 min = 2.17 max = 3.04 avg = 2.59\n mnasnet min = 2.19 max = 2.98 avg = 2.69\n proxylessnasnet min = 2.12 max = 4.08 avg = 2.62\n efficientnet_b0 min = 3.62 max = 5.27 avg = 4.21\n efficientnetv2_b0 min = 6.09 max = 7.15 avg = 6.49\n regnety_400m min = 2.55 max = 3.82 avg = 3.00\n blazeface min = 1.93 max = 2.56 avg = 2.28\n googlenet min = 3.35 max = 4.46 avg = 3.75\n googlenet_int8 min = 8.02 max = 12.84 avg = 9.15\n resnet18 min = 2.46 max = 3.14 avg = 2.84\n resnet18_int8 min = 6.37 max = 9.15 avg = 7.30\n alexnet min = 2.31 max = 2.91 avg = 2.69\n vgg16 min = 4.76 max = 5.79 avg = 5.24\n vgg16_int8 min = 35.94 max = 46.27 avg = 39.05\n resnet50 min = 3.25 max = 4.09 avg = 3.75\n resnet50_int8 min = 12.04 max = 20.53 avg = 14.61\n squeezenet_ssd min = 3.03 max = 5.31 avg = 3.66\n squeezenet_ssd_int8 min = 9.74 max = 13.46 avg = 10.42\n mobilenet_ssd min = 2.82 max = 4.75 avg = 3.39\n mobilenet_ssd_int8 min = 4.67 max = 6.76 avg = 5.30\n mobilenet_yolo min = 3.01 max = 3.67 avg = 3.34\n mobilenetv2_yolov3 min = 4.04 max = 6.46 avg = 4.55\n yolov4-tiny min = 5.75 max = 8.05 avg = 6.52\n nanodet_m min = 10.16 max = 14.97 avg = 13.11\n yolo-fastest-1.1 min = 2.36 max = 3.80 avg = 2.88\n yolo-fastestv2 min = 2.24 max = 3.19 avg = 2.80\n vision_transformer min = 20.43 max = 25.06 avg = 21.07\n FastestDet min = 2.49 max = 3.18 avg = 2.93\n```\n\n### NVIDIA GeForce RTX 3060 Ti of Desktop[2023-10-12]\n```\nE:\\github\\ncnn\\build-ncnn-vs2019\\benchmark\\Release>benchncnn.exe 100 16 0 0 0\n[0 NVIDIA GeForce RTX 3060 Ti] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA GeForce RTX 3060 Ti] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA GeForce RTX 3060 Ti] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA GeForce RTX 3060 Ti] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA GeForce RTX 3060 Ti] fp16-matrix-16_8_8/16_8_16/16_16_16=1/1/1\n[1 Intel(R) UHD Graphics 770] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 Intel(R) UHD Graphics 770] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Intel(R) UHD Graphics 770] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 Intel(R) UHD Graphics 770] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[1 Intel(R) UHD Graphics 770] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 100\nnum_threads = 16\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 0.80 max = 2.51 avg = 0.89\n squeezenet_int8 min = 2.81 max = 3.51 avg = 2.96\n mobilenet min = 0.70 max = 0.79 avg = 0.71\n mobilenet_int8 min = 2.95 max = 3.44 avg = 3.03\n mobilenet_v2 min = 1.09 max = 1.25 avg = 1.12\n mobilenet_v3 min = 1.33 max = 2.04 avg = 1.56\n shufflenet min = 1.20 max = 1.39 avg = 1.27\n shufflenet_v2 min = 1.50 max = 1.66 avg = 1.57\n mnasnet min = 1.11 max = 1.22 avg = 1.15\n proxylessnasnet min = 1.20 max = 1.63 avg = 1.24\n efficientnet_b0 min = 2.38 max = 3.21 avg = 2.61\n efficientnetv2_b0 min = 9.16 max = 11.35 avg = 9.63\n regnety_400m min = 1.86 max = 2.03 avg = 1.94\n blazeface min = 0.70 max = 1.10 avg = 0.76\n googlenet min = 2.11 max = 2.40 avg = 2.30\n googlenet_int8 min = 6.91 max = 7.88 avg = 7.17\n resnet18 min = 1.14 max = 1.47 avg = 1.19\n resnet18_int8 min = 4.96 max = 6.82 avg = 5.40\n alexnet min = 1.10 max = 1.85 avg = 1.19\n vgg16 min = 2.27 max = 3.97 avg = 2.46\n vgg16_int8 min = 19.02 max = 22.20 avg = 20.28\n resnet50 min = 2.00 max = 2.99 avg = 2.10\n resnet50_int8 min = 10.66 max = 13.30 avg = 11.29\n squeezenet_ssd min = 2.74 max = 3.44 avg = 2.90\n squeezenet_ssd_int8 min = 6.93 max = 7.95 avg = 7.19\n mobilenet_ssd min = 1.86 max = 2.07 avg = 1.96\n mobilenet_ssd_int8 min = 5.92 max = 6.48 avg = 6.09\n mobilenet_yolo min = 1.65 max = 2.58 avg = 1.78\n mobilenetv2_yolov3 min = 3.85 max = 4.11 avg = 3.96\n yolov4-tiny min = 6.54 max = 7.05 avg = 6.69\n nanodet_m min = 2.38 max = 3.28 avg = 2.72\n yolo-fastest-1.1 min = 1.73 max = 2.07 avg = 1.83\n yolo-fastestv2 min = 1.72 max = 1.92 avg = 1.80\n vision_transformer min = 53.91 max = 56.59 avg = 55.27\n FastestDet min = 1.48 max = 1.83 avg = 1.69\n```\n\n### Intel(R) UHD Graphics 770 of Desktop[2023-10-12]\n```\nE:\\github\\ncnn\\build-ncnn-vs2019\\benchmark\\Release>benchncnn.exe 100 16 0 1 0\n[0 NVIDIA GeForce RTX 3060 Ti] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 NVIDIA GeForce RTX 3060 Ti] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 NVIDIA GeForce RTX 3060 Ti] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 NVIDIA GeForce RTX 3060 Ti] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 NVIDIA GeForce RTX 3060 Ti] fp16-matrix-16_8_8/16_8_16/16_16_16=1/1/1\n[1 Intel(R) UHD Graphics 770] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 Intel(R) UHD Graphics 770] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 Intel(R) UHD Graphics 770] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[1 Intel(R) UHD Graphics 770] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[1 Intel(R) UHD Graphics 770] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 100\nnum_threads = 16\npowersave = 0\ngpu_device = 1\ncooling_down = 0\n squeezenet min = 3.11 max = 4.47 avg = 3.45\n squeezenet_int8 min = 1.89 max = 2.84 avg = 2.23\n mobilenet min = 4.98 max = 5.67 avg = 5.18\n mobilenet_int8 min = 2.54 max = 3.17 avg = 2.98\n mobilenet_v2 min = 4.03 max = 4.89 avg = 4.37\n mobilenet_v3 min = 4.45 max = 5.68 avg = 4.86\n shufflenet min = 3.42 max = 4.42 avg = 3.79\n shufflenet_v2 min = 3.00 max = 4.01 avg = 3.30\n mnasnet min = 4.21 max = 5.12 avg = 4.51\n proxylessnasnet min = 4.62 max = 5.64 avg = 4.90\n efficientnet_b0 min = 7.82 max = 8.63 avg = 8.10\n efficientnetv2_b0 min = 34.52 max = 36.34 avg = 35.29\n regnety_400m min = 6.07 max = 7.31 avg = 6.44\n blazeface min = 1.54 max = 1.67 avg = 1.59\n googlenet min = 11.53 max = 12.64 avg = 11.89\n googlenet_int8 min = 13.71 max = 15.52 avg = 14.38\n resnet18 min = 10.75 max = 12.94 avg = 11.07\n resnet18_int8 min = 9.04 max = 11.05 avg = 9.53\n alexnet min = 13.64 max = 14.37 avg = 13.98\n vgg16 min = 38.53 max = 40.16 avg = 39.22\n vgg16_int8 min = 16.04 max = 21.16 avg = 19.35\n resnet50 min = 25.61 max = 28.22 avg = 26.62\n resnet50_int8 min = 7.72 max = 12.83 avg = 10.29\n squeezenet_ssd min = 10.34 max = 15.88 avg = 14.75\n squeezenet_ssd_int8 min = 4.63 max = 7.13 avg = 5.66\n mobilenet_ssd min = 11.35 max = 13.06 avg = 12.44\n mobilenet_ssd_int8 min = 4.21 max = 6.31 avg = 5.32\n mobilenet_yolo min = 20.14 max = 22.92 avg = 21.94\n mobilenetv2_yolov3 min = 12.58 max = 14.88 avg = 14.21\n yolov4-tiny min = 20.62 max = 25.58 avg = 24.39\n nanodet_m min = 7.75 max = 12.49 avg = 11.42\n yolo-fastest-1.1 min = 3.68 max = 6.49 avg = 5.54\n yolo-fastestv2 min = 4.32 max = 5.39 avg = 4.51\n vision_transformer min = 796.51 max = 805.29 avg = 802.39\n FastestDet min = 2.89 max = 4.83 avg = 3.95\n```\n\n### Intel® Core™ i7-13700K of Desktop[2023-10-12]\n```\nE:\\github\\ncnn\\build-ncnn-vs2019\\benchmark\\Release>benchncnn.exe 100 16 0 -1 0\nloop_count = 100\nnum_threads = 16\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 1.69 max = 2.63 avg = 2.12\n squeezenet_int8 min = 1.83 max = 3.03 avg = 2.26\n mobilenet min = 1.69 max = 2.64 avg = 2.24\n mobilenet_int8 min = 2.47 max = 3.06 avg = 2.84\n mobilenet_v2 min = 1.94 max = 3.47 avg = 2.47\n mobilenet_v3 min = 1.49 max = 2.74 avg = 1.87\n shufflenet min = 1.57 max = 3.00 avg = 1.82\n shufflenet_v2 min = 1.41 max = 1.72 avg = 1.51\n mnasnet min = 1.73 max = 2.94 avg = 2.13\n proxylessnasnet min = 2.08 max = 3.31 avg = 2.69\n efficientnet_b0 min = 3.20 max = 4.99 avg = 3.78\n efficientnetv2_b0 min = 3.51 max = 5.16 avg = 4.08\n regnety_400m min = 4.51 max = 10.29 avg = 6.18\n blazeface min = 0.52 max = 0.92 avg = 0.59\n googlenet min = 5.49 max = 7.48 avg = 6.26\n googlenet_int8 min = 4.83 max = 7.54 avg = 5.90\n resnet18 min = 4.05 max = 6.61 avg = 4.83\n resnet18_int8 min = 3.77 max = 5.70 avg = 4.57\n alexnet min = 3.60 max = 5.09 avg = 4.26\n vgg16 min = 25.19 max = 28.79 avg = 26.81\n vgg16_int8 min = 17.52 max = 21.79 avg = 19.80\n resnet50 min = 9.23 max = 13.15 avg = 11.34\n resnet50_int8 min = 7.77 max = 12.00 avg = 10.18\n squeezenet_ssd min = 4.33 max = 6.73 avg = 4.96\n squeezenet_ssd_int8 min = 4.77 max = 7.62 avg = 5.71\n mobilenet_ssd min = 3.70 max = 6.43 avg = 4.53\n mobilenet_ssd_int8 min = 4.16 max = 6.53 avg = 5.38\n mobilenet_yolo min = 11.27 max = 14.93 avg = 12.90\n mobilenetv2_yolov3 min = 7.41 max = 11.52 avg = 9.11\n yolov4-tiny min = 12.05 max = 18.96 avg = 14.15\n nanodet_m min = 3.39 max = 5.77 avg = 4.07\n yolo-fastest-1.1 min = 1.95 max = 3.85 avg = 2.30\n yolo-fastestv2 min = 1.91 max = 3.52 avg = 2.27\n vision_transformer min = 79.50 max = 99.93 avg = 88.91\n FastestDet min = 1.92 max = 2.72 avg = 2.19\n```\n\n### Amlogic S805 (Cortex-A5, 4 × 1.536GHz)\n\n- Platform: Xunlei OneCloud (玩客云)\n- OS: Armbian buster (20.12) armv7l\n- Compiler: gcc version 8.3.0 (Debian 8.3.0-6)\n- ncnn tag: 20240102\n\n```\nmizu-bai@aml-s812:~/ncnn-20240102/benchmark$ ../build/benchmark/benchncnn\nloop_count = 4\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 376.45 max = 445.48 avg = 408.08\n squeezenet_int8 min = 247.06 max = 340.34 avg = 281.40\n mobilenet min = 696.71 max = 745.63 avg = 718.49\n mobilenet_int8 min = 355.78 max = 472.06 avg = 401.17\n mobilenet_v2 min = 428.86 max = 491.25 avg = 458.45\n mobilenet_v3 min = 361.78 max = 425.90 avg = 396.94\n shufflenet min = 245.90 max = 333.41 avg = 293.46\n shufflenet_v2 min = 210.69 max = 329.51 avg = 260.73\n mnasnet min = 418.49 max = 493.40 avg = 448.95\n proxylessnasnet min = 542.20 max = 566.65 avg = 554.75\n efficientnet_b0 min = 727.72 max = 785.47 avg = 750.72\n efficientnetv2_b0 min = 805.70 max = 874.57 avg = 843.87\n regnety_400m min = 627.74 max = 686.57 avg = 660.60\n blazeface min = 62.14 max = 121.32 avg = 82.10\n googlenet min = 1295.31 max = 1411.88 avg = 1342.26\n googlenet_int8 min = 796.39 max = 860.28 avg = 823.76\n resnet18 min = 1076.93 max = 1125.12 avg = 1099.37\n resnet18_int8 min = 587.12 max = 634.97 avg = 605.29\n alexnet min = 701.70 max = 729.68 avg = 718.99\n vgg16 min = 5584.13 max = 5748.84 avg = 5660.70\n vgg16_int8 min = 3107.89 max = 3138.78 avg = 3121.28\n resnet50 min = 3378.84 max = 3461.61 avg = 3425.38\n resnet50_int8 min = 2044.93 max = 2067.70 avg = 2061.38\n squeezenet_ssd min = 908.77 max = 972.68 avg = 939.98\n squeezenet_ssd_int8 min = 609.58 max = 703.88 avg = 662.43\n mobilenet_ssd min = 1524.69 max = 1589.79 avg = 1552.12\n mobilenet_ssd_int8 min = 817.70 max = 885.45 avg = 840.30\n mobilenet_yolo min = 3497.13 max = 3605.83 avg = 3543.72\n mobilenetv2_yolov3 min = 1734.10 max = 1824.98 avg = 1795.42\n yolov4-tiny min = 2093.70 max = 2163.44 avg = 2128.30\n nanodet_m min = 593.75 max = 647.03 avg = 608.03\n yolo-fastest-1.1 min = 228.68 max = 318.40 avg = 265.74\n yolo-fastestv2 min = 194.29 max = 258.78 avg = 219.82\n vision_transformer min = 14836.43 max = 15238.27 avg = 15125.26\n FastestDet min = 215.60 max = 264.69 avg = 239.85\n```\n\n### Qualcomm SM8550-AB Snapdragon 8 Gen 2 (Kyro 3.20 GHz + 2.8 GHz x 2 + 2.80 GHz x 2 + 2.00 GHz * 3 + Adreno 740)\n```\n./benchncnn 4 1 2 -1 1\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 8.44 max = 8.51 avg = 8.47\n squeezenet_int8 min = 6.91 max = 7.13 avg = 7.00\n mobilenet min = 15.45 max = 15.53 avg = 15.49\n mobilenet_int8 min = 8.76 max = 9.03 avg = 8.88\n mobilenet_v2 min = 9.52 max = 10.71 avg = 10.02\n mobilenet_v3 min = 7.89 max = 8.02 avg = 7.93\n shufflenet min = 5.07 max = 5.61 avg = 5.25\n shufflenet_v2 min = 5.28 max = 5.41 avg = 5.37\n mnasnet min = 9.52 max = 9.58 avg = 9.54\n proxylessnasnet min = 11.26 max = 11.41 avg = 11.36\n efficientnet_b0 min = 18.84 max = 18.91 avg = 18.88\n efficientnetv2_b0 min = 28.60 max = 28.73 avg = 28.66\n regnety_400m min = 12.35 max = 12.39 avg = 12.37\n blazeface min = 1.83 max = 2.23 avg = 1.94\n googlenet min = 32.07 max = 37.37 avg = 35.59\n googlenet_int8 min = 28.50 max = 28.57 avg = 28.53\n resnet18 min = 21.88 max = 22.05 avg = 21.94\n resnet18_int8 min = 24.43 max = 40.52 avg = 32.04\n alexnet min = 23.69 max = 24.22 avg = 23.98\n vgg16 min = 91.85 max = 100.71 avg = 94.80\n vgg16_int8 min = 206.66 max = 325.74 avg = 258.40\n resnet50 min = 53.59 max = 54.20 avg = 53.96\n resnet50_int8 min = 44.39 max = 45.11 avg = 44.74\n squeezenet_ssd min = 23.80 max = 24.12 avg = 23.94\n squeezenet_ssd_int8 min = 30.17 max = 30.42 avg = 30.31\n mobilenet_ssd min = 33.49 max = 33.69 avg = 33.59\n mobilenet_ssd_int8 min = 19.37 max = 19.76 avg = 19.56\n mobilenet_yolo min = 72.63 max = 73.00 avg = 72.77\n mobilenetv2_yolov3 min = 36.86 max = 37.40 avg = 37.08\n yolov4-tiny min = 44.94 max = 45.46 avg = 45.22\n nanodet_m min = 13.65 max = 13.99 avg = 13.82\n yolo-fastest-1.1 min = 3.84 max = 3.93 avg = 3.89\n yolo-fastestv2 min = 4.78 max = 4.93 avg = 4.84\n vision_transformer min = 1042.50 max = 1043.06 avg = 1042.80\n FastestDet min = 4.67 max = 4.75 avg = 4.70\n./benchncnn 4 4 2 -1 1\nloop_count = 4\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 2.60 max = 2.66 avg = 2.64\n squeezenet_int8 min = 2.38 max = 2.43 avg = 2.40\n mobilenet min = 4.17 max = 4.25 avg = 4.21\n mobilenet_int8 min = 2.59 max = 2.60 avg = 2.60\n mobilenet_v2 min = 3.13 max = 3.44 avg = 3.23\n mobilenet_v3 min = 2.90 max = 5.07 avg = 3.46\n shufflenet min = 2.34 max = 2.44 avg = 2.38\n shufflenet_v2 min = 2.06 max = 2.15 avg = 2.11\n mnasnet min = 3.19 max = 3.20 avg = 3.20\n proxylessnasnet min = 3.53 max = 3.61 avg = 3.57\n efficientnet_b0 min = 5.72 max = 5.75 avg = 5.74\n efficientnetv2_b0 min = 8.61 max = 8.67 avg = 8.64\n regnety_400m min = 6.22 max = 6.27 avg = 6.25\n blazeface min = 0.82 max = 0.92 avg = 0.86\n googlenet min = 10.62 max = 14.39 avg = 11.59\n googlenet_int8 min = 8.84 max = 8.99 avg = 8.92\n resnet18 min = 6.61 max = 6.66 avg = 6.63\n resnet18_int8 min = 21.41 max = 23.48 avg = 22.57\n alexnet min = 8.18 max = 8.24 avg = 8.21\n vgg16 min = 36.99 max = 39.65 avg = 37.75\n vgg16_int8 min = 86.21 max = 89.00 avg = 86.95\n resnet50 min = 18.90 max = 18.98 avg = 18.94\n resnet50_int8 min = 19.18 max = 19.28 avg = 19.22\n squeezenet_ssd min = 8.26 max = 8.42 avg = 8.32\n squeezenet_ssd_int8 min = 21.02 max = 21.15 avg = 21.09\n mobilenet_ssd min = 9.29 max = 9.42 avg = 9.34\n mobilenet_ssd_int8 min = 5.85 max = 5.91 avg = 5.87\n mobilenet_yolo min = 21.64 max = 21.71 avg = 21.69\n mobilenetv2_yolov3 min = 11.50 max = 11.62 avg = 11.57\n yolov4-tiny min = 14.91 max = 14.99 avg = 14.95\n nanodet_m min = 4.93 max = 5.02 avg = 4.98\n yolo-fastest-1.1 min = 2.19 max = 2.26 avg = 2.21\n yolo-fastestv2 min = 2.29 max = 2.44 avg = 2.39\n vision_transformer min = 242.50 max = 301.91 avg = 271.32\n FastestDet min = 2.01 max = 2.12 avg = 2.05\n./benchncnn 4 8 0 -1 1\nloop_count = 4\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 4.53 max = 6.34 avg = 5.48\n squeezenet_int8 min = 5.48 max = 7.02 avg = 6.14\n mobilenet min = 6.89 max = 8.44 avg = 7.61\n mobilenet_int8 min = 4.89 max = 6.39 avg = 5.43\n mobilenet_v2 min = 6.01 max = 7.28 avg = 6.53\n mobilenet_v3 min = 4.85 max = 12.13 avg = 7.16\n shufflenet min = 4.41 max = 6.20 avg = 5.25\n shufflenet_v2 min = 3.50 max = 4.34 avg = 3.74\n mnasnet min = 5.52 max = 7.03 avg = 6.18\n proxylessnasnet min = 6.21 max = 7.76 avg = 6.94\n efficientnet_b0 min = 9.49 max = 10.57 avg = 9.94\n efficientnetv2_b0 min = 15.26 max = 19.50 avg = 17.42\n regnety_400m min = 9.89 max = 14.30 avg = 12.02\n blazeface min = 2.25 max = 3.44 avg = 2.66\n googlenet min = 18.98 max = 23.38 avg = 21.07\n googlenet_int8 min = 17.99 max = 20.47 avg = 19.45\n resnet18 min = 34.98 max = 84.52 avg = 69.50\n resnet18_int8 min = 14.58 max = 15.43 avg = 15.04\n alexnet min = 13.56 max = 15.05 avg = 14.29\n vgg16 min = 63.32 max = 73.69 avg = 67.01\n vgg16_int8 min = 91.17 max = 99.80 avg = 94.81\n resnet50 min = 32.01 max = 42.22 avg = 36.06\n resnet50_int8 min = 30.16 max = 32.25 avg = 30.72\n squeezenet_ssd min = 14.72 max = 21.45 avg = 17.51\n squeezenet_ssd_int8 min = 18.21 max = 23.93 avg = 21.45\n mobilenet_ssd min = 16.38 max = 17.92 avg = 16.97\n mobilenet_ssd_int8 min = 10.15 max = 15.88 avg = 12.92\n mobilenet_yolo min = 35.88 max = 37.10 avg = 36.26\n mobilenetv2_yolov3 min = 21.92 max = 27.60 avg = 24.12\n yolov4-tiny min = 32.03 max = 34.45 avg = 33.51\n nanodet_m min = 9.49 max = 14.35 avg = 11.20\n yolo-fastest-1.1 min = 3.97 max = 5.16 avg = 4.40\n yolo-fastestv2 min = 5.13 max = 7.84 avg = 6.18\n vision_transformer min = 364.37 max = 391.13 avg = 374.55\n FastestDet min = 3.01 max = 7.36 avg = 4.76\n./benchncnn 4 1 2 0 0\n[0 Adreno (TM) 740] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 740] bugsbn1=1 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Adreno (TM) 740] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Adreno (TM) 740] subgroup=64 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 4\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 9.73 max = 11.72 avg = 10.55\n squeezenet_int8 min = 7.21 max = 7.34 avg = 7.27\n mobilenet min = 10.87 max = 13.09 avg = 12.01\n mobilenet_int8 min = 8.82 max = 9.23 avg = 9.11\n mobilenet_v2 min = 15.77 max = 16.21 avg = 15.96\n mobilenet_v3 min = 18.04 max = 18.68 avg = 18.40\n shufflenet min = 9.82 max = 11.92 avg = 10.79\n shufflenet_v2 min = 14.41 max = 15.41 avg = 14.96\n mnasnet min = 16.01 max = 16.43 avg = 16.27\n proxylessnasnet min = 14.18 max = 16.28 avg = 15.51\n efficientnet_b0 min = 36.38 max = 37.06 avg = 36.83\n efficientnetv2_b0 min = 55.98 max = 66.59 avg = 59.54\n regnety_400m min = 21.94 max = 22.46 avg = 22.30\n blazeface min = 3.92 max = 4.47 avg = 4.08\n googlenet min = 31.79 max = 35.63 avg = 33.04\n googlenet_int8 min = 23.21 max = 29.38 avg = 26.60\n resnet18 min = 22.61 max = 24.05 avg = 23.09\n resnet18_int8 min = 24.56 max = 24.78 avg = 24.62\n alexnet min = 25.98 max = 27.05 avg = 26.49\n vgg16 min = 39.00 max = 39.82 avg = 39.29\n vgg16_int8 min = 207.47 max = 208.56 avg = 207.90\n resnet50 min = 44.07 max = 44.43 avg = 44.29\n resnet50_int8 min = 44.77 max = 47.04 avg = 45.44\n squeezenet_ssd min = 33.71 max = 34.27 avg = 34.09\n squeezenet_ssd_int8 min = 22.53 max = 30.33 avg = 25.07\n mobilenet_ssd min = 26.91 max = 28.35 avg = 27.42\n mobilenet_ssd_int8 min = 19.43 max = 19.82 avg = 19.69\n mobilenet_yolo min = 28.03 max = 29.19 avg = 28.65\n mobilenetv2_yolov3 min = 33.54 max = 34.65 avg = 34.31\n yolov4-tiny min = 49.77 max = 51.21 avg = 50.55\n nanodet_m min = 17.35 max = 18.83 avg = 18.06\n yolo-fastest-1.1 min = 9.45 max = 9.59 avg = 9.51\n yolo-fastestv2 min = 13.13 max = 13.63 avg = 13.36\n vision_transformer min = 671.13 max = 679.90 avg = 675.27\n FastestDet min = 8.62 max = 9.01 avg = 8.86\n./benchncnn 64 1 2 0 0\n[0 Adreno (TM) 740] queueC=0[3] queueG=0[3] queueT=0[3]\n[0 Adreno (TM) 740] bugsbn1=1 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Adreno (TM) 740] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Adreno (TM) 740] subgroup=64 basic=1 vote=1 ballot=1 shuffle=1\nloop_count = 64\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 9.56 max = 12.14 avg = 11.48\n squeezenet_int8 min = 6.78 max = 8.47 avg = 7.04\n mobilenet min = 11.59 max = 12.90 avg = 12.44\n mobilenet_int8 min = 8.69 max = 9.42 avg = 8.90\n mobilenet_v2 min = 14.00 max = 16.08 avg = 15.12\n mobilenet_v3 min = 16.66 max = 19.62 avg = 18.51\n shufflenet min = 8.72 max = 13.02 avg = 11.86\n shufflenet_v2 min = 12.82 max = 14.66 avg = 14.03\n mnasnet min = 15.06 max = 17.55 avg = 16.12\n proxylessnasnet min = 15.42 max = 17.28 avg = 16.59\n efficientnet_b0 min = 35.96 max = 41.24 avg = 37.89\n efficientnetv2_b0 min = 46.11 max = 65.75 avg = 58.52\n regnety_400m min = 22.07 max = 26.40 avg = 24.43\n blazeface min = 3.61 max = 6.26 avg = 4.53\n googlenet min = 32.60 max = 37.05 avg = 34.55\n googlenet_int8 min = 21.79 max = 30.65 avg = 24.84\n resnet18 min = 19.46 max = 24.26 avg = 22.76\n resnet18_int8 min = 38.09 max = 40.42 avg = 38.44\n alexnet min = 20.80 max = 28.44 avg = 26.86\n vgg16 min = 36.00 max = 44.01 avg = 39.18\n vgg16_int8 min = 201.54 max = 209.87 avg = 207.06\n resnet50 min = 42.50 max = 46.82 avg = 44.26\n resnet50_int8 min = 44.63 max = 47.47 avg = 45.15\n squeezenet_ssd min = 33.19 max = 36.74 avg = 34.62\n squeezenet_ssd_int8 min = 22.40 max = 31.99 avg = 25.65\n mobilenet_ssd min = 26.35 max = 29.79 avg = 28.09\n mobilenet_ssd_int8 min = 19.15 max = 20.86 avg = 19.48\n mobilenet_yolo min = 28.42 max = 31.16 avg = 29.06\n mobilenetv2_yolov3 min = 33.86 max = 36.54 avg = 35.36\n yolov4-tiny min = 46.51 max = 49.29 avg = 48.01\n nanodet_m min = 17.14 max = 19.79 avg = 18.49\n yolo-fastest-1.1 min = 9.49 max = 15.00 avg = 13.59\n yolo-fastestv2 min = 11.65 max = 15.61 avg = 14.36\n vision_transformer min = 650.85 max = 696.67 avg = 671.13\n FastestDet min = 8.63 max = 13.12 avg = 11.39\n```\n\n### MediaTek Dimensity 9300 (MT6989) (Cortex-X4 3.25 GHz + 2.85 GHz x 3 + Cortex-A720 2.0 GHz x 4 + Mali-G720-Immortalis MC12)\n```\nk6989v1_64:/data/local/tmp/benchmark # ../build-android/benchmark/benchncnn 8 8 0 -1 1 \nloop_count = 8\nnum_threads = 8\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 1.87 max = 2.18 avg = 2.01\n squeezenet_int8 min = 1.52 max = 1.98 avg = 1.77\n mobilenet min = 3.02 max = 3.34 avg = 3.15\n mobilenet_int8 min = 1.90 max = 2.27 avg = 2.04\n mobilenet_v2 min = 2.72 max = 3.13 avg = 2.89\n mobilenet_v3 min = 2.20 max = 3.82 avg = 2.78\n shufflenet min = 1.97 max = 2.56 avg = 2.20\n shufflenet_v2 min = 1.77 max = 2.29 avg = 1.96\n mnasnet min = 2.61 max = 3.48 avg = 2.90\n proxylessnasnet min = 2.72 max = 3.06 avg = 2.89\n efficientnet_b0 min = 4.57 max = 5.17 avg = 4.89\n efficientnetv2_b0 min = 5.24 max = 6.72 avg = 5.81\n regnety_400m min = 4.94 max = 6.78 avg = 5.70\n blazeface min = 0.80 max = 1.02 avg = 0.91\n googlenet min = 7.76 max = 8.53 avg = 8.12\n googlenet_int8 min = 5.68 max = 6.62 avg = 6.19\n resnet18 min = 5.35 max = 6.06 avg = 5.61\n resnet18_int8 min = 4.20 max = 4.40 avg = 4.29\n alexnet min = 5.96 max = 7.30 avg = 6.77\n vgg16 min = 29.27 max = 30.58 avg = 29.93\n vgg16_int8 min = 26.72 max = 28.12 avg = 27.27\n resnet50 min = 15.21 max = 19.16 avg = 16.09\n resnet50_int8 min = 8.57 max = 9.16 avg = 8.91\n squeezenet_ssd min = 6.29 max = 7.56 avg = 6.82\n squeezenet_ssd_int8 min = 5.57 max = 6.96 avg = 6.12\n mobilenet_ssd min = 6.90 max = 8.90 avg = 7.55\n mobilenet_ssd_int8 min = 4.53 max = 5.22 avg = 4.86\n mobilenet_yolo min = 16.88 max = 19.71 avg = 17.88\n mobilenetv2_yolov3 min = 10.51 max = 14.19 avg = 11.95\n yolov4-tiny min = 12.81 max = 16.23 avg = 14.22\n nanodet_m min = 4.38 max = 5.96 avg = 5.19\n yolo-fastest-1.1 min = 2.22 max = 3.08 avg = 2.73\n yolo-fastestv2 min = 2.09 max = 2.73 avg = 2.41\n vision_transformer min = 193.39 max = 203.13 avg = 198.32\n FastestDet min = 1.98 max = 2.35 avg = 2.16\nk6989v1_64:/data/local/tmp/benchmark # ../build-android/benchmark/benchncnn 8 4 2 -1 1 \nloop_count = 8\nnum_threads = 4\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 2.23 max = 2.31 avg = 2.27\n squeezenet_int8 min = 1.68 max = 1.73 avg = 1.70\n mobilenet min = 3.76 max = 3.86 avg = 3.81\n mobilenet_int8 min = 2.07 max = 2.16 avg = 2.11\n mobilenet_v2 min = 2.72 max = 2.95 avg = 2.80\n mobilenet_v3 min = 2.43 max = 2.51 avg = 2.47\n shufflenet min = 1.78 max = 1.87 avg = 1.81\n shufflenet_v2 min = 1.61 max = 1.66 avg = 1.63\n mnasnet min = 2.69 max = 2.82 avg = 2.76\n proxylessnasnet min = 2.95 max = 3.13 avg = 3.05\n efficientnet_b0 min = 4.99 max = 5.29 avg = 5.08\n efficientnetv2_b0 min = 5.73 max = 5.86 avg = 5.79\n regnety_400m min = 4.97 max = 5.04 avg = 5.00\n blazeface min = 1.07 max = 1.17 avg = 1.10\n googlenet min = 8.51 max = 9.43 avg = 8.75\n googlenet_int8 min = 6.01 max = 6.13 avg = 6.07\n resnet18 min = 6.72 max = 7.04 avg = 6.95\n resnet18_int8 min = 4.31 max = 4.40 avg = 4.34\n alexnet min = 7.41 max = 7.71 avg = 7.57\n vgg16 min = 33.77 max = 34.68 avg = 34.08\n vgg16_int8 min = 32.61 max = 33.83 avg = 33.12\n resnet50 min = 18.76 max = 19.53 avg = 19.05\n resnet50_int8 min = 9.56 max = 9.70 avg = 9.61\n squeezenet_ssd min = 6.86 max = 7.26 avg = 7.01\n squeezenet_ssd_int8 min = 5.42 max = 6.17 avg = 5.64\n mobilenet_ssd min = 8.38 max = 9.14 avg = 8.62\n mobilenet_ssd_int8 min = 4.60 max = 4.90 avg = 4.69\n mobilenet_yolo min = 19.59 max = 20.06 avg = 19.78\n mobilenetv2_yolov3 min = 10.46 max = 11.01 avg = 10.70\n yolov4-tiny min = 13.46 max = 14.18 avg = 13.86\n nanodet_m min = 4.52 max = 4.59 avg = 4.55\n yolo-fastest-1.1 min = 1.88 max = 1.94 avg = 1.91\n yolo-fastestv2 min = 1.73 max = 1.79 avg = 1.76\n vision_transformer min = 220.32 max = 229.49 avg = 223.92\n FastestDet min = 1.67 max = 1.73 avg = 1.70\nk6989v1_64:/data/local/tmp/benchmark # ../build-android/benchmark/benchncnn 8 4 1 -1 1 \nloop_count = 8\nnum_threads = 4\npowersave = 1\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 3.42 max = 4.25 avg = 3.62\n squeezenet_int8 min = 2.63 max = 2.78 avg = 2.73\n mobilenet min = 5.66 max = 6.25 avg = 5.82\n mobilenet_int8 min = 3.13 max = 5.66 avg = 3.58\n mobilenet_v2 min = 4.40 max = 4.46 avg = 4.42\n mobilenet_v3 min = 3.74 max = 4.07 avg = 3.94\n shufflenet min = 2.77 max = 2.86 avg = 2.82\n shufflenet_v2 min = 2.52 max = 2.62 avg = 2.57\n mnasnet min = 4.24 max = 4.37 avg = 4.28\n proxylessnasnet min = 4.65 max = 4.91 avg = 4.74\n efficientnet_b0 min = 7.71 max = 10.00 avg = 8.08\n efficientnetv2_b0 min = 9.24 max = 10.34 avg = 9.87\n regnety_400m min = 7.87 max = 8.35 avg = 8.02\n blazeface min = 2.38 max = 2.46 avg = 2.40\n googlenet min = 13.21 max = 13.78 avg = 13.40\n googlenet_int8 min = 10.23 max = 10.65 avg = 10.36\n resnet18 min = 9.25 max = 9.68 avg = 9.49\n resnet18_int8 min = 6.86 max = 6.97 avg = 6.91\n alexnet min = 9.73 max = 10.53 avg = 9.97\n vgg16 min = 47.43 max = 48.12 avg = 47.78\n vgg16_int8 min = 47.08 max = 48.18 avg = 47.46\n resnet50 min = 26.82 max = 27.14 avg = 26.99\n resnet50_int8 min = 15.01 max = 15.57 avg = 15.20\n squeezenet_ssd min = 9.96 max = 12.66 avg = 10.83\n squeezenet_ssd_int8 min = 8.47 max = 9.26 avg = 8.88\n mobilenet_ssd min = 12.54 max = 13.25 avg = 12.82\n mobilenet_ssd_int8 min = 7.03 max = 10.91 avg = 7.94\n mobilenet_yolo min = 29.73 max = 30.45 avg = 30.23\n mobilenetv2_yolov3 min = 16.64 max = 17.71 avg = 17.13\n yolov4-tiny min = 22.25 max = 22.65 avg = 22.45\n nanodet_m min = 7.56 max = 7.86 avg = 7.69\n yolo-fastest-1.1 min = 3.32 max = 3.45 avg = 3.39\n yolo-fastestv2 min = 2.76 max = 2.96 avg = 2.84\n vision_transformer min = 328.11 max = 337.26 avg = 332.12\n FastestDet min = 2.66 max = 2.77 avg = 2.71\nk6989v1_64:/data/local/tmp/benchmark # ../build-android/benchmark/benchncnn 8 1 2 -1 1 \nloop_count = 8\nnum_threads = 1\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 5.27 max = 5.35 avg = 5.32\n squeezenet_int8 min = 3.06 max = 3.22 avg = 3.16\n mobilenet min = 9.59 max = 9.85 avg = 9.74\n mobilenet_int8 min = 4.29 max = 4.45 avg = 4.37\n mobilenet_v2 min = 5.14 max = 5.33 avg = 5.20\n mobilenet_v3 min = 4.28 max = 4.54 avg = 4.42\n shufflenet min = 3.18 max = 3.34 avg = 3.27\n shufflenet_v2 min = 2.78 max = 3.23 avg = 3.05\n mnasnet min = 5.01 max = 5.38 avg = 5.19\n proxylessnasnet min = 6.11 max = 6.30 avg = 6.21\n efficientnet_b0 min = 11.53 max = 11.78 avg = 11.66\n efficientnetv2_b0 min = 13.88 max = 14.28 avg = 14.13\n regnety_400m min = 8.11 max = 8.18 avg = 8.16\n blazeface min = 0.99 max = 1.08 avg = 1.01\n googlenet min = 19.68 max = 20.71 avg = 20.25\n googlenet_int8 min = 13.42 max = 13.86 avg = 13.60\n resnet18 min = 18.10 max = 18.84 avg = 18.53\n resnet18_int8 min = 9.67 max = 10.17 avg = 9.99\n alexnet min = 15.76 max = 16.35 avg = 16.03\n vgg16 min = 70.22 max = 72.85 avg = 71.58\n vgg16_int8 min = 76.83 max = 79.70 avg = 78.45\n resnet50 min = 39.73 max = 41.24 avg = 40.30\n resnet50_int8 min = 20.76 max = 21.54 avg = 21.27\n squeezenet_ssd min = 12.63 max = 18.67 avg = 15.20\n squeezenet_ssd_int8 min = 10.29 max = 16.13 avg = 14.13\n mobilenet_ssd min = 17.21 max = 18.43 avg = 17.68\n mobilenet_ssd_int8 min = 8.92 max = 9.49 avg = 9.07\n mobilenet_yolo min = 37.45 max = 38.29 avg = 37.88\n mobilenetv2_yolov3 min = 19.18 max = 19.83 avg = 19.58\n yolov4-tiny min = 27.06 max = 27.86 avg = 27.45\n nanodet_m min = 9.33 max = 9.50 avg = 9.42\n yolo-fastest-1.1 min = 3.48 max = 3.59 avg = 3.54\n yolo-fastestv2 min = 2.29 max = 2.37 avg = 2.33\n vision_transformer min = 730.38 max = 739.99 avg = 735.77\n FastestDet min = 2.40 max = 2.48 avg = 2.43\nk6989v1_64:/data/local/tmp/benchmark # ../build-android/benchmark/benchncnn 64 1 2 0 0 \n[0 Mali-G720-Immortalis MC12] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 Mali-G720-Immortalis MC12] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Mali-G720-Immortalis MC12] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1\n[0 Mali-G720-Immortalis MC12] subgroup=16 basic/vote/ballot/shuffle=1/1/1/1\n[0 Mali-G720-Immortalis MC12] fp16-matrix-16_8_8/16_8_16/16_16_16=0/0/0\nloop_count = 64\nnum_threads = 1\npowersave = 2\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 11.26 max = 13.58 avg = 12.32\n squeezenet_int8 min = 3.08 max = 3.29 avg = 3.17\n mobilenet min = 11.96 max = 14.52 avg = 13.48\n mobilenet_int8 min = 4.20 max = 4.58 avg = 4.34\n mobilenet_v2 min = 13.62 max = 16.46 avg = 14.62\n mobilenet_v3 min = 13.98 max = 17.16 avg = 15.25\n shufflenet min = 10.22 max = 11.82 avg = 11.07\n shufflenet_v2 min = 12.42 max = 15.39 avg = 14.35\n mnasnet min = 12.94 max = 16.30 avg = 14.91\n proxylessnasnet min = 13.18 max = 16.55 avg = 15.05\n efficientnet_b0 min = 16.70 max = 20.35 avg = 18.27\n efficientnetv2_b0 min = 54.09 max = 70.05 avg = 58.68\n regnety_400m min = 16.20 max = 18.42 avg = 17.27\n blazeface min = 6.50 max = 7.86 avg = 6.93\n googlenet min = 15.29 max = 17.54 avg = 16.19\n googlenet_int8 min = 20.38 max = 22.08 avg = 20.98\n resnet18 min = 12.22 max = 15.63 avg = 14.27\n resnet18_int8 min = 9.50 max = 10.46 avg = 9.75\n alexnet min = 12.00 max = 16.09 avg = 13.65\n vgg16 min = 31.06 max = 32.77 avg = 31.85\n vgg16_int8 min = 115.72 max = 123.71 avg = 118.23\n resnet50 min = 15.74 max = 16.53 avg = 16.10\n resnet50_int8 min = 32.43 max = 33.78 avg = 33.07\n squeezenet_ssd min = 17.24 max = 21.80 avg = 20.68\n squeezenet_ssd_int8 min = 9.69 max = 10.52 avg = 9.97\n mobilenet_ssd min = 15.32 max = 17.63 avg = 16.62\n mobilenet_ssd_int8 min = 8.84 max = 9.54 avg = 9.05\n mobilenet_yolo min = 16.67 max = 18.21 avg = 17.25\n mobilenetv2_yolov3 min = 20.08 max = 25.40 avg = 23.12\n yolov4-tiny min = 21.98 max = 29.67 avg = 24.75\n nanodet_m min = 23.19 max = 29.95 avg = 25.69\n yolo-fastest-1.1 min = 15.07 max = 17.78 avg = 16.49\n yolo-fastestv2 min = 14.67 max = 16.07 avg = 15.44\n vision_transformer min = 768.04 max = 801.48 avg = 786.79\n FastestDet min = 8.33 max = 16.07 avg = 14.38\n```\n\n### Xeon Phi 3120A (1.10 GHz 57-core 228-thread)\n\n- Host: CentOS 7.9\n- Compiler: icc & icpc (ICC) 17.0.2 20170213\n- ncnn tag: 20240102\n\nBuild command\n\n```bash\n$ CC=icc CXX=icpc CFLAGS=\"-mmic\" CXXFLAGS=\"-mmic\" cmake .. -DCMAKE_BUILD_TYPE=Release -DNCNN_SSE2=OFF -DNCNN_AVX=OFF -DNCNN_AVX2=OFF\n```\n\nCopy the whole `ncnn` directory and libraries in `/opt/intel/compilers_and_libraries_2017/linux/lib/mic/lib` to `mic0`, then set the `LD_LIBRARY_PATH` environment variable. Some tools cannot be built, but `benchncnn` should work. The built `benchncnn` is for Intel Xeon Phi coprocessor (k1om).\n\n```bash\n[mizu-bai@DESKTOP-1D9QDE1-mic0 benchmark]$ file benchncnn \nbenchncnn: ELF 64-bit LSB executable, Intel Xeon Phi coprocessor (k1om), version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.32, not stripped\n```\n\nThe benchmark is run in the native mode, ssh into the Xeon Phi by `ssh user@mic0`, then run `benckncnn` as under general linux systems.\n\n```\n[mizu-bai@DESKTOP-1D9QDE1-mic0 benchmark]$ KMP_AFFINITY=scatter ../build/benchmark/benchncnn 4 56 0 -1 1\nloop_count = 4\nnum_threads = 56\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 43.42 max = 44.20 avg = 43.64\n squeezenet_int8 min = 161.92 max = 162.41 avg = 162.15\n mobilenet min = 44.49 max = 46.90 avg = 45.68\n mobilenet_int8 min = 230.47 max = 232.40 avg = 231.77\n mobilenet_v2 min = 57.22 max = 62.03 avg = 59.42\n mobilenet_v3 min = 301.16 max = 306.62 avg = 303.90\n shufflenet min = 65.80 max = 70.18 avg = 67.70\n shufflenet_v2 min = 49.54 max = 53.17 avg = 51.22\n mnasnet min = 521.87 max = 527.76 avg = 524.63\n proxylessnasnet min = 745.79 max = 748.55 avg = 746.92\n efficientnet_b0 min = 582.21 max = 584.64 avg = 583.34\n efficientnetv2_b0 min = 84.13 max = 86.13 avg = 85.19\n regnety_400m min = 209.67 max = 214.84 avg = 212.39\n blazeface min = 26.33 max = 27.39 avg = 26.74\n googlenet min = 124.14 max = 125.72 avg = 124.83\n googlenet_int8 min = 498.36 max = 502.37 avg = 500.29\n resnet18 min = 87.86 max = 88.83 avg = 88.35\n resnet18_int8 min = 359.50 max = 360.71 avg = 360.11\n alexnet min = 49.87 max = 51.25 avg = 50.76\n vgg16 min = 341.87 max = 343.92 avg = 342.42\n vgg16_int8 min = 1649.34 max = 1655.37 avg = 1652.98\n resnet50 min = 198.91 max = 202.32 avg = 200.58\n resnet50_int8 min = 983.48 max = 988.73 avg = 986.22\n squeezenet_ssd min = 108.33 max = 111.45 avg = 110.18\n squeezenet_ssd_int8 min = 368.96 max = 370.30 avg = 369.54\n mobilenet_ssd min = 98.29 max = 101.49 avg = 99.99\n mobilenet_ssd_int8 min = 462.18 max = 466.20 avg = 464.85\n mobilenet_yolo min = 262.42 max = 266.84 avg = 263.91\n mobilenetv2_yolov3 min = 159.20 max = 161.58 avg = 160.66\n yolov4-tiny min = 229.22 max = 230.48 avg = 229.87\n nanodet_m min = 115.10 max = 116.78 avg = 115.86\n yolo-fastest-1.1 min = 154.48 max = 155.33 avg = 154.79\n yolo-fastestv2 min = 161.10 max = 163.98 avg = 161.88\n vision_transformer min = 848.51 max = 863.03 avg = 854.92\n FastestDet min = 251.64 max = 253.22 avg = 252.38\n[mizu-bai@DESKTOP-1D9QDE1-mic0 benchmark]$ KMP_AFFINITY=scatter ../build/benchmark/benchncnn 4 112 0 -1 1\nloop_count = 4\nnum_threads = 112\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 41.07 max = 41.19 avg = 41.12\n squeezenet_int8 min = 161.73 max = 163.90 avg = 162.74\n mobilenet min = 36.82 max = 37.53 avg = 37.11\n mobilenet_int8 min = 231.50 max = 233.81 avg = 232.65\n mobilenet_v2 min = 53.12 max = 55.87 avg = 54.44\n mobilenet_v3 min = 277.82 max = 280.61 avg = 279.66\n shufflenet min = 64.11 max = 64.92 avg = 64.63\n shufflenet_v2 min = 48.23 max = 50.00 avg = 49.19\n mnasnet min = 532.09 max = 534.73 avg = 533.34\n proxylessnasnet min = 760.43 max = 763.94 avg = 762.34\n efficientnet_b0 min = 534.29 max = 547.51 avg = 541.29\n efficientnetv2_b0 min = 75.94 max = 76.88 avg = 76.39\n regnety_400m min = 226.37 max = 227.81 avg = 227.23\n blazeface min = 26.03 max = 26.93 avg = 26.51\n googlenet min = 106.53 max = 107.54 avg = 107.06\n googlenet_int8 min = 503.01 max = 505.16 avg = 504.13\n resnet18 min = 73.63 max = 76.61 avg = 75.11\n resnet18_int8 min = 358.18 max = 359.50 avg = 358.99\n alexnet min = 37.40 max = 38.17 avg = 37.83\n vgg16 min = 244.95 max = 250.05 avg = 247.24\n vgg16_int8 min = 1511.89 max = 1512.66 avg = 1512.35\n resnet50 min = 151.99 max = 154.66 avg = 153.37\n resnet50_int8 min = 954.16 max = 957.63 avg = 956.55\n squeezenet_ssd min = 91.46 max = 97.18 avg = 94.00\n squeezenet_ssd_int8 min = 368.03 max = 375.96 avg = 370.99\n mobilenet_ssd min = 79.61 max = 81.38 avg = 80.33\n mobilenet_ssd_int8 min = 458.93 max = 463.41 avg = 461.63\n mobilenet_yolo min = 234.59 max = 236.91 avg = 235.43\n mobilenetv2_yolov3 min = 145.82 max = 146.92 avg = 146.23\n yolov4-tiny min = 219.22 max = 220.51 avg = 219.83\n nanodet_m min = 109.43 max = 113.94 avg = 112.20\n yolo-fastest-1.1 min = 158.13 max = 160.59 avg = 159.20\n yolo-fastestv2 min = 162.05 max = 162.80 avg = 162.47\n vision_transformer min = 615.14 max = 625.35 avg = 618.47\n FastestDet min = 279.98 max = 282.49 avg = 281.14\n[mizu-bai@DESKTOP-1D9QDE1-mic0 benchmark]$ KMP_AFFINITY=scatter ../build/benchmark/benchncnn 4 224 0 -1 1\nloop_count = 4\nnum_threads = 224\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 45.54 max = 46.81 avg = 46.13\n squeezenet_int8 min = 186.81 max = 187.14 avg = 186.97\n mobilenet min = 38.33 max = 39.11 avg = 38.64\n mobilenet_int8 min = 251.06 max = 251.91 avg = 251.40\n mobilenet_v2 min = 56.57 max = 57.15 avg = 56.88\n mobilenet_v3 min = 365.04 max = 366.87 avg = 365.94\n shufflenet min = 71.16 max = 72.02 avg = 71.68\n shufflenet_v2 min = 52.14 max = 53.60 avg = 52.92\n mnasnet min = 596.37 max = 603.62 avg = 600.50\n proxylessnasnet min = 911.84 max = 912.23 avg = 912.04\n efficientnet_b0 min = 611.77 max = 614.32 avg = 612.69\n efficientnetv2_b0 min = 82.16 max = 83.05 avg = 82.62\n regnety_400m min = 253.43 max = 255.79 avg = 254.66\n blazeface min = 30.54 max = 30.91 avg = 30.70\n googlenet min = 111.68 max = 112.65 avg = 112.11\n googlenet_int8 min = 594.07 max = 597.09 avg = 596.03\n resnet18 min = 78.14 max = 79.12 avg = 78.75\n resnet18_int8 min = 412.69 max = 413.92 avg = 413.46\n alexnet min = 40.93 max = 41.43 avg = 41.17\n vgg16 min = 242.45 max = 244.46 avg = 243.47\n vgg16_int8 min = 1545.61 max = 1548.72 avg = 1547.47\n resnet50 min = 147.73 max = 148.56 avg = 148.07\n resnet50_int8 min = 1034.47 max = 1042.31 avg = 1038.41\n squeezenet_ssd min = 107.82 max = 110.53 avg = 108.98\n squeezenet_ssd_int8 min = 423.30 max = 426.91 avg = 425.67\n mobilenet_ssd min = 74.54 max = 77.13 avg = 75.97\n mobilenet_ssd_int8 min = 510.95 max = 513.33 avg = 512.40\n mobilenet_yolo min = 238.83 max = 239.64 avg = 239.27\n mobilenetv2_yolov3 min = 159.80 max = 160.31 avg = 160.04\n yolov4-tiny min = 233.89 max = 237.41 avg = 236.22\n nanodet_m min = 122.39 max = 123.42 avg = 122.89\n yolo-fastest-1.1 min = 194.49 max = 195.25 avg = 194.94\n yolo-fastestv2 min = 193.06 max = 195.03 avg = 194.05\n vision_transformer min = 547.36 max = 554.17 avg = 549.99\n FastestDet min = 317.76 max = 321.38 avg = 320.18\n```\n\n### PhytiumPi, Phytium E2000 (FTC664@1.8GHz x2 + FTC310@1.5GHz x2)\n```\nloop_count = 4\nnum_threads = 2\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 43.84 max = 43.95 avg = 43.88\n squeezenet_int8 min = 35.48 max = 35.77 avg = 35.66\n mobilenet min = 69.31 max = 70.03 avg = 69.66\n mobilenet_int8 min = 42.30 max = 42.40 avg = 42.35\n mobilenet_v2 min = 59.07 max = 59.35 avg = 59.19\n mobilenet_v3 min = 46.02 max = 46.37 avg = 46.19\n shufflenet min = 31.52 max = 31.61 avg = 31.56\n shufflenet_v2 min = 23.99 max = 24.07 avg = 24.04\n mnasnet min = 49.40 max = 50.45 avg = 49.92\n proxylessnasnet min = 53.24 max = 53.85 avg = 53.53\n efficientnet_b0 min = 77.49 max = 77.84 avg = 77.62\n efficientnetv2_b0 min = 88.51 max = 88.92 avg = 88.69\n regnety_400m min = 66.99 max = 67.05 avg = 67.03\n blazeface min = 7.74 max = 8.14 avg = 7.98\n googlenet min = 126.62 max = 127.23 avg = 126.91\n googlenet_int8 min = 102.87 max = 103.16 avg = 103.01\n resnet18 min = 102.28 max = 102.63 avg = 102.48\n resnet18_int8 min = 72.01 max = 72.45 avg = 72.29\n alexnet min = 76.00 max = 124.61 avg = 88.24\n vgg16 min = 597.75 max = 601.99 avg = 599.44\n vgg16_int8 min = 421.40 max = 423.83 avg = 423.01\n resnet50 min = 278.16 max = 280.64 avg = 279.37\n resnet50_int8 min = 207.26 max = 207.47 avg = 207.36\n squeezenet_ssd min = 108.69 max = 109.26 avg = 108.99\n squeezenet_ssd_int8 min = 84.05 max = 84.60 avg = 84.28\n mobilenet_ssd min = 141.65 max = 142.46 avg = 142.14\n mobilenet_ssd_int8 min = 84.43 max = 84.99 avg = 84.73\n mobilenet_yolo min = 322.53 max = 325.15 avg = 323.51\n mobilenetv2_yolov3 min = 194.84 max = 196.98 avg = 196.07\n yolov4-tiny min = 208.29 max = 213.26 avg = 210.77\n nanodet_m min = 64.78 max = 65.38 avg = 65.08\n yolo-fastest-1.1 min = 37.89 max = 38.23 avg = 38.07\n yolo-fastestv2 min = 29.75 max = 30.33 avg = 30.09\n vision_transformer min = 4257.71 max = 4263.73 avg = 4260.60\n FastestDet min = 30.86 max = 44.67 avg = 34.41\n```\n\n### AMD EPYC 7742 (2.25GHz) ubuntu 22.04 AOCC_4.2.0-Build#89\n\nsingle core\n\n```\n# nice -20 ../build-host-aocc-linux/benchmark/benchncnn 100 1 0 -1 0\nloop_count = 100\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 9.26 max = 10.05 avg = 9.45\n squeezenet_int8 min = 9.54 max = 13.35 avg = 9.67\n mobilenet min = 16.20 max = 16.83 avg = 16.35\n mobilenet_int8 min = 16.79 max = 17.28 avg = 16.89\n mobilenet_v2 min = 10.69 max = 11.13 avg = 10.78\n mobilenet_v3 min = 8.87 max = 14.09 avg = 9.03\n shufflenet min = 4.99 max = 5.29 avg = 5.06\n shufflenet_v2 min = 5.61 max = 7.14 avg = 5.66\n mnasnet min = 11.94 max = 12.39 avg = 12.05\n proxylessnasnet min = 13.48 max = 16.57 avg = 13.62\n efficientnet_b0 min = 19.58 max = 20.34 avg = 19.73\n efficientnetv2_b0 min = 22.66 max = 23.63 avg = 22.89\n regnety_400m min = 14.89 max = 18.76 avg = 15.11\n blazeface min = 1.45 max = 1.59 avg = 1.51\n googlenet min = 35.38 max = 36.94 avg = 35.79\n googlenet_int8 min = 30.55 max = 42.18 avg = 30.88\n resnet18 min = 34.73 max = 48.15 avg = 35.43\n resnet18_int8 min = 27.39 max = 28.22 avg = 27.61\n alexnet min = 31.42 max = 32.26 avg = 31.64\n vgg16 min = 160.38 max = 172.02 avg = 162.52\n vgg16_int8 min = 134.03 max = 153.69 avg = 135.12\n resnet50 min = 85.47 max = 87.90 avg = 86.21\n resnet50_int8 min = 71.18 max = 80.37 avg = 71.70\n squeezenet_ssd min = 24.66 max = 25.71 avg = 24.84\n squeezenet_ssd_int8 min = 23.61 max = 24.28 avg = 23.78\n mobilenet_ssd min = 34.48 max = 35.69 avg = 34.64\n mobilenet_ssd_int8 min = 33.26 max = 34.32 avg = 33.45\n mobilenet_yolo min = 77.25 max = 86.54 avg = 77.73\n mobilenetv2_yolov3 min = 41.72 max = 42.92 avg = 42.02\n yolov4-tiny min = 57.61 max = 59.49 avg = 58.46\n nanodet_m min = 12.92 max = 13.39 avg = 13.03\n yolo-fastest-1.1 min = 5.02 max = 5.26 avg = 5.11\n yolo-fastestv2 min = 5.06 max = 5.20 avg = 5.09\n vision_transformer min = 637.63 max = 670.46 avg = 640.60\n FastestDet min = 5.59 max = 5.82 avg = 5.66\n```\n\n64 cores\n\n```\n# nice -20 ../build-host-aocc-linux/benchmark/benchncnn 300 64 0 -1 0\nloop_count = 300\nnum_threads = 64\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 4.19 max = 13.94 avg = 5.06\n squeezenet_int8 min = 4.93 max = 13.59 avg = 5.14\n mobilenet min = 3.29 max = 5.28 avg = 3.39\n mobilenet_int8 min = 2.32 max = 3.32 avg = 2.40\n mobilenet_v2 min = 4.58 max = 8.64 avg = 4.76\n mobilenet_v3 min = 4.11 max = 6.89 avg = 4.88\n shufflenet min = 5.67 max = 8.60 avg = 5.92\n shufflenet_v2 min = 4.83 max = 6.29 avg = 5.02\n mnasnet min = 4.08 max = 12.75 avg = 4.29\n proxylessnasnet min = 4.46 max = 7.28 avg = 4.68\n efficientnet_b0 min = 5.51 max = 11.67 avg = 6.33\n efficientnetv2_b0 min = 7.50 max = 11.30 avg = 9.34\n regnety_400m min = 12.50 max = 20.88 avg = 12.76\n blazeface min = 1.67 max = 3.37 avg = 1.76\n googlenet min = 10.64 max = 11.59 avg = 10.87\n googlenet_int8 min = 8.49 max = 17.88 avg = 9.90\n resnet18 min = 6.36 max = 6.88 avg = 6.48\n resnet18_int8 min = 4.65 max = 13.03 avg = 4.77\n alexnet min = 3.88 max = 4.62 avg = 3.97\n vgg16 min = 26.00 max = 36.86 avg = 27.25\n vgg16_int8 min = 17.75 max = 19.63 avg = 18.42\n resnet50 min = 13.94 max = 23.10 avg = 14.17\n resnet50_int8 min = 8.73 max = 18.32 avg = 8.92\n squeezenet_ssd min = 10.39 max = 12.10 avg = 10.77\n squeezenet_ssd_int8 min = 11.53 max = 20.24 avg = 12.01\n mobilenet_ssd min = 6.80 max = 8.16 avg = 6.96\n mobilenet_ssd_int8 min = 4.98 max = 5.21 avg = 5.07\n mobilenet_yolo min = 17.75 max = 30.34 avg = 18.29\n mobilenetv2_yolov3 min = 13.74 max = 15.69 avg = 14.18\n yolov4-tiny min = 21.27 max = 29.53 avg = 22.81\n nanodet_m min = 10.22 max = 12.25 avg = 10.89\n yolo-fastest-1.1 min = 5.56 max = 6.03 avg = 5.66\n yolo-fastestv2 min = 5.61 max = 5.78 avg = 5.67\n vision_transformer min = 69.07 max = 508.15 avg = 71.73\n FastestDet min = 5.74 max = 6.83 avg = 5.81\n```\n\n### NVIDIA Tesla V100-PCIE-32GB (GV100 SM x 80 + Tensor Core x 640)\n\n```\n# ../build-host-gcc-vk-linux/benchmark/benchncnn 300 1 0 0 0\n[0 Tesla V100-PCIE-32GB] queueC=2[8] queueG=0[16] queueT=1[2]\n[0 Tesla V100-PCIE-32GB] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 Tesla V100-PCIE-32GB] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[0 Tesla V100-PCIE-32GB] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[0 Tesla V100-PCIE-32GB] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\n[1 llvmpipe (LLVM 15.0.7, 256 bits)] queueC=0[1] queueG=0[1] queueT=0[1]\n[1 llvmpipe (LLVM 15.0.7, 256 bits)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[1 llvmpipe (LLVM 15.0.7, 256 bits)] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[1 llvmpipe (LLVM 15.0.7, 256 bits)] subgroup=8 basic/vote/ballot/shuffle=1/1/1/1\n[1 llvmpipe (LLVM 15.0.7, 256 bits)] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\n[2 Tesla V100-PCIE-32GB] queueC=2[8] queueG=0[16] queueT=1[2]\n[2 Tesla V100-PCIE-32GB] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[2 Tesla V100-PCIE-32GB] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[2 Tesla V100-PCIE-32GB] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[2 Tesla V100-PCIE-32GB] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\n[3 Tesla V100-PCIE-32GB] queueC=2[8] queueG=0[16] queueT=1[2]\n[3 Tesla V100-PCIE-32GB] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[3 Tesla V100-PCIE-32GB] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[3 Tesla V100-PCIE-32GB] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[3 Tesla V100-PCIE-32GB] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\n[4 Tesla V100-PCIE-32GB] queueC=2[8] queueG=0[16] queueT=1[2]\n[4 Tesla V100-PCIE-32GB] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[4 Tesla V100-PCIE-32GB] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[4 Tesla V100-PCIE-32GB] subgroup=32 basic/vote/ballot/shuffle=1/1/1/1\n[4 Tesla V100-PCIE-32GB] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 300\nnum_threads = 1\npowersave = 0\ngpu_device = 0\ncooling_down = 0\n squeezenet min = 1.16 max = 16.79 avg = 1.64\n squeezenet_int8 min = 9.03 max = 10.06 avg = 9.15\n mobilenet min = 1.05 max = 2.60 avg = 1.25\n mobilenet_int8 min = 16.78 max = 19.89 avg = 16.93\n mobilenet_v2 min = 1.60 max = 3.29 avg = 1.76\n mobilenet_v3 min = 1.84 max = 8.43 avg = 2.04\n shufflenet min = 1.35 max = 3.73 avg = 1.54\n shufflenet_v2 min = 1.66 max = 8.02 avg = 1.93\n mnasnet min = 1.69 max = 3.31 avg = 1.82\n proxylessnasnet min = 1.74 max = 3.70 avg = 1.89\n efficientnet_b0 min = 2.86 max = 5.21 avg = 3.02\n efficientnetv2_b0 min = 60.41 max = 80.28 avg = 69.51\n regnety_400m min = 2.38 max = 6.84 avg = 2.57\n blazeface min = 0.85 max = 3.50 avg = 0.96\n googlenet min = 3.69 max = 16.66 avg = 4.10\n googlenet_int8 min = 33.66 max = 47.27 avg = 34.32\n resnet18 min = 1.76 max = 7.58 avg = 1.95\n resnet18_int8 min = 27.12 max = 36.43 avg = 27.62\n alexnet min = 1.33 max = 2.97 avg = 1.49\n vgg16 min = 2.98 max = 4.60 avg = 3.17\n vgg16_int8 min = 133.97 max = 154.41 avg = 136.22\n resnet50 min = 3.42 max = 17.05 avg = 3.72\n resnet50_int8 min = 70.53 max = 93.57 avg = 71.96\n squeezenet_ssd min = 16.88 max = 22.55 avg = 18.49\n squeezenet_ssd_int8 min = 23.12 max = 30.45 avg = 23.50\n mobilenet_ssd min = 5.44 max = 7.09 avg = 5.93\n mobilenet_ssd_int8 min = 33.28 max = 38.92 avg = 33.62\n mobilenet_yolo min = 5.67 max = 7.66 avg = 6.26\n mobilenetv2_yolov3 min = 6.33 max = 7.89 avg = 6.67\n yolov4-tiny min = 14.66 max = 17.29 avg = 15.57\n nanodet_m min = 5.36 max = 16.11 avg = 5.95\n yolo-fastest-1.1 min = 5.60 max = 7.45 avg = 6.13\n yolo-fastestv2 min = 3.48 max = 5.29 avg = 3.96\n vision_transformer min = 153.75 max = 198.81 avg = 165.58\n FastestDet min = 3.01 max = 5.01 avg = 3.29\n```\n\n### AXERA AX630C (Cortex-A53 1.2GHz * 2)\n\n```\n# ~/ncnn/build-aarch64-linux-gnu/benchmark # ./benchncnn 4 1 0 -1 0\nloop_count = 4\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 129.78 max = 130.30 avg = 130.09\n squeezenet_int8 min = 123.08 max = 123.48 avg = 123.22\n mobilenet min = 211.46 max = 221.68 avg = 214.14\n mobilenet_int8 min = 196.00 max = 212.73 avg = 200.23\n mobilenet_v2 min = 149.15 max = 149.21 avg = 149.17\n mobilenet_v3 min = 124.70 max = 125.54 avg = 125.08\n shufflenet min = 80.75 max = 80.88 avg = 80.81\n shufflenet_v2 min = 74.30 max = 74.50 avg = 74.37\n mnasnet min = 148.87 max = 165.85 avg = 153.26\n proxylessnasnet min = 203.05 max = 213.50 avg = 205.82\n efficientnet_b0 min = 270.39 max = 280.59 avg = 273.13\n efficientnetv2_b0 min = 302.93 max = 318.07 avg = 307.30\n regnety_400m min = 187.47 max = 187.90 avg = 187.60\n blazeface min = 22.64 max = 22.78 avg = 22.72\n googlenet min = 487.36 max = 503.50 avg = 493.93\n googlenet_int8 min = 418.16 max = 434.44 avg = 426.09\n resnet18_int8 min = 290.39 max = 301.90 avg = 293.70\n resnet50_int8 min = 888.81 max = 898.34 avg = 895.92\n squeezenet_ssd min = 320.78 max = 330.33 avg = 323.54\n squeezenet_ssd_int8 min = 281.52 max = 299.11 avg = 286.89\n mobilenet_ssd min = 435.79 max = 452.66 avg = 444.19\n mobilenet_ssd_int8 min = 394.38 max = 411.09 avg = 398.65\n mobilenet_yolo min = 955.48 max = 972.38 avg = 967.52\n mobilenetv2_yolov3 min = 519.47 max = 536.58 avg = 524.25\n yolo-fastestv2 min = 73.94 max = 74.15 avg = 74.05\n FastestDet min = 81.89 max = 82.07 avg = 81.98\n \n# ~/ncnn/build-aarch64-linux-gnu/benchmark # ./benchncnn 4 2 0 -1 0\nloop_count = 4\nnum_threads = 2\npowersave = 0\ngpu_device = -1\ncooling_down = 0\n squeezenet min = 75.14 max = 88.89 avg = 79.06\n squeezenet_int8 min = 70.11 max = 85.48 avg = 74.32\n mobilenet min = 112.72 max = 124.85 avg = 115.87\n mobilenet_int8 min = 100.35 max = 100.58 avg = 100.49\n mobilenet_v2 min = 85.92 max = 86.20 avg = 86.03\n mobilenet_v3 min = 73.94 max = 74.34 avg = 74.20\n shufflenet min = 53.99 max = 66.11 avg = 57.63\n shufflenet_v2 min = 47.47 max = 47.72 avg = 47.59\n mnasnet min = 85.96 max = 86.27 avg = 86.13\n proxylessnasnet min = 111.15 max = 121.84 avg = 113.92\n efficientnet_b0 min = 149.72 max = 150.00 avg = 149.85\n efficientnetv2_b0 min = 168.84 max = 170.57 avg = 169.35\n regnety_400m min = 120.42 max = 135.50 avg = 124.26\n blazeface min = 14.27 max = 14.48 avg = 14.39\n googlenet min = 263.82 max = 274.74 avg = 266.84\n googlenet_int8 min = 226.91 max = 227.36 avg = 227.23\n resnet18_int8 min = 157.66 max = 168.11 avg = 160.57\n resnet50_int8 min = 469.84 max = 484.00 avg = 476.59\n squeezenet_ssd min = 190.23 max = 204.41 avg = 193.99\n squeezenet_ssd_int8 min = 162.73 max = 174.30 avg = 165.79\n mobilenet_ssd min = 236.26 max = 251.16 avg = 240.34\n mobilenet_ssd_int8 min = 203.22 max = 212.01 avg = 206.00\n mobilenet_yolo min = 522.45 max = 537.99 avg = 529.95\n mobilenetv2_yolov3 min = 300.33 max = 316.59 avg = 304.89\n yolo-fastestv2 min = 50.27 max = 50.62 avg = 50.43\n FastestDet min = 53.34 max = 53.64 avg = 53.51\n```\n\n### Spacemit MUSE Pi Pro Spacemit M1 (Spacemit X60 *8 + PowerVR B-Series BXE-2-32 MC1)\n```\nroot@spacemit-k1-x-MUSE-Pi-Pro-board:/home/yingxi/ncnn/build/benchmark# ./benchncnn 4 8 2 -1 1\nloop_count = 4\nnum_threads = 8\npowersave = 2\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 192.55 max = 203.73 avg = 195.61\n squeezenet_int8 min = 863.38 max = 875.44 avg = 867.96\n mobilenet min = 260.32 max = 274.70 avg = 266.42\n mobilenet_int8 min = 1287.80 max = 1606.98 avg = 1461.52\n mobilenet_v2 min = 168.08 max = 173.99 avg = 169.97\n mobilenet_v3 min = 141.06 max = 166.83 avg = 147.74\n shufflenet min = 82.91 max = 92.83 avg = 85.57\n shufflenet_v2 min = 83.11 max = 83.35 avg = 83.26\n mnasnet min = 168.99 max = 180.35 avg = 171.95\n proxylessnasnet min = 186.14 max = 194.56 avg = 188.91\n efficientnet_b0 min = 257.93 max = 263.18 avg = 259.94\n efficientnetv2_b0 min = 385.35 max = 394.09 avg = 388.57\n regnety_400m min = 228.02 max = 229.55 avg = 228.88\n blazeface min = 26.78 max = 27.43 avg = 26.97\n googlenet min = 781.12 max = 796.37 avg = 788.60\n googlenet_int8 min = 2422.82 max = 2441.75 avg = 2432.78\n resnet18 min = 864.67 max = 874.15 avg = 869.32\n resnet18_int8 min = 2409.34 max = 2728.57 avg = 2530.44\n alexnet min = 389.93 max = 393.67 avg = 391.77\n vgg16 min = 8213.96 max = 8957.49 avg = 8405.27\n vgg16_int8 min = 34268.94 max = 36044.89 avg = 35244.72\n resnet50 min = 1798.75 max = 1859.80 avg = 1825.00\n resnet50_int8 min = 7364.21 max = 7500.24 avg = 7428.21\n squeezenet_ssd min = 693.59 max = 701.68 avg = 697.60\n squeezenet_ssd_int8 min = 1447.64 max = 1461.21 avg = 1455.02\n mobilenet_ssd min = 530.90 max = 542.81 avg = 534.42\n mobilenet_ssd_int8 min = 4347.45 max = 4391.44 avg = 4377.68\n mobilenet_yolo min = 1285.07 max = 1369.59 avg = 1312.64\n mobilenetv2_yolov3 min = 605.19 max = 628.05 avg = 616.37\n yolov4-tiny min = 1743.00 max = 1751.39 avg = 1748.09\n nanodet_m min = 201.46 max = 202.80 avg = 202.03\n yolo-fastest-1.1 min = 97.02 max = 98.29 avg = 97.71\n yolo-fastestv2 min = 75.53 max = 76.62 avg = 76.20\n vision_transformer min = 11328.10 max = 11334.80 avg = 11332.34\n FastestDet min = 85.01 max = 86.04 avg = 85.45\n\nroot@spacemit-k1-x-MUSE-Pi-Pro-board:/home/yingxi/ncnn/build/benchmark# ./benchncnn 4 8 2 0 1\n[0 PowerVR B-Series BXE-2-32 MC1] queueC=0[2] queueG=0[2] queueT=0[2]\n[0 PowerVR B-Series BXE-2-32 MC1] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0\n[0 PowerVR B-Series BXE-2-32 MC1] fp16-p/s/u/a=1/1/1/1 int8-p/s/u/a=1/1/1/1\n[0 PowerVR B-Series BXE-2-32 MC1] subgroup=1(1~1) ops=1/1/1/1/1/1/0/0/1/1\n[0 PowerVR B-Series BXE-2-32 MC1] fp16-8x8x16/16x8x8/16x8x16/16x16x16=0/0/0/0\nloop_count = 4\nnum_threads = 8\npowersave = 2\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 381.51 max = 382.05 avg = 381.73\n squeezenet_int8 min = 862.26 max = 890.38 avg = 879.94\n mobilenet min = 795.29 max = 796.41 avg = 795.80\n mobilenet_int8 min = 1284.16 max = 1298.86 avg = 1290.31\n mobilenet_v2 min = 512.00 max = 512.59 avg = 512.19\n mobilenet_v3 min = 428.55 max = 428.95 avg = 428.76\n shufflenet min = 198.17 max = 198.83 avg = 198.39\n shufflenet_v2 min = 272.36 max = 272.73 avg = 272.55\n mnasnet min = 526.92 max = 527.44 avg = 527.12\n proxylessnasnet min = 601.43 max = 602.65 avg = 602.05\n efficientnet_b0 min = 704.94 max = 705.23 avg = 705.13\n efficientnetv2_b0 min = 854.83 max = 866.51 avg = 859.85\n regnety_400m min = 526.46 max = 527.04 avg = 526.65\n blazeface min = 69.74 max = 69.84 avg = 69.80\n googlenet min = 1230.07 max = 1231.04 avg = 1230.53\n googlenet_int8 min = 2409.25 max = 2423.38 avg = 2416.76\n resnet18 min = 1134.72 max = 1136.35 avg = 1135.44\n resnet18_int8 min = 2431.48 max = 2552.62 avg = 2473.90\n alexnet min = 692.35 max = 697.08 avg = 695.61\n vgg16 min = 5790.33 max = 5805.37 avg = 5796.20\n vgg16_int8 min = 34057.43 max = 35714.99 avg = 35080.62\n resnet50 min = 3426.54 max = 3429.97 avg = 3427.94\n resnet50_int8 min = 7370.03 max = 7409.63 avg = 7390.83\n squeezenet_ssd min = 1057.50 max = 1061.42 avg = 1059.26\n squeezenet_ssd_int8 min = 1454.99 max = 1469.47 avg = 1462.61\n mobilenet_ssd min = 1670.02 max = 1673.22 avg = 1671.34\n mobilenet_ssd_int8 min = 4372.23 max = 4424.18 avg = 4400.11\n mobilenet_yolo min = 3794.02 max = 3796.52 avg = 3795.21\n mobilenetv2_yolov3 min = 1841.86 max = 1844.70 avg = 1843.49\n yolov4-tiny min = 2099.86 max = 2104.18 avg = 2102.34\n nanodet_m min = 646.19 max = 647.41 avg = 646.69\n yolo-fastest-1.1 min = 322.08 max = 323.71 avg = 323.22\n yolo-fastestv2 min = 209.42 max = 209.72 avg = 209.56\n vision_transformer min = 26499.86 max = 26548.73 avg = 26528.54\n FastestDet min = 251.68 max = 252.52 avg = 252.14\n```\n\n### Arduino UNO Q - QRB2210 (ARM Cortex-A53 @ 2.0GHz x 4)\n```\narduino@noivis-uno-q:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 4 0 -1 -1\nloop_count = 10\nnum_threads = 4\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 35.57 max = 111.57 avg = 43.99\n squeezenet_int8 min = 31.61 max = 32.34 avg = 31.91\n mobilenet min = 47.82 max = 133.12 avg = 56.77\n mobilenet_int8 min = 33.96 max = 102.49 avg = 44.91\n mobilenet_v2 min = 42.62 max = 119.38 avg = 51.88\n mobilenet_v3 min = 34.53 max = 35.91 avg = 35.27\n shufflenet min = 26.18 max = 26.47 avg = 26.32\n shufflenet_v2 min = 22.02 max = 88.82 avg = 30.98\n mnasnet min = 38.96 max = 92.30 avg = 50.98\n proxylessnasnet min = 47.04 max = 137.34 avg = 56.91\n efficientnet_b0 min = 58.75 max = 141.67 avg = 76.36\n efficientnetv2_b0 min = 79.72 max = 175.06 avg = 99.54\n regnety_400m min = 65.97 max = 184.19 avg = 96.94\n blazeface min = 6.43 max = 7.84 avg = 6.76\n googlenet min = 105.37 max = 197.46 avg = 130.49\n googlenet_int8 min = 89.68 max = 179.01 avg = 107.28\n resnet18 min = 86.52 max = 166.67 avg = 102.49\n resnet18_int8 min = 57.96 max = 107.52 avg = 66.63\n alexnet min = 56.77 max = 127.20 avg = 67.50\n vgg16 min = 463.45 max = 557.00 avg = 511.24\n vgg16_int8 min = 323.15 max = 415.10 avg = 367.00\n resnet50 min = 219.89 max = 298.83 avg = 250.55\n resnet50_int8 min = 177.14 max = 261.74 avg = 208.69\n squeezenet_ssd min = 96.95 max = 195.33 avg = 123.10\n squeezenet_ssd_int8 min = 79.66 max = 179.98 avg = 97.71\n mobilenet_ssd min = 100.40 max = 191.42 avg = 119.07\n mobilenet_ssd_int8 min = 71.88 max = 173.69 avg = 92.27\n mobilenet_yolo min = 216.49 max = 301.24 avg = 248.78\n mobilenetv2_yolov3 min = 154.69 max = 245.76 avg = 179.31\n yolov4-tiny min = 191.17 max = 261.76 avg = 218.64\n nanodet_m min = 57.66 max = 113.14 avg = 67.66\n yolo-fastest-1.1 min = 34.72 max = 131.85 avg = 49.81\n yolo-fastestv2 min = 26.91 max = 28.23 avg = 27.46\n vision_transformer min = 2529.77 max = 2703.20 avg = 2601.17\n FastestDet min = 28.09 max = 29.11 avg = 28.48\n\narduino@noivis-uno-q:~/ncnn/benchmark$ ../build/benchmark/benchncnn 10 1 0 -1 -1\nloop_count = 10\nnum_threads = 1\npowersave = 0\ngpu_device = -1\ncooling_down = 1\n squeezenet min = 94.15 max = 111.95 avg = 99.15\n squeezenet_int8 min = 78.23 max = 86.76 avg = 80.23\n mobilenet min = 146.45 max = 165.20 avg = 153.61\n mobilenet_int8 min = 123.70 max = 133.75 avg = 126.28\n mobilenet_v2 min = 99.85 max = 108.01 avg = 103.90\n mobilenet_v3 min = 93.31 max = 102.90 avg = 96.41\n shufflenet min = 61.80 max = 79.39 avg = 65.28\n shufflenet_v2 min = 47.57 max = 56.28 avg = 49.89\n mnasnet min = 106.41 max = 119.18 avg = 109.83\n proxylessnasnet min = 143.93 max = 164.33 avg = 151.37\n efficientnet_b0 min = 164.14 max = 173.38 avg = 167.91\n efficientnetv2_b0 min = 206.05 max = 225.26 avg = 211.93\n regnety_400m min = 133.84 max = 144.94 avg = 137.26\n blazeface min = 13.90 max = 14.97 avg = 14.25\n googlenet min = 337.11 max = 364.05 avg = 347.30\n googlenet_int8 min = 281.64 max = 293.46 avg = 288.34\n resnet18 min = 276.23 max = 304.36 avg = 289.94\n resnet18_int8 min = 190.11 max = 217.07 avg = 199.87\n alexnet min = 196.14 max = 203.26 avg = 198.63\n vgg16 min = 1391.13 max = 1626.54 avg = 1502.86\n vgg16_int8 min = 1128.65 max = 1290.60 avg = 1200.60\n resnet50 min = 739.44 max = 774.68 avg = 750.76\n resnet50_int8 min = 591.32 max = 612.44 avg = 603.38\n squeezenet_ssd min = 245.57 max = 280.32 avg = 262.18\n squeezenet_ssd_int8 min = 182.86 max = 228.61 avg = 199.68\n mobilenet_ssd min = 308.26 max = 320.81 avg = 314.58\n mobilenet_ssd_int8 min = 246.33 max = 265.22 avg = 253.05\n mobilenet_yolo min = 682.76 max = 703.99 avg = 696.30\n mobilenetv2_yolov3 min = 346.53 max = 365.76 avg = 355.41\n yolov4-tiny min = 527.86 max = 558.38 avg = 542.25\n nanodet_m min = 135.87 max = 153.99 avg = 145.11\n yolo-fastest-1.1 min = 58.92 max = 76.24 avg = 65.08\n yolo-fastestv2 min = 48.54 max = 59.97 avg = 53.21\n vision_transformer min = 9218.64 max = 10723.27 avg = 10253.49\n FastestDet min = 51.52 max = 62.65 avg = 55.04\n```\n\n" }, { "path": "benchmark/RankCards/CMakeLists.txt", "content": "cmake_minimum_required(VERSION 3.10)\n\nproject(RankCards CXX)\n\nset(CMAKE_CXX_STANDARD 11)\nset(CMAKE_CXX_STANDARD_REQUIRED ON)\nset(EXECUTABLE_OUTPUT_PATH \"../\")\n\nadd_executable(RankCards main.cpp)\n" }, { "path": "benchmark/RankCards/README.md", "content": "### Rank the boards.\nThe table below is generated by RankCards, using the timings found in the /ncnn/benchmark/README.md file.
\nFirst, the best set of timings is selected from each board.
\nThe set is then compared to a reference set by calculating the ratio of each model one by one and averaging all results.
\nFinally, the boards are ranked from fast to slow.
\n| | Board | Ratio | \n| :--: | :---- | :--- | \n| 1 | NVIDIA Quadro RTX 8000 (TU102 SM x 72 + Tensor Core x 576) | 0.147 | \n| 2 | nVIDIA RTX2080 of Desktop | 0.15 | \n| 3 | NVIDIA GeForce RTX 3060 Ti of Desktop[2023-10-12] | 0.18 | \n| 4 | nVIDIA RTX2060 of Notebook | 0.198 | \n| 5 | Intel® Core™ i7-13700K of Desktop[2023-10-12] | 0.255 | \n| 6 | AMD Radeon RX 6900 XT of Desktop[2023-10-12] | 0.275 | \n| 7 | NVIDIA RTX3090 (GA102 SM x 82 + Tensor Core 328) | 0.277 | \n| 8 | MediaTek Dimensity 9300 (MT6989) (Cortex-X4 3.25 GHz + 2.85 GHz x 3 + Cortex-A720 2.0 GHz x 4 + Mali-G720-Immortalis MC12) | 0.309 | \n| 9 | MacBook Pro (13-inch, M1, 2020) | 0.346 | \n| 10 | AWS c5.4xlarge Instance | 0.418 | \n| 11 | AMD Ryzen 9 5950X 16-Core of Desktop[2023-10-12] | 0.427 | \n| 12 | Qualcomm SM8550-AB Snapdragon 8 Gen 2 (Kyro 3.20 GHz + 2.8 GHz x 2 + 2.80 GHz x 2 + 2.00 GHz * 3 + Adreno 740) | 0.45 | \n| 13 | AMD Ryzen 5700g (Zen3 3.8 GHz ~ 4.6 GHz x 8) | 0.478 | \n| 14 | HUAWEI KunPeng 920 3211K (x24 cores) | 0.482 | \n| 15 | NVIDIA Jetson AGX Orin (Cortex-A78AE 2.2 GHz x 12 + Ampere@1.3 GHz Tensor Cores 64) | 0.485 | \n| 16 | HUAWEI KunPeng 920 2251K (x8 cores) | 0.54 | \n| 17 | nVIDIA RTX A3000 of Notebook (6GB) | 0.577 | \n| 18 | Intel(R) UHD Graphics 770 of Desktop[2023-10-12] | 0.593 | \n| 19 | OrangePi5, Rockchip RK3588s (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz) | 0.642 | \n| 20 | Qualcomm SM8150-AC Snapdragon 855+ (Kyro485 2.96 GHz + 2.42 GHz x 3 + 1.80 GHz x 4 + Adreno 640) | 0.665 | \n| 21 | Rockchip RK3588 (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz) | 0.753 | \n| 22 | NVIDIA Jetson Orin Nano | 0.819 | \n| 23 | Raspberry Pi 5 Broadcom BCM2712, Cortex-A76 (ARMv8) (2.4GHz x 4) | 1 | \n| 24 | Station-M3/ROC-RK3588S-PC, Rockchip RK3588S (Quad Core A76 2.4GHz + Quad Core A55 1.8GHz + Mali-G610) StationOS (Android) | 1 | \n| 25 | NVIDIA Jetson AGX Xavier (Carmel 2.2 GHz x 8 + Volta Tensor Cores 64) | 1.05 | \n| 26 | Loongson 3A6000 (LA664 2.5GHz * 4+4) | 1.11 | \n| 27 | Hyper-V Linux Guest with GPU-PV enabled (Intel Core i7-11800H, NVIDIA GeForce RTX 3070 Laptop GPU) | 1.19 | \n| 28 | Rockchip RK3588 (Cortex-A76 2.4GHz x 4 + Cortex-A55 1.8GHz x 4) | 1.35 | \n| 29 | NVIDIA Jetson TX2 NX(NV-Denver2 2.0Ghz x 2 + Cortex-A57 2.0Ghz x 4 + 256-core NVIDIA Pascal iGPU) | 1.59 | \n| 30 | Hyper-V Linux Guest with GPU-PV enabled (Intel Core i7-7700K, NVIDIA GeForce GTX 1050 Ti) | 1.66 | \n| 31 | Phytium FT-2000+/64 (FTC662 armv8 2.4GHz x 8) | 1.75 | \n| 32 | AMD Ryzen Threadripper 3970X (Zen2 3.7 GHz ~ 4.5 GHz x 32) | 2.19 | \n| 33 | AMD Ryzen Embedded V1605B (Zen 2.0 GHz ~ 3.6 GHz x 4 + Radeon Vega 8 1.1GHz 8CU) | 2.23 | \n| 34 | Avaota Aim T527, Allwinner T527 (Cortex-A55 2.2GHz x 4 + Cortex-A55 1.8GHz x 4) | 2.28 | \n| 35 | Loongson 3A5000 (LA464 2.5GHz * 4) | 2.31 | \n| 36 | Qualcomm MSM8996 Pro Snapdragon 821 (Kyro 2.35GHz x 2 + Kyro 2.19GHz x 2) | 2.37 | \n| 37 | NVIDIA Jetson Nano | 2.44 | \n| 38 | Intel Celeron N5105 | 2.8 | \n| 39 | Loongson 3A4000 (GS464V 1.8GHz * 4 with MSA128) | 3.24 | \n| 40 | Khadas VIM3, Amlogic A311D (Cortex-A73 2.2GHz x 4 + Cortex-A53 1.8GHz x 2) | 3.48 | \n| 41 | Kirin 970 (Cortex-A73 2.4GHz x 4 + Cortex-A53 1.8GHz x 4) | 3.58 | \n| 42 | Qualcomm MSM8998 Snapdragon 835 (Kyro 2.45GHz x 4 + Kyro 1.9GHz x 4 + Adreno 540) | 3.63 | \n| 43 | MacBook Pro (15-inch, 2019) - 2.6GHz six cores Intel Core i7 && Radeon Pro 555X 4GB && Intel UHD Graphics 630 1536MB | 3.75 | \n| 44 | Qualcomm MSM6150 Snapdragon 675 (Kyro460 2.0GHz x 2 + Kyro460 1.7GHz x 6 + Adreno 612) | 3.75 | \n| 45 | Qualcomm MSM8994 Snapdragon 810 (Cortex-A57 2.0GHz x 4 + Cortex-A53 1.55GHz x 4) | 3.82 | \n| 46 | Station P2, Rockchip RK3568 (Cortex-A55 2.0GHz x 4) | 3.85 | \n| 47 | Rock3A, Rockchip RK3568 (Cortex-A55 2.0GHz x 4) ubuntu 20.04 | 3.86 | \n| 48 | Loongson 3A4000 (GS464V 1.8GHz * 4 with MSA128) | 4.08 | \n| 49 | Radxa Zero 3W, Cortex-A55 (ARMv82) (1.416 GHz x 4) | 4.5 | \n| 50 | Raspberry Pi 4 Model B Broadcom BCM2711B0, Cortex-A72 (ARMv8) (1.8GHz x 4) | 4.95 | \n| 51 | OrangePi4 LTS, Rockchip RK3399 (Cortex-A72 1.8GHz x 2 + Cortex-A53 1.5GHz x 4) | 5.11 | \n| 52 | Rockchip RK3399 (Cortex-A72 1.8GHz x 2 + Cortex-A53 1.5GHz x 4) | 5.16 | \n| 53 | PhytiumPi, Phytium E2000 (FTC664@1.8GHz x2 + FTC310@1.5GHz x2) | 5.16 | \n| 54 | Qualcomm SDM660 Snapdragon 660 (Kyro260 2.2GHz x 4 + Kyro260 1.84GHz x 4 + Adreno 512) | 5.26 | \n| 55 | Phytium FT-2000/4 (FTC663 armv8 2.2GHz x 4) | 5.27 | \n| 56 | RDK X3 Module (Cortex-A53 1.5GHz x 4) aarch64 | 5.88 | \n| 57 | Station-M2/ROC-RK3566-PC, Rockchip RK3566 (Cortex-A55 1.8GHz x 4 + Mali-G52) StationOS (Android) | 6.51 | \n| 58 | Rockchip RK3288-CG.W (Cortex-A17 1.8GHz x 4) | 6.66 | \n| 59 | Qualcomm MSM8916 Snapdragon 410 (Cortex-A53 1.2GHz x 4) | 7.63 | \n| 60 | NanoPi R2S, Rockchip RK3328 (Cortex-A53 1.3GHz x 4) Armbian focal (21.05.1) aarch64 | 7.66 | \n| 61 | Intel Atom x5-Z8350 | 7.74 | \n| 62 | Loongson 2K2000 (LA364 1.5GHz * 2 with lsx) | 8.23 | \n| 63 | EAIDK 310, Rockchip RK3228H (Cortex-A53 1.3GHz x 4) fedora-28 aarch64 | 8.34 | \n| 64 | OrangePi Zero 2, Allwinner H616 (Cortex-A53 1.5GHz x 4) | 9.51 | \n| 65 | Raspberry Pi 3 Model B+ Broadcom BCM2837B0, Cortex-A53 (ARMv8) (1.4GHz x 4) | 9.87 | \n| 66 | iPhone 5S (Apple A7 1.3GHz x 2) | 11 | \n| 67 | MYIR RemiPi,Renesas RZG2L(Cortex-A55 1.5GHz x 2) | 11.9 | \n| 68 | Raspberry Pi 5 Broadcom BCM2712, VideoCore VII Graphics (Vulkan 1.2) | 12.5 | \n| 69 | Raspberry Pi Zero 2 W Broadcom BCM2710A1, Cortex-A53 (ARMv8) (1.0GHz x 4) | 13.7 | \n| 70 | Xeon Phi 3120A (1.10 GHz 57-core 228-thread) | 15.1 | \n| 71 | Loongson 3A3000 (GS464E 1.45GHz * 4) | 16.3 | \n| 72 | AXERA AX620A (Cortex-A7 1.0GHz * 4) | 18.8 | \n| 73 | Loongson 2K1000LA (LA264 1.0GHz * 2) | 24.4 | \n| 74 | Loongson 2K1000 (GS264 1.0GHz x 2) | 24.8 | \n| 75 | Freescale i.MX7 Dual (Cortex A7 1.0GHz x 2) | 26.7 | \n| 76 | Banana Pi M2 Zero 2 AllWinner H2+, Cortex-A7 (ARMv7-A) (1.2GHz x 4) | 26.8 | \n| 77 | HiSilicon Hi3519V101 (Cortex-A17 1.2GHz x 1) | 36.2 | \n| 78 | Sunway SW831 (sw_64 2.5GHz * 8) | 40.7 | \n| 79 | Z7-Lite 7020 XC7Z020CLG400-2 (Cortex-A9 766MHz x 2) | 43.2 | \n| 80 | Intel Celeron M 420 (Yonah 1.60 GHz x 1) | 43.9 | \n| 81 | Amlogic S805 (Cortex-A5, 4 × 1.536GHz) | 45.9 | \n| 82 | VisionFive2 , JH7110 (SiFive-U74(RV64GC) 1.5GHz x 4) riscv64 with PowerVR B-Series BXE-4-32 | 72.4 | \n| 83 | T-Head TH1520 (C910V, 1.848 GHz x 4 + BXM-4-64 PowerVR) | 83.3 | \n| 84 | Sunway SW421 (sw_64 1.7GHz * 4) | 116 | \n| 85 | Ingenic T40XP Xburst2 Core X2 1.4Ghz (without MSA) | 165 | \n" }, { "path": "benchmark/RankCards/Rcards.h", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n#ifndef RCARDS_H\n#define RCARDS_H\n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\n//---------------------------------------------------------------------------\n// Global hardcoded parameters\n//---------------------------------------------------------------------------\n// LERP(a,b,c) = linear interpolation macro, is 'a' when c == 0.0 and 'b' when c == 1.0 */\n#define MIN(a, b) ((a) > (b) ? (b) : (a))\n#define MAX(a, b) ((a) < (b) ? (b) : (a))\n#define LIM(a, b, c) (((a) > (c)) ? (c) : ((a) < (b)) ? (b) : (a))\n#define LERP(a, b, c) (((b) - (a)) * (c) + (a))\n#define ROUND(a) (static_cast((a) + 0.5))\n#define EUCLIDEAN(x1, y1, x2, y2) sqrt(((x1) - (x2)) * ((x1) - (x2)) + ((y1) - (y2)) * ((y1) - (y2)))\n//---------------------------------------------------------------------------\nstruct TModel\n{\n std::string Name;\n float AvrTime{0.0};\n};\n//---------------------------------------------------------------------------\nstruct TModelSet\n{\n std::vector Mset;\n\n //use push_back to prevent issues with CMake\n inline TModelSet(void)\n {\n TModel model;\n model.Name = \"squeezenet\";\n Mset.push_back(model);\n model.Name = \"squeezenet_int8\";\n Mset.push_back(model);\n model.Name = \"mobilenet\";\n Mset.push_back(model);\n model.Name = \"mobilenet_int8\";\n Mset.push_back(model);\n model.Name = \"mobilenet_v2\";\n Mset.push_back(model);\n model.Name = \"mobilenet_v3\";\n Mset.push_back(model);\n model.Name = \"shufflenet\";\n Mset.push_back(model);\n model.Name = \"shufflenet_v2\";\n Mset.push_back(model);\n model.Name = \"mnasnet\";\n Mset.push_back(model);\n model.Name = \"proxylessnasnet\";\n Mset.push_back(model);\n model.Name = \"efficientnet_b0\";\n Mset.push_back(model);\n model.Name = \"efficientnetv2_b0\";\n Mset.push_back(model);\n model.Name = \"regnety_400m\";\n Mset.push_back(model);\n model.Name = \"blazeface\";\n Mset.push_back(model);\n model.Name = \"googlenet\";\n Mset.push_back(model);\n model.Name = \"googlenet_int8\";\n Mset.push_back(model);\n model.Name = \"resnet18\";\n Mset.push_back(model);\n model.Name = \"resnet18_int8\";\n Mset.push_back(model);\n model.Name = \"alexnet\";\n Mset.push_back(model);\n model.Name = \"vgg16\";\n Mset.push_back(model);\n model.Name = \"vgg16_int8\";\n Mset.push_back(model);\n model.Name = \"resnet50\";\n Mset.push_back(model);\n model.Name = \"resnet50_int8\";\n Mset.push_back(model);\n model.Name = \"squeezenet_ssd\";\n Mset.push_back(model);\n model.Name = \"squeezenet_ssd_int8\";\n Mset.push_back(model);\n model.Name = \"mobilenet_ssd\";\n Mset.push_back(model);\n model.Name = \"mobilenet_ssd_int8\";\n Mset.push_back(model);\n model.Name = \"mobilenet_yolo\";\n Mset.push_back(model);\n model.Name = \"mobilenetv2_yolov3\";\n Mset.push_back(model);\n model.Name = \"yolov4-tiny\";\n Mset.push_back(model);\n model.Name = \"nanodet_m\";\n Mset.push_back(model);\n model.Name = \"yolo-fastest-1.1\";\n Mset.push_back(model);\n model.Name = \"yolo-fastestv2\";\n Mset.push_back(model);\n model.Name = \"vision_transformer\";\n Mset.push_back(model);\n model.Name = \"FastestDet\";\n Mset.push_back(model);\n }\n\n void Store(const TModel& model)\n {\n for (size_t i = 0; i < Mset.size(); i++)\n {\n if (Mset[i].Name == model.Name)\n {\n Mset[i].AvrTime = model.AvrTime;\n break;\n }\n }\n }\n\n float Sum(void)\n {\n float t = 0;\n\n for (size_t i = 0; i < Mset.size(); i++) t += Mset[i].AvrTime;\n\n return t;\n }\n\n float Ratio(const TModelSet& Rset)\n {\n float w;\n float s = 0;\n float t = 0;\n\n for (size_t r = 0; r < Rset.Mset.size(); r++)\n {\n if (Rset.Mset[r].AvrTime > 0.0)\n {\n for (size_t i = 0; i < Mset.size(); i++)\n {\n if (Mset[i].AvrTime > 0.0)\n {\n if (Mset[i].Name == Rset.Mset[r].Name)\n {\n w = log(Rset.Mset[r].AvrTime);\n s += w * (Mset[i].AvrTime / Rset.Mset[r].AvrTime);\n t += w;\n }\n }\n }\n }\n }\n if (t > 0) s /= t;\n return s;\n }\n};\n//---------------------------------------------------------------------------\nstruct TBoard\n{\n std::string Name;\n size_t StartLine;\n size_t EndLine;\n std::vector BenchSet;\n int BestSet;\n float Ratio;\n};\n//---------------------------------------------------------------------------\ninline bool FileExists(const std::string& name)\n{\n struct stat buffer;\n return (stat(name.c_str(), &buffer) == 0);\n}\n//---------------------------------------------------------------------------\ninline void FileCopy(const std::string& Src, const std::string& Dst)\n{\n std::ifstream src(Src, std::ios::binary);\n std::ofstream dst(Dst, std::ios::binary);\n\n dst << src.rdbuf();\n}\n//---------------------------------------------------------------------------\n// to lower case\nstatic inline void lcase(std::string& s)\n{\n std::transform(s.begin(), s.end(), s.begin(),\n [](unsigned char c) {\n return std::tolower(c);\n });\n}\n//---------------------------------------------------------------------------\n// to lower case (copying)\nstatic inline std::string lcase_copy(std::string s)\n{\n lcase(s);\n return s;\n}\n//---------------------------------------------------------------------------\n// trim from start (in place)\nstatic inline void ltrim(std::string& s)\n{\n s.erase(s.begin(), std::find_if(s.begin(), s.end(), [](int ch) {\n return !std::isspace(ch);\n }));\n}\n//---------------------------------------------------------------------------\n// trim from end (in place)\nstatic inline void rtrim(std::string& s)\n{\n s.erase(std::find_if(s.rbegin(), s.rend(), [](int ch) {\n return !std::isspace(ch);\n }).base(),\n s.end());\n}\n//---------------------------------------------------------------------------\n// trim from both ends (in place)\nstatic inline void trim(std::string& s)\n{\n ltrim(s);\n rtrim(s);\n}\n//---------------------------------------------------------------------------\n// trim from start (copying)\nstatic inline std::string ltrim_copy(std::string s)\n{\n ltrim(s);\n return s;\n}\n//---------------------------------------------------------------------------\n// trim from end (copying)\nstatic inline std::string rtrim_copy(std::string s)\n{\n rtrim(s);\n return s;\n}\n//---------------------------------------------------------------------------\n// trim from both ends (copying)\nstatic inline std::string trim_copy(std::string s)\n{\n trim(s);\n return s;\n}\n//---------------------------------------------------------------------------\nstatic inline void GetNameAver(std::string line, TModel& model)\n{\n // line example: squeezenet min = 46.28 max = 46.91 avg = 46.65\n\n size_t p = line.find(\"min =\");\n\n if (p != std::string::npos)\n {\n model.Name = trim_copy(line.substr(0, p));\n p = line.find(\"avg =\");\n if (p != std::string::npos)\n {\n try\n {\n model.AvrTime = std::stof(trim_copy(line.substr(p + 5, line.length() - p - 5)));\n }\n catch (...)\n {\n }\n }\n else\n model.AvrTime = 0.0;\n }\n else\n {\n model.Name = \"\";\n model.AvrTime = 0.0;\n }\n}\n//---------------------------------------------------------------------------\n#endif // RCARDS_H\n" }, { "path": "benchmark/RankCards/main.cpp", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \"Rcards.h\"\n//---------------------------------------------------------------------------\nusing namespace std;\n//---------------------------------------------------------------------------\n#define REF_BOARD \"Raspberry Pi 5 Broadcom BCM2712, Cortex-A76 (ARMv8)\"\n//---------------------------------------------------------------------------\n// Define a custom comparator function for sorting based on Ratio\nbool compareByRatio(const TBoard& a, const TBoard& b)\n{\n return a.Ratio < b.Ratio;\n}\n//---------------------------------------------------------------------------\nint main(int argc, char** argv)\n{\n size_t i, t, n, r;\n int RefBoard;\n float f, x;\n string Line;\n TModel Model;\n vector Lines; // Vector to store strings\n vector Boards; // Vector to store boards\n ifstream inputFile;\n\n // Check existence of the ../README.md file\n inputFile.open(\"../README.md\");\n if (!inputFile.is_open())\n {\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: ./RankCards \\n\");\n return -1;\n }\n const char* imagepath = argv[1];\n // Open the file given as argument\n inputFile.open(imagepath);\n // Check if the file is open\n if (!inputFile.is_open())\n {\n cerr << \"Error opening file\" << endl;\n return 1; // Return an error code\n }\n }\n\n // Read each Line from the file and add it to the vector\n while (std::getline(inputFile, Line))\n {\n Lines.push_back(Line);\n }\n // Close the file\n inputFile.close();\n\n // Get the boards.\n for (i = 0; i < Lines.size(); i++)\n {\n TBoard Brd;\n if (Lines[i].find(\"###\") != string::npos)\n {\n Brd.Name = Lines[i].substr(4, Lines[i].length() - 4);\n Brd.StartLine = i + 1;\n Boards.push_back(Brd);\n }\n }\n // Get the boards end Line.\n for (t = 0; t < Boards.size() - 1; t++)\n {\n Boards[t].EndLine = Boards[t + 1].StartLine;\n }\n Boards[t].EndLine = Lines.size();\n\n // Get the bench sets (must always start with squeezenet)\n for (t = 0; t < Boards.size(); t++)\n {\n TModelSet MdSet;\n bool FirstSet = true;\n for (n = Boards[t].StartLine; n < Boards[t].EndLine; n++)\n {\n GetNameAver(Lines[n], Model);\n MdSet.Store(Model);\n\n if (Model.Name == \"squeezenet\")\n {\n //start of new set, check if it is the first set\n if (FirstSet)\n FirstSet = false;\n else\n Boards[t].BenchSet.push_back(MdSet);\n }\n }\n Boards[t].BenchSet.push_back(MdSet);\n }\n\n // Get the total AvrTime of the bench sets and set the lowest as best set\n for (t = 0; t < Boards.size(); t++)\n {\n x = FLT_MAX;\n for (n = 0; n < Boards[t].BenchSet.size(); n++)\n {\n f = Boards[t].BenchSet[n].Sum();\n if (f < x)\n {\n x = f;\n Boards[t].BestSet = n;\n }\n }\n }\n\n // Get the reference set\n RefBoard = -1;\n for (t = 0; t < Boards.size(); t++)\n {\n if (Boards[t].Name.find(REF_BOARD) != string::npos)\n {\n RefBoard = static_cast(t);\n }\n }\n if (RefBoard == -1)\n {\n cerr << \"Error finding reference board :\" << endl;\n cerr << REF_BOARD << endl;\n return 1; // Return an error code\n }\n\n // Get the ratios between the best bench sets and reference\n r = Boards[RefBoard].BestSet;\n for (t = 0; t < Boards.size(); t++)\n {\n n = Boards[t].BestSet;\n Boards[t].Ratio = Boards[t].BenchSet[n].Ratio(Boards[RefBoard].BenchSet[r]);\n }\n\n // Sort the vector using the custom comparator\n std::sort(Boards.begin(), Boards.end(), compareByRatio);\n\n // Open an output README.md file\n std::ofstream outputFile(\"README.md\");\n\n // Check if the file is successfully opened\n if (outputFile.is_open())\n {\n outputFile << \"### Rank the boards.\" << endl;\n outputFile << \"The table below is generated by RankCards, using the timings found in the /ncnn/benchmark/README.md file.
\" << endl;\n outputFile << \"First, the best set of timings is selected from each board.
\" << endl;\n outputFile << \"The set is then compared to a reference set by calculating the ratio of each model one by one and averaging all results.
\" << endl;\n outputFile << \"Finally, the boards are ranked from fast to slow.
\" << endl;\n outputFile << \"| | Board | Ratio | \" << endl;\n outputFile << \"| :--: | :---- | :--- | \" << endl;\n // Write the sorted vector to the file\n for (t = 0; t < Boards.size(); t++)\n {\n outputFile << \"| \" << t + 1 << \" | \" << Boards[t].Name << \" | \" << setprecision(3) << Boards[t].Ratio << \" | \" << endl;\n }\n // Close the file stream\n outputFile.close();\n cout << \"Sorted data has been written to README.md\" << endl;\n }\n else\n {\n cerr << \"Error opening the file.\" << endl;\n return 1; // Return an error code\n }\n\n return 0; // Return success\n}\n//---------------------------------------------------------------------------\n" }, { "path": "benchmark/alexnet.param", "content": "7767517\n15 15\nInput data 0 1 data -23330=4,3,227,227,3 0=227 1=227 2=3\nConvolution conv1 1 1 data conv1_relu1 -23330=4,3,55,55,96 0=96 1=11 3=4 5=1 6=34848 9=1\nLRN norm1 1 1 conv1_relu1 norm1 -23330=4,3,55,55,96 2=1.000000e-04\nPooling pool1 1 1 norm1 pool1 -23330=4,3,27,27,96 1=3 2=2\nConvolutionDepthWise conv2 1 1 pool1 conv2_relu2 -23330=4,3,27,27,256 0=256 1=5 4=2 5=1 6=307200 7=2 9=1\nLRN norm2 1 1 conv2_relu2 norm2 -23330=4,3,27,27,256 2=1.000000e-04\nPooling pool2 1 1 norm2 pool2 -23330=4,3,13,13,256 1=3 2=2\nConvolution conv3 1 1 pool2 conv3_relu3 -23330=4,3,13,13,384 0=384 1=3 4=1 5=1 6=884736 9=1\nConvolutionDepthWise conv4 1 1 conv3_relu3 conv4_relu4 -23330=4,3,13,13,384 0=384 1=3 4=1 5=1 6=663552 7=2 9=1\nConvolutionDepthWise conv5 1 1 conv4_relu4 conv5_relu5 -23330=4,3,13,13,256 0=256 1=3 4=1 5=1 6=442368 7=2 9=1\nPooling pool5 1 1 conv5_relu5 pool5 -23330=4,3,6,6,256 1=3 2=2\nInnerProduct fc6 1 1 pool5 fc6_drop6 -23330=4,1,4096,1,1 0=4096 1=1 2=37748736 9=1\nInnerProduct fc7 1 1 fc6_drop6 fc7_drop7 -23330=4,1,4096,1,1 0=4096 1=1 2=16777216 9=1\nInnerProduct fc8 1 1 fc7_drop7 fc8 -23330=4,1,1000,1,1 0=1000 1=1 2=4096000\nSoftmax prob 1 1 fc8 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/benchncnn.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \n#include \n#include \n\n#ifdef __EMSCRIPTEN__\n#include \n#endif\n\n#include \"benchmark.h\"\n#include \"cpu.h\"\n#include \"datareader.h\"\n#include \"net.h\"\n#include \"gpu.h\"\n\n#include \"benchncnn_param_data.h\"\n\n#ifndef NCNN_SIMPLESTL\n#include \n#endif\n\nclass DataReaderFromEmpty : public ncnn::DataReader\n{\npublic:\n virtual int scan(const char* format, void* p) const\n {\n return 0;\n }\n virtual size_t read(void* buf, size_t size) const\n {\n memset(buf, 0, size);\n return size;\n }\n};\n\nstatic int g_warmup_loop_count = 8;\nstatic int g_loop_count = 4;\nstatic bool g_enable_cooling_down = true;\n\nstatic ncnn::UnlockedPoolAllocator g_blob_pool_allocator;\nstatic ncnn::PoolAllocator g_workspace_pool_allocator;\n\n#if NCNN_VULKAN\nstatic ncnn::VulkanDevice* g_vkdev = 0;\nstatic ncnn::VkAllocator* g_blob_vkallocator = 0;\nstatic ncnn::VkAllocator* g_staging_vkallocator = 0;\n#endif // NCNN_VULKAN\n\nvoid benchmark(const char* comment, const std::vector& _in, const ncnn::Option& opt, const char* model_param_data = NULL)\n{\n // Skip if int8 model name and using GPU\n if (opt.use_vulkan_compute && strstr(comment, \"int8\") != NULL)\n {\n if (!model_param_data)\n fprintf(stderr, \"%20s skipped (int8+GPU not supported)\\n\", comment);\n return;\n }\n\n g_blob_pool_allocator.clear();\n g_workspace_pool_allocator.clear();\n\n#if NCNN_VULKAN\n if (opt.use_vulkan_compute)\n {\n g_blob_vkallocator->clear();\n g_staging_vkallocator->clear();\n }\n#endif // NCNN_VULKAN\n\n ncnn::Net net;\n\n net.opt = opt;\n\n#if NCNN_VULKAN\n if (net.opt.use_vulkan_compute)\n {\n net.set_vulkan_device(g_vkdev);\n }\n#endif // NCNN_VULKAN\n\n if (model_param_data)\n {\n net.load_param_mem(model_param_data);\n }\n else\n {\n net.load_param(comment);\n }\n\n DataReaderFromEmpty dr;\n net.load_model(dr);\n\n const std::vector& input_names = net.input_names();\n const std::vector& output_names = net.output_names();\n\n if (g_enable_cooling_down)\n {\n // sleep 10 seconds for cooling down SOC :(\n ncnn::sleep(10 * 1000);\n }\n\n if (input_names.size() > _in.size())\n {\n fprintf(stderr, \"input %zu tensors while model has %zu inputs\\n\", _in.size(), input_names.size());\n return;\n }\n\n // initialize input\n for (size_t j = 0; j < input_names.size(); ++j)\n {\n ncnn::Mat in = _in[j];\n in.fill(0.01f);\n }\n\n // warm up\n for (int i = 0; i < g_warmup_loop_count; i++)\n {\n ncnn::Extractor ex = net.create_extractor();\n for (size_t j = 0; j < input_names.size(); ++j)\n {\n ncnn::Mat in = _in[j];\n ex.input(input_names[j], in);\n }\n\n for (size_t j = 0; j < output_names.size(); ++j)\n {\n ncnn::Mat out;\n ex.extract(output_names[j], out);\n }\n }\n\n double time_min = DBL_MAX;\n double time_max = -DBL_MAX;\n double time_avg = 0;\n\n for (int i = 0; i < g_loop_count; i++)\n {\n double start = ncnn::get_current_time();\n {\n ncnn::Extractor ex = net.create_extractor();\n for (size_t j = 0; j < input_names.size(); ++j)\n {\n ncnn::Mat in = _in[j];\n ex.input(input_names[j], in);\n }\n\n for (size_t j = 0; j < output_names.size(); ++j)\n {\n ncnn::Mat out;\n ex.extract(output_names[j], out);\n }\n }\n\n double end = ncnn::get_current_time();\n\n double time = end - start;\n\n time_min = std::min(time_min, time);\n time_max = std::max(time_max, time);\n time_avg += time;\n }\n\n time_avg /= g_loop_count;\n\n fprintf(stderr, \"%20s min = %7.2f max = %7.2f avg = %7.2f\\n\", comment, time_min, time_max, time_avg);\n}\n\nvoid benchmark(const char* comment, const ncnn::Mat& _in, const ncnn::Option& opt, const char* model_param_data = NULL)\n{\n std::vector inputs;\n inputs.push_back(_in);\n return benchmark(comment, inputs, opt, model_param_data);\n}\n\nvoid show_usage()\n{\n fprintf(stderr, \"Usage: benchncnn [loop count] [num threads] [powersave] [gpu device] [cooling down] [(key=value)...]\\n\");\n fprintf(stderr, \" param=model.param\\n\");\n fprintf(stderr, \" shape=[227,227,3],...\\n\");\n}\n\nstatic std::vector parse_shape_list(char* s)\n{\n std::vector > shapes;\n std::vector mats;\n\n char* pch = strtok(s, \"[]\");\n while (pch != NULL)\n {\n // parse a,b,c\n int v;\n int nconsumed = 0;\n int nscan = sscanf(pch, \"%d%n\", &v, &nconsumed);\n if (nscan == 1)\n {\n // ok we get shape\n pch += nconsumed;\n\n std::vector s;\n s.push_back(v);\n\n nscan = sscanf(pch, \",%d%n\", &v, &nconsumed);\n while (nscan == 1)\n {\n pch += nconsumed;\n\n s.push_back(v);\n\n nscan = sscanf(pch, \",%d%n\", &v, &nconsumed);\n }\n\n // shape end\n shapes.push_back(s);\n }\n\n pch = strtok(NULL, \"[]\");\n }\n\n for (size_t i = 0; i < shapes.size(); ++i)\n {\n const std::vector& shape = shapes[i];\n switch (shape.size())\n {\n case 4:\n mats.push_back(ncnn::Mat(shape[0], shape[1], shape[2], shape[3]));\n break;\n case 3:\n mats.push_back(ncnn::Mat(shape[0], shape[1], shape[2]));\n break;\n case 2:\n mats.push_back(ncnn::Mat(shape[0], shape[1]));\n break;\n case 1:\n mats.push_back(ncnn::Mat(shape[0]));\n break;\n default:\n fprintf(stderr, \"unsupported input shape size %zu\\n\", shape.size());\n break;\n }\n }\n return mats;\n}\n\nint main(int argc, char** argv)\n{\n int loop_count = 4;\n int num_threads = ncnn::get_physical_big_cpu_count();\n int powersave = 2;\n int gpu_device = -1;\n int cooling_down = 1;\n char* model = 0;\n std::vector inputs;\n\n for (int i = 1; i < argc; i++)\n {\n if (argv[i][0] == '-' && argv[i][1] == 'h')\n {\n show_usage();\n return -1;\n }\n\n if (strcmp(argv[i], \"--help\") == 0)\n {\n show_usage();\n return -1;\n }\n }\n\n if (argc >= 2)\n {\n loop_count = atoi(argv[1]);\n }\n if (argc >= 3)\n {\n num_threads = atoi(argv[2]);\n }\n if (argc >= 4)\n {\n powersave = atoi(argv[3]);\n }\n if (argc >= 5)\n {\n gpu_device = atoi(argv[4]);\n }\n if (argc >= 6)\n {\n cooling_down = atoi(argv[5]);\n }\n\n for (int i = 6; i < argc; i++)\n {\n // key=value\n char* kv = argv[i];\n\n char* eqs = strchr(kv, '=');\n if (eqs == NULL)\n {\n fprintf(stderr, \"unrecognized arg %s\\n\", kv);\n continue;\n }\n\n // split k v\n eqs[0] = '\\0';\n const char* key = kv;\n char* value = eqs + 1;\n\n if (strcmp(key, \"param\") == 0)\n model = value;\n if (strcmp(key, \"shape\") == 0)\n inputs = parse_shape_list(value);\n }\n\n if (model && inputs.empty())\n {\n fprintf(stderr, \"input tensor shape empty!\\n\");\n return -1;\n }\n\n#ifdef __EMSCRIPTEN__\n EM_ASM(\n FS.mkdir('/working');\n FS.mount(NODEFS, {root: '.'}, '/working'););\n#endif // __EMSCRIPTEN__\n\n bool use_vulkan_compute = gpu_device != -1;\n\n g_enable_cooling_down = cooling_down != 0;\n\n g_loop_count = loop_count;\n\n g_blob_pool_allocator.set_size_compare_ratio(0.f);\n g_workspace_pool_allocator.set_size_compare_ratio(0.f);\n\n#if NCNN_VULKAN\n if (use_vulkan_compute)\n {\n g_warmup_loop_count = 10;\n\n g_vkdev = ncnn::get_gpu_device(gpu_device);\n\n g_blob_vkallocator = new ncnn::VkBlobAllocator(g_vkdev);\n g_staging_vkallocator = new ncnn::VkStagingAllocator(g_vkdev);\n }\n#endif // NCNN_VULKAN\n\n ncnn::set_cpu_powersave(powersave);\n\n ncnn::set_omp_dynamic(0);\n ncnn::set_omp_num_threads(num_threads);\n\n // default option\n ncnn::Option opt;\n opt.lightmode = true;\n opt.num_threads = num_threads;\n opt.blob_allocator = &g_blob_pool_allocator;\n opt.workspace_allocator = &g_workspace_pool_allocator;\n#if NCNN_VULKAN\n opt.blob_vkallocator = g_blob_vkallocator;\n opt.workspace_vkallocator = g_blob_vkallocator;\n opt.staging_vkallocator = g_staging_vkallocator;\n#endif // NCNN_VULKAN\n opt.use_winograd_convolution = true;\n opt.use_sgemm_convolution = true;\n opt.use_int8_inference = true;\n opt.use_vulkan_compute = use_vulkan_compute;\n opt.use_fp16_packed = true;\n opt.use_fp16_storage = true;\n opt.use_fp16_arithmetic = true;\n opt.use_int8_storage = true;\n opt.use_int8_arithmetic = true;\n opt.use_packing_layout = true;\n\n fprintf(stderr, \"loop_count = %d\\n\", g_loop_count);\n fprintf(stderr, \"num_threads = %d\\n\", num_threads);\n fprintf(stderr, \"powersave = %d\\n\", ncnn::get_cpu_powersave());\n fprintf(stderr, \"gpu_device = %d\\n\", gpu_device);\n fprintf(stderr, \"cooling_down = %d\\n\", (int)g_enable_cooling_down);\n\n if (model != 0)\n {\n // run user defined benchmark\n benchmark(model, inputs, opt);\n }\n else\n {\n // run default cases\n benchmark(\"squeezenet\", ncnn::Mat(227, 227, 3), opt, squeezenet_param_data);\n\n benchmark(\"squeezenet_int8\", ncnn::Mat(227, 227, 3), opt, squeezenet_int8_param_data);\n\n benchmark(\"mobilenet\", ncnn::Mat(224, 224, 3), opt, mobilenet_param_data);\n\n benchmark(\"mobilenet_int8\", ncnn::Mat(224, 224, 3), opt, mobilenet_int8_param_data);\n\n benchmark(\"mobilenet_v2\", ncnn::Mat(224, 224, 3), opt, mobilenet_v2_param_data);\n\n // benchmark(\"mobilenet_v2_int8\", ncnn::Mat(224, 224, 3), opt, mobilenet_v2_int8_param_data);\n\n benchmark(\"mobilenet_v3\", ncnn::Mat(224, 224, 3), opt, mobilenet_v3_param_data);\n\n benchmark(\"shufflenet\", ncnn::Mat(224, 224, 3), opt, shufflenet_param_data);\n\n benchmark(\"shufflenet_v2\", ncnn::Mat(224, 224, 3), opt, shufflenet_v2_param_data);\n\n benchmark(\"mnasnet\", ncnn::Mat(224, 224, 3), opt, mnasnet_param_data);\n\n benchmark(\"proxylessnasnet\", ncnn::Mat(224, 224, 3), opt, proxylessnasnet_param_data);\n\n benchmark(\"efficientnet_b0\", ncnn::Mat(224, 224, 3), opt, efficientnet_b0_param_data);\n\n benchmark(\"efficientnetv2_b0\", ncnn::Mat(224, 224, 3), opt, efficientnetv2_b0_param_data);\n\n benchmark(\"regnety_400m\", ncnn::Mat(224, 224, 3), opt, regnety_400m_param_data);\n\n benchmark(\"blazeface\", ncnn::Mat(128, 128, 3), opt, blazeface_param_data);\n\n benchmark(\"googlenet\", ncnn::Mat(224, 224, 3), opt, googlenet_param_data);\n\n benchmark(\"googlenet_int8\", ncnn::Mat(224, 224, 3), opt, googlenet_int8_param_data);\n\n benchmark(\"resnet18\", ncnn::Mat(224, 224, 3), opt, resnet18_param_data);\n\n benchmark(\"resnet18_int8\", ncnn::Mat(224, 224, 3), opt, resnet18_int8_param_data);\n\n benchmark(\"alexnet\", ncnn::Mat(227, 227, 3), opt, alexnet_param_data);\n\n benchmark(\"vgg16\", ncnn::Mat(224, 224, 3), opt, vgg16_param_data);\n\n benchmark(\"vgg16_int8\", ncnn::Mat(224, 224, 3), opt, vgg16_int8_param_data);\n\n benchmark(\"resnet50\", ncnn::Mat(224, 224, 3), opt, resnet50_param_data);\n\n benchmark(\"resnet50_int8\", ncnn::Mat(224, 224, 3), opt, resnet50_int8_param_data);\n\n benchmark(\"squeezenet_ssd\", ncnn::Mat(300, 300, 3), opt, squeezenet_ssd_param_data);\n\n benchmark(\"squeezenet_ssd_int8\", ncnn::Mat(300, 300, 3), opt, squeezenet_ssd_int8_param_data);\n\n benchmark(\"mobilenet_ssd\", ncnn::Mat(300, 300, 3), opt, mobilenet_ssd_param_data);\n\n benchmark(\"mobilenet_ssd_int8\", ncnn::Mat(300, 300, 3), opt, mobilenet_ssd_int8_param_data);\n\n benchmark(\"mobilenet_yolo\", ncnn::Mat(416, 416, 3), opt, mobilenet_yolo_param_data);\n\n benchmark(\"mobilenetv2_yolov3\", ncnn::Mat(352, 352, 3), opt, mobilenetv2_yolov3_param_data);\n\n benchmark(\"yolov4-tiny\", ncnn::Mat(416, 416, 3), opt, yolov4_tiny_param_data);\n\n benchmark(\"nanodet_m\", ncnn::Mat(320, 320, 3), opt, nanodet_m_param_data);\n\n benchmark(\"yolo-fastest-1.1\", ncnn::Mat(320, 320, 3), opt, yolo_fastest_1_1_param_data);\n\n benchmark(\"yolo-fastestv2\", ncnn::Mat(352, 352, 3), opt, yolo_fastestv2_param_data);\n\n benchmark(\"vision_transformer\", ncnn::Mat(384, 384, 3), opt, vision_transformer_param_data);\n\n benchmark(\"FastestDet\", ncnn::Mat(352, 352, 3), opt, FastestDet_param_data);\n }\n#if NCNN_VULKAN\n delete g_blob_vkallocator;\n delete g_staging_vkallocator;\n#endif // NCNN_VULKAN\n\n return 0;\n}\n" }, { "path": "benchmark/benchncnn_param_data.h.in", "content": "// Benchncnn Param Data header\n//\n// This file is auto-generated by cmake, don't edit it.\n\n@param_header_data@\n" }, { "path": "benchmark/blazeface.param", "content": "7767517\n101 117\nInput data 0 1 data 0=128 1=128 2=3\nPadding 75 1 1 data 75 0=1 1=2 2=1 3=2 4=0 5=0.000000e+00 7=0 8=0\nConvolution 76 1 1 75 76 0=24 1=5 11=5 2=1 12=1 3=2 13=2 4=0 14=0 15=0 16=0 5=1 6=1800\nReLU 77 1 1 76 77\nSplit splitncnn_0 1 2 77 77_splitncnn_0 77_splitncnn_1\nConvolutionDepthWise 78 1 1 77_splitncnn_1 78 0=24 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=216 7=24\nConvolution 79 1 1 78 79 0=24 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=576\nBinaryOp 80 2 1 79 77_splitncnn_0 80 0=0\nReLU 81 1 1 80 81\nSplit splitncnn_1 1 2 81 81_splitncnn_0 81_splitncnn_1\nPadding 82 1 1 81_splitncnn_1 82 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=4\nConvolutionDepthWise 83 1 1 81_splitncnn_0 83 0=24 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=216 7=24\nConvolution 84 1 1 83 84 0=28 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=672\nBinaryOp 85 2 1 84 82 85 0=0\nReLU 86 1 1 85 86\nSplit splitncnn_2 1 2 86 86_splitncnn_0 86_splitncnn_1\nPadding 87 1 1 86_splitncnn_1 87 0=0 1=2 2=0 3=2 4=0 5=0.000000e+00 7=0 8=0\nPooling 88 1 1 86_splitncnn_0 88 0=0 1=2 11=2 2=2 12=2 3=0 13=0 14=0 15=0 5=1\nPadding 89 1 1 88 89 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=4\nConvolutionDepthWise 90 1 1 87 90 0=28 1=3 11=3 2=1 12=1 3=2 13=2 4=0 14=0 15=0 16=0 5=1 6=252 7=28\nConvolution 91 1 1 90 91 0=32 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=896\nBinaryOp 92 2 1 91 89 92 0=0\nReLU 93 1 1 92 93\nSplit splitncnn_3 1 2 93 93_splitncnn_0 93_splitncnn_1\nPadding 94 1 1 93_splitncnn_1 94 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=4\nConvolutionDepthWise 95 1 1 93_splitncnn_0 95 0=32 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=288 7=32\nConvolution 96 1 1 95 96 0=36 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=1152\nBinaryOp 97 2 1 96 94 97 0=0\nReLU 98 1 1 97 98\nSplit splitncnn_4 1 2 98 98_splitncnn_0 98_splitncnn_1\nPadding 99 1 1 98_splitncnn_1 99 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=6\nConvolutionDepthWise 100 1 1 98_splitncnn_0 100 0=36 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=324 7=36\nConvolution 101 1 1 100 101 0=42 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=1512\nBinaryOp 102 2 1 101 99 102 0=0\nReLU 103 1 1 102 103\nSplit splitncnn_5 1 2 103 103_splitncnn_0 103_splitncnn_1\nPadding 104 1 1 103_splitncnn_1 104 0=0 1=2 2=0 3=2 4=0 5=0.000000e+00 7=0 8=0\nPooling 105 1 1 103_splitncnn_0 105 0=0 1=2 11=2 2=2 12=2 3=0 13=0 14=0 15=0 5=1\nPadding 106 1 1 105 106 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=6\nConvolutionDepthWise 107 1 1 104 107 0=42 1=3 11=3 2=1 12=1 3=2 13=2 4=0 14=0 15=0 16=0 5=1 6=378 7=42\nConvolution 108 1 1 107 108 0=48 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=2016\nBinaryOp 109 2 1 108 106 109 0=0\nReLU 110 1 1 109 110\nSplit splitncnn_6 1 2 110 110_splitncnn_0 110_splitncnn_1\nPadding 111 1 1 110_splitncnn_1 111 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 112 1 1 110_splitncnn_0 112 0=48 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=432 7=48\nConvolution 113 1 1 112 113 0=56 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=2688\nBinaryOp 114 2 1 113 111 114 0=0\nReLU 115 1 1 114 115\nSplit splitncnn_7 1 2 115 115_splitncnn_0 115_splitncnn_1\nPadding 116 1 1 115_splitncnn_1 116 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 117 1 1 115_splitncnn_0 117 0=56 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=504 7=56\nConvolution 118 1 1 117 118 0=64 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=3584\nBinaryOp 119 2 1 118 116 119 0=0\nReLU 120 1 1 119 120\nSplit splitncnn_8 1 2 120 120_splitncnn_0 120_splitncnn_1\nPadding 121 1 1 120_splitncnn_1 121 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 122 1 1 120_splitncnn_0 122 0=64 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=576 7=64\nConvolution 123 1 1 122 123 0=72 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=4608\nBinaryOp 124 2 1 123 121 124 0=0\nReLU 125 1 1 124 125\nSplit splitncnn_9 1 2 125 125_splitncnn_0 125_splitncnn_1\nPadding 126 1 1 125_splitncnn_1 126 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 127 1 1 125_splitncnn_0 127 0=72 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=648 7=72\nConvolution 128 1 1 127 128 0=80 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=5760\nBinaryOp 129 2 1 128 126 129 0=0\nReLU 130 1 1 129 130\nSplit splitncnn_10 1 2 130 130_splitncnn_0 130_splitncnn_1\nPadding 131 1 1 130_splitncnn_1 131 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 132 1 1 130_splitncnn_0 132 0=80 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=720 7=80\nConvolution 133 1 1 132 133 0=88 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=7040\nBinaryOp 134 2 1 133 131 134 0=0\nReLU 135 1 1 134 135\nSplit splitncnn_11 1 2 135 135_splitncnn_0 135_splitncnn_1\nPadding 136 1 1 135_splitncnn_1 136 0=0 1=2 2=0 3=2 4=0 5=0.000000e+00 7=0 8=0\nPooling 137 1 1 135_splitncnn_0 137 0=0 1=2 11=2 2=2 12=2 3=0 13=0 14=0 15=0 5=1\nPadding 138 1 1 137 138 0=0 1=0 2=0 3=0 4=0 5=0.000000e+00 7=0 8=8\nConvolutionDepthWise 139 1 1 136 139 0=88 1=3 11=3 2=1 12=1 3=2 13=2 4=0 14=0 15=0 16=0 5=1 6=792 7=88\nConvolution 140 1 1 139 140 0=96 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=8448\nBinaryOp 141 2 1 140 138 141 0=0\nReLU 142 1 1 141 142\nSplit splitncnn_12 1 2 142 142_splitncnn_0 142_splitncnn_1\nConvolutionDepthWise 143 1 1 142_splitncnn_1 143 0=96 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=864 7=96\nConvolution 144 1 1 143 144 0=96 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=9216\nBinaryOp 145 2 1 144 142_splitncnn_0 145 0=0\nReLU 146 1 1 145 146\nSplit splitncnn_13 1 2 146 146_splitncnn_0 146_splitncnn_1\nConvolutionDepthWise 147 1 1 146_splitncnn_1 147 0=96 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=864 7=96\nConvolution 148 1 1 147 148 0=96 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=9216\nBinaryOp 149 2 1 148 146_splitncnn_0 149 0=0\nReLU 150 1 1 149 150\nSplit splitncnn_14 1 2 150 150_splitncnn_0 150_splitncnn_1\nConvolutionDepthWise 151 1 1 150_splitncnn_1 151 0=96 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=864 7=96\nConvolution 152 1 1 151 152 0=96 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=9216\nBinaryOp 153 2 1 152 150_splitncnn_0 153 0=0\nReLU 154 1 1 153 154\nSplit splitncnn_15 1 2 154 154_splitncnn_0 154_splitncnn_1\nConvolutionDepthWise 155 1 1 154_splitncnn_1 155 0=96 1=3 11=3 2=1 12=1 3=1 13=1 4=1 14=1 15=1 16=1 5=1 6=864 7=96\nConvolution 156 1 1 155 156 0=96 1=1 11=1 2=1 12=1 3=1 13=1 4=0 14=0 15=0 16=0 5=1 6=9216\nBinaryOp 157 2 1 156 154_splitncnn_0 157 0=0\nReLU output 1 1 157 output\n" }, { "path": "benchmark/efficientnet_b0.param", "content": "7767517\n200 225\nInput input.1 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution Conv_0 1 1 data 362 -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864\nSwish Mul_3 1 1 362 364 -23330=4,3,112,112,32\nConvolutionDepthWise Conv_4 1 1 364 366 -23330=4,3,112,112,32 0=32 1=3 4=1 5=1 6=288 7=32\nSwish Mul_7 1 1 366 368 -23330=4,3,112,112,32\nSplit splitncnn_0 1 2 368 368_splitncnn_0 368_splitncnn_1 -23330=8,3,112,112,32,3,112,112,32\nPooling GlobalAveragePool_8 1 1 368_splitncnn_1 369 -23330=4,1,32,1,1 0=1 4=1\nInnerProduct Conv_9 1 1 369 370 -23330=4,1,8,1,1 0=8 1=1 2=256\nSwish Mul_11 1 1 370 372 -23330=4,1,8,1,1\nConvolution Conv_12 1 1 372 374 -23330=4,1,32,1,1 0=32 1=1 5=1 6=256 9=4\nBinaryOp Mul_14 2 1 368_splitncnn_0 374 375 -23330=4,3,112,112,32 0=2\nConvolution Conv_15 1 1 375 377 -23330=4,3,112,112,16 0=16 1=1 5=1 6=512\nConvolution Conv_17 1 1 377 379 -23330=4,3,112,112,96 0=96 1=1 5=1 6=1536\nSwish Mul_20 1 1 379 381 -23330=4,3,112,112,96\nConvolutionDepthWise Conv_21 1 1 381 383 -23330=4,3,56,56,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96\nSwish Mul_24 1 1 383 385 -23330=4,3,56,56,96\nSplit splitncnn_1 1 2 385 385_splitncnn_0 385_splitncnn_1 -23330=8,3,56,56,96,3,56,56,96\nPooling GlobalAveragePool_25 1 1 385_splitncnn_1 386 -23330=4,1,96,1,1 0=1 4=1\nInnerProduct Conv_26 1 1 386 387 -23330=4,1,4,1,1 0=4 1=1 2=384\nSwish Mul_28 1 1 387 389 -23330=4,1,4,1,1\nConvolution Conv_29 1 1 389 391 -23330=4,1,96,1,1 0=96 1=1 5=1 6=384 9=4\nBinaryOp Mul_31 2 1 385_splitncnn_0 391 392 -23330=4,3,56,56,96 0=2\nConvolution Conv_32 1 1 392 394 -23330=4,3,56,56,24 0=24 1=1 5=1 6=2304\nSplit splitncnn_2 1 2 394 394_splitncnn_0 394_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolution Conv_34 1 1 394_splitncnn_1 396 -23330=4,3,56,56,144 0=144 1=1 5=1 6=3456\nSwish Mul_37 1 1 396 398 -23330=4,3,56,56,144\nConvolutionDepthWise Conv_38 1 1 398 400 -23330=4,3,56,56,144 0=144 1=3 4=1 5=1 6=1296 7=144\nSwish Mul_41 1 1 400 402 -23330=4,3,56,56,144\nSplit splitncnn_3 1 2 402 402_splitncnn_0 402_splitncnn_1 -23330=8,3,56,56,144,3,56,56,144\nPooling GlobalAveragePool_42 1 1 402_splitncnn_1 403 -23330=4,1,144,1,1 0=1 4=1\nInnerProduct Conv_43 1 1 403 404 -23330=4,1,6,1,1 0=6 1=1 2=864\nSwish Mul_45 1 1 404 406 -23330=4,1,6,1,1\nConvolution Conv_46 1 1 406 408 -23330=4,1,144,1,1 0=144 1=1 5=1 6=864 9=4\nBinaryOp Mul_48 2 1 402_splitncnn_0 408 409 -23330=4,3,56,56,144 0=2\nConvolution Conv_49 1 1 409 411 -23330=4,3,56,56,24 0=24 1=1 5=1 6=3456\nBinaryOp Add_51 2 1 394_splitncnn_0 411 412 -23330=4,3,56,56,24\nConvolution Conv_52 1 1 412 414 -23330=4,3,56,56,144 0=144 1=1 5=1 6=3456\nSwish Mul_55 1 1 414 416 -23330=4,3,56,56,144\nConvolutionDepthWise Conv_56 1 1 416 418 -23330=4,3,28,28,144 0=144 1=5 3=2 4=2 5=1 6=3600 7=144\nSwish Mul_59 1 1 418 420 -23330=4,3,28,28,144\nSplit splitncnn_4 1 2 420 420_splitncnn_0 420_splitncnn_1 -23330=8,3,28,28,144,3,28,28,144\nPooling GlobalAveragePool_60 1 1 420_splitncnn_1 421 -23330=4,1,144,1,1 0=1 4=1\nInnerProduct Conv_61 1 1 421 422 -23330=4,1,6,1,1 0=6 1=1 2=864\nSwish Mul_63 1 1 422 424 -23330=4,1,6,1,1\nConvolution Conv_64 1 1 424 426 -23330=4,1,144,1,1 0=144 1=1 5=1 6=864 9=4\nBinaryOp Mul_66 2 1 420_splitncnn_0 426 427 -23330=4,3,28,28,144 0=2\nConvolution Conv_67 1 1 427 429 -23330=4,3,28,28,40 0=40 1=1 5=1 6=5760\nSplit splitncnn_5 1 2 429 429_splitncnn_0 429_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution Conv_69 1 1 429_splitncnn_1 431 -23330=4,3,28,28,240 0=240 1=1 5=1 6=9600\nSwish Mul_72 1 1 431 433 -23330=4,3,28,28,240\nConvolutionDepthWise Conv_73 1 1 433 435 -23330=4,3,28,28,240 0=240 1=5 4=2 5=1 6=6000 7=240\nSwish Mul_76 1 1 435 437 -23330=4,3,28,28,240\nSplit splitncnn_6 1 2 437 437_splitncnn_0 437_splitncnn_1 -23330=8,3,28,28,240,3,28,28,240\nPooling GlobalAveragePool_77 1 1 437_splitncnn_1 438 -23330=4,1,240,1,1 0=1 4=1\nInnerProduct Conv_78 1 1 438 439 -23330=4,1,10,1,1 0=10 1=1 2=2400\nSwish Mul_80 1 1 439 441 -23330=4,1,10,1,1\nConvolution Conv_81 1 1 441 443 -23330=4,1,240,1,1 0=240 1=1 5=1 6=2400 9=4\nBinaryOp Mul_83 2 1 437_splitncnn_0 443 444 -23330=4,3,28,28,240 0=2\nConvolution Conv_84 1 1 444 446 -23330=4,3,28,28,40 0=40 1=1 5=1 6=9600\nBinaryOp Add_86 2 1 429_splitncnn_0 446 447 -23330=4,3,28,28,40\nConvolution Conv_87 1 1 447 449 -23330=4,3,28,28,240 0=240 1=1 5=1 6=9600\nSwish Mul_90 1 1 449 451 -23330=4,3,28,28,240\nConvolutionDepthWise Conv_91 1 1 451 453 -23330=4,3,14,14,240 0=240 1=3 3=2 4=1 5=1 6=2160 7=240\nSwish Mul_94 1 1 453 455 -23330=4,3,14,14,240\nSplit splitncnn_7 1 2 455 455_splitncnn_0 455_splitncnn_1 -23330=8,3,14,14,240,3,14,14,240\nPooling GlobalAveragePool_95 1 1 455_splitncnn_1 456 -23330=4,1,240,1,1 0=1 4=1\nInnerProduct Conv_96 1 1 456 457 -23330=4,1,10,1,1 0=10 1=1 2=2400\nSwish Mul_98 1 1 457 459 -23330=4,1,10,1,1\nConvolution Conv_99 1 1 459 461 -23330=4,1,240,1,1 0=240 1=1 5=1 6=2400 9=4\nBinaryOp Mul_101 2 1 455_splitncnn_0 461 462 -23330=4,3,14,14,240 0=2\nConvolution Conv_102 1 1 462 464 -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nSplit splitncnn_8 1 2 464 464_splitncnn_0 464_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution Conv_104 1 1 464_splitncnn_1 466 -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400\nSwish Mul_107 1 1 466 468 -23330=4,3,14,14,480\nConvolutionDepthWise Conv_108 1 1 468 470 -23330=4,3,14,14,480 0=480 1=3 4=1 5=1 6=4320 7=480\nSwish Mul_111 1 1 470 472 -23330=4,3,14,14,480\nSplit splitncnn_9 1 2 472 472_splitncnn_0 472_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nPooling GlobalAveragePool_112 1 1 472_splitncnn_1 473 -23330=4,1,480,1,1 0=1 4=1\nInnerProduct Conv_113 1 1 473 474 -23330=4,1,20,1,1 0=20 1=1 2=9600\nSwish Mul_115 1 1 474 476 -23330=4,1,20,1,1\nConvolution Conv_116 1 1 476 478 -23330=4,1,480,1,1 0=480 1=1 5=1 6=9600 9=4\nBinaryOp Mul_118 2 1 472_splitncnn_0 478 479 -23330=4,3,14,14,480 0=2\nConvolution Conv_119 1 1 479 481 -23330=4,3,14,14,80 0=80 1=1 5=1 6=38400\nBinaryOp Add_121 2 1 464_splitncnn_0 481 482 -23330=4,3,14,14,80\nSplit splitncnn_10 1 2 482 482_splitncnn_0 482_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution Conv_122 1 1 482_splitncnn_1 484 -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400\nSwish Mul_125 1 1 484 486 -23330=4,3,14,14,480\nConvolutionDepthWise Conv_126 1 1 486 488 -23330=4,3,14,14,480 0=480 1=3 4=1 5=1 6=4320 7=480\nSwish Mul_129 1 1 488 490 -23330=4,3,14,14,480\nSplit splitncnn_11 1 2 490 490_splitncnn_0 490_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nPooling GlobalAveragePool_130 1 1 490_splitncnn_1 491 -23330=4,1,480,1,1 0=1 4=1\nInnerProduct Conv_131 1 1 491 492 -23330=4,1,20,1,1 0=20 1=1 2=9600\nSwish Mul_133 1 1 492 494 -23330=4,1,20,1,1\nConvolution Conv_134 1 1 494 496 -23330=4,1,480,1,1 0=480 1=1 5=1 6=9600 9=4\nBinaryOp Mul_136 2 1 490_splitncnn_0 496 497 -23330=4,3,14,14,480 0=2\nConvolution Conv_137 1 1 497 499 -23330=4,3,14,14,80 0=80 1=1 5=1 6=38400\nBinaryOp Add_139 2 1 482_splitncnn_0 499 500 -23330=4,3,14,14,80\nConvolution Conv_140 1 1 500 502 -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400\nSwish Mul_143 1 1 502 504 -23330=4,3,14,14,480\nConvolutionDepthWise Conv_144 1 1 504 506 -23330=4,3,14,14,480 0=480 1=5 4=2 5=1 6=12000 7=480\nSwish Mul_147 1 1 506 508 -23330=4,3,14,14,480\nSplit splitncnn_12 1 2 508 508_splitncnn_0 508_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nPooling GlobalAveragePool_148 1 1 508_splitncnn_1 509 -23330=4,1,480,1,1 0=1 4=1\nInnerProduct Conv_149 1 1 509 510 -23330=4,1,20,1,1 0=20 1=1 2=9600\nSwish Mul_151 1 1 510 512 -23330=4,1,20,1,1\nConvolution Conv_152 1 1 512 514 -23330=4,1,480,1,1 0=480 1=1 5=1 6=9600 9=4\nBinaryOp Mul_154 2 1 508_splitncnn_0 514 515 -23330=4,3,14,14,480 0=2\nConvolution Conv_155 1 1 515 517 -23330=4,3,14,14,112 0=112 1=1 5=1 6=53760\nSplit splitncnn_13 1 2 517 517_splitncnn_0 517_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution Conv_157 1 1 517_splitncnn_1 519 -23330=4,3,14,14,672 0=672 1=1 5=1 6=75264\nSwish Mul_160 1 1 519 521 -23330=4,3,14,14,672\nConvolutionDepthWise Conv_161 1 1 521 523 -23330=4,3,14,14,672 0=672 1=5 4=2 5=1 6=16800 7=672\nSwish Mul_164 1 1 523 525 -23330=4,3,14,14,672\nSplit splitncnn_14 1 2 525 525_splitncnn_0 525_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nPooling GlobalAveragePool_165 1 1 525_splitncnn_1 526 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct Conv_166 1 1 526 527 -23330=4,1,28,1,1 0=28 1=1 2=18816\nSwish Mul_168 1 1 527 529 -23330=4,1,28,1,1\nConvolution Conv_169 1 1 529 531 -23330=4,1,672,1,1 0=672 1=1 5=1 6=18816 9=4\nBinaryOp Mul_171 2 1 525_splitncnn_0 531 532 -23330=4,3,14,14,672 0=2\nConvolution Conv_172 1 1 532 534 -23330=4,3,14,14,112 0=112 1=1 5=1 6=75264\nBinaryOp Add_174 2 1 517_splitncnn_0 534 535 -23330=4,3,14,14,112\nSplit splitncnn_15 1 2 535 535_splitncnn_0 535_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution Conv_175 1 1 535_splitncnn_1 537 -23330=4,3,14,14,672 0=672 1=1 5=1 6=75264\nSwish Mul_178 1 1 537 539 -23330=4,3,14,14,672\nConvolutionDepthWise Conv_179 1 1 539 541 -23330=4,3,14,14,672 0=672 1=5 4=2 5=1 6=16800 7=672\nSwish Mul_182 1 1 541 543 -23330=4,3,14,14,672\nSplit splitncnn_16 1 2 543 543_splitncnn_0 543_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nPooling GlobalAveragePool_183 1 1 543_splitncnn_1 544 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct Conv_184 1 1 544 545 -23330=4,1,28,1,1 0=28 1=1 2=18816\nSwish Mul_186 1 1 545 547 -23330=4,1,28,1,1\nConvolution Conv_187 1 1 547 549 -23330=4,1,672,1,1 0=672 1=1 5=1 6=18816 9=4\nBinaryOp Mul_189 2 1 543_splitncnn_0 549 550 -23330=4,3,14,14,672 0=2\nConvolution Conv_190 1 1 550 552 -23330=4,3,14,14,112 0=112 1=1 5=1 6=75264\nBinaryOp Add_192 2 1 535_splitncnn_0 552 553 -23330=4,3,14,14,112\nConvolution Conv_193 1 1 553 555 -23330=4,3,14,14,672 0=672 1=1 5=1 6=75264\nSwish Mul_196 1 1 555 557 -23330=4,3,14,14,672\nConvolutionDepthWise Conv_197 1 1 557 559 -23330=4,3,7,7,672 0=672 1=5 3=2 4=2 5=1 6=16800 7=672\nSwish Mul_200 1 1 559 561 -23330=4,3,7,7,672\nSplit splitncnn_17 1 2 561 561_splitncnn_0 561_splitncnn_1 -23330=8,3,7,7,672,3,7,7,672\nPooling GlobalAveragePool_201 1 1 561_splitncnn_1 562 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct Conv_202 1 1 562 563 -23330=4,1,28,1,1 0=28 1=1 2=18816\nSwish Mul_204 1 1 563 565 -23330=4,1,28,1,1\nConvolution Conv_205 1 1 565 567 -23330=4,1,672,1,1 0=672 1=1 5=1 6=18816 9=4\nBinaryOp Mul_207 2 1 561_splitncnn_0 567 568 -23330=4,3,7,7,672 0=2\nConvolution Conv_208 1 1 568 570 -23330=4,3,7,7,192 0=192 1=1 5=1 6=129024\nSplit splitncnn_18 1 2 570 570_splitncnn_0 570_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution Conv_210 1 1 570_splitncnn_1 572 -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184\nSwish Mul_213 1 1 572 574 -23330=4,3,7,7,1152\nConvolutionDepthWise Conv_214 1 1 574 576 -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152\nSwish Mul_217 1 1 576 578 -23330=4,3,7,7,1152\nSplit splitncnn_19 1 2 578 578_splitncnn_0 578_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nPooling GlobalAveragePool_218 1 1 578_splitncnn_1 579 -23330=4,1,1152,1,1 0=1 4=1\nInnerProduct Conv_219 1 1 579 580 -23330=4,1,48,1,1 0=48 1=1 2=55296\nSwish Mul_221 1 1 580 582 -23330=4,1,48,1,1\nConvolution Conv_222 1 1 582 584 -23330=4,1,1152,1,1 0=1152 1=1 5=1 6=55296 9=4\nBinaryOp Mul_224 2 1 578_splitncnn_0 584 585 -23330=4,3,7,7,1152 0=2\nConvolution Conv_225 1 1 585 587 -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp Add_227 2 1 570_splitncnn_0 587 588 -23330=4,3,7,7,192\nSplit splitncnn_20 1 2 588 588_splitncnn_0 588_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution Conv_228 1 1 588_splitncnn_1 590 -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184\nSwish Mul_231 1 1 590 592 -23330=4,3,7,7,1152\nConvolutionDepthWise Conv_232 1 1 592 594 -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152\nSwish Mul_235 1 1 594 596 -23330=4,3,7,7,1152\nSplit splitncnn_21 1 2 596 596_splitncnn_0 596_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nPooling GlobalAveragePool_236 1 1 596_splitncnn_1 597 -23330=4,1,1152,1,1 0=1 4=1\nInnerProduct Conv_237 1 1 597 598 -23330=4,1,48,1,1 0=48 1=1 2=55296\nSwish Mul_239 1 1 598 600 -23330=4,1,48,1,1\nConvolution Conv_240 1 1 600 602 -23330=4,1,1152,1,1 0=1152 1=1 5=1 6=55296 9=4\nBinaryOp Mul_242 2 1 596_splitncnn_0 602 603 -23330=4,3,7,7,1152 0=2\nConvolution Conv_243 1 1 603 605 -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp Add_245 2 1 588_splitncnn_0 605 606 -23330=4,3,7,7,192\nSplit splitncnn_22 1 2 606 606_splitncnn_0 606_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution Conv_246 1 1 606_splitncnn_1 608 -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184\nSwish Mul_249 1 1 608 610 -23330=4,3,7,7,1152\nConvolutionDepthWise Conv_250 1 1 610 612 -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152\nSwish Mul_253 1 1 612 614 -23330=4,3,7,7,1152\nSplit splitncnn_23 1 2 614 614_splitncnn_0 614_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nPooling GlobalAveragePool_254 1 1 614_splitncnn_1 615 -23330=4,1,1152,1,1 0=1 4=1\nInnerProduct Conv_255 1 1 615 616 -23330=4,1,48,1,1 0=48 1=1 2=55296\nSwish Mul_257 1 1 616 618 -23330=4,1,48,1,1\nConvolution Conv_258 1 1 618 620 -23330=4,1,1152,1,1 0=1152 1=1 5=1 6=55296 9=4\nBinaryOp Mul_260 2 1 614_splitncnn_0 620 621 -23330=4,3,7,7,1152 0=2\nConvolution Conv_261 1 1 621 623 -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp Add_263 2 1 606_splitncnn_0 623 624 -23330=4,3,7,7,192\nConvolution Conv_264 1 1 624 626 -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184\nSwish Mul_267 1 1 626 628 -23330=4,3,7,7,1152\nConvolutionDepthWise Conv_268 1 1 628 630 -23330=4,3,7,7,1152 0=1152 1=3 4=1 5=1 6=10368 7=1152\nSwish Mul_271 1 1 630 632 -23330=4,3,7,7,1152\nSplit splitncnn_24 1 2 632 632_splitncnn_0 632_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nPooling GlobalAveragePool_272 1 1 632_splitncnn_1 633 -23330=4,1,1152,1,1 0=1 4=1\nInnerProduct Conv_273 1 1 633 634 -23330=4,1,48,1,1 0=48 1=1 2=55296\nSwish Mul_275 1 1 634 636 -23330=4,1,48,1,1\nConvolution Conv_276 1 1 636 638 -23330=4,1,1152,1,1 0=1152 1=1 5=1 6=55296 9=4\nBinaryOp Mul_278 2 1 632_splitncnn_0 638 639 -23330=4,3,7,7,1152 0=2\nConvolution Conv_279 1 1 639 641 -23330=4,3,7,7,320 0=320 1=1 5=1 6=368640\nConvolution Conv_281 1 1 641 643 -23330=4,3,7,7,1280 0=1280 1=1 5=1 6=409600\nSwish Mul_284 1 1 643 645 -23330=4,3,7,7,1280\nPooling GlobalAveragePool_285 1 1 645 654 -23330=4,1,1280,1,1 0=1 4=1\nInnerProduct Gemm_292 1 1 654 655 -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\nSoftmax prob 1 1 655 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/efficientnetv2_b0.param", "content": "7767517\n257 288\nMemoryData 110:12 0 1 110:12 -23330=4,1,112,1,1 0=112\nMemoryData 133:12 0 1 133:12 -23330=4,1,192,1,1 0=192\nMemoryData 144:12 0 1 144:12 -23330=4,1,192,1,1 0=192\nMemoryData 14:11 0 1 14:11 -23330=4,1,32,1,1 0=32\nMemoryData 155:12 0 1 155:12 -23330=4,1,192,1,1 0=192\nMemoryData 166:12 0 1 166:12 -23330=4,1,192,1,1 0=192\nMemoryData 177:12 0 1 177:12 -23330=4,1,192,1,1 0=192\nMemoryData 188:12 0 1 188:12 -23330=4,1,192,1,1 0=192\nMemoryData 199:12 0 1 199:12 -23330=4,1,192,1,1 0=192\nMemoryData 22:11 0 1 22:11 -23330=4,1,48,1,1 0=48\nMemoryData 33:11 0 1 33:11 -23330=4,1,112,1,1 0=112\nMemoryData 44:11 0 1 44:11 -23330=4,1,112,1,1 0=112\nMemoryData 55:11 0 1 55:11 -23330=4,1,112,1,1 0=112\nMemoryData 77:11 0 1 77:11 -23330=4,1,96,1,1 0=96\nMemoryData 88:11 0 1 88:11 -23330=4,1,96,1,1 0=96\nInput op_201 0 1 204:12 -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution op_202 1 1 204:12 206:12 -23330=4,3,112,112,32 0=32 1=3 3=2 4=-233 5=1 6=864\nSwish op_203 1 1 206:12 208:12 -23330=4,3,112,112,32\nConvolution op_204 1 1 208:12 210:12 -23330=4,3,112,112,16 0=16 1=3 4=-233 5=1 6=4608\nSwish op_205 1 1 210:12 212:12_splitncnn_0 -23330=4,3,112,112,16\nConvolution op_207 1 1 212:12_splitncnn_0 215:12 -23330=4,3,56,56,64 0=64 1=3 3=2 4=-233 5=1 6=9216\nSwish op_208 1 1 215:12 217:12 -23330=4,3,56,56,64\nConvolution op_209 1 1 217:12 219:12 -23330=4,3,56,56,32 0=32 1=1 4=-233 5=1 6=2048\nSplit splitncnn_1 1 2 219:12 219:12_splitncnn_0 219:12_splitncnn_1 -23330=8,3,56,56,32,3,56,56,32\nConvolution op_210 1 1 219:12_splitncnn_1 221:12 -23330=4,3,56,56,128 0=128 1=3 4=-233 5=1 6=36864\nSwish op_211 1 1 221:12 223:12 -23330=4,3,56,56,128\nConvolution op_212 1 1 223:12 224:12 -23330=4,3,56,56,32 0=32 1=1 4=-233 6=4096\nEltwise op_213 2 1 219:12_splitncnn_0 224:12 225:12 -23330=4,3,56,56,32 0=1\nBinaryOp op_214 2 1 225:12 14:11 226:12_splitncnn_0 -23330=4,3,56,56,32\nConvolution op_216 1 1 226:12_splitncnn_0 229:12 -23330=4,3,28,28,128 0=128 1=3 3=2 4=-233 5=1 6=36864\nSwish op_217 1 1 229:12 231:12 -23330=4,3,28,28,128\nConvolution op_218 1 1 231:12 233:12 -23330=4,3,28,28,48 0=48 1=1 4=-233 5=1 6=6144\nSplit splitncnn_3 1 2 233:12 233:12_splitncnn_0 233:12_splitncnn_1 -23330=8,3,28,28,48,3,28,28,48\nConvolution op_219 1 1 233:12_splitncnn_1 235:12 -23330=4,3,28,28,192 0=192 1=3 4=-233 5=1 6=82944\nSwish op_220 1 1 235:12 237:12 -23330=4,3,28,28,192\nConvolution op_221 1 1 237:12 238:12 -23330=4,3,28,28,48 0=48 1=1 4=-233 6=9216\nEltwise op_222 2 1 233:12_splitncnn_0 238:12 239:12 -23330=4,3,28,28,48 0=1\nBinaryOp op_223 2 1 239:12 22:11 240:12_splitncnn_0 -23330=4,3,28,28,48\nConvolution op_225 1 1 240:12_splitncnn_0 243:12 -23330=4,3,28,28,192 0=192 1=1 4=-233 5=1 6=9216\nSwish op_226 1 1 243:12 245:12 -23330=4,3,28,28,192\nConvolutionDepthWise op_227 1 1 245:12 248:12 -23330=4,3,14,14,192 0=192 1=3 3=2 4=-233 5=1 6=1728 7=192\nSwish op_229 1 1 248:12 250:12 -23330=4,3,14,14,192\nSplit splitncnn_5 1 2 250:12 250:12_splitncnn_0 250:12_splitncnn_1 -23330=8,3,14,14,192,3,14,14,192\nReduction op_230 1 1 250:12_splitncnn_1 251:12 -23330=4,3,1,1,192 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_231 1 1 251:12 253:12 -23330=4,3,1,1,12 0=12 1=1 4=-233 5=1 6=2304\nSwish op_232 1 1 253:12 255:12 -23330=4,3,1,1,12\nConvolution op_233 1 1 255:12 258:12 -23330=4,3,1,1,192 0=192 1=1 4=-233 5=1 6=2304 9=4\nBinaryOp op_235 2 1 250:12_splitncnn_0 258:12 259:12 -23330=4,3,14,14,192 0=2\nConvolution op_236 1 1 259:12 261:12 -23330=4,3,14,14,96 0=96 1=1 4=-233 5=1 6=18432\nSplit splitncnn_6 1 2 261:12 261:12_splitncnn_0 261:12_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution op_237 1 1 261:12_splitncnn_1 263:12 -23330=4,3,14,14,384 0=384 1=1 4=-233 5=1 6=36864\nSwish op_238 1 1 263:12 265:12 -23330=4,3,14,14,384\nConvolutionDepthWise op_239 1 1 265:12 268:12 -23330=4,3,14,14,384 0=384 1=3 4=-233 5=1 6=3456 7=384\nSwish op_241 1 1 268:12 270:12 -23330=4,3,14,14,384\nSplit splitncnn_7 1 2 270:12 270:12_splitncnn_0 270:12_splitncnn_1 -23330=8,3,14,14,384,3,14,14,384\nReduction op_242 1 1 270:12_splitncnn_1 271:12 -23330=4,3,1,1,384 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_243 1 1 271:12 273:12 -23330=4,3,1,1,24 0=24 1=1 4=-233 5=1 6=9216\nSwish op_244 1 1 273:12 275:12 -23330=4,3,1,1,24\nConvolution op_245 1 1 275:12 278:12 -23330=4,3,1,1,384 0=384 1=1 4=-233 5=1 6=9216 9=4\nBinaryOp op_247 2 1 270:12_splitncnn_0 278:12 279:12 -23330=4,3,14,14,384 0=2\nConvolution op_248 1 1 279:12 280:12 -23330=4,3,14,14,96 0=96 1=1 4=-233 6=36864\nEltwise op_249 2 1 261:12_splitncnn_0 280:12 281:12 -23330=4,3,14,14,96 0=1\nBinaryOp op_250 2 1 281:12 77:11 282:12 -23330=4,3,14,14,96\nSplit splitncnn_8 1 2 282:12 282:12_splitncnn_0 282:12_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution op_251 1 1 282:12_splitncnn_1 284:12 -23330=4,3,14,14,384 0=384 1=1 4=-233 5=1 6=36864\nSwish op_252 1 1 284:12 286:12 -23330=4,3,14,14,384\nConvolutionDepthWise op_253 1 1 286:12 289:12 -23330=4,3,14,14,384 0=384 1=3 4=-233 5=1 6=3456 7=384\nSwish op_255 1 1 289:12 291:12 -23330=4,3,14,14,384\nSplit splitncnn_9 1 2 291:12 291:12_splitncnn_0 291:12_splitncnn_1 -23330=8,3,14,14,384,3,14,14,384\nReduction op_256 1 1 291:12_splitncnn_1 292:12 -23330=4,3,1,1,384 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_257 1 1 292:12 294:12 -23330=4,3,1,1,24 0=24 1=1 4=-233 5=1 6=9216\nSwish op_258 1 1 294:12 296:12 -23330=4,3,1,1,24\nConvolution op_259 1 1 296:12 299:12 -23330=4,3,1,1,384 0=384 1=1 4=-233 5=1 6=9216 9=4\nBinaryOp op_261 2 1 291:12_splitncnn_0 299:12 300:12 -23330=4,3,14,14,384 0=2\nConvolution op_262 1 1 300:12 301:12 -23330=4,3,14,14,96 0=96 1=1 4=-233 6=36864\nEltwise op_263 2 1 282:12_splitncnn_0 301:12 302:12 -23330=4,3,14,14,96 0=1\nBinaryOp op_264 2 1 302:12 88:11 303:12 -23330=4,3,14,14,96\nConvolution op_265 1 1 303:12 305:12 -23330=4,3,14,14,576 0=576 1=1 4=-233 5=1 6=55296\nSwish op_266 1 1 305:12 307:12 -23330=4,3,14,14,576\nConvolutionDepthWise op_267 1 1 307:12 310:12 -23330=4,3,14,14,576 0=576 1=3 4=-233 5=1 6=5184 7=576\nSwish op_269 1 1 310:12 312:12 -23330=4,3,14,14,576\nSplit splitncnn_10 1 2 312:12 312:12_splitncnn_0 312:12_splitncnn_1 -23330=8,3,14,14,576,3,14,14,576\nReduction op_270 1 1 312:12_splitncnn_1 313:12 -23330=4,3,1,1,576 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_271 1 1 313:12 315:12 -23330=4,3,1,1,24 0=24 1=1 4=-233 5=1 6=13824\nSwish op_272 1 1 315:12 317:12 -23330=4,3,1,1,24\nConvolution op_273 1 1 317:12 320:12 -23330=4,3,1,1,576 0=576 1=1 4=-233 5=1 6=13824 9=4\nBinaryOp op_275 2 1 312:12_splitncnn_0 320:12 321:12 -23330=4,3,14,14,576 0=2\nConvolution op_276 1 1 321:12 323:12 -23330=4,3,14,14,112 0=112 1=1 4=-233 5=1 6=64512\nSplit splitncnn_11 1 2 323:12 323:12_splitncnn_0 323:12_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution op_277 1 1 323:12_splitncnn_1 325:12 -23330=4,3,14,14,672 0=672 1=1 4=-233 5=1 6=75264\nSwish op_278 1 1 325:12 327:12 -23330=4,3,14,14,672\nConvolutionDepthWise op_279 1 1 327:12 330:12 -23330=4,3,14,14,672 0=672 1=3 4=-233 5=1 6=6048 7=672\nSwish op_281 1 1 330:12 332:12 -23330=4,3,14,14,672\nSplit splitncnn_12 1 2 332:12 332:12_splitncnn_0 332:12_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nReduction op_282 1 1 332:12_splitncnn_1 333:12 -23330=4,3,1,1,672 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_283 1 1 333:12 335:12 -23330=4,3,1,1,28 0=28 1=1 4=-233 5=1 6=18816\nSwish op_284 1 1 335:12 337:12 -23330=4,3,1,1,28\nConvolution op_285 1 1 337:12 340:12 -23330=4,3,1,1,672 0=672 1=1 4=-233 5=1 6=18816 9=4\nBinaryOp op_287 2 1 332:12_splitncnn_0 340:12 341:12 -23330=4,3,14,14,672 0=2\nConvolution op_288 1 1 341:12 342:12 -23330=4,3,14,14,112 0=112 1=1 4=-233 6=75264\nEltwise op_289 2 1 323:12_splitncnn_0 342:12 343:12 -23330=4,3,14,14,112 0=1\nBinaryOp op_290 2 1 343:12 110:12 344:12 -23330=4,3,14,14,112\nSplit splitncnn_13 1 2 344:12 344:12_splitncnn_0 344:12_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution op_291 1 1 344:12_splitncnn_1 346:12 -23330=4,3,14,14,672 0=672 1=1 4=-233 5=1 6=75264\nSwish op_292 1 1 346:12 348:12 -23330=4,3,14,14,672\nConvolutionDepthWise op_293 1 1 348:12 351:12 -23330=4,3,14,14,672 0=672 1=3 4=-233 5=1 6=6048 7=672\nSwish op_295 1 1 351:12 353:12 -23330=4,3,14,14,672\nSplit splitncnn_14 1 2 353:12 353:12_splitncnn_0 353:12_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nReduction op_296 1 1 353:12_splitncnn_1 354:12 -23330=4,3,1,1,672 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_297 1 1 354:12 356:12 -23330=4,3,1,1,28 0=28 1=1 4=-233 5=1 6=18816\nSwish op_298 1 1 356:12 358:12 -23330=4,3,1,1,28\nConvolution op_299 1 1 358:12 361:12 -23330=4,3,1,1,672 0=672 1=1 4=-233 5=1 6=18816 9=4\nBinaryOp op_301 2 1 353:12_splitncnn_0 361:12 362:12 -23330=4,3,14,14,672 0=2\nConvolution op_302 1 1 362:12 363:12 -23330=4,3,14,14,112 0=112 1=1 4=-233 6=75264\nEltwise op_303 2 1 363:12 344:12_splitncnn_0 364:12 -23330=4,3,14,14,112 0=1\nBinaryOp op_304 2 1 364:12 33:11 365:12 -23330=4,3,14,14,112\nSplit splitncnn_15 1 2 365:12 365:12_splitncnn_0 365:12_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution op_305 1 1 365:12_splitncnn_1 367:12 -23330=4,3,14,14,672 0=672 1=1 4=-233 5=1 6=75264\nSwish op_306 1 1 367:12 369:12 -23330=4,3,14,14,672\nConvolutionDepthWise op_307 1 1 369:12 372:12 -23330=4,3,14,14,672 0=672 1=3 4=-233 5=1 6=6048 7=672\nSwish op_309 1 1 372:12 374:12 -23330=4,3,14,14,672\nSplit splitncnn_16 1 2 374:12 374:12_splitncnn_0 374:12_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nReduction op_310 1 1 374:12_splitncnn_1 375:12 -23330=4,3,1,1,672 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_311 1 1 375:12 377:12 -23330=4,3,1,1,28 0=28 1=1 4=-233 5=1 6=18816\nSwish op_312 1 1 377:12 379:12 -23330=4,3,1,1,28\nConvolution op_313 1 1 379:12 382:12 -23330=4,3,1,1,672 0=672 1=1 4=-233 5=1 6=18816 9=4\nBinaryOp op_315 2 1 374:12_splitncnn_0 382:12 383:12 -23330=4,3,14,14,672 0=2\nConvolution op_316 1 1 383:12 384:12 -23330=4,3,14,14,112 0=112 1=1 4=-233 6=75264\nEltwise op_317 2 1 365:12_splitncnn_0 384:12 385:12 -23330=4,3,14,14,112 0=1\nBinaryOp op_318 2 1 385:12 44:11 386:12 -23330=4,3,14,14,112\nSplit splitncnn_17 1 2 386:12 386:12_splitncnn_0 386:12_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution op_319 1 1 386:12_splitncnn_1 388:12 -23330=4,3,14,14,672 0=672 1=1 4=-233 5=1 6=75264\nSwish op_320 1 1 388:12 390:12 -23330=4,3,14,14,672\nConvolutionDepthWise op_321 1 1 390:12 393:12 -23330=4,3,14,14,672 0=672 1=3 4=-233 5=1 6=6048 7=672\nSwish op_323 1 1 393:12 395:12 -23330=4,3,14,14,672\nSplit splitncnn_18 1 2 395:12 395:12_splitncnn_0 395:12_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nReduction op_324 1 1 395:12_splitncnn_1 396:12 -23330=4,3,1,1,672 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_325 1 1 396:12 398:12 -23330=4,3,1,1,28 0=28 1=1 4=-233 5=1 6=18816\nSwish op_326 1 1 398:12 400:12 -23330=4,3,1,1,28\nConvolution op_327 1 1 400:12 403:12 -23330=4,3,1,1,672 0=672 1=1 4=-233 5=1 6=18816 9=4\nBinaryOp op_329 2 1 395:12_splitncnn_0 403:12 404:12 -23330=4,3,14,14,672 0=2\nConvolution op_330 1 1 404:12 405:12 -23330=4,3,14,14,112 0=112 1=1 4=-233 6=75264\nEltwise op_331 2 1 386:12_splitncnn_0 405:12 406:12 -23330=4,3,14,14,112 0=1\nBinaryOp op_332 2 1 406:12 55:11 407:12_splitncnn_0 -23330=4,3,14,14,112\nConvolution op_334 1 1 407:12_splitncnn_0 410:12 -23330=4,3,14,14,672 0=672 1=1 4=-233 5=1 6=75264\nSwish op_335 1 1 410:12 412:12 -23330=4,3,14,14,672\nConvolutionDepthWise op_336 1 1 412:12 415:12 -23330=4,3,7,7,672 0=672 1=3 3=2 4=-233 5=1 6=6048 7=672\nSwish op_338 1 1 415:12 417:12 -23330=4,3,7,7,672\nSplit splitncnn_20 1 2 417:12 417:12_splitncnn_0 417:12_splitncnn_1 -23330=8,3,7,7,672,3,7,7,672\nReduction op_339 1 1 417:12_splitncnn_1 418:12 -23330=4,3,1,1,672 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_340 1 1 418:12 420:12 -23330=4,3,1,1,28 0=28 1=1 4=-233 5=1 6=18816\nSwish op_341 1 1 420:12 422:12 -23330=4,3,1,1,28\nConvolution op_342 1 1 422:12 425:12 -23330=4,3,1,1,672 0=672 1=1 4=-233 5=1 6=18816 9=4\nBinaryOp op_344 2 1 417:12_splitncnn_0 425:12 426:12 -23330=4,3,7,7,672 0=2\nConvolution op_345 1 1 426:12 428:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 5=1 6=129024\nSplit splitncnn_21 1 2 428:12 428:12_splitncnn_0 428:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_346 1 1 428:12_splitncnn_1 430:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_347 1 1 430:12 432:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_348 1 1 432:12 435:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_350 1 1 435:12 437:12 -23330=4,3,7,7,1152\nSplit splitncnn_22 1 2 437:12 437:12_splitncnn_0 437:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_351 1 1 437:12_splitncnn_1 438:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_352 1 1 438:12 440:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_353 1 1 440:12 442:12 -23330=4,3,1,1,48\nConvolution op_354 1 1 442:12 445:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_356 2 1 437:12_splitncnn_0 445:12 446:12 -23330=4,3,7,7,1152 0=2\nConvolution op_357 1 1 446:12 447:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_358 2 1 428:12_splitncnn_0 447:12 448:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_359 2 1 448:12 133:12 449:12 -23330=4,3,7,7,192\nSplit splitncnn_23 1 2 449:12 449:12_splitncnn_0 449:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_360 1 1 449:12_splitncnn_1 451:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_361 1 1 451:12 453:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_362 1 1 453:12 456:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_364 1 1 456:12 458:12 -23330=4,3,7,7,1152\nSplit splitncnn_24 1 2 458:12 458:12_splitncnn_0 458:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_365 1 1 458:12_splitncnn_1 459:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_366 1 1 459:12 461:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_367 1 1 461:12 463:12 -23330=4,3,1,1,48\nConvolution op_368 1 1 463:12 466:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_370 2 1 458:12_splitncnn_0 466:12 467:12 -23330=4,3,7,7,1152 0=2\nConvolution op_371 1 1 467:12 468:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_372 2 1 449:12_splitncnn_0 468:12 469:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_373 2 1 469:12 144:12 470:12 -23330=4,3,7,7,192\nSplit splitncnn_25 1 2 470:12 470:12_splitncnn_0 470:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_374 1 1 470:12_splitncnn_1 472:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_375 1 1 472:12 474:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_376 1 1 474:12 477:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_378 1 1 477:12 479:12 -23330=4,3,7,7,1152\nSplit splitncnn_26 1 2 479:12 479:12_splitncnn_0 479:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_379 1 1 479:12_splitncnn_1 480:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_380 1 1 480:12 482:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_381 1 1 482:12 484:12 -23330=4,3,1,1,48\nConvolution op_382 1 1 484:12 487:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_384 2 1 479:12_splitncnn_0 487:12 488:12 -23330=4,3,7,7,1152 0=2\nConvolution op_385 1 1 488:12 489:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_386 2 1 470:12_splitncnn_0 489:12 490:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_387 2 1 490:12 155:12 491:12 -23330=4,3,7,7,192\nSplit splitncnn_27 1 2 491:12 491:12_splitncnn_0 491:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_388 1 1 491:12_splitncnn_1 493:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_389 1 1 493:12 495:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_390 1 1 495:12 498:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_392 1 1 498:12 500:12 -23330=4,3,7,7,1152\nSplit splitncnn_28 1 2 500:12 500:12_splitncnn_0 500:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_393 1 1 500:12_splitncnn_1 501:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_394 1 1 501:12 503:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_395 1 1 503:12 505:12 -23330=4,3,1,1,48\nConvolution op_396 1 1 505:12 508:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_398 2 1 500:12_splitncnn_0 508:12 509:12 -23330=4,3,7,7,1152 0=2\nConvolution op_399 1 1 509:12 510:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_400 2 1 491:12_splitncnn_0 510:12 511:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_401 2 1 511:12 166:12 512:12 -23330=4,3,7,7,192\nSplit splitncnn_29 1 2 512:12 512:12_splitncnn_0 512:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_402 1 1 512:12_splitncnn_1 514:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_403 1 1 514:12 516:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_404 1 1 516:12 519:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_406 1 1 519:12 521:12 -23330=4,3,7,7,1152\nSplit splitncnn_30 1 2 521:12 521:12_splitncnn_0 521:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_407 1 1 521:12_splitncnn_1 522:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_408 1 1 522:12 524:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_409 1 1 524:12 526:12 -23330=4,3,1,1,48\nConvolution op_410 1 1 526:12 529:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_412 2 1 521:12_splitncnn_0 529:12 530:12 -23330=4,3,7,7,1152 0=2\nConvolution op_413 1 1 530:12 531:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_414 2 1 512:12_splitncnn_0 531:12 532:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_415 2 1 532:12 177:12 533:12 -23330=4,3,7,7,192\nSplit splitncnn_31 1 2 533:12 533:12_splitncnn_0 533:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_416 1 1 533:12_splitncnn_1 535:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_417 1 1 535:12 537:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_418 1 1 537:12 540:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_420 1 1 540:12 542:12 -23330=4,3,7,7,1152\nSplit splitncnn_32 1 2 542:12 542:12_splitncnn_0 542:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_421 1 1 542:12_splitncnn_1 543:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_422 1 1 543:12 545:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_423 1 1 545:12 547:12 -23330=4,3,1,1,48\nConvolution op_424 1 1 547:12 550:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_426 2 1 542:12_splitncnn_0 550:12 551:12 -23330=4,3,7,7,1152 0=2\nConvolution op_427 1 1 551:12 552:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_428 2 1 533:12_splitncnn_0 552:12 553:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_429 2 1 553:12 188:12 554:12 -23330=4,3,7,7,192\nSplit splitncnn_33 1 2 554:12 554:12_splitncnn_0 554:12_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution op_430 1 1 554:12_splitncnn_1 556:12 -23330=4,3,7,7,1152 0=1152 1=1 4=-233 5=1 6=221184\nSwish op_431 1 1 556:12 558:12 -23330=4,3,7,7,1152\nConvolutionDepthWise op_432 1 1 558:12 561:12 -23330=4,3,7,7,1152 0=1152 1=3 4=-233 5=1 6=10368 7=1152\nSwish op_434 1 1 561:12 563:12 -23330=4,3,7,7,1152\nSplit splitncnn_34 1 2 563:12 563:12_splitncnn_0 563:12_splitncnn_1 -23330=8,3,7,7,1152,3,7,7,1152\nReduction op_435 1 1 563:12_splitncnn_1 564:12 -23330=4,3,1,1,1152 0=3 1=0 -23303=2,1,2 4=1 5=1\nConvolution op_436 1 1 564:12 566:12 -23330=4,3,1,1,48 0=48 1=1 4=-233 5=1 6=55296\nSwish op_437 1 1 566:12 568:12 -23330=4,3,1,1,48\nConvolution op_438 1 1 568:12 571:12 -23330=4,3,1,1,1152 0=1152 1=1 4=-233 5=1 6=55296 9=4\nBinaryOp op_440 2 1 563:12_splitncnn_0 571:12 572:12 -23330=4,3,7,7,1152 0=2\nConvolution op_441 1 1 572:12 573:12 -23330=4,3,7,7,192 0=192 1=1 4=-233 6=221184\nEltwise op_442 2 1 554:12_splitncnn_0 573:12 574:12 -23330=4,3,7,7,192 0=1\nBinaryOp op_443 2 1 574:12 199:12 575:12_splitncnn_0 -23330=4,3,7,7,192\nConvolution op_445 1 1 575:12_splitncnn_0 578:12 -23330=4,3,7,7,1280 0=1280 1=1 4=-233 5=1 6=245760\nSwish op_446 1 1 578:12 580:12 -23330=4,3,7,7,1280\nPooling op_447 1 1 580:12 581:12 -23330=4,1,1280,1,1 0=1 4=1\nInnerProduct op_448 1 1 581:12 584:12 -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\n" }, { "path": "benchmark/googlenet.param", "content": "7767517\n94 121\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1/7x7_s2 1 1 data conv1/7x7_s2_conv1/relu_7x7 -23330=4,3,112,112,64 0=64 1=7 3=2 4=3 5=1 6=9408 9=1\nPooling pool1/3x3_s2 1 1 conv1/7x7_s2_conv1/relu_7x7 pool1/3x3_s2 -23330=4,3,56,56,64 1=3 2=2\nLRN pool1/norm1 1 1 pool1/3x3_s2 pool1/norm1 -23330=4,3,56,56,64 2=1.000000e-04\nConvolution conv2/3x3_reduce 1 1 pool1/norm1 conv2/3x3_reduce_conv2/relu_3x3_reduce -23330=4,3,56,56,64 0=64 1=1 5=1 6=4096 9=1\nConvolution conv2/3x3 1 1 conv2/3x3_reduce_conv2/relu_3x3_reduce conv2/3x3_conv2/relu_3x3 -23330=4,3,56,56,192 0=192 1=3 4=1 5=1 6=110592 9=1\nLRN conv2/norm2 1 1 conv2/3x3_conv2/relu_3x3 conv2/norm2 -23330=4,3,56,56,192 2=1.000000e-04\nPooling pool2/3x3_s2 1 1 conv2/norm2 pool2/3x3_s2 -23330=4,3,28,28,192 1=3 2=2\nSplit splitncnn_0 1 4 pool2/3x3_s2 pool2/3x3_s2_splitncnn_0 pool2/3x3_s2_splitncnn_1 pool2/3x3_s2_splitncnn_2 pool2/3x3_s2_splitncnn_3 -23330=16,3,28,28,192,3,28,28,192,3,28,28,192,3,28,28,192\nConvolution inception_3a/1x1 1 1 pool2/3x3_s2_splitncnn_3 inception_3a/1x1_inception_3a/relu_1x1 -23330=4,3,28,28,64 0=64 1=1 5=1 6=12288 9=1\nConvolution inception_3a/3x3_reduce 1 1 pool2/3x3_s2_splitncnn_2 inception_3a/3x3_reduce_inception_3a/relu_3x3_reduce -23330=4,3,28,28,96 0=96 1=1 5=1 6=18432 9=1\nConvolution inception_3a/3x3 1 1 inception_3a/3x3_reduce_inception_3a/relu_3x3_reduce inception_3a/3x3_inception_3a/relu_3x3 -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=110592 9=1\nConvolution inception_3a/5x5_reduce 1 1 pool2/3x3_s2_splitncnn_1 inception_3a/5x5_reduce_inception_3a/relu_5x5_reduce -23330=4,3,28,28,16 0=16 1=1 5=1 6=3072 9=1\nConvolution inception_3a/5x5 1 1 inception_3a/5x5_reduce_inception_3a/relu_5x5_reduce inception_3a/5x5_inception_3a/relu_5x5 -23330=4,3,28,28,32 0=32 1=5 4=2 5=1 6=12800 9=1\nPooling inception_3a/pool 1 1 pool2/3x3_s2_splitncnn_0 inception_3a/pool -23330=4,3,28,28,192 1=3 3=1\nConvolution inception_3a/pool_proj 1 1 inception_3a/pool inception_3a/pool_proj_inception_3a/relu_pool_proj -23330=4,3,28,28,32 0=32 1=1 5=1 6=6144 9=1\nConcat inception_3a/output 4 1 inception_3a/1x1_inception_3a/relu_1x1 inception_3a/3x3_inception_3a/relu_3x3 inception_3a/5x5_inception_3a/relu_5x5 inception_3a/pool_proj_inception_3a/relu_pool_proj inception_3a/output -23330=4,3,28,28,256\nSplit splitncnn_1 1 4 inception_3a/output inception_3a/output_splitncnn_0 inception_3a/output_splitncnn_1 inception_3a/output_splitncnn_2 inception_3a/output_splitncnn_3 -23330=16,3,28,28,256,3,28,28,256,3,28,28,256,3,28,28,256\nConvolution inception_3b/1x1 1 1 inception_3a/output_splitncnn_3 inception_3b/1x1_inception_3b/relu_1x1 -23330=4,3,28,28,128 0=128 1=1 5=1 6=32768 9=1\nConvolution inception_3b/3x3_reduce 1 1 inception_3a/output_splitncnn_2 inception_3b/3x3_reduce_inception_3b/relu_3x3_reduce -23330=4,3,28,28,128 0=128 1=1 5=1 6=32768 9=1\nConvolution inception_3b/3x3 1 1 inception_3b/3x3_reduce_inception_3b/relu_3x3_reduce inception_3b/3x3_inception_3b/relu_3x3 -23330=4,3,28,28,192 0=192 1=3 4=1 5=1 6=221184 9=1\nConvolution inception_3b/5x5_reduce 1 1 inception_3a/output_splitncnn_1 inception_3b/5x5_reduce_inception_3b/relu_5x5_reduce -23330=4,3,28,28,32 0=32 1=1 5=1 6=8192 9=1\nConvolution inception_3b/5x5 1 1 inception_3b/5x5_reduce_inception_3b/relu_5x5_reduce inception_3b/5x5_inception_3b/relu_5x5 -23330=4,3,28,28,96 0=96 1=5 4=2 5=1 6=76800 9=1\nPooling inception_3b/pool 1 1 inception_3a/output_splitncnn_0 inception_3b/pool -23330=4,3,28,28,256 1=3 3=1\nConvolution inception_3b/pool_proj 1 1 inception_3b/pool inception_3b/pool_proj_inception_3b/relu_pool_proj -23330=4,3,28,28,64 0=64 1=1 5=1 6=16384 9=1\nConcat inception_3b/output 4 1 inception_3b/1x1_inception_3b/relu_1x1 inception_3b/3x3_inception_3b/relu_3x3 inception_3b/5x5_inception_3b/relu_5x5 inception_3b/pool_proj_inception_3b/relu_pool_proj inception_3b/output -23330=4,3,28,28,480\nPooling pool3/3x3_s2 1 1 inception_3b/output pool3/3x3_s2 -23330=4,3,14,14,480 1=3 2=2\nSplit splitncnn_2 1 4 pool3/3x3_s2 pool3/3x3_s2_splitncnn_0 pool3/3x3_s2_splitncnn_1 pool3/3x3_s2_splitncnn_2 pool3/3x3_s2_splitncnn_3 -23330=16,3,14,14,480,3,14,14,480,3,14,14,480,3,14,14,480\nConvolution inception_4a/1x1 1 1 pool3/3x3_s2_splitncnn_3 inception_4a/1x1_inception_4a/relu_1x1 -23330=4,3,14,14,192 0=192 1=1 5=1 6=92160 9=1\nConvolution inception_4a/3x3_reduce 1 1 pool3/3x3_s2_splitncnn_2 inception_4a/3x3_reduce_inception_4a/relu_3x3_reduce -23330=4,3,14,14,96 0=96 1=1 5=1 6=46080 9=1\nConvolution inception_4a/3x3 1 1 inception_4a/3x3_reduce_inception_4a/relu_3x3_reduce inception_4a/3x3_inception_4a/relu_3x3 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=179712 9=1\nConvolution inception_4a/5x5_reduce 1 1 pool3/3x3_s2_splitncnn_1 inception_4a/5x5_reduce_inception_4a/relu_5x5_reduce -23330=4,3,14,14,16 0=16 1=1 5=1 6=7680 9=1\nConvolution inception_4a/5x5 1 1 inception_4a/5x5_reduce_inception_4a/relu_5x5_reduce inception_4a/5x5_inception_4a/relu_5x5 -23330=4,3,14,14,48 0=48 1=5 4=2 5=1 6=19200 9=1\nPooling inception_4a/pool 1 1 pool3/3x3_s2_splitncnn_0 inception_4a/pool -23330=4,3,14,14,480 1=3 3=1\nConvolution inception_4a/pool_proj 1 1 inception_4a/pool inception_4a/pool_proj_inception_4a/relu_pool_proj -23330=4,3,14,14,64 0=64 1=1 5=1 6=30720 9=1\nConcat inception_4a/output 4 1 inception_4a/1x1_inception_4a/relu_1x1 inception_4a/3x3_inception_4a/relu_3x3 inception_4a/5x5_inception_4a/relu_5x5 inception_4a/pool_proj_inception_4a/relu_pool_proj inception_4a/output -23330=4,3,14,14,512\nSplit splitncnn_3 1 4 inception_4a/output inception_4a/output_splitncnn_0 inception_4a/output_splitncnn_1 inception_4a/output_splitncnn_2 inception_4a/output_splitncnn_3 -23330=16,3,14,14,512,3,14,14,512,3,14,14,512,3,14,14,512\nConvolution inception_4b/1x1 1 1 inception_4a/output_splitncnn_3 inception_4b/1x1_inception_4b/relu_1x1 -23330=4,3,14,14,160 0=160 1=1 5=1 6=81920 9=1\nConvolution inception_4b/3x3_reduce 1 1 inception_4a/output_splitncnn_2 inception_4b/3x3_reduce_inception_4b/relu_3x3_reduce -23330=4,3,14,14,112 0=112 1=1 5=1 6=57344 9=1\nConvolution inception_4b/3x3 1 1 inception_4b/3x3_reduce_inception_4b/relu_3x3_reduce inception_4b/3x3_inception_4b/relu_3x3 -23330=4,3,14,14,224 0=224 1=3 4=1 5=1 6=225792 9=1\nConvolution inception_4b/5x5_reduce 1 1 inception_4a/output_splitncnn_1 inception_4b/5x5_reduce_inception_4b/relu_5x5_reduce -23330=4,3,14,14,24 0=24 1=1 5=1 6=12288 9=1\nConvolution inception_4b/5x5 1 1 inception_4b/5x5_reduce_inception_4b/relu_5x5_reduce inception_4b/5x5_inception_4b/relu_5x5 -23330=4,3,14,14,64 0=64 1=5 4=2 5=1 6=38400 9=1\nPooling inception_4b/pool 1 1 inception_4a/output_splitncnn_0 inception_4b/pool -23330=4,3,14,14,512 1=3 3=1\nConvolution inception_4b/pool_proj 1 1 inception_4b/pool inception_4b/pool_proj_inception_4b/relu_pool_proj -23330=4,3,14,14,64 0=64 1=1 5=1 6=32768 9=1\nConcat inception_4b/output 4 1 inception_4b/1x1_inception_4b/relu_1x1 inception_4b/3x3_inception_4b/relu_3x3 inception_4b/5x5_inception_4b/relu_5x5 inception_4b/pool_proj_inception_4b/relu_pool_proj inception_4b/output -23330=4,3,14,14,512\nSplit splitncnn_4 1 4 inception_4b/output inception_4b/output_splitncnn_0 inception_4b/output_splitncnn_1 inception_4b/output_splitncnn_2 inception_4b/output_splitncnn_3 -23330=16,3,14,14,512,3,14,14,512,3,14,14,512,3,14,14,512\nConvolution inception_4c/1x1 1 1 inception_4b/output_splitncnn_3 inception_4c/1x1_inception_4c/relu_1x1 -23330=4,3,14,14,128 0=128 1=1 5=1 6=65536 9=1\nConvolution inception_4c/3x3_reduce 1 1 inception_4b/output_splitncnn_2 inception_4c/3x3_reduce_inception_4c/relu_3x3_reduce -23330=4,3,14,14,128 0=128 1=1 5=1 6=65536 9=1\nConvolution inception_4c/3x3 1 1 inception_4c/3x3_reduce_inception_4c/relu_3x3_reduce inception_4c/3x3_inception_4c/relu_3x3 -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=294912 9=1\nConvolution inception_4c/5x5_reduce 1 1 inception_4b/output_splitncnn_1 inception_4c/5x5_reduce_inception_4c/relu_5x5_reduce -23330=4,3,14,14,24 0=24 1=1 5=1 6=12288 9=1\nConvolution inception_4c/5x5 1 1 inception_4c/5x5_reduce_inception_4c/relu_5x5_reduce inception_4c/5x5_inception_4c/relu_5x5 -23330=4,3,14,14,64 0=64 1=5 4=2 5=1 6=38400 9=1\nPooling inception_4c/pool 1 1 inception_4b/output_splitncnn_0 inception_4c/pool -23330=4,3,14,14,512 1=3 3=1\nConvolution inception_4c/pool_proj 1 1 inception_4c/pool inception_4c/pool_proj_inception_4c/relu_pool_proj -23330=4,3,14,14,64 0=64 1=1 5=1 6=32768 9=1\nConcat inception_4c/output 4 1 inception_4c/1x1_inception_4c/relu_1x1 inception_4c/3x3_inception_4c/relu_3x3 inception_4c/5x5_inception_4c/relu_5x5 inception_4c/pool_proj_inception_4c/relu_pool_proj inception_4c/output -23330=4,3,14,14,512\nSplit splitncnn_5 1 4 inception_4c/output inception_4c/output_splitncnn_0 inception_4c/output_splitncnn_1 inception_4c/output_splitncnn_2 inception_4c/output_splitncnn_3 -23330=16,3,14,14,512,3,14,14,512,3,14,14,512,3,14,14,512\nConvolution inception_4d/1x1 1 1 inception_4c/output_splitncnn_3 inception_4d/1x1_inception_4d/relu_1x1 -23330=4,3,14,14,112 0=112 1=1 5=1 6=57344 9=1\nConvolution inception_4d/3x3_reduce 1 1 inception_4c/output_splitncnn_2 inception_4d/3x3_reduce_inception_4d/relu_3x3_reduce -23330=4,3,14,14,144 0=144 1=1 5=1 6=73728 9=1\nConvolution inception_4d/3x3 1 1 inception_4d/3x3_reduce_inception_4d/relu_3x3_reduce inception_4d/3x3_inception_4d/relu_3x3 -23330=4,3,14,14,288 0=288 1=3 4=1 5=1 6=373248 9=1\nConvolution inception_4d/5x5_reduce 1 1 inception_4c/output_splitncnn_1 inception_4d/5x5_reduce_inception_4d/relu_5x5_reduce -23330=4,3,14,14,32 0=32 1=1 5=1 6=16384 9=1\nConvolution inception_4d/5x5 1 1 inception_4d/5x5_reduce_inception_4d/relu_5x5_reduce inception_4d/5x5_inception_4d/relu_5x5 -23330=4,3,14,14,64 0=64 1=5 4=2 5=1 6=51200 9=1\nPooling inception_4d/pool 1 1 inception_4c/output_splitncnn_0 inception_4d/pool -23330=4,3,14,14,512 1=3 3=1\nConvolution inception_4d/pool_proj 1 1 inception_4d/pool inception_4d/pool_proj_inception_4d/relu_pool_proj -23330=4,3,14,14,64 0=64 1=1 5=1 6=32768 9=1\nConcat inception_4d/output 4 1 inception_4d/1x1_inception_4d/relu_1x1 inception_4d/3x3_inception_4d/relu_3x3 inception_4d/5x5_inception_4d/relu_5x5 inception_4d/pool_proj_inception_4d/relu_pool_proj inception_4d/output -23330=4,3,14,14,528\nSplit splitncnn_6 1 4 inception_4d/output inception_4d/output_splitncnn_0 inception_4d/output_splitncnn_1 inception_4d/output_splitncnn_2 inception_4d/output_splitncnn_3 -23330=16,3,14,14,528,3,14,14,528,3,14,14,528,3,14,14,528\nConvolution inception_4e/1x1 1 1 inception_4d/output_splitncnn_3 inception_4e/1x1_inception_4e/relu_1x1 -23330=4,3,14,14,256 0=256 1=1 5=1 6=135168 9=1\nConvolution inception_4e/3x3_reduce 1 1 inception_4d/output_splitncnn_2 inception_4e/3x3_reduce_inception_4e/relu_3x3_reduce -23330=4,3,14,14,160 0=160 1=1 5=1 6=84480 9=1\nConvolution inception_4e/3x3 1 1 inception_4e/3x3_reduce_inception_4e/relu_3x3_reduce inception_4e/3x3_inception_4e/relu_3x3 -23330=4,3,14,14,320 0=320 1=3 4=1 5=1 6=460800 9=1\nConvolution inception_4e/5x5_reduce 1 1 inception_4d/output_splitncnn_1 inception_4e/5x5_reduce_inception_4e/relu_5x5_reduce -23330=4,3,14,14,32 0=32 1=1 5=1 6=16896 9=1\nConvolution inception_4e/5x5 1 1 inception_4e/5x5_reduce_inception_4e/relu_5x5_reduce inception_4e/5x5_inception_4e/relu_5x5 -23330=4,3,14,14,128 0=128 1=5 4=2 5=1 6=102400 9=1\nPooling inception_4e/pool 1 1 inception_4d/output_splitncnn_0 inception_4e/pool -23330=4,3,14,14,528 1=3 3=1\nConvolution inception_4e/pool_proj 1 1 inception_4e/pool inception_4e/pool_proj_inception_4e/relu_pool_proj -23330=4,3,14,14,128 0=128 1=1 5=1 6=67584 9=1\nConcat inception_4e/output 4 1 inception_4e/1x1_inception_4e/relu_1x1 inception_4e/3x3_inception_4e/relu_3x3 inception_4e/5x5_inception_4e/relu_5x5 inception_4e/pool_proj_inception_4e/relu_pool_proj inception_4e/output -23330=4,3,14,14,832\nPooling pool4/3x3_s2 1 1 inception_4e/output pool4/3x3_s2 -23330=4,3,7,7,832 1=3 2=2\nSplit splitncnn_7 1 4 pool4/3x3_s2 pool4/3x3_s2_splitncnn_0 pool4/3x3_s2_splitncnn_1 pool4/3x3_s2_splitncnn_2 pool4/3x3_s2_splitncnn_3 -23330=16,3,7,7,832,3,7,7,832,3,7,7,832,3,7,7,832\nConvolution inception_5a/1x1 1 1 pool4/3x3_s2_splitncnn_3 inception_5a/1x1_inception_5a/relu_1x1 -23330=4,3,7,7,256 0=256 1=1 5=1 6=212992 9=1\nConvolution inception_5a/3x3_reduce 1 1 pool4/3x3_s2_splitncnn_2 inception_5a/3x3_reduce_inception_5a/relu_3x3_reduce -23330=4,3,7,7,160 0=160 1=1 5=1 6=133120 9=1\nConvolution inception_5a/3x3 1 1 inception_5a/3x3_reduce_inception_5a/relu_3x3_reduce inception_5a/3x3_inception_5a/relu_3x3 -23330=4,3,7,7,320 0=320 1=3 4=1 5=1 6=460800 9=1\nConvolution inception_5a/5x5_reduce 1 1 pool4/3x3_s2_splitncnn_1 inception_5a/5x5_reduce_inception_5a/relu_5x5_reduce -23330=4,3,7,7,32 0=32 1=1 5=1 6=26624 9=1\nConvolution inception_5a/5x5 1 1 inception_5a/5x5_reduce_inception_5a/relu_5x5_reduce inception_5a/5x5_inception_5a/relu_5x5 -23330=4,3,7,7,128 0=128 1=5 4=2 5=1 6=102400 9=1\nPooling inception_5a/pool 1 1 pool4/3x3_s2_splitncnn_0 inception_5a/pool -23330=4,3,7,7,832 1=3 3=1\nConvolution inception_5a/pool_proj 1 1 inception_5a/pool inception_5a/pool_proj_inception_5a/relu_pool_proj -23330=4,3,7,7,128 0=128 1=1 5=1 6=106496 9=1\nConcat inception_5a/output 4 1 inception_5a/1x1_inception_5a/relu_1x1 inception_5a/3x3_inception_5a/relu_3x3 inception_5a/5x5_inception_5a/relu_5x5 inception_5a/pool_proj_inception_5a/relu_pool_proj inception_5a/output -23330=4,3,7,7,832\nSplit splitncnn_8 1 4 inception_5a/output inception_5a/output_splitncnn_0 inception_5a/output_splitncnn_1 inception_5a/output_splitncnn_2 inception_5a/output_splitncnn_3 -23330=16,3,7,7,832,3,7,7,832,3,7,7,832,3,7,7,832\nConvolution inception_5b/1x1 1 1 inception_5a/output_splitncnn_3 inception_5b/1x1_inception_5b/relu_1x1 -23330=4,3,7,7,384 0=384 1=1 5=1 6=319488 9=1\nConvolution inception_5b/3x3_reduce 1 1 inception_5a/output_splitncnn_2 inception_5b/3x3_reduce_inception_5b/relu_3x3_reduce -23330=4,3,7,7,192 0=192 1=1 5=1 6=159744 9=1\nConvolution inception_5b/3x3 1 1 inception_5b/3x3_reduce_inception_5b/relu_3x3_reduce inception_5b/3x3_inception_5b/relu_3x3 -23330=4,3,7,7,384 0=384 1=3 4=1 5=1 6=663552 9=1\nConvolution inception_5b/5x5_reduce 1 1 inception_5a/output_splitncnn_1 inception_5b/5x5_reduce_inception_5b/relu_5x5_reduce -23330=4,3,7,7,48 0=48 1=1 5=1 6=39936 9=1\nConvolution inception_5b/5x5 1 1 inception_5b/5x5_reduce_inception_5b/relu_5x5_reduce inception_5b/5x5_inception_5b/relu_5x5 -23330=4,3,7,7,128 0=128 1=5 4=2 5=1 6=153600 9=1\nPooling inception_5b/pool 1 1 inception_5a/output_splitncnn_0 inception_5b/pool -23330=4,3,7,7,832 1=3 3=1\nConvolution inception_5b/pool_proj 1 1 inception_5b/pool inception_5b/pool_proj_inception_5b/relu_pool_proj -23330=4,3,7,7,128 0=128 1=1 5=1 6=106496 9=1\nConcat inception_5b/output 4 1 inception_5b/1x1_inception_5b/relu_1x1 inception_5b/3x3_inception_5b/relu_3x3 inception_5b/5x5_inception_5b/relu_5x5 inception_5b/pool_proj_inception_5b/relu_pool_proj inception_5b/output -23330=4,3,7,7,1024\nPooling pool5/7x7_s1 1 1 inception_5b/output pool5/7x7_s1_pool5/drop_7x7_s1 -23330=4,3,1,1,1024 0=1 1=7\nInnerProduct loss3/classifier 1 1 pool5/7x7_s1_pool5/drop_7x7_s1 loss3/classifier -23330=4,1,1000,1,1 0=1000 1=1 2=1024000\nSoftmax prob 1 1 loss3/classifier output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/googlenet_int8.param", "content": "7767517\n94 121\nInput data 0 1 data 0=224 1=224 2=3\nConvolution conv1/7x7_s2 1 1 data conv1/7x7_s2_conv1/relu_7x7 0=64 1=7 3=2 4=3 5=1 6=9408 8=2 9=1\nPooling pool1/3x3_s2 1 1 conv1/7x7_s2_conv1/relu_7x7 pool1/3x3_s2 1=3 2=2\nLRN pool1/norm1 1 1 pool1/3x3_s2 pool1/norm1 2=0.000100\nConvolution conv2/3x3_reduce 1 1 pool1/norm1 conv2/3x3_reduce_conv2/relu_3x3_reduce 0=64 1=1 5=1 6=4096 8=102 9=1\nConvolution conv2/3x3 1 1 conv2/3x3_reduce_conv2/relu_3x3_reduce conv2/3x3_conv2/relu_3x3 0=192 1=3 4=1 5=1 6=110592 8=2 9=1\nLRN conv2/norm2 1 1 conv2/3x3_conv2/relu_3x3 conv2/norm2 2=0.000100\nPooling pool2/3x3_s2 1 1 conv2/norm2 pool2/3x3_s2 1=3 2=2\nSplit splitncnn_0 1 4 pool2/3x3_s2 pool2/3x3_s2_splitncnn_0 pool2/3x3_s2_splitncnn_1 pool2/3x3_s2_splitncnn_2 pool2/3x3_s2_splitncnn_3\nConvolution inception_3a/1x1 1 1 pool2/3x3_s2_splitncnn_3 inception_3a/1x1_inception_3a/relu_1x1 0=64 1=1 5=1 6=12288 8=2 9=1\nConvolution inception_3a/3x3_reduce 1 1 pool2/3x3_s2_splitncnn_2 inception_3a/3x3_reduce_inception_3a/relu_3x3_reduce 0=96 1=1 5=1 6=18432 8=102 9=1\nConvolution inception_3a/3x3 1 1 inception_3a/3x3_reduce_inception_3a/relu_3x3_reduce inception_3a/3x3_inception_3a/relu_3x3 0=128 1=3 4=1 5=1 6=110592 8=2 9=1\nConvolution inception_3a/5x5_reduce 1 1 pool2/3x3_s2_splitncnn_1 inception_3a/5x5_reduce_inception_3a/relu_5x5_reduce 0=16 1=1 5=1 6=3072 8=102 9=1\nConvolution inception_3a/5x5 1 1 inception_3a/5x5_reduce_inception_3a/relu_5x5_reduce inception_3a/5x5_inception_3a/relu_5x5 0=32 1=5 4=2 5=1 6=12800 8=2 9=1\nPooling inception_3a/pool 1 1 pool2/3x3_s2_splitncnn_0 inception_3a/pool 1=3 3=1\nConvolution inception_3a/pool_proj 1 1 inception_3a/pool inception_3a/pool_proj_inception_3a/relu_pool_proj 0=32 1=1 5=1 6=6144 8=2 9=1\nConcat inception_3a/output 4 1 inception_3a/1x1_inception_3a/relu_1x1 inception_3a/3x3_inception_3a/relu_3x3 inception_3a/5x5_inception_3a/relu_5x5 inception_3a/pool_proj_inception_3a/relu_pool_proj inception_3a/output\nSplit splitncnn_1 1 4 inception_3a/output inception_3a/output_splitncnn_0 inception_3a/output_splitncnn_1 inception_3a/output_splitncnn_2 inception_3a/output_splitncnn_3\nConvolution inception_3b/1x1 1 1 inception_3a/output_splitncnn_3 inception_3b/1x1_inception_3b/relu_1x1 0=128 1=1 5=1 6=32768 8=2 9=1\nConvolution inception_3b/3x3_reduce 1 1 inception_3a/output_splitncnn_2 inception_3b/3x3_reduce_inception_3b/relu_3x3_reduce 0=128 1=1 5=1 6=32768 8=102 9=1\nConvolution inception_3b/3x3 1 1 inception_3b/3x3_reduce_inception_3b/relu_3x3_reduce inception_3b/3x3_inception_3b/relu_3x3 0=192 1=3 4=1 5=1 6=221184 8=2 9=1\nConvolution inception_3b/5x5_reduce 1 1 inception_3a/output_splitncnn_1 inception_3b/5x5_reduce_inception_3b/relu_5x5_reduce 0=32 1=1 5=1 6=8192 8=102 9=1\nConvolution inception_3b/5x5 1 1 inception_3b/5x5_reduce_inception_3b/relu_5x5_reduce inception_3b/5x5_inception_3b/relu_5x5 0=96 1=5 4=2 5=1 6=76800 8=2 9=1\nPooling inception_3b/pool 1 1 inception_3a/output_splitncnn_0 inception_3b/pool 1=3 3=1\nConvolution inception_3b/pool_proj 1 1 inception_3b/pool inception_3b/pool_proj_inception_3b/relu_pool_proj 0=64 1=1 5=1 6=16384 8=2 9=1\nConcat inception_3b/output 4 1 inception_3b/1x1_inception_3b/relu_1x1 inception_3b/3x3_inception_3b/relu_3x3 inception_3b/5x5_inception_3b/relu_5x5 inception_3b/pool_proj_inception_3b/relu_pool_proj inception_3b/output\nPooling pool3/3x3_s2 1 1 inception_3b/output pool3/3x3_s2 1=3 2=2\nSplit splitncnn_2 1 4 pool3/3x3_s2 pool3/3x3_s2_splitncnn_0 pool3/3x3_s2_splitncnn_1 pool3/3x3_s2_splitncnn_2 pool3/3x3_s2_splitncnn_3\nConvolution inception_4a/1x1 1 1 pool3/3x3_s2_splitncnn_3 inception_4a/1x1_inception_4a/relu_1x1 0=192 1=1 5=1 6=92160 8=2 9=1\nConvolution inception_4a/3x3_reduce 1 1 pool3/3x3_s2_splitncnn_2 inception_4a/3x3_reduce_inception_4a/relu_3x3_reduce 0=96 1=1 5=1 6=46080 8=102 9=1\nConvolution inception_4a/3x3 1 1 inception_4a/3x3_reduce_inception_4a/relu_3x3_reduce inception_4a/3x3_inception_4a/relu_3x3 0=208 1=3 4=1 5=1 6=179712 8=2 9=1\nConvolution inception_4a/5x5_reduce 1 1 pool3/3x3_s2_splitncnn_1 inception_4a/5x5_reduce_inception_4a/relu_5x5_reduce 0=16 1=1 5=1 6=7680 8=102 9=1\nConvolution inception_4a/5x5 1 1 inception_4a/5x5_reduce_inception_4a/relu_5x5_reduce inception_4a/5x5_inception_4a/relu_5x5 0=48 1=5 4=2 5=1 6=19200 8=2 9=1\nPooling inception_4a/pool 1 1 pool3/3x3_s2_splitncnn_0 inception_4a/pool 1=3 3=1\nConvolution inception_4a/pool_proj 1 1 inception_4a/pool inception_4a/pool_proj_inception_4a/relu_pool_proj 0=64 1=1 5=1 6=30720 8=2 9=1\nConcat inception_4a/output 4 1 inception_4a/1x1_inception_4a/relu_1x1 inception_4a/3x3_inception_4a/relu_3x3 inception_4a/5x5_inception_4a/relu_5x5 inception_4a/pool_proj_inception_4a/relu_pool_proj inception_4a/output\nSplit splitncnn_3 1 4 inception_4a/output inception_4a/output_splitncnn_0 inception_4a/output_splitncnn_1 inception_4a/output_splitncnn_2 inception_4a/output_splitncnn_3\nConvolution inception_4b/1x1 1 1 inception_4a/output_splitncnn_3 inception_4b/1x1_inception_4b/relu_1x1 0=160 1=1 5=1 6=81920 8=2 9=1\nConvolution inception_4b/3x3_reduce 1 1 inception_4a/output_splitncnn_2 inception_4b/3x3_reduce_inception_4b/relu_3x3_reduce 0=112 1=1 5=1 6=57344 8=102 9=1\nConvolution inception_4b/3x3 1 1 inception_4b/3x3_reduce_inception_4b/relu_3x3_reduce inception_4b/3x3_inception_4b/relu_3x3 0=224 1=3 4=1 5=1 6=225792 8=2 9=1\nConvolution inception_4b/5x5_reduce 1 1 inception_4a/output_splitncnn_1 inception_4b/5x5_reduce_inception_4b/relu_5x5_reduce 0=24 1=1 5=1 6=12288 8=102 9=1\nConvolution inception_4b/5x5 1 1 inception_4b/5x5_reduce_inception_4b/relu_5x5_reduce inception_4b/5x5_inception_4b/relu_5x5 0=64 1=5 4=2 5=1 6=38400 8=2 9=1\nPooling inception_4b/pool 1 1 inception_4a/output_splitncnn_0 inception_4b/pool 1=3 3=1\nConvolution inception_4b/pool_proj 1 1 inception_4b/pool inception_4b/pool_proj_inception_4b/relu_pool_proj 0=64 1=1 5=1 6=32768 8=2 9=1\nConcat inception_4b/output 4 1 inception_4b/1x1_inception_4b/relu_1x1 inception_4b/3x3_inception_4b/relu_3x3 inception_4b/5x5_inception_4b/relu_5x5 inception_4b/pool_proj_inception_4b/relu_pool_proj inception_4b/output\nSplit splitncnn_4 1 4 inception_4b/output inception_4b/output_splitncnn_0 inception_4b/output_splitncnn_1 inception_4b/output_splitncnn_2 inception_4b/output_splitncnn_3\nConvolution inception_4c/1x1 1 1 inception_4b/output_splitncnn_3 inception_4c/1x1_inception_4c/relu_1x1 0=128 1=1 5=1 6=65536 8=2 9=1\nConvolution inception_4c/3x3_reduce 1 1 inception_4b/output_splitncnn_2 inception_4c/3x3_reduce_inception_4c/relu_3x3_reduce 0=128 1=1 5=1 6=65536 8=102 9=1\nConvolution inception_4c/3x3 1 1 inception_4c/3x3_reduce_inception_4c/relu_3x3_reduce inception_4c/3x3_inception_4c/relu_3x3 0=256 1=3 4=1 5=1 6=294912 8=2 9=1\nConvolution inception_4c/5x5_reduce 1 1 inception_4b/output_splitncnn_1 inception_4c/5x5_reduce_inception_4c/relu_5x5_reduce 0=24 1=1 5=1 6=12288 8=102 9=1\nConvolution inception_4c/5x5 1 1 inception_4c/5x5_reduce_inception_4c/relu_5x5_reduce inception_4c/5x5_inception_4c/relu_5x5 0=64 1=5 4=2 5=1 6=38400 8=2 9=1\nPooling inception_4c/pool 1 1 inception_4b/output_splitncnn_0 inception_4c/pool 1=3 3=1\nConvolution inception_4c/pool_proj 1 1 inception_4c/pool inception_4c/pool_proj_inception_4c/relu_pool_proj 0=64 1=1 5=1 6=32768 8=2 9=1\nConcat inception_4c/output 4 1 inception_4c/1x1_inception_4c/relu_1x1 inception_4c/3x3_inception_4c/relu_3x3 inception_4c/5x5_inception_4c/relu_5x5 inception_4c/pool_proj_inception_4c/relu_pool_proj inception_4c/output\nSplit splitncnn_5 1 4 inception_4c/output inception_4c/output_splitncnn_0 inception_4c/output_splitncnn_1 inception_4c/output_splitncnn_2 inception_4c/output_splitncnn_3\nConvolution inception_4d/1x1 1 1 inception_4c/output_splitncnn_3 inception_4d/1x1_inception_4d/relu_1x1 0=112 1=1 5=1 6=57344 8=2 9=1\nConvolution inception_4d/3x3_reduce 1 1 inception_4c/output_splitncnn_2 inception_4d/3x3_reduce_inception_4d/relu_3x3_reduce 0=144 1=1 5=1 6=73728 8=102 9=1\nConvolution inception_4d/3x3 1 1 inception_4d/3x3_reduce_inception_4d/relu_3x3_reduce inception_4d/3x3_inception_4d/relu_3x3 0=288 1=3 4=1 5=1 6=373248 8=2 9=1\nConvolution inception_4d/5x5_reduce 1 1 inception_4c/output_splitncnn_1 inception_4d/5x5_reduce_inception_4d/relu_5x5_reduce 0=32 1=1 5=1 6=16384 8=102 9=1\nConvolution inception_4d/5x5 1 1 inception_4d/5x5_reduce_inception_4d/relu_5x5_reduce inception_4d/5x5_inception_4d/relu_5x5 0=64 1=5 4=2 5=1 6=51200 8=2 9=1\nPooling inception_4d/pool 1 1 inception_4c/output_splitncnn_0 inception_4d/pool 1=3 3=1\nConvolution inception_4d/pool_proj 1 1 inception_4d/pool inception_4d/pool_proj_inception_4d/relu_pool_proj 0=64 1=1 5=1 6=32768 8=2 9=1\nConcat inception_4d/output 4 1 inception_4d/1x1_inception_4d/relu_1x1 inception_4d/3x3_inception_4d/relu_3x3 inception_4d/5x5_inception_4d/relu_5x5 inception_4d/pool_proj_inception_4d/relu_pool_proj inception_4d/output\nSplit splitncnn_6 1 4 inception_4d/output inception_4d/output_splitncnn_0 inception_4d/output_splitncnn_1 inception_4d/output_splitncnn_2 inception_4d/output_splitncnn_3\nConvolution inception_4e/1x1 1 1 inception_4d/output_splitncnn_3 inception_4e/1x1_inception_4e/relu_1x1 0=256 1=1 5=1 6=135168 8=2 9=1\nConvolution inception_4e/3x3_reduce 1 1 inception_4d/output_splitncnn_2 inception_4e/3x3_reduce_inception_4e/relu_3x3_reduce 0=160 1=1 5=1 6=84480 8=102 9=1\nConvolution inception_4e/3x3 1 1 inception_4e/3x3_reduce_inception_4e/relu_3x3_reduce inception_4e/3x3_inception_4e/relu_3x3 0=320 1=3 4=1 5=1 6=460800 8=2 9=1\nConvolution inception_4e/5x5_reduce 1 1 inception_4d/output_splitncnn_1 inception_4e/5x5_reduce_inception_4e/relu_5x5_reduce 0=32 1=1 5=1 6=16896 8=102 9=1\nConvolution inception_4e/5x5 1 1 inception_4e/5x5_reduce_inception_4e/relu_5x5_reduce inception_4e/5x5_inception_4e/relu_5x5 0=128 1=5 4=2 5=1 6=102400 8=2 9=1\nPooling inception_4e/pool 1 1 inception_4d/output_splitncnn_0 inception_4e/pool 1=3 3=1\nConvolution inception_4e/pool_proj 1 1 inception_4e/pool inception_4e/pool_proj_inception_4e/relu_pool_proj 0=128 1=1 5=1 6=67584 8=2 9=1\nConcat inception_4e/output 4 1 inception_4e/1x1_inception_4e/relu_1x1 inception_4e/3x3_inception_4e/relu_3x3 inception_4e/5x5_inception_4e/relu_5x5 inception_4e/pool_proj_inception_4e/relu_pool_proj inception_4e/output\nPooling pool4/3x3_s2 1 1 inception_4e/output pool4/3x3_s2 1=3 2=2\nSplit splitncnn_7 1 4 pool4/3x3_s2 pool4/3x3_s2_splitncnn_0 pool4/3x3_s2_splitncnn_1 pool4/3x3_s2_splitncnn_2 pool4/3x3_s2_splitncnn_3\nConvolution inception_5a/1x1 1 1 pool4/3x3_s2_splitncnn_3 inception_5a/1x1_inception_5a/relu_1x1 0=256 1=1 5=1 6=212992 8=2 9=1\nConvolution inception_5a/3x3_reduce 1 1 pool4/3x3_s2_splitncnn_2 inception_5a/3x3_reduce_inception_5a/relu_3x3_reduce 0=160 1=1 5=1 6=133120 8=102 9=1\nConvolution inception_5a/3x3 1 1 inception_5a/3x3_reduce_inception_5a/relu_3x3_reduce inception_5a/3x3_inception_5a/relu_3x3 0=320 1=3 4=1 5=1 6=460800 8=2 9=1\nConvolution inception_5a/5x5_reduce 1 1 pool4/3x3_s2_splitncnn_1 inception_5a/5x5_reduce_inception_5a/relu_5x5_reduce 0=32 1=1 5=1 6=26624 8=102 9=1\nConvolution inception_5a/5x5 1 1 inception_5a/5x5_reduce_inception_5a/relu_5x5_reduce inception_5a/5x5_inception_5a/relu_5x5 0=128 1=5 4=2 5=1 6=102400 8=2 9=1\nPooling inception_5a/pool 1 1 pool4/3x3_s2_splitncnn_0 inception_5a/pool 1=3 3=1\nConvolution inception_5a/pool_proj 1 1 inception_5a/pool inception_5a/pool_proj_inception_5a/relu_pool_proj 0=128 1=1 5=1 6=106496 8=2 9=1\nConcat inception_5a/output 4 1 inception_5a/1x1_inception_5a/relu_1x1 inception_5a/3x3_inception_5a/relu_3x3 inception_5a/5x5_inception_5a/relu_5x5 inception_5a/pool_proj_inception_5a/relu_pool_proj inception_5a/output\nSplit splitncnn_8 1 4 inception_5a/output inception_5a/output_splitncnn_0 inception_5a/output_splitncnn_1 inception_5a/output_splitncnn_2 inception_5a/output_splitncnn_3\nConvolution inception_5b/1x1 1 1 inception_5a/output_splitncnn_3 inception_5b/1x1_inception_5b/relu_1x1 0=384 1=1 5=1 6=319488 8=2 9=1\nConvolution inception_5b/3x3_reduce 1 1 inception_5a/output_splitncnn_2 inception_5b/3x3_reduce_inception_5b/relu_3x3_reduce 0=192 1=1 5=1 6=159744 8=102 9=1\nConvolution inception_5b/3x3 1 1 inception_5b/3x3_reduce_inception_5b/relu_3x3_reduce inception_5b/3x3_inception_5b/relu_3x3 0=384 1=3 4=1 5=1 6=663552 8=2 9=1\nConvolution inception_5b/5x5_reduce 1 1 inception_5a/output_splitncnn_1 inception_5b/5x5_reduce_inception_5b/relu_5x5_reduce 0=48 1=1 5=1 6=39936 8=102 9=1\nConvolution inception_5b/5x5 1 1 inception_5b/5x5_reduce_inception_5b/relu_5x5_reduce inception_5b/5x5_inception_5b/relu_5x5 0=128 1=5 4=2 5=1 6=153600 8=2 9=1\nPooling inception_5b/pool 1 1 inception_5a/output_splitncnn_0 inception_5b/pool 1=3 3=1\nConvolution inception_5b/pool_proj 1 1 inception_5b/pool inception_5b/pool_proj_inception_5b/relu_pool_proj 0=128 1=1 5=1 6=106496 8=2 9=1\nConcat inception_5b/output 4 1 inception_5b/1x1_inception_5b/relu_1x1 inception_5b/3x3_inception_5b/relu_3x3 inception_5b/5x5_inception_5b/relu_5x5 inception_5b/pool_proj_inception_5b/relu_pool_proj inception_5b/output\nPooling pool5/7x7_s1 1 1 inception_5b/output pool5/7x7_s1_pool5/drop_7x7_s1 0=1 1=7\nInnerProduct loss3/classifier 1 1 pool5/7x7_s1_pool5/drop_7x7_s1 loss3/classifier 0=1000 1=1 2=1024000\nSoftmax prob 1 1 loss3/classifier output\n" }, { "path": "benchmark/mnasnet.param", "content": "7767517\n76 86\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution first-3x3-conv 1 1 data first-3x3-conv_relu -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise A0_dw 1 1 first-3x3-conv_relu A0_dw_relu -23330=4,3,112,112,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution A0_linear 1 1 A0_dw_relu A0_linear_bn -23330=4,3,112,112,16 0=16 1=1 5=1 6=512\nConvolution B0_expand 1 1 A0_linear_bn B0_expand_relu -23330=4,3,112,112,48 0=48 1=1 5=1 6=768 9=1\nConvolutionDepthWise B0_dw 1 1 B0_expand_relu B0_dw_relu -23330=4,3,56,56,48 0=48 1=3 3=2 4=1 5=1 6=432 7=48 9=1\nConvolution B0_linear 1 1 B0_dw_relu B0_linear_bn -23330=4,3,56,56,24 0=24 1=1 5=1 6=1152\nSplit splitncnn_0 1 2 B0_linear_bn B0_linear_bn_splitncnn_0 B0_linear_bn_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolution B1_expand 1 1 B0_linear_bn_splitncnn_1 B1_expand_relu -23330=4,3,56,56,72 0=72 1=1 5=1 6=1728 9=1\nConvolutionDepthWise B1_dw 1 1 B1_expand_relu B1_dw_relu -23330=4,3,56,56,72 0=72 1=3 4=1 5=1 6=648 7=72 9=1\nConvolution B1_linear 1 1 B1_dw_relu B1_linear_bn -23330=4,3,56,56,24 0=24 1=1 5=1 6=1728\nBinaryOp unknownncnn_0 2 1 B0_linear_bn_splitncnn_0 B1_linear_bn unknownncnn_0 -23330=4,3,56,56,24\nSplit splitncnn_1 1 2 unknownncnn_0 unknownncnn_0_splitncnn_0 unknownncnn_0_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolution B2_expand 1 1 unknownncnn_0_splitncnn_1 B2_expand_relu -23330=4,3,56,56,72 0=72 1=1 5=1 6=1728 9=1\nConvolutionDepthWise B2_dw 1 1 B2_expand_relu B2_dw_relu -23330=4,3,56,56,72 0=72 1=3 4=1 5=1 6=648 7=72 9=1\nConvolution B2_linear 1 1 B2_dw_relu B2_linear_bn -23330=4,3,56,56,24 0=24 1=1 5=1 6=1728\nBinaryOp unknownncnn_1 2 1 unknownncnn_0_splitncnn_0 B2_linear_bn unknownncnn_1 -23330=4,3,56,56,24\nConvolution C0_expand 1 1 unknownncnn_1 C0_expand_relu -23330=4,3,56,56,72 0=72 1=1 5=1 6=1728 9=1\nConvolutionDepthWise C0_dw 1 1 C0_expand_relu C0_dw_relu -23330=4,3,28,28,72 0=72 1=5 3=2 4=2 5=1 6=1800 7=72 9=1\nConvolution C0_linear 1 1 C0_dw_relu C0_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=2880\nSplit splitncnn_2 1 2 C0_linear_bn C0_linear_bn_splitncnn_0 C0_linear_bn_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution C1_expand 1 1 C0_linear_bn_splitncnn_1 C1_expand_relu -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise C1_dw 1 1 C1_expand_relu C1_dw_relu -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=1\nConvolution C1_linear 1 1 C1_dw_relu C1_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp unknownncnn_2 2 1 C0_linear_bn_splitncnn_0 C1_linear_bn unknownncnn_2 -23330=4,3,28,28,40\nSplit splitncnn_3 1 2 unknownncnn_2 unknownncnn_2_splitncnn_0 unknownncnn_2_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution C2_expand 1 1 unknownncnn_2_splitncnn_1 C2_expand_relu -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise C2_dw 1 1 C2_expand_relu C2_dw_relu -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=1\nConvolution C2_linear 1 1 C2_dw_relu C2_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp unknownncnn_3 2 1 unknownncnn_2_splitncnn_0 C2_linear_bn unknownncnn_3 -23330=4,3,28,28,40\nConvolution D0_expand 1 1 unknownncnn_3 D0_expand_relu -23330=4,3,28,28,240 0=240 1=1 5=1 6=9600 9=1\nConvolutionDepthWise D0_dw 1 1 D0_expand_relu D0_dw_relu -23330=4,3,14,14,240 0=240 1=5 3=2 4=2 5=1 6=6000 7=240 9=1\nConvolution D0_linear 1 1 D0_dw_relu D0_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nSplit splitncnn_4 1 2 D0_linear_bn D0_linear_bn_splitncnn_0 D0_linear_bn_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution D1_expand 1 1 D0_linear_bn_splitncnn_1 D1_expand_relu -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400 9=1\nConvolutionDepthWise D1_dw 1 1 D1_expand_relu D1_dw_relu -23330=4,3,14,14,480 0=480 1=5 4=2 5=1 6=12000 7=480 9=1\nConvolution D1_linear 1 1 D1_dw_relu D1_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=38400\nBinaryOp unknownncnn_4 2 1 D0_linear_bn_splitncnn_0 D1_linear_bn unknownncnn_4 -23330=4,3,14,14,80\nSplit splitncnn_5 1 2 unknownncnn_4 unknownncnn_4_splitncnn_0 unknownncnn_4_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution D2_expand 1 1 unknownncnn_4_splitncnn_1 D2_expand_relu -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400 9=1\nConvolutionDepthWise D2_dw 1 1 D2_expand_relu D2_dw_relu -23330=4,3,14,14,480 0=480 1=5 4=2 5=1 6=12000 7=480 9=1\nConvolution D2_linear 1 1 D2_dw_relu D2_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=38400\nBinaryOp unknownncnn_5 2 1 unknownncnn_4_splitncnn_0 D2_linear_bn unknownncnn_5 -23330=4,3,14,14,80\nConvolution E0_expand 1 1 unknownncnn_5 E0_expand_relu -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400 9=1\nConvolutionDepthWise E0_dw 1 1 E0_expand_relu E0_dw_relu -23330=4,3,14,14,480 0=480 1=3 4=1 5=1 6=4320 7=480 9=1\nConvolution E0_linear 1 1 E0_dw_relu E0_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=46080\nSplit splitncnn_6 1 2 E0_linear_bn E0_linear_bn_splitncnn_0 E0_linear_bn_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution E1_expand 1 1 E0_linear_bn_splitncnn_1 E1_expand_relu -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise E1_dw 1 1 E1_expand_relu E1_dw_relu -23330=4,3,14,14,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution E1_linear 1 1 E1_dw_relu E1_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=55296\nBinaryOp unknownncnn_6 2 1 E0_linear_bn_splitncnn_0 E1_linear_bn unknownncnn_6 -23330=4,3,14,14,96\nConvolution F0_expand 1 1 unknownncnn_6 F0_expand_relu -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise F0_dw 1 1 F0_expand_relu F0_dw_relu -23330=4,3,7,7,576 0=576 1=5 3=2 4=2 5=1 6=14400 7=576 9=1\nConvolution F0_linear 1 1 F0_dw_relu F0_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=110592\nSplit splitncnn_7 1 2 F0_linear_bn F0_linear_bn_splitncnn_0 F0_linear_bn_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F1_expand 1 1 F0_linear_bn_splitncnn_1 F1_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise F1_dw 1 1 F1_expand_relu F1_dw_relu -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152 9=1\nConvolution F1_linear 1 1 F1_dw_relu F1_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp unknownncnn_7 2 1 F0_linear_bn_splitncnn_0 F1_linear_bn unknownncnn_7 -23330=4,3,7,7,192\nSplit splitncnn_8 1 2 unknownncnn_7 unknownncnn_7_splitncnn_0 unknownncnn_7_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F2_expand 1 1 unknownncnn_7_splitncnn_1 F2_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise F2_dw 1 1 F2_expand_relu F2_dw_relu -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152 9=1\nConvolution F2_linear 1 1 F2_dw_relu F2_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp unknownncnn_8 2 1 unknownncnn_7_splitncnn_0 F2_linear_bn unknownncnn_8 -23330=4,3,7,7,192\nSplit splitncnn_9 1 2 unknownncnn_8 unknownncnn_8_splitncnn_0 unknownncnn_8_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F3_expand 1 1 unknownncnn_8_splitncnn_1 F3_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise F3_dw 1 1 F3_expand_relu F3_dw_relu -23330=4,3,7,7,1152 0=1152 1=5 4=2 5=1 6=28800 7=1152 9=1\nConvolution F3_linear 1 1 F3_dw_relu F3_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp unknownncnn_9 2 1 unknownncnn_8_splitncnn_0 F3_linear_bn unknownncnn_9 -23330=4,3,7,7,192\nConvolution G0_expand 1 1 unknownncnn_9 G0_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise G0_dw 1 1 G0_expand_relu G0_dw_relu -23330=4,3,7,7,1152 0=1152 1=3 4=1 5=1 6=10368 7=1152 9=1\nConvolution G0_linear 1 1 G0_dw_relu G0_linear_bn -23330=4,3,7,7,320 0=320 1=1 5=1 6=368640\nConvolution last-1x1-conv 1 1 G0_linear_bn last-1x1-conv_relu -23330=4,3,7,7,1280 0=1280 1=1 5=1 6=409600 9=1\nPooling avgpool 1 1 last-1x1-conv_relu flatten -23330=4,1,1280,1,1 0=1 1=7 4=1 5=1\nInnerProduct fc 1 1 flatten fc -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\nSoftmax prob 1 1 fc output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/mobilenet.param", "content": "7767517\n31 31\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_relu1 -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise conv2_1/dw 1 1 conv1_relu1 conv2_1/dw_relu2_1/dw -23330=4,3,112,112,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution conv2_1/sep 1 1 conv2_1/dw_relu2_1/dw conv2_1/sep_relu2_1/sep -23330=4,3,112,112,64 0=64 1=1 5=1 6=2048 9=1\nConvolutionDepthWise conv2_2/dw 1 1 conv2_1/sep_relu2_1/sep conv2_2/dw_relu2_2/dw -23330=4,3,56,56,64 0=64 1=3 3=2 4=1 5=1 6=576 7=64 9=1\nConvolution conv2_2/sep 1 1 conv2_2/dw_relu2_2/dw conv2_2/sep_relu2_2/sep -23330=4,3,56,56,128 0=128 1=1 5=1 6=8192 9=1\nConvolutionDepthWise conv3_1/dw 1 1 conv2_2/sep_relu2_2/sep conv3_1/dw_relu3_1/dw -23330=4,3,56,56,128 0=128 1=3 4=1 5=1 6=1152 7=128 9=1\nConvolution conv3_1/sep 1 1 conv3_1/dw_relu3_1/dw conv3_1/sep_relu3_1/sep -23330=4,3,56,56,128 0=128 1=1 5=1 6=16384 9=1\nConvolutionDepthWise conv3_2/dw 1 1 conv3_1/sep_relu3_1/sep conv3_2/dw_relu3_2/dw -23330=4,3,28,28,128 0=128 1=3 3=2 4=1 5=1 6=1152 7=128 9=1\nConvolution conv3_2/sep 1 1 conv3_2/dw_relu3_2/dw conv3_2/sep_relu3_2/sep -23330=4,3,28,28,256 0=256 1=1 5=1 6=32768 9=1\nConvolutionDepthWise conv4_1/dw 1 1 conv3_2/sep_relu3_2/sep conv4_1/dw_relu4_1/dw -23330=4,3,28,28,256 0=256 1=3 4=1 5=1 6=2304 7=256 9=1\nConvolution conv4_1/sep 1 1 conv4_1/dw_relu4_1/dw conv4_1/sep_relu4_1/sep -23330=4,3,28,28,256 0=256 1=1 5=1 6=65536 9=1\nConvolutionDepthWise conv4_2/dw 1 1 conv4_1/sep_relu4_1/sep conv4_2/dw_relu4_2/dw -23330=4,3,14,14,256 0=256 1=3 3=2 4=1 5=1 6=2304 7=256 9=1\nConvolution conv4_2/sep 1 1 conv4_2/dw_relu4_2/dw conv4_2/sep_relu4_2/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=131072 9=1\nConvolutionDepthWise conv5_1/dw 1 1 conv4_2/sep_relu4_2/sep conv5_1/dw_relu5_1/dw -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_1/sep 1 1 conv5_1/dw_relu5_1/dw conv5_1/sep_relu5_1/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv5_2/dw 1 1 conv5_1/sep_relu5_1/sep conv5_2/dw_relu5_2/dw -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_2/sep 1 1 conv5_2/dw_relu5_2/dw conv5_2/sep_relu5_2/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv5_3/dw 1 1 conv5_2/sep_relu5_2/sep conv5_3/dw_relu5_3/dw -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_3/sep 1 1 conv5_3/dw_relu5_3/dw conv5_3/sep_relu5_3/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv5_4/dw 1 1 conv5_3/sep_relu5_3/sep conv5_4/dw_relu5_4/dw -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_4/sep 1 1 conv5_4/dw_relu5_4/dw conv5_4/sep_relu5_4/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv5_5/dw 1 1 conv5_4/sep_relu5_4/sep conv5_5/dw_relu5_5/dw -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_5/sep 1 1 conv5_5/dw_relu5_5/dw conv5_5/sep_relu5_5/sep -23330=4,3,14,14,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv5_6/dw 1 1 conv5_5/sep_relu5_5/sep conv5_6/dw_relu5_6/dw -23330=4,3,7,7,512 0=512 1=3 3=2 4=1 5=1 6=4608 7=512 9=1\nConvolution conv5_6/sep 1 1 conv5_6/dw_relu5_6/dw conv5_6/sep_relu5_6/sep -23330=4,3,7,7,1024 0=1024 1=1 5=1 6=524288 9=1\nConvolutionDepthWise conv6/dw 1 1 conv5_6/sep_relu5_6/sep conv6/dw_relu6/dw -23330=4,3,7,7,1024 0=1024 1=3 4=1 5=1 6=9216 7=1024 9=1\nConvolution conv6/sep 1 1 conv6/dw_relu6/dw conv6/sep_relu6/sep -23330=4,3,7,7,1024 0=1024 1=1 5=1 6=1048576 9=1\nPooling pool6 1 1 conv6/sep_relu6/sep pool6 -23330=4,1,1024,1,1 0=1 4=1\nInnerProduct fc7 1 1 pool6 fc7 -23330=4,1,1000,1,1 0=1000 1=1 2=1024000\nSoftmax prob 1 1 fc7 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/mobilenet_int8.param", "content": "7767517\n31 31\nInput data 0 1 data 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_relu1 0=32 1=3 3=2 4=1 5=1 6=864 8=102 9=1\nConvolutionDepthWise conv2_1/dw 1 1 conv1_relu1 conv2_1/dw_relu2_1/dw 0=32 1=3 4=1 5=1 6=288 7=32 8=101 9=1\nConvolution conv2_1/sep 1 1 conv2_1/dw_relu2_1/dw conv2_1/sep_relu2_1/sep 0=64 1=1 5=1 6=2048 8=102 9=1\nConvolutionDepthWise conv2_2/dw 1 1 conv2_1/sep_relu2_1/sep conv2_2/dw_relu2_2/dw 0=64 1=3 3=2 4=1 5=1 6=576 7=64 8=101 9=1\nConvolution conv2_2/sep 1 1 conv2_2/dw_relu2_2/dw conv2_2/sep_relu2_2/sep 0=128 1=1 5=1 6=8192 8=102 9=1\nConvolutionDepthWise conv3_1/dw 1 1 conv2_2/sep_relu2_2/sep conv3_1/dw_relu3_1/dw 0=128 1=3 4=1 5=1 6=1152 7=128 8=101 9=1\nConvolution conv3_1/sep 1 1 conv3_1/dw_relu3_1/dw conv3_1/sep_relu3_1/sep 0=128 1=1 5=1 6=16384 8=102 9=1\nConvolutionDepthWise conv3_2/dw 1 1 conv3_1/sep_relu3_1/sep conv3_2/dw_relu3_2/dw 0=128 1=3 3=2 4=1 5=1 6=1152 7=128 8=101 9=1\nConvolution conv3_2/sep 1 1 conv3_2/dw_relu3_2/dw conv3_2/sep_relu3_2/sep 0=256 1=1 5=1 6=32768 8=102 9=1\nConvolutionDepthWise conv4_1/dw 1 1 conv3_2/sep_relu3_2/sep conv4_1/dw_relu4_1/dw 0=256 1=3 4=1 5=1 6=2304 7=256 8=101 9=1\nConvolution conv4_1/sep 1 1 conv4_1/dw_relu4_1/dw conv4_1/sep_relu4_1/sep 0=256 1=1 5=1 6=65536 8=102 9=1\nConvolutionDepthWise conv4_2/dw 1 1 conv4_1/sep_relu4_1/sep conv4_2/dw_relu4_2/dw 0=256 1=3 3=2 4=1 5=1 6=2304 7=256 8=101 9=1\nConvolution conv4_2/sep 1 1 conv4_2/dw_relu4_2/dw conv4_2/sep_relu4_2/sep 0=512 1=1 5=1 6=131072 8=102 9=1\nConvolutionDepthWise conv5_1/dw 1 1 conv4_2/sep_relu4_2/sep conv5_1/dw_relu5_1/dw 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_1/sep 1 1 conv5_1/dw_relu5_1/dw conv5_1/sep_relu5_1/sep 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv5_2/dw 1 1 conv5_1/sep_relu5_1/sep conv5_2/dw_relu5_2/dw 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_2/sep 1 1 conv5_2/dw_relu5_2/dw conv5_2/sep_relu5_2/sep 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv5_3/dw 1 1 conv5_2/sep_relu5_2/sep conv5_3/dw_relu5_3/dw 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_3/sep 1 1 conv5_3/dw_relu5_3/dw conv5_3/sep_relu5_3/sep 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv5_4/dw 1 1 conv5_3/sep_relu5_3/sep conv5_4/dw_relu5_4/dw 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_4/sep 1 1 conv5_4/dw_relu5_4/dw conv5_4/sep_relu5_4/sep 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv5_5/dw 1 1 conv5_4/sep_relu5_4/sep conv5_5/dw_relu5_5/dw 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_5/sep 1 1 conv5_5/dw_relu5_5/dw conv5_5/sep_relu5_5/sep 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv5_6/dw 1 1 conv5_5/sep_relu5_5/sep conv5_6/dw_relu5_6/dw 0=512 1=3 3=2 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv5_6/sep 1 1 conv5_6/dw_relu5_6/dw conv5_6/sep_relu5_6/sep 0=1024 1=1 5=1 6=524288 8=102 9=1\nConvolutionDepthWise conv6/dw 1 1 conv5_6/sep_relu5_6/sep conv6/dw_relu6/dw 0=1024 1=3 4=1 5=1 6=9216 7=1024 8=101 9=1\nConvolution conv6/sep 1 1 conv6/dw_relu6/dw conv6/sep_relu6/sep 0=1024 1=1 5=1 6=1048576 8=2 9=1\nPooling pool6 1 1 conv6/sep_relu6/sep pool6 0=1 4=1\nInnerProduct fc7 1 1 pool6 fc7 0=1000 1=1 2=1024000 8=2\nSoftmax prob 1 1 fc7 output\n" }, { "path": "benchmark/mobilenet_ssd.param", "content": "7767517\n92 115\nInput input 0 1 data -23330=4,3,300,300,3 0=300 1=300 2=3\nSplit splitncnn_0 1 7 data data_splitncnn_0 data_splitncnn_1 data_splitncnn_2 data_splitncnn_3 data_splitncnn_4 data_splitncnn_5 data_splitncnn_6 -23330=28,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3\nConvolution conv0 1 1 data_splitncnn_6 conv0_conv0/relu -23330=4,3,150,150,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise conv1/dw 1 1 conv0_conv0/relu conv1/dw_conv1/dw/relu -23330=4,3,150,150,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution conv1 1 1 conv1/dw_conv1/dw/relu conv1_conv1/relu -23330=4,3,150,150,64 0=64 1=1 5=1 6=2048 9=1\nConvolutionDepthWise conv2/dw 1 1 conv1_conv1/relu conv2/dw_conv2/dw/relu -23330=4,3,75,75,64 0=64 1=3 3=2 4=1 5=1 6=576 7=64 9=1\nConvolution conv2 1 1 conv2/dw_conv2/dw/relu conv2_conv2/relu -23330=4,3,75,75,128 0=128 1=1 5=1 6=8192 9=1\nConvolutionDepthWise conv3/dw 1 1 conv2_conv2/relu conv3/dw_conv3/dw/relu -23330=4,3,75,75,128 0=128 1=3 4=1 5=1 6=1152 7=128 9=1\nConvolution conv3 1 1 conv3/dw_conv3/dw/relu conv3_conv3/relu -23330=4,3,75,75,128 0=128 1=1 5=1 6=16384 9=1\nConvolutionDepthWise conv4/dw 1 1 conv3_conv3/relu conv4/dw_conv4/dw/relu -23330=4,3,38,38,128 0=128 1=3 3=2 4=1 5=1 6=1152 7=128 9=1\nConvolution conv4 1 1 conv4/dw_conv4/dw/relu conv4_conv4/relu -23330=4,3,38,38,256 0=256 1=1 5=1 6=32768 9=1\nConvolutionDepthWise conv5/dw 1 1 conv4_conv4/relu conv5/dw_conv5/dw/relu -23330=4,3,38,38,256 0=256 1=3 4=1 5=1 6=2304 7=256 9=1\nConvolution conv5 1 1 conv5/dw_conv5/dw/relu conv5_conv5/relu -23330=4,3,38,38,256 0=256 1=1 5=1 6=65536 9=1\nConvolutionDepthWise conv6/dw 1 1 conv5_conv5/relu conv6/dw_conv6/dw/relu -23330=4,3,19,19,256 0=256 1=3 3=2 4=1 5=1 6=2304 7=256 9=1\nConvolution conv6 1 1 conv6/dw_conv6/dw/relu conv6_conv6/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=131072 9=1\nConvolutionDepthWise conv7/dw 1 1 conv6_conv6/relu conv7/dw_conv7/dw/relu -23330=4,3,19,19,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv7 1 1 conv7/dw_conv7/dw/relu conv7_conv7/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv8/dw 1 1 conv7_conv7/relu conv8/dw_conv8/dw/relu -23330=4,3,19,19,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv8 1 1 conv8/dw_conv8/dw/relu conv8_conv8/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv9/dw 1 1 conv8_conv8/relu conv9/dw_conv9/dw/relu -23330=4,3,19,19,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv9 1 1 conv9/dw_conv9/dw/relu conv9_conv9/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv10/dw 1 1 conv9_conv9/relu conv10/dw_conv10/dw/relu -23330=4,3,19,19,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv10 1 1 conv10/dw_conv10/dw/relu conv10_conv10/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv11/dw 1 1 conv10_conv10/relu conv11/dw_conv11/dw/relu -23330=4,3,19,19,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv11 1 1 conv11/dw_conv11/dw/relu conv11_conv11/relu -23330=4,3,19,19,512 0=512 1=1 5=1 6=262144 9=1\nSplit splitncnn_1 1 4 conv11_conv11/relu conv11_conv11/relu_splitncnn_0 conv11_conv11/relu_splitncnn_1 conv11_conv11/relu_splitncnn_2 conv11_conv11/relu_splitncnn_3 -23330=16,3,19,19,512,3,19,19,512,3,19,19,512,3,19,19,512\nConvolutionDepthWise conv12/dw 1 1 conv11_conv11/relu_splitncnn_3 conv12/dw_conv12/dw/relu -23330=4,3,10,10,512 0=512 1=3 3=2 4=1 5=1 6=4608 7=512 9=1\nConvolution conv12 1 1 conv12/dw_conv12/dw/relu conv12_conv12/relu -23330=4,3,10,10,1024 0=1024 1=1 5=1 6=524288 9=1\nConvolutionDepthWise conv13/dw 1 1 conv12_conv12/relu conv13/dw_conv13/dw/relu -23330=4,3,10,10,1024 0=1024 1=3 4=1 5=1 6=9216 7=1024 9=1\nConvolution conv13 1 1 conv13/dw_conv13/dw/relu conv13_conv13/relu -23330=4,3,10,10,1024 0=1024 1=1 5=1 6=1048576 9=1\nSplit splitncnn_2 1 4 conv13_conv13/relu conv13_conv13/relu_splitncnn_0 conv13_conv13/relu_splitncnn_1 conv13_conv13/relu_splitncnn_2 conv13_conv13/relu_splitncnn_3 -23330=16,3,10,10,1024,3,10,10,1024,3,10,10,1024,3,10,10,1024\nConvolution conv14_1 1 1 conv13_conv13/relu_splitncnn_3 conv14_1_conv14_1/relu -23330=4,3,10,10,256 0=256 1=1 5=1 6=262144 9=1\nConvolution conv14_2 1 1 conv14_1_conv14_1/relu conv14_2_conv14_2/relu -23330=4,3,5,5,512 0=512 1=3 3=2 4=1 5=1 6=1179648 9=1\nSplit splitncnn_3 1 4 conv14_2_conv14_2/relu conv14_2_conv14_2/relu_splitncnn_0 conv14_2_conv14_2/relu_splitncnn_1 conv14_2_conv14_2/relu_splitncnn_2 conv14_2_conv14_2/relu_splitncnn_3 -23330=16,3,5,5,512,3,5,5,512,3,5,5,512,3,5,5,512\nConvolution conv15_1 1 1 conv14_2_conv14_2/relu_splitncnn_3 conv15_1_conv15_1/relu -23330=4,3,5,5,128 0=128 1=1 5=1 6=65536 9=1\nConvolution conv15_2 1 1 conv15_1_conv15_1/relu conv15_2_conv15_2/relu -23330=4,3,3,3,256 0=256 1=3 3=2 4=1 5=1 6=294912 9=1\nSplit splitncnn_4 1 4 conv15_2_conv15_2/relu conv15_2_conv15_2/relu_splitncnn_0 conv15_2_conv15_2/relu_splitncnn_1 conv15_2_conv15_2/relu_splitncnn_2 conv15_2_conv15_2/relu_splitncnn_3 -23330=16,3,3,3,256,3,3,3,256,3,3,3,256,3,3,3,256\nConvolution conv16_1 1 1 conv15_2_conv15_2/relu_splitncnn_3 conv16_1_conv16_1/relu -23330=4,3,3,3,128 0=128 1=1 5=1 6=32768 9=1\nConvolution conv16_2 1 1 conv16_1_conv16_1/relu conv16_2_conv16_2/relu -23330=4,3,2,2,256 0=256 1=3 3=2 4=1 5=1 6=294912 9=1\nSplit splitncnn_5 1 4 conv16_2_conv16_2/relu conv16_2_conv16_2/relu_splitncnn_0 conv16_2_conv16_2/relu_splitncnn_1 conv16_2_conv16_2/relu_splitncnn_2 conv16_2_conv16_2/relu_splitncnn_3 -23330=16,3,2,2,256,3,2,2,256,3,2,2,256,3,2,2,256\nConvolution conv17_1 1 1 conv16_2_conv16_2/relu_splitncnn_3 conv17_1_conv17_1/relu -23330=4,3,2,2,64 0=64 1=1 5=1 6=16384 9=1\nConvolution conv17_2 1 1 conv17_1_conv17_1/relu conv17_2_conv17_2/relu -23330=4,3,1,1,128 0=128 1=3 3=2 4=1 5=1 6=73728 9=1\nSplit splitncnn_6 1 3 conv17_2_conv17_2/relu conv17_2_conv17_2/relu_splitncnn_0 conv17_2_conv17_2/relu_splitncnn_1 conv17_2_conv17_2/relu_splitncnn_2 -23330=12,3,1,1,128,3,1,1,128,3,1,1,128\nConvolution conv11_mbox_loc 1 1 conv11_conv11/relu_splitncnn_2 conv11_mbox_loc -23330=4,3,19,19,12 0=12 1=1 5=1 6=6144\nPermute conv11_mbox_loc_perm 1 1 conv11_mbox_loc conv11_mbox_loc_perm -23330=4,3,12,19,19 0=3\nFlatten conv11_mbox_loc_flat 1 1 conv11_mbox_loc_perm conv11_mbox_loc_flat -23330=4,1,4332,1,1\nConvolution conv11_mbox_conf 1 1 conv11_conv11/relu_splitncnn_1 conv11_mbox_conf -23330=4,3,19,19,63 0=63 1=1 5=1 6=32256\nPermute conv11_mbox_conf_perm 1 1 conv11_mbox_conf conv11_mbox_conf_perm -23330=4,3,63,19,19 0=3\nFlatten conv11_mbox_conf_flat 1 1 conv11_mbox_conf_perm conv11_mbox_conf_flat -23330=4,1,22743,1,1\nPriorBox conv11_mbox_priorbox 2 1 conv11_conv11/relu_splitncnn_0 data_splitncnn_5 conv11_mbox_priorbox -23330=4,2,4332,2,1 -23300=1,6.000000e+01 -23302=1,2.000000e+00 9=-233 10=-233 13=5.000000e-01\nConvolution conv13_mbox_loc 1 1 conv13_conv13/relu_splitncnn_2 conv13_mbox_loc -23330=4,3,10,10,24 0=24 1=1 5=1 6=24576\nPermute conv13_mbox_loc_perm 1 1 conv13_mbox_loc conv13_mbox_loc_perm -23330=4,3,24,10,10 0=3\nFlatten conv13_mbox_loc_flat 1 1 conv13_mbox_loc_perm conv13_mbox_loc_flat -23330=4,1,2400,1,1\nConvolution conv13_mbox_conf 1 1 conv13_conv13/relu_splitncnn_1 conv13_mbox_conf -23330=4,3,10,10,126 0=126 1=1 5=1 6=129024\nPermute conv13_mbox_conf_perm 1 1 conv13_mbox_conf conv13_mbox_conf_perm -23330=4,3,126,10,10 0=3\nFlatten conv13_mbox_conf_flat 1 1 conv13_mbox_conf_perm conv13_mbox_conf_flat -23330=4,1,12600,1,1\nPriorBox conv13_mbox_priorbox 2 1 conv13_conv13/relu_splitncnn_0 data_splitncnn_4 conv13_mbox_priorbox -23330=4,2,2400,2,1 -23300=1,1.050000e+02 -23301=1,1.500000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 13=5.000000e-01\nConvolution conv14_2_mbox_loc 1 1 conv14_2_conv14_2/relu_splitncnn_2 conv14_2_mbox_loc -23330=4,3,5,5,24 0=24 1=1 5=1 6=12288\nPermute conv14_2_mbox_loc_perm 1 1 conv14_2_mbox_loc conv14_2_mbox_loc_perm -23330=4,3,24,5,5 0=3\nFlatten conv14_2_mbox_loc_flat 1 1 conv14_2_mbox_loc_perm conv14_2_mbox_loc_flat -23330=4,1,600,1,1\nConvolution conv14_2_mbox_conf 1 1 conv14_2_conv14_2/relu_splitncnn_1 conv14_2_mbox_conf -23330=4,3,5,5,126 0=126 1=1 5=1 6=64512\nPermute conv14_2_mbox_conf_perm 1 1 conv14_2_mbox_conf conv14_2_mbox_conf_perm -23330=4,3,126,5,5 0=3\nFlatten conv14_2_mbox_conf_flat 1 1 conv14_2_mbox_conf_perm conv14_2_mbox_conf_flat -23330=4,1,3150,1,1\nPriorBox conv14_2_mbox_priorbox 2 1 conv14_2_conv14_2/relu_splitncnn_0 data_splitncnn_3 conv14_2_mbox_priorbox -23330=4,2,600,2,1 -23300=1,1.500000e+02 -23301=1,1.950000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 13=5.000000e-01\nConvolution conv15_2_mbox_loc 1 1 conv15_2_conv15_2/relu_splitncnn_2 conv15_2_mbox_loc -23330=4,3,3,3,24 0=24 1=1 5=1 6=6144\nPermute conv15_2_mbox_loc_perm 1 1 conv15_2_mbox_loc conv15_2_mbox_loc_perm -23330=4,3,24,3,3 0=3\nFlatten conv15_2_mbox_loc_flat 1 1 conv15_2_mbox_loc_perm conv15_2_mbox_loc_flat -23330=4,1,216,1,1\nConvolution conv15_2_mbox_conf 1 1 conv15_2_conv15_2/relu_splitncnn_1 conv15_2_mbox_conf -23330=4,3,3,3,126 0=126 1=1 5=1 6=32256\nPermute conv15_2_mbox_conf_perm 1 1 conv15_2_mbox_conf conv15_2_mbox_conf_perm -23330=4,3,126,3,3 0=3\nFlatten conv15_2_mbox_conf_flat 1 1 conv15_2_mbox_conf_perm conv15_2_mbox_conf_flat -23330=4,1,1134,1,1\nPriorBox conv15_2_mbox_priorbox 2 1 conv15_2_conv15_2/relu_splitncnn_0 data_splitncnn_2 conv15_2_mbox_priorbox -23330=4,2,216,2,1 -23300=1,1.950000e+02 -23301=1,2.400000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 13=5.000000e-01\nConvolution conv16_2_mbox_loc 1 1 conv16_2_conv16_2/relu_splitncnn_2 conv16_2_mbox_loc -23330=4,3,2,2,24 0=24 1=1 5=1 6=6144\nPermute conv16_2_mbox_loc_perm 1 1 conv16_2_mbox_loc conv16_2_mbox_loc_perm -23330=4,3,24,2,2 0=3\nFlatten conv16_2_mbox_loc_flat 1 1 conv16_2_mbox_loc_perm conv16_2_mbox_loc_flat -23330=4,1,96,1,1\nConvolution conv16_2_mbox_conf 1 1 conv16_2_conv16_2/relu_splitncnn_1 conv16_2_mbox_conf -23330=4,3,2,2,126 0=126 1=1 5=1 6=32256\nPermute conv16_2_mbox_conf_perm 1 1 conv16_2_mbox_conf conv16_2_mbox_conf_perm -23330=4,3,126,2,2 0=3\nFlatten conv16_2_mbox_conf_flat 1 1 conv16_2_mbox_conf_perm conv16_2_mbox_conf_flat -23330=4,1,504,1,1\nPriorBox conv16_2_mbox_priorbox 2 1 conv16_2_conv16_2/relu_splitncnn_0 data_splitncnn_1 conv16_2_mbox_priorbox -23330=4,2,96,2,1 -23300=1,2.400000e+02 -23301=1,2.850000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 13=5.000000e-01\nConvolution conv17_2_mbox_loc 1 1 conv17_2_conv17_2/relu_splitncnn_2 conv17_2_mbox_loc -23330=4,3,1,1,24 0=24 1=1 5=1 6=3072\nPermute conv17_2_mbox_loc_perm 1 1 conv17_2_mbox_loc conv17_2_mbox_loc_perm -23330=4,3,24,1,1 0=3\nFlatten conv17_2_mbox_loc_flat 1 1 conv17_2_mbox_loc_perm conv17_2_mbox_loc_flat -23330=4,1,24,1,1\nConvolution conv17_2_mbox_conf 1 1 conv17_2_conv17_2/relu_splitncnn_1 conv17_2_mbox_conf -23330=4,3,1,1,126 0=126 1=1 5=1 6=16128\nPermute conv17_2_mbox_conf_perm 1 1 conv17_2_mbox_conf conv17_2_mbox_conf_perm -23330=4,3,126,1,1 0=3\nFlatten conv17_2_mbox_conf_flat 1 1 conv17_2_mbox_conf_perm conv17_2_mbox_conf_flat -23330=4,1,126,1,1\nPriorBox conv17_2_mbox_priorbox 2 1 conv17_2_conv17_2/relu_splitncnn_0 data_splitncnn_0 conv17_2_mbox_priorbox -23330=4,2,24,2,1 -23300=1,2.850000e+02 -23301=1,3.000000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 13=5.000000e-01\nConcat mbox_loc 6 1 conv11_mbox_loc_flat conv13_mbox_loc_flat conv14_2_mbox_loc_flat conv15_2_mbox_loc_flat conv16_2_mbox_loc_flat conv17_2_mbox_loc_flat mbox_loc -23330=4,1,7668,1,1\nConcat mbox_conf 6 1 conv11_mbox_conf_flat conv13_mbox_conf_flat conv14_2_mbox_conf_flat conv15_2_mbox_conf_flat conv16_2_mbox_conf_flat conv17_2_mbox_conf_flat mbox_conf -23330=4,1,40257,1,1\nConcat mbox_priorbox 6 1 conv11_mbox_priorbox conv13_mbox_priorbox conv14_2_mbox_priorbox conv15_2_mbox_priorbox conv16_2_mbox_priorbox conv17_2_mbox_priorbox mbox_priorbox -23330=4,2,7668,2,1 0=1\nReshape mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape -23330=4,2,21,1917,1 0=21 1=-1\nSoftmax mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax -23330=4,2,21,1917,1 0=1 1=1\nFlatten mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten -23330=4,1,40257,1,1\nDetectionOutput detection_out 3 1 mbox_loc mbox_conf_flatten mbox_priorbox output 0=21 1=4.500000e-01 2=100 4=2.500000e-01\n" }, { "path": "benchmark/mobilenet_ssd_int8.param", "content": "7767517\n92 115\nInput input 0 1 data 0=300 1=300 2=3\nSplit splitncnn_0 1 7 data data_splitncnn_0 data_splitncnn_1 data_splitncnn_2 data_splitncnn_3 data_splitncnn_4 data_splitncnn_5 data_splitncnn_6\nConvolution conv0 1 1 data_splitncnn_6 conv0_conv0/relu 0=32 1=3 3=2 4=1 5=1 6=864 8=102 9=1\nConvolutionDepthWise conv1/dw 1 1 conv0_conv0/relu conv1/dw_conv1/dw/relu 0=32 1=3 4=1 5=1 6=288 7=32 8=101 9=1\nConvolution conv1 1 1 conv1/dw_conv1/dw/relu conv1_conv1/relu 0=64 1=1 5=1 6=2048 8=102 9=1\nConvolutionDepthWise conv2/dw 1 1 conv1_conv1/relu conv2/dw_conv2/dw/relu 0=64 1=3 3=2 4=1 5=1 6=576 7=64 8=101 9=1\nConvolution conv2 1 1 conv2/dw_conv2/dw/relu conv2_conv2/relu 0=128 1=1 5=1 6=8192 8=102 9=1\nConvolutionDepthWise conv3/dw 1 1 conv2_conv2/relu conv3/dw_conv3/dw/relu 0=128 1=3 4=1 5=1 6=1152 7=128 8=101 9=1\nConvolution conv3 1 1 conv3/dw_conv3/dw/relu conv3_conv3/relu 0=128 1=1 5=1 6=16384 8=102 9=1\nConvolutionDepthWise conv4/dw 1 1 conv3_conv3/relu conv4/dw_conv4/dw/relu 0=128 1=3 3=2 4=1 5=1 6=1152 7=128 8=101 9=1\nConvolution conv4 1 1 conv4/dw_conv4/dw/relu conv4_conv4/relu 0=256 1=1 5=1 6=32768 8=102 9=1\nConvolutionDepthWise conv5/dw 1 1 conv4_conv4/relu conv5/dw_conv5/dw/relu 0=256 1=3 4=1 5=1 6=2304 7=256 8=101 9=1\nConvolution conv5 1 1 conv5/dw_conv5/dw/relu conv5_conv5/relu 0=256 1=1 5=1 6=65536 8=102 9=1\nConvolutionDepthWise conv6/dw 1 1 conv5_conv5/relu conv6/dw_conv6/dw/relu 0=256 1=3 3=2 4=1 5=1 6=2304 7=256 8=101 9=1\nConvolution conv6 1 1 conv6/dw_conv6/dw/relu conv6_conv6/relu 0=512 1=1 5=1 6=131072 8=102 9=1\nConvolutionDepthWise conv7/dw 1 1 conv6_conv6/relu conv7/dw_conv7/dw/relu 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv7 1 1 conv7/dw_conv7/dw/relu conv7_conv7/relu 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv8/dw 1 1 conv7_conv7/relu conv8/dw_conv8/dw/relu 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv8 1 1 conv8/dw_conv8/dw/relu conv8_conv8/relu 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv9/dw 1 1 conv8_conv8/relu conv9/dw_conv9/dw/relu 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv9 1 1 conv9/dw_conv9/dw/relu conv9_conv9/relu 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv10/dw 1 1 conv9_conv9/relu conv10/dw_conv10/dw/relu 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv10 1 1 conv10/dw_conv10/dw/relu conv10_conv10/relu 0=512 1=1 5=1 6=262144 8=102 9=1\nConvolutionDepthWise conv11/dw 1 1 conv10_conv10/relu conv11/dw_conv11/dw/relu 0=512 1=3 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv11 1 1 conv11/dw_conv11/dw/relu conv11_conv11/relu 0=512 1=1 5=1 6=262144 8=2 9=1\nSplit splitncnn_1 1 4 conv11_conv11/relu conv11_conv11/relu_splitncnn_0 conv11_conv11/relu_splitncnn_1 conv11_conv11/relu_splitncnn_2 conv11_conv11/relu_splitncnn_3\nConvolutionDepthWise conv12/dw 1 1 conv11_conv11/relu_splitncnn_3 conv12/dw_conv12/dw/relu 0=512 1=3 3=2 4=1 5=1 6=4608 7=512 8=101 9=1\nConvolution conv12 1 1 conv12/dw_conv12/dw/relu conv12_conv12/relu 0=1024 1=1 5=1 6=524288 8=102 9=1\nConvolutionDepthWise conv13/dw 1 1 conv12_conv12/relu conv13/dw_conv13/dw/relu 0=1024 1=3 4=1 5=1 6=9216 7=1024 8=101 9=1\nConvolution conv13 1 1 conv13/dw_conv13/dw/relu conv13_conv13/relu 0=1024 1=1 5=1 6=1048576 8=2 9=1\nSplit splitncnn_2 1 4 conv13_conv13/relu conv13_conv13/relu_splitncnn_0 conv13_conv13/relu_splitncnn_1 conv13_conv13/relu_splitncnn_2 conv13_conv13/relu_splitncnn_3\nConvolution conv14_1 1 1 conv13_conv13/relu_splitncnn_3 conv14_1_conv14_1/relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution conv14_2 1 1 conv14_1_conv14_1/relu conv14_2_conv14_2/relu 0=512 1=3 3=2 4=1 5=1 6=1179648 8=2 9=1\nSplit splitncnn_3 1 4 conv14_2_conv14_2/relu conv14_2_conv14_2/relu_splitncnn_0 conv14_2_conv14_2/relu_splitncnn_1 conv14_2_conv14_2/relu_splitncnn_2 conv14_2_conv14_2/relu_splitncnn_3\nConvolution conv15_1 1 1 conv14_2_conv14_2/relu_splitncnn_3 conv15_1_conv15_1/relu 0=128 1=1 5=1 6=65536 8=102 9=1\nConvolution conv15_2 1 1 conv15_1_conv15_1/relu conv15_2_conv15_2/relu 0=256 1=3 3=2 4=1 5=1 6=294912 8=2 9=1\nSplit splitncnn_4 1 4 conv15_2_conv15_2/relu conv15_2_conv15_2/relu_splitncnn_0 conv15_2_conv15_2/relu_splitncnn_1 conv15_2_conv15_2/relu_splitncnn_2 conv15_2_conv15_2/relu_splitncnn_3\nConvolution conv16_1 1 1 conv15_2_conv15_2/relu_splitncnn_3 conv16_1_conv16_1/relu 0=128 1=1 5=1 6=32768 8=102 9=1\nConvolution conv16_2 1 1 conv16_1_conv16_1/relu conv16_2_conv16_2/relu 0=256 1=3 3=2 4=1 5=1 6=294912 8=2 9=1\nSplit splitncnn_5 1 4 conv16_2_conv16_2/relu conv16_2_conv16_2/relu_splitncnn_0 conv16_2_conv16_2/relu_splitncnn_1 conv16_2_conv16_2/relu_splitncnn_2 conv16_2_conv16_2/relu_splitncnn_3\nConvolution conv17_1 1 1 conv16_2_conv16_2/relu_splitncnn_3 conv17_1_conv17_1/relu 0=64 1=1 5=1 6=16384 8=102 9=1\nConvolution conv17_2 1 1 conv17_1_conv17_1/relu conv17_2_conv17_2/relu 0=128 1=3 3=2 4=1 5=1 6=73728 8=2 9=1\nSplit splitncnn_6 1 3 conv17_2_conv17_2/relu conv17_2_conv17_2/relu_splitncnn_0 conv17_2_conv17_2/relu_splitncnn_1 conv17_2_conv17_2/relu_splitncnn_2\nConvolution conv11_mbox_loc 1 1 conv11_conv11/relu_splitncnn_2 conv11_mbox_loc 0=12 1=1 5=1 6=6144 8=2\nPermute conv11_mbox_loc_perm 1 1 conv11_mbox_loc conv11_mbox_loc_perm 0=3\nFlatten conv11_mbox_loc_flat 1 1 conv11_mbox_loc_perm conv11_mbox_loc_flat\nConvolution conv11_mbox_conf 1 1 conv11_conv11/relu_splitncnn_1 conv11_mbox_conf 0=63 1=1 5=1 6=32256 8=2\nPermute conv11_mbox_conf_perm 1 1 conv11_mbox_conf conv11_mbox_conf_perm 0=3\nFlatten conv11_mbox_conf_flat 1 1 conv11_mbox_conf_perm conv11_mbox_conf_flat\nPriorBox conv11_mbox_priorbox 2 1 conv11_conv11/relu_splitncnn_0 data_splitncnn_5 conv11_mbox_priorbox -23300=1,60.000000 -23302=1,2.000000 9=-233 10=-233 13=0.500000\nConvolution conv13_mbox_loc 1 1 conv13_conv13/relu_splitncnn_2 conv13_mbox_loc 0=24 1=1 5=1 6=24576 8=2\nPermute conv13_mbox_loc_perm 1 1 conv13_mbox_loc conv13_mbox_loc_perm 0=3\nFlatten conv13_mbox_loc_flat 1 1 conv13_mbox_loc_perm conv13_mbox_loc_flat\nConvolution conv13_mbox_conf 1 1 conv13_conv13/relu_splitncnn_1 conv13_mbox_conf 0=126 1=1 5=1 6=129024 8=2\nPermute conv13_mbox_conf_perm 1 1 conv13_mbox_conf conv13_mbox_conf_perm 0=3\nFlatten conv13_mbox_conf_flat 1 1 conv13_mbox_conf_perm conv13_mbox_conf_flat\nPriorBox conv13_mbox_priorbox 2 1 conv13_conv13/relu_splitncnn_0 data_splitncnn_4 conv13_mbox_priorbox -23300=1,105.000000 -23301=1,150.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 13=0.500000\nConvolution conv14_2_mbox_loc 1 1 conv14_2_conv14_2/relu_splitncnn_2 conv14_2_mbox_loc 0=24 1=1 5=1 6=12288 8=2\nPermute conv14_2_mbox_loc_perm 1 1 conv14_2_mbox_loc conv14_2_mbox_loc_perm 0=3\nFlatten conv14_2_mbox_loc_flat 1 1 conv14_2_mbox_loc_perm conv14_2_mbox_loc_flat\nConvolution conv14_2_mbox_conf 1 1 conv14_2_conv14_2/relu_splitncnn_1 conv14_2_mbox_conf 0=126 1=1 5=1 6=64512 8=2\nPermute conv14_2_mbox_conf_perm 1 1 conv14_2_mbox_conf conv14_2_mbox_conf_perm 0=3\nFlatten conv14_2_mbox_conf_flat 1 1 conv14_2_mbox_conf_perm conv14_2_mbox_conf_flat\nPriorBox conv14_2_mbox_priorbox 2 1 conv14_2_conv14_2/relu_splitncnn_0 data_splitncnn_3 conv14_2_mbox_priorbox -23300=1,150.000000 -23301=1,195.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 13=0.500000\nConvolution conv15_2_mbox_loc 1 1 conv15_2_conv15_2/relu_splitncnn_2 conv15_2_mbox_loc 0=24 1=1 5=1 6=6144 8=2\nPermute conv15_2_mbox_loc_perm 1 1 conv15_2_mbox_loc conv15_2_mbox_loc_perm 0=3\nFlatten conv15_2_mbox_loc_flat 1 1 conv15_2_mbox_loc_perm conv15_2_mbox_loc_flat\nConvolution conv15_2_mbox_conf 1 1 conv15_2_conv15_2/relu_splitncnn_1 conv15_2_mbox_conf 0=126 1=1 5=1 6=32256 8=2\nPermute conv15_2_mbox_conf_perm 1 1 conv15_2_mbox_conf conv15_2_mbox_conf_perm 0=3\nFlatten conv15_2_mbox_conf_flat 1 1 conv15_2_mbox_conf_perm conv15_2_mbox_conf_flat\nPriorBox conv15_2_mbox_priorbox 2 1 conv15_2_conv15_2/relu_splitncnn_0 data_splitncnn_2 conv15_2_mbox_priorbox -23300=1,195.000000 -23301=1,240.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 13=0.500000\nConvolution conv16_2_mbox_loc 1 1 conv16_2_conv16_2/relu_splitncnn_2 conv16_2_mbox_loc 0=24 1=1 5=1 6=6144 8=2\nPermute conv16_2_mbox_loc_perm 1 1 conv16_2_mbox_loc conv16_2_mbox_loc_perm 0=3\nFlatten conv16_2_mbox_loc_flat 1 1 conv16_2_mbox_loc_perm conv16_2_mbox_loc_flat\nConvolution conv16_2_mbox_conf 1 1 conv16_2_conv16_2/relu_splitncnn_1 conv16_2_mbox_conf 0=126 1=1 5=1 6=32256 8=2\nPermute conv16_2_mbox_conf_perm 1 1 conv16_2_mbox_conf conv16_2_mbox_conf_perm 0=3\nFlatten conv16_2_mbox_conf_flat 1 1 conv16_2_mbox_conf_perm conv16_2_mbox_conf_flat\nPriorBox conv16_2_mbox_priorbox 2 1 conv16_2_conv16_2/relu_splitncnn_0 data_splitncnn_1 conv16_2_mbox_priorbox -23300=1,240.000000 -23301=1,285.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 13=0.500000\nConvolution conv17_2_mbox_loc 1 1 conv17_2_conv17_2/relu_splitncnn_2 conv17_2_mbox_loc 0=24 1=1 5=1 6=3072 8=2\nPermute conv17_2_mbox_loc_perm 1 1 conv17_2_mbox_loc conv17_2_mbox_loc_perm 0=3\nFlatten conv17_2_mbox_loc_flat 1 1 conv17_2_mbox_loc_perm conv17_2_mbox_loc_flat\nConvolution conv17_2_mbox_conf 1 1 conv17_2_conv17_2/relu_splitncnn_1 conv17_2_mbox_conf 0=126 1=1 5=1 6=16128 8=2\nPermute conv17_2_mbox_conf_perm 1 1 conv17_2_mbox_conf conv17_2_mbox_conf_perm 0=3\nFlatten conv17_2_mbox_conf_flat 1 1 conv17_2_mbox_conf_perm conv17_2_mbox_conf_flat\nPriorBox conv17_2_mbox_priorbox 2 1 conv17_2_conv17_2/relu_splitncnn_0 data_splitncnn_0 conv17_2_mbox_priorbox -23300=1,285.000000 -23301=1,300.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 13=0.500000\nConcat mbox_loc 6 1 conv11_mbox_loc_flat conv13_mbox_loc_flat conv14_2_mbox_loc_flat conv15_2_mbox_loc_flat conv16_2_mbox_loc_flat conv17_2_mbox_loc_flat mbox_loc\nConcat mbox_conf 6 1 conv11_mbox_conf_flat conv13_mbox_conf_flat conv14_2_mbox_conf_flat conv15_2_mbox_conf_flat conv16_2_mbox_conf_flat conv17_2_mbox_conf_flat mbox_conf\nConcat mbox_priorbox 6 1 conv11_mbox_priorbox conv13_mbox_priorbox conv14_2_mbox_priorbox conv15_2_mbox_priorbox conv16_2_mbox_priorbox conv17_2_mbox_priorbox mbox_priorbox 0=1\nReshape mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1\nSoftmax mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax 0=1 1=1\nFlatten mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten\nDetectionOutput detection_out 3 1 mbox_loc mbox_conf_flatten mbox_priorbox output 0=21 1=0.450000 2=100 4=0.250000\n" }, { "path": "benchmark/mobilenet_v2.param", "content": "7767517\n77 87\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1/bn_relu1 -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolution conv2_1/expand 1 1 conv1/bn_relu1 conv2_1/expand/bn_relu2_1/expand -23330=4,3,112,112,32 0=32 1=1 5=1 6=1024 9=1\nConvolutionDepthWise conv2_1/dwise 1 1 conv2_1/expand/bn_relu2_1/expand conv2_1/dwise/bn_relu2_1/dwise -23330=4,3,112,112,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution conv2_1/linear 1 1 conv2_1/dwise/bn_relu2_1/dwise conv2_1/linear/bn_conv2_1/linear/scale -23330=4,3,112,112,16 0=16 1=1 5=1 6=512\nConvolution conv2_2/expand 1 1 conv2_1/linear/bn_conv2_1/linear/scale conv2_2/expand/bn_relu2_2/expand -23330=4,3,112,112,96 0=96 1=1 5=1 6=1536 9=1\nConvolutionDepthWise conv2_2/dwise 1 1 conv2_2/expand/bn_relu2_2/expand conv2_2/dwise/bn_relu2_2/dwise -23330=4,3,56,56,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96 9=1\nConvolution conv2_2/linear 1 1 conv2_2/dwise/bn_relu2_2/dwise conv2_2/linear/bn_conv2_2/linear/scale -23330=4,3,56,56,24 0=24 1=1 5=1 6=2304\nSplit splitncnn_0 1 2 conv2_2/linear/bn_conv2_2/linear/scale conv2_2/linear/bn_conv2_2/linear/scale_splitncnn_0 conv2_2/linear/bn_conv2_2/linear/scale_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolution conv3_1/expand 1 1 conv2_2/linear/bn_conv2_2/linear/scale_splitncnn_1 conv3_1/expand/bn_relu3_1/expand -23330=4,3,56,56,144 0=144 1=1 5=1 6=3456 9=1\nConvolutionDepthWise conv3_1/dwise 1 1 conv3_1/expand/bn_relu3_1/expand conv3_1/dwise/bn_relu3_1/dwise -23330=4,3,56,56,144 0=144 1=3 4=1 5=1 6=1296 7=144 9=1\nConvolution conv3_1/linear 1 1 conv3_1/dwise/bn_relu3_1/dwise conv3_1/linear/bn_conv3_1/linear/scale -23330=4,3,56,56,24 0=24 1=1 5=1 6=3456\nEltwise block_3_1 2 1 conv2_2/linear/bn_conv2_2/linear/scale_splitncnn_0 conv3_1/linear/bn_conv3_1/linear/scale block_3_1 -23330=4,3,56,56,24 0=1\nConvolution conv3_2/expand 1 1 block_3_1 conv3_2/expand/bn_relu3_2/expand -23330=4,3,56,56,144 0=144 1=1 5=1 6=3456 9=1\nConvolutionDepthWise conv3_2/dwise 1 1 conv3_2/expand/bn_relu3_2/expand conv3_2/dwise/bn_relu3_2/dwise -23330=4,3,28,28,144 0=144 1=3 3=2 4=1 5=1 6=1296 7=144 9=1\nConvolution conv3_2/linear 1 1 conv3_2/dwise/bn_relu3_2/dwise conv3_2/linear/bn_conv3_2/linear/scale -23330=4,3,28,28,32 0=32 1=1 5=1 6=4608\nSplit splitncnn_1 1 2 conv3_2/linear/bn_conv3_2/linear/scale conv3_2/linear/bn_conv3_2/linear/scale_splitncnn_0 conv3_2/linear/bn_conv3_2/linear/scale_splitncnn_1 -23330=8,3,28,28,32,3,28,28,32\nConvolution conv4_1/expand 1 1 conv3_2/linear/bn_conv3_2/linear/scale_splitncnn_1 conv4_1/expand/bn_relu4_1/expand -23330=4,3,28,28,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv4_1/dwise 1 1 conv4_1/expand/bn_relu4_1/expand conv4_1/dwise/bn_relu4_1/dwise -23330=4,3,28,28,192 0=192 1=3 4=1 5=1 6=1728 7=192 9=1\nConvolution conv4_1/linear 1 1 conv4_1/dwise/bn_relu4_1/dwise conv4_1/linear/bn_conv4_1/linear/scale -23330=4,3,28,28,32 0=32 1=1 5=1 6=6144\nEltwise block_4_1 2 1 conv3_2/linear/bn_conv3_2/linear/scale_splitncnn_0 conv4_1/linear/bn_conv4_1/linear/scale block_4_1 -23330=4,3,28,28,32 0=1\nSplit splitncnn_2 1 2 block_4_1 block_4_1_splitncnn_0 block_4_1_splitncnn_1 -23330=8,3,28,28,32,3,28,28,32\nConvolution conv4_2/expand 1 1 block_4_1_splitncnn_1 conv4_2/expand/bn_relu4_2/expand -23330=4,3,28,28,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv4_2/dwise 1 1 conv4_2/expand/bn_relu4_2/expand conv4_2/dwise/bn_relu4_2/dwise -23330=4,3,28,28,192 0=192 1=3 4=1 5=1 6=1728 7=192 9=1\nConvolution conv4_2/linear 1 1 conv4_2/dwise/bn_relu4_2/dwise conv4_2/linear/bn_conv4_2/linear/scale -23330=4,3,28,28,32 0=32 1=1 5=1 6=6144\nEltwise block_4_2 2 1 block_4_1_splitncnn_0 conv4_2/linear/bn_conv4_2/linear/scale block_4_2 -23330=4,3,28,28,32 0=1\nConvolution conv4_3/expand 1 1 block_4_2 conv4_3/expand/bn_relu4_3/expand -23330=4,3,28,28,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv4_3/dwise 1 1 conv4_3/expand/bn_relu4_3/expand conv4_3/dwise/bn_relu4_3/dwise -23330=4,3,14,14,192 0=192 1=3 3=2 4=1 5=1 6=1728 7=192 9=1\nConvolution conv4_3/linear 1 1 conv4_3/dwise/bn_relu4_3/dwise conv4_3/linear/bn_conv4_3/linear/scale -23330=4,3,14,14,64 0=64 1=1 5=1 6=12288\nSplit splitncnn_3 1 2 conv4_3/linear/bn_conv4_3/linear/scale conv4_3/linear/bn_conv4_3/linear/scale_splitncnn_0 conv4_3/linear/bn_conv4_3/linear/scale_splitncnn_1 -23330=8,3,14,14,64,3,14,14,64\nConvolution conv4_4/expand 1 1 conv4_3/linear/bn_conv4_3/linear/scale_splitncnn_1 conv4_4/expand/bn_relu4_4/expand -23330=4,3,14,14,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv4_4/dwise 1 1 conv4_4/expand/bn_relu4_4/expand conv4_4/dwise/bn_relu4_4/dwise -23330=4,3,14,14,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv4_4/linear 1 1 conv4_4/dwise/bn_relu4_4/dwise conv4_4/linear/bn_conv4_4/linear/scale -23330=4,3,14,14,64 0=64 1=1 5=1 6=24576\nEltwise block_4_4 2 1 conv4_3/linear/bn_conv4_3/linear/scale_splitncnn_0 conv4_4/linear/bn_conv4_4/linear/scale block_4_4 -23330=4,3,14,14,64 0=1\nSplit splitncnn_4 1 2 block_4_4 block_4_4_splitncnn_0 block_4_4_splitncnn_1 -23330=8,3,14,14,64,3,14,14,64\nConvolution conv4_5/expand 1 1 block_4_4_splitncnn_1 conv4_5/expand/bn_relu4_5/expand -23330=4,3,14,14,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv4_5/dwise 1 1 conv4_5/expand/bn_relu4_5/expand conv4_5/dwise/bn_relu4_5/dwise -23330=4,3,14,14,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv4_5/linear 1 1 conv4_5/dwise/bn_relu4_5/dwise conv4_5/linear/bn_conv4_5/linear/scale -23330=4,3,14,14,64 0=64 1=1 5=1 6=24576\nEltwise block_4_5 2 1 block_4_4_splitncnn_0 conv4_5/linear/bn_conv4_5/linear/scale block_4_5 -23330=4,3,14,14,64 0=1\nSplit splitncnn_5 1 2 block_4_5 block_4_5_splitncnn_0 block_4_5_splitncnn_1 -23330=8,3,14,14,64,3,14,14,64\nConvolution conv4_6/expand 1 1 block_4_5_splitncnn_1 conv4_6/expand/bn_relu4_6/expand -23330=4,3,14,14,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv4_6/dwise 1 1 conv4_6/expand/bn_relu4_6/expand conv4_6/dwise/bn_relu4_6/dwise -23330=4,3,14,14,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv4_6/linear 1 1 conv4_6/dwise/bn_relu4_6/dwise conv4_6/linear/bn_conv4_6/linear/scale -23330=4,3,14,14,64 0=64 1=1 5=1 6=24576\nEltwise block_4_6 2 1 block_4_5_splitncnn_0 conv4_6/linear/bn_conv4_6/linear/scale block_4_6 -23330=4,3,14,14,64 0=1\nConvolution conv4_7/expand 1 1 block_4_6 conv4_7/expand/bn_relu4_7/expand -23330=4,3,14,14,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv4_7/dwise 1 1 conv4_7/expand/bn_relu4_7/expand conv4_7/dwise/bn_relu4_7/dwise -23330=4,3,14,14,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv4_7/linear 1 1 conv4_7/dwise/bn_relu4_7/dwise conv4_7/linear/bn_conv4_7/linear/scale -23330=4,3,14,14,96 0=96 1=1 5=1 6=36864\nSplit splitncnn_6 1 2 conv4_7/linear/bn_conv4_7/linear/scale conv4_7/linear/bn_conv4_7/linear/scale_splitncnn_0 conv4_7/linear/bn_conv4_7/linear/scale_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution conv5_1/expand 1 1 conv4_7/linear/bn_conv4_7/linear/scale_splitncnn_1 conv5_1/expand/bn_relu5_1/expand -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise conv5_1/dwise 1 1 conv5_1/expand/bn_relu5_1/expand conv5_1/dwise/bn_relu5_1/dwise -23330=4,3,14,14,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution conv5_1/linear 1 1 conv5_1/dwise/bn_relu5_1/dwise conv5_1/linear/bn_conv5_1/linear/scale -23330=4,3,14,14,96 0=96 1=1 5=1 6=55296\nEltwise block_5_1 2 1 conv4_7/linear/bn_conv4_7/linear/scale_splitncnn_0 conv5_1/linear/bn_conv5_1/linear/scale block_5_1 -23330=4,3,14,14,96 0=1\nSplit splitncnn_7 1 2 block_5_1 block_5_1_splitncnn_0 block_5_1_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution conv5_2/expand 1 1 block_5_1_splitncnn_1 conv5_2/expand/bn_relu5_2/expand -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise conv5_2/dwise 1 1 conv5_2/expand/bn_relu5_2/expand conv5_2/dwise/bn_relu5_2/dwise -23330=4,3,14,14,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution conv5_2/linear 1 1 conv5_2/dwise/bn_relu5_2/dwise conv5_2/linear/bn_conv5_2/linear/scale -23330=4,3,14,14,96 0=96 1=1 5=1 6=55296\nEltwise block_5_2 2 1 block_5_1_splitncnn_0 conv5_2/linear/bn_conv5_2/linear/scale block_5_2 -23330=4,3,14,14,96 0=1\nConvolution conv5_3/expand 1 1 block_5_2 conv5_3/expand/bn_relu5_3/expand -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise conv5_3/dwise 1 1 conv5_3/expand/bn_relu5_3/expand conv5_3/dwise/bn_relu5_3/dwise -23330=4,3,7,7,576 0=576 1=3 3=2 4=1 5=1 6=5184 7=576 9=1\nConvolution conv5_3/linear 1 1 conv5_3/dwise/bn_relu5_3/dwise conv5_3/linear/bn_conv5_3/linear/scale -23330=4,3,7,7,160 0=160 1=1 5=1 6=92160\nSplit splitncnn_8 1 2 conv5_3/linear/bn_conv5_3/linear/scale conv5_3/linear/bn_conv5_3/linear/scale_splitncnn_0 conv5_3/linear/bn_conv5_3/linear/scale_splitncnn_1 -23330=8,3,7,7,160,3,7,7,160\nConvolution conv6_1/expand 1 1 conv5_3/linear/bn_conv5_3/linear/scale_splitncnn_1 conv6_1/expand/bn_relu6_1/expand -23330=4,3,7,7,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv6_1/dwise 1 1 conv6_1/expand/bn_relu6_1/expand conv6_1/dwise/bn_relu6_1/dwise -23330=4,3,7,7,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv6_1/linear 1 1 conv6_1/dwise/bn_relu6_1/dwise conv6_1/linear/bn_conv6_1/linear/scale -23330=4,3,7,7,160 0=160 1=1 5=1 6=153600\nEltwise block_6_1 2 1 conv5_3/linear/bn_conv5_3/linear/scale_splitncnn_0 conv6_1/linear/bn_conv6_1/linear/scale block_6_1 -23330=4,3,7,7,160 0=1\nSplit splitncnn_9 1 2 block_6_1 block_6_1_splitncnn_0 block_6_1_splitncnn_1 -23330=8,3,7,7,160,3,7,7,160\nConvolution conv6_2/expand 1 1 block_6_1_splitncnn_1 conv6_2/expand/bn_relu6_2/expand -23330=4,3,7,7,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv6_2/dwise 1 1 conv6_2/expand/bn_relu6_2/expand conv6_2/dwise/bn_relu6_2/dwise -23330=4,3,7,7,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv6_2/linear 1 1 conv6_2/dwise/bn_relu6_2/dwise conv6_2/linear/bn_conv6_2/linear/scale -23330=4,3,7,7,160 0=160 1=1 5=1 6=153600\nEltwise block_6_2 2 1 block_6_1_splitncnn_0 conv6_2/linear/bn_conv6_2/linear/scale block_6_2 -23330=4,3,7,7,160 0=1\nConvolution conv6_3/expand 1 1 block_6_2 conv6_3/expand/bn_relu6_3/expand -23330=4,3,7,7,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv6_3/dwise 1 1 conv6_3/expand/bn_relu6_3/expand conv6_3/dwise/bn_relu6_3/dwise -23330=4,3,7,7,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv6_3/linear 1 1 conv6_3/dwise/bn_relu6_3/dwise conv6_3/linear/bn_conv6_3/linear/scale -23330=4,3,7,7,320 0=320 1=1 5=1 6=307200\nConvolution conv6_4 1 1 conv6_3/linear/bn_conv6_3/linear/scale conv6_4/bn_relu6_4 -23330=4,3,7,7,1280 0=1280 1=1 5=1 6=409600 9=1\nPooling pool6 1 1 conv6_4/bn_relu6_4 pool6 -23330=4,1,1280,1,1 0=1 4=1\nInnerProduct fc7 1 1 pool6 fc7 -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\nSoftmax prob 1 1 fc7 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/mobilenet_v3.param", "content": "7767517\n145 163\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution 313 1 1 data 313 -23330=4,3,112,112,16 0=16 1=3 3=2 4=1 5=1 6=432\nSplit splitncnn_0 1 2 313 313_splitncnn_0 313_splitncnn_1 -23330=8,3,112,112,16,3,112,112,16\nHardSigmoid 319 1 1 313_splitncnn_1 319 -23330=4,3,112,112,16\nBinaryOp 320 2 1 313_splitncnn_0 319 320 -23330=4,3,112,112,16 0=2\nSplit splitncnn_1 1 2 320 320_splitncnn_0 320_splitncnn_1 -23330=8,3,112,112,16,3,112,112,16\nConvolutionDepthWise 321 1 1 320_splitncnn_1 323 -23330=4,3,112,112,16 0=16 1=3 4=1 5=1 6=144 7=16 9=1\nConvolution 324 1 1 323 324 -23330=4,3,112,112,16 0=16 1=1 5=1 6=256\nBinaryOp 326 2 1 320_splitncnn_0 324 326 -23330=4,3,112,112,16\nConvolution 327 1 1 326 329 -23330=4,3,112,112,64 0=64 1=1 5=1 6=1024 9=1\nConvolutionDepthWise 330 1 1 329 332 -23330=4,3,56,56,64 0=64 1=3 3=2 4=1 5=1 6=576 7=64 9=1\nConvolution 333 1 1 332 333 -23330=4,3,56,56,24 0=24 1=1 5=1 6=1536\nSplit splitncnn_2 1 2 333 333_splitncnn_0 333_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolution 335 1 1 333_splitncnn_1 337 -23330=4,3,56,56,72 0=72 1=1 5=1 6=1728 9=1\nConvolutionDepthWise 338 1 1 337 340 -23330=4,3,56,56,72 0=72 1=3 4=1 5=1 6=648 7=72 9=1\nConvolution 341 1 1 340 341 -23330=4,3,56,56,24 0=24 1=1 5=1 6=1728\nBinaryOp 343 2 1 333_splitncnn_0 341 343 -23330=4,3,56,56,24\nConvolution 344 1 1 343 346 -23330=4,3,56,56,72 0=72 1=1 5=1 6=1728 9=1\nConvolutionDepthWise 347 1 1 346 347 -23330=4,3,28,28,72 0=72 1=5 3=2 4=2 5=1 6=1800 7=72\nSplit splitncnn_3 1 2 347 347_splitncnn_0 347_splitncnn_1 -23330=8,3,28,28,72,3,28,28,72\nPooling 355 1 1 347_splitncnn_1 359 -23330=4,1,72,1,1 0=1 4=1\nInnerProduct 360 1 1 359 361 -23330=4,1,18,1,1 0=18 1=1 2=1296 9=1\nInnerProduct 362 1 1 361 362 -23330=4,1,72,1,1 0=72 1=1 2=1296\nHardSigmoid 367 1 1 362 367 -23330=4,1,72,1,1\nBinaryOp 376 2 1 347_splitncnn_0 367 376 -23330=4,3,28,28,72 0=2\nReLU 377 1 1 376 377 -23330=4,3,28,28,72\nConvolution 378 1 1 377 378 -23330=4,3,28,28,40 0=40 1=1 5=1 6=2880\nSplit splitncnn_4 1 2 378 378_splitncnn_0 378_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution 380 1 1 378_splitncnn_1 382 -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise 383 1 1 382 383 -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120\nSplit splitncnn_5 1 2 383 383_splitncnn_0 383_splitncnn_1 -23330=8,3,28,28,120,3,28,28,120\nPooling 391 1 1 383_splitncnn_1 395 -23330=4,1,120,1,1 0=1 4=1\nInnerProduct 396 1 1 395 397 -23330=4,1,30,1,1 0=30 1=1 2=3600 9=1\nInnerProduct 398 1 1 397 398 -23330=4,1,120,1,1 0=120 1=1 2=3600\nHardSigmoid 403 1 1 398 403 -23330=4,1,120,1,1\nBinaryOp 412 2 1 383_splitncnn_0 403 412 -23330=4,3,28,28,120 0=2\nReLU 413 1 1 412 413 -23330=4,3,28,28,120\nConvolution 414 1 1 413 414 -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp 416 2 1 378_splitncnn_0 414 416 -23330=4,3,28,28,40\nSplit splitncnn_6 1 2 416 416_splitncnn_0 416_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution 417 1 1 416_splitncnn_1 419 -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise 420 1 1 419 420 -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120\nSplit splitncnn_7 1 2 420 420_splitncnn_0 420_splitncnn_1 -23330=8,3,28,28,120,3,28,28,120\nPooling 428 1 1 420_splitncnn_1 432 -23330=4,1,120,1,1 0=1 4=1\nInnerProduct 433 1 1 432 434 -23330=4,1,30,1,1 0=30 1=1 2=3600 9=1\nInnerProduct 435 1 1 434 435 -23330=4,1,120,1,1 0=120 1=1 2=3600\nHardSigmoid 440 1 1 435 440 -23330=4,1,120,1,1\nBinaryOp 449 2 1 420_splitncnn_0 440 449 -23330=4,3,28,28,120 0=2\nReLU 450 1 1 449 450 -23330=4,3,28,28,120\nConvolution 451 1 1 450 451 -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp 453 2 1 416_splitncnn_0 451 453 -23330=4,3,28,28,40\nConvolution 454 1 1 453 454 -23330=4,3,28,28,240 0=240 1=1 5=1 6=9600\nHardSwish 461 1 1 454 461 -23330=4,3,28,28,240\nConvolutionDepthWise 462 1 1 461 462 -23330=4,3,14,14,240 0=240 1=3 3=2 4=1 5=1 6=2160 7=240\nHardSwish 469 1 1 462 469 -23330=4,3,14,14,240\nConvolution 470 1 1 469 470 -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nSplit splitncnn_8 1 2 470 470_splitncnn_0 470_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution 472 1 1 470_splitncnn_1 472 -23330=4,3,14,14,200 0=200 1=1 5=1 6=16000\nHardSwish 479 1 1 472 479 -23330=4,3,14,14,200\nConvolutionDepthWise 480 1 1 479 480 -23330=4,3,14,14,200 0=200 1=3 4=1 5=1 6=1800 7=200\nHardSwish 487 1 1 480 487 -23330=4,3,14,14,200\nConvolution 488 1 1 487 488 -23330=4,3,14,14,80 0=80 1=1 5=1 6=16000\nBinaryOp 490 2 1 470_splitncnn_0 488 490 -23330=4,3,14,14,80\nSplit splitncnn_9 1 2 490 490_splitncnn_0 490_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution 491 1 1 490_splitncnn_1 491 -23330=4,3,14,14,184 0=184 1=1 5=1 6=14720\nHardSwish 498 1 1 491 498 -23330=4,3,14,14,184\nConvolutionDepthWise 499 1 1 498 499 -23330=4,3,14,14,184 0=184 1=3 4=1 5=1 6=1656 7=184\nHardSwish 506 1 1 499 506 -23330=4,3,14,14,184\nConvolution 507 1 1 506 507 -23330=4,3,14,14,80 0=80 1=1 5=1 6=14720\nBinaryOp 509 2 1 490_splitncnn_0 507 509 -23330=4,3,14,14,80\nSplit splitncnn_10 1 2 509 509_splitncnn_0 509_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution 510 1 1 509_splitncnn_1 510 -23330=4,3,14,14,184 0=184 1=1 5=1 6=14720\nHardSwish 517 1 1 510 517 -23330=4,3,14,14,184\nConvolutionDepthWise 518 1 1 517 518 -23330=4,3,14,14,184 0=184 1=3 4=1 5=1 6=1656 7=184\nHardSwish 525 1 1 518 525 -23330=4,3,14,14,184\nConvolution 526 1 1 525 526 -23330=4,3,14,14,80 0=80 1=1 5=1 6=14720\nBinaryOp 528 2 1 509_splitncnn_0 526 528 -23330=4,3,14,14,80\nConvolution 529 1 1 528 529 -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400\nHardSwish 536 1 1 529 536 -23330=4,3,14,14,480\nConvolutionDepthWise 537 1 1 536 537 -23330=4,3,14,14,480 0=480 1=3 4=1 5=1 6=4320 7=480\nSplit splitncnn_11 1 2 537 537_splitncnn_0 537_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nPooling 545 1 1 537_splitncnn_1 549 -23330=4,1,480,1,1 0=1 4=1\nInnerProduct 550 1 1 549 551 -23330=4,1,120,1,1 0=120 1=1 2=57600 9=1\nInnerProduct 552 1 1 551 552 -23330=4,1,480,1,1 0=480 1=1 2=57600\nHardSigmoid 557 1 1 552 557 -23330=4,1,480,1,1\nBinaryOp 566 2 1 537_splitncnn_0 557 566 -23330=4,3,14,14,480 0=2\nHardSwish 572 1 1 566 572 -23330=4,3,14,14,480\nConvolution 573 1 1 572 573 -23330=4,3,14,14,112 0=112 1=1 5=1 6=53760\nSplit splitncnn_12 1 2 573 573_splitncnn_0 573_splitncnn_1 -23330=8,3,14,14,112,3,14,14,112\nConvolution 575 1 1 573_splitncnn_1 575 -23330=4,3,14,14,672 0=672 1=1 5=1 6=75264\nHardSwish 582 1 1 575 582 -23330=4,3,14,14,672\nConvolutionDepthWise 583 1 1 582 583 -23330=4,3,14,14,672 0=672 1=3 4=1 5=1 6=6048 7=672\nSplit splitncnn_13 1 2 583 583_splitncnn_0 583_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nPooling 591 1 1 583_splitncnn_1 595 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct 596 1 1 595 597 -23330=4,1,168,1,1 0=168 1=1 2=112896 9=1\nInnerProduct 598 1 1 597 598 -23330=4,1,672,1,1 0=672 1=1 2=112896\nHardSigmoid 603 1 1 598 603 -23330=4,1,672,1,1\nBinaryOp 612 2 1 583_splitncnn_0 603 612 -23330=4,3,14,14,672 0=2\nHardSwish 618 1 1 612 618 -23330=4,3,14,14,672\nConvolution 619 1 1 618 619 -23330=4,3,14,14,112 0=112 1=1 5=1 6=75264\nBinaryOp 621 2 1 573_splitncnn_0 619 621 -23330=4,3,14,14,112\nConvolution 622 1 1 621 622 -23330=4,3,14,14,672 0=672 1=1 5=1 6=75264\nHardSwish 629 1 1 622 629 -23330=4,3,14,14,672\nConvolutionDepthWise 630 1 1 629 630 -23330=4,3,14,14,672 0=672 1=5 4=2 5=1 6=16800 7=672\nSplit splitncnn_14 1 2 630 630_splitncnn_0 630_splitncnn_1 -23330=8,3,14,14,672,3,14,14,672\nPooling 638 1 1 630_splitncnn_1 642 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct 643 1 1 642 644 -23330=4,1,168,1,1 0=168 1=1 2=112896 9=1\nInnerProduct 645 1 1 644 645 -23330=4,1,672,1,1 0=672 1=1 2=112896\nHardSigmoid 650 1 1 645 650 -23330=4,1,672,1,1\nBinaryOp 659 2 1 630_splitncnn_0 650 659 -23330=4,3,14,14,672 0=2\nHardSwish 665 1 1 659 665 -23330=4,3,14,14,672\nConvolution 666 1 1 665 666 -23330=4,3,14,14,160 0=160 1=1 5=1 6=107520\nConvolution 668 1 1 666 668 -23330=4,3,14,14,672 0=672 1=1 5=1 6=107520\nHardSwish 675 1 1 668 675 -23330=4,3,14,14,672\nConvolutionDepthWise 676 1 1 675 676 -23330=4,3,7,7,672 0=672 1=5 3=2 4=2 5=1 6=16800 7=672\nSplit splitncnn_15 1 2 676 676_splitncnn_0 676_splitncnn_1 -23330=8,3,7,7,672,3,7,7,672\nPooling 684 1 1 676_splitncnn_1 688 -23330=4,1,672,1,1 0=1 4=1\nInnerProduct 689 1 1 688 690 -23330=4,1,168,1,1 0=168 1=1 2=112896 9=1\nInnerProduct 691 1 1 690 691 -23330=4,1,672,1,1 0=672 1=1 2=112896\nHardSigmoid 696 1 1 691 696 -23330=4,1,672,1,1\nBinaryOp 705 2 1 676_splitncnn_0 696 705 -23330=4,3,7,7,672 0=2\nHardSwish 711 1 1 705 711 -23330=4,3,7,7,672\nConvolution 712 1 1 711 712 -23330=4,3,7,7,160 0=160 1=1 5=1 6=107520\nSplit splitncnn_16 1 2 712 712_splitncnn_0 712_splitncnn_1 -23330=8,3,7,7,160,3,7,7,160\nConvolution 714 1 1 712_splitncnn_1 714 -23330=4,3,7,7,960 0=960 1=1 5=1 6=153600\nHardSwish 721 1 1 714 721 -23330=4,3,7,7,960\nConvolutionDepthWise 722 1 1 721 722 -23330=4,3,7,7,960 0=960 1=5 4=2 5=1 6=24000 7=960\nSplit splitncnn_17 1 2 722 722_splitncnn_0 722_splitncnn_1 -23330=8,3,7,7,960,3,7,7,960\nPooling 730 1 1 722_splitncnn_1 734 -23330=4,1,960,1,1 0=1 4=1\nInnerProduct 735 1 1 734 736 -23330=4,1,240,1,1 0=240 1=1 2=230400 9=1\nInnerProduct 737 1 1 736 737 -23330=4,1,960,1,1 0=960 1=1 2=230400\nHardSigmoid 742 1 1 737 742 -23330=4,1,960,1,1\nBinaryOp 751 2 1 722_splitncnn_0 742 751 -23330=4,3,7,7,960 0=2\nHardSwish 757 1 1 751 757 -23330=4,3,7,7,960\nConvolution 758 1 1 757 758 -23330=4,3,7,7,160 0=160 1=1 5=1 6=153600\nBinaryOp 760 2 1 712_splitncnn_0 758 760 -23330=4,3,7,7,160\nConvolution 761 1 1 760 761 -23330=4,3,7,7,960 0=960 1=1 5=1 6=153600\nHardSwish 768 1 1 761 768 -23330=4,3,7,7,960\nPooling 769 1 1 768 769 -23330=4,1,960,1,1 0=1 4=1\nHardSwish 775 1 1 769 775 -23330=4,1,960,1,1\nReshape 783 1 1 775 783 -23330=4,1,960,1,1 0=-1\nInnerProduct 784 1 1 783 784 -23330=4,1,1280,1,1 0=1280 1=1 2=1228800\nHardSwish 790 1 1 784 790 -23330=4,1,1280,1,1\nInnerProduct 791 1 1 790 791 -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\nSoftmax prob 1 1 791 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/mobilenet_yolo.param", "content": "7767517\n39 41\nInput data 0 1 data -23330=4,3,416,416,3 0=416 1=416 2=3\nConvolution conv0 1 1 data conv0_conv0/relu -23330=4,3,208,208,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise conv1/dw 1 1 conv0_conv0/relu conv1/dw_conv1/dw/relu -23330=4,3,208,208,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution conv1 1 1 conv1/dw_conv1/dw/relu conv1_conv1/relu -23330=4,3,208,208,64 0=64 1=1 5=1 6=2048 9=1\nConvolutionDepthWise conv2/dw 1 1 conv1_conv1/relu conv2/dw_conv2/dw/relu -23330=4,3,104,104,64 0=64 1=3 3=2 4=1 5=1 6=576 7=64 9=1\nConvolution conv2 1 1 conv2/dw_conv2/dw/relu conv2_conv2/relu -23330=4,3,104,104,128 0=128 1=1 5=1 6=8192 9=1\nConvolutionDepthWise conv3/dw 1 1 conv2_conv2/relu conv3/dw_conv3/dw/relu -23330=4,3,104,104,128 0=128 1=3 4=1 5=1 6=1152 7=128 9=1\nConvolution conv3 1 1 conv3/dw_conv3/dw/relu conv3_conv3/relu -23330=4,3,104,104,128 0=128 1=1 5=1 6=16384 9=1\nConvolutionDepthWise conv4/dw 1 1 conv3_conv3/relu conv4/dw_conv4/dw/relu -23330=4,3,52,52,128 0=128 1=3 3=2 4=1 5=1 6=1152 7=128 9=1\nConvolution conv4 1 1 conv4/dw_conv4/dw/relu conv4_conv4/relu -23330=4,3,52,52,256 0=256 1=1 5=1 6=32768 9=1\nConvolutionDepthWise conv5/dw 1 1 conv4_conv4/relu conv5/dw_conv5/dw/relu -23330=4,3,52,52,256 0=256 1=3 4=1 5=1 6=2304 7=256 9=1\nConvolution conv5 1 1 conv5/dw_conv5/dw/relu conv5_conv5/relu -23330=4,3,52,52,256 0=256 1=1 5=1 6=65536 9=1\nConvolutionDepthWise conv6/dw 1 1 conv5_conv5/relu conv6/dw_conv6/dw/relu -23330=4,3,26,26,256 0=256 1=3 3=2 4=1 5=1 6=2304 7=256 9=1\nConvolution conv6 1 1 conv6/dw_conv6/dw/relu conv6_conv6/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=131072 9=1\nConvolutionDepthWise conv7/dw 1 1 conv6_conv6/relu conv7/dw_conv7/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv7 1 1 conv7/dw_conv7/dw/relu conv7_conv7/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv8/dw 1 1 conv7_conv7/relu conv8/dw_conv8/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv8 1 1 conv8/dw_conv8/dw/relu conv8_conv8/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv9/dw 1 1 conv8_conv8/relu conv9/dw_conv9/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv9 1 1 conv9/dw_conv9/dw/relu conv9_conv9/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv10/dw 1 1 conv9_conv9/relu conv10/dw_conv10/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv10 1 1 conv10/dw_conv10/dw/relu conv10_conv10/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=262144 9=1\nConvolutionDepthWise conv11/dw 1 1 conv10_conv10/relu conv11/dw_conv11/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv11 1 1 conv11/dw_conv11/dw/relu conv11_conv11/relu -23330=4,3,26,26,512 0=512 1=1 5=1 6=262144 9=1\nSplit splitncnn_0 1 2 conv11_conv11/relu conv11_conv11/relu_splitncnn_0 conv11_conv11/relu_splitncnn_1 -23330=8,3,26,26,512,3,26,26,512\nConvolutionDepthWise conv12/dw 1 1 conv11_conv11/relu_splitncnn_1 conv12/dw_conv12/dw/relu -23330=4,3,13,13,512 0=512 1=3 3=2 4=1 5=1 6=4608 7=512 9=1\nConvolution conv12 1 1 conv12/dw_conv12/dw/relu conv12_conv12/relu -23330=4,3,13,13,1024 0=1024 1=1 5=1 6=524288 9=1\nConvolutionDepthWise conv13/dw 1 1 conv12_conv12/relu conv13/dw_conv13/dw/relu -23330=4,3,13,13,1024 0=1024 1=3 4=1 5=1 6=9216 7=1024 9=1\nConvolution conv13 1 1 conv13/dw_conv13/dw/relu conv13_conv13/relu -23330=4,3,13,13,1024 0=1024 1=1 5=1 6=1048576 9=1\nConvolutionDepthWise conv16/dw 1 1 conv13_conv13/relu conv16/dw_conv16/dw/relu -23330=4,3,13,13,1024 0=1024 1=3 4=1 5=1 6=9216 7=1024 9=1\nConvolution conv17 1 1 conv16/dw_conv16/dw/relu conv17_conv17/relu -23330=4,3,13,13,1024 0=1024 1=1 5=1 6=1048576 9=1\nSplit splitncnn_1 1 2 conv17_conv17/relu conv17_conv17/relu_splitncnn_0 conv17_conv17/relu_splitncnn_1 -23330=8,3,13,13,1024,3,13,13,1024\nDeconvolutionDepthWise upsample 1 1 conv17_conv17/relu_splitncnn_1 upsample -23330=4,3,26,26,512 0=512 1=4 3=2 4=1 6=16384 7=512\nEltwise conv_18/sum 2 1 conv11_conv11/relu_splitncnn_0 upsample conv_18/sum -23330=4,3,26,26,512 0=1\nConvolutionDepthWise conv19/dw 1 1 conv_18/sum conv19/dw_conv19/dw/relu -23330=4,3,26,26,512 0=512 1=3 4=1 5=1 6=4608 7=512 9=1\nConvolution conv20 1 1 conv19/dw_conv19/dw/relu conv20_conv20/relu -23330=4,3,26,26,1024 0=1024 1=1 5=1 6=524288 9=1\nConvolution conv22_indoor 1 1 conv17_conv17/relu_splitncnn_0 conv22 -23330=4,3,13,13,125 0=125 1=1 5=1 6=128000\nConvolution conv23_indoor 1 1 conv20_conv20/relu conv23 -23330=4,3,26,26,125 0=125 1=1 5=1 6=128000\nYoloDetectionOutput detection_out 2 1 conv22 conv23 output -23330=4,3,13,13,125 2=4.000000e-01 -23304=10,1.080000e+00,1.190000e+00,3.420000e+00,4.410000e+00,6.630000e+00,1.138000e+01,9.420000e+00,5.110000e+00,1.662000e+01,1.052000e+01\n" }, { "path": "benchmark/mobilenetv2_yolov3.param", "content": "7767517\n87 99\nInput data 0 1 data -23330=4,3,352,352,3 0=352 1=352 2=3\nConvolution conv1 1 1 data conv1_relu1 -23330=4,3,176,176,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise conv2 1 1 conv1_relu1 conv2_relu2 -23330=4,3,176,176,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution conv3 1 1 conv2_relu2 conv3 -23330=4,3,176,176,16 0=16 1=1 5=1 6=512\nConvolution conv4 1 1 conv3 conv4_relu3 -23330=4,3,176,176,96 0=96 1=1 5=1 6=1536 9=1\nConvolutionDepthWise conv5 1 1 conv4_relu3 conv5_relu4 -23330=4,3,88,88,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96 9=1\nConvolution conv6 1 1 conv5_relu4 conv6 -23330=4,3,88,88,24 0=24 1=1 5=1 6=2304\nSplit splitncnn_0 1 2 conv6 conv6_splitncnn_0 conv6_splitncnn_1 -23330=8,3,88,88,24,3,88,88,24\nConvolution conv7 1 1 conv6_splitncnn_1 conv7_relu5 -23330=4,3,88,88,144 0=144 1=1 5=1 6=3456 9=1\nConvolutionDepthWise conv8 1 1 conv7_relu5 conv8_relu6 -23330=4,3,88,88,144 0=144 1=3 4=1 5=1 6=1296 7=144 9=1\nConvolution conv9 1 1 conv8_relu6 conv9 -23330=4,3,88,88,24 0=24 1=1 5=1 6=3456\nEltwise add1 2 1 conv6_splitncnn_0 conv9 add1 -23330=4,3,88,88,24 0=1\nConvolution conv10 1 1 add1 conv10_relu7 -23330=4,3,88,88,144 0=144 1=1 5=1 6=3456 9=1\nConvolutionDepthWise conv11 1 1 conv10_relu7 conv11_relu8 -23330=4,3,44,44,144 0=144 1=3 3=2 4=1 5=1 6=1296 7=144 9=1\nConvolution conv12 1 1 conv11_relu8 conv12 -23330=4,3,44,44,32 0=32 1=1 5=1 6=4608\nSplit splitncnn_1 1 2 conv12 conv12_splitncnn_0 conv12_splitncnn_1 -23330=8,3,44,44,32,3,44,44,32\nConvolution conv13 1 1 conv12_splitncnn_1 conv13_relu9 -23330=4,3,44,44,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv14 1 1 conv13_relu9 conv14_relu10 -23330=4,3,44,44,192 0=192 1=3 4=1 5=1 6=1728 7=192 9=1\nConvolution conv15 1 1 conv14_relu10 conv15 -23330=4,3,44,44,32 0=32 1=1 5=1 6=6144\nEltwise add2 2 1 conv12_splitncnn_0 conv15 add2 -23330=4,3,44,44,32 0=1\nSplit splitncnn_2 1 2 add2 add2_splitncnn_0 add2_splitncnn_1 -23330=8,3,44,44,32,3,44,44,32\nConvolution conv16 1 1 add2_splitncnn_1 conv16_relu11 -23330=4,3,44,44,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv17 1 1 conv16_relu11 conv17_relu12 -23330=4,3,44,44,192 0=192 1=3 4=1 5=1 6=1728 7=192 9=1\nConvolution conv18 1 1 conv17_relu12 conv18 -23330=4,3,44,44,32 0=32 1=1 5=1 6=6144\nEltwise add3 2 1 add2_splitncnn_0 conv18 add3 -23330=4,3,44,44,32 0=1\nConvolution conv19 1 1 add3 conv19_relu13 -23330=4,3,44,44,192 0=192 1=1 5=1 6=6144 9=1\nConvolutionDepthWise conv20 1 1 conv19_relu13 conv20_relu14 -23330=4,3,22,22,192 0=192 1=3 3=2 4=1 5=1 6=1728 7=192 9=1\nConvolution conv21 1 1 conv20_relu14 conv21 -23330=4,3,22,22,64 0=64 1=1 5=1 6=12288\nSplit splitncnn_3 1 2 conv21 conv21_splitncnn_0 conv21_splitncnn_1 -23330=8,3,22,22,64,3,22,22,64\nConvolution conv22 1 1 conv21_splitncnn_1 conv22_relu15 -23330=4,3,22,22,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv23 1 1 conv22_relu15 conv23_relu16 -23330=4,3,22,22,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv24 1 1 conv23_relu16 conv24 -23330=4,3,22,22,64 0=64 1=1 5=1 6=24576\nEltwise add4 2 1 conv21_splitncnn_0 conv24 add4 -23330=4,3,22,22,64 0=1\nSplit splitncnn_4 1 2 add4 add4_splitncnn_0 add4_splitncnn_1 -23330=8,3,22,22,64,3,22,22,64\nConvolution conv25 1 1 add4_splitncnn_1 conv25_relu17 -23330=4,3,22,22,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv26 1 1 conv25_relu17 conv26_relu18 -23330=4,3,22,22,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv27 1 1 conv26_relu18 conv27 -23330=4,3,22,22,64 0=64 1=1 5=1 6=24576\nEltwise add5 2 1 add4_splitncnn_0 conv27 add5 -23330=4,3,22,22,64 0=1\nSplit splitncnn_5 1 2 add5 add5_splitncnn_0 add5_splitncnn_1 -23330=8,3,22,22,64,3,22,22,64\nConvolution conv28 1 1 add5_splitncnn_1 conv28_relu19 -23330=4,3,22,22,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv29 1 1 conv28_relu19 conv29_relu20 -23330=4,3,22,22,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv30 1 1 conv29_relu20 conv30 -23330=4,3,22,22,64 0=64 1=1 5=1 6=24576\nEltwise add6 2 1 add5_splitncnn_0 conv30 add6 -23330=4,3,22,22,64 0=1\nConvolution conv31 1 1 add6 conv31_relu21 -23330=4,3,22,22,384 0=384 1=1 5=1 6=24576 9=1\nConvolutionDepthWise conv32 1 1 conv31_relu21 conv32_relu22 -23330=4,3,22,22,384 0=384 1=3 4=1 5=1 6=3456 7=384 9=1\nConvolution conv33 1 1 conv32_relu22 conv33 -23330=4,3,22,22,96 0=96 1=1 5=1 6=36864\nSplit splitncnn_6 1 2 conv33 conv33_splitncnn_0 conv33_splitncnn_1 -23330=8,3,22,22,96,3,22,22,96\nConvolution conv34 1 1 conv33_splitncnn_1 conv34_relu23 -23330=4,3,22,22,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise conv35 1 1 conv34_relu23 conv35_relu24 -23330=4,3,22,22,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution conv36 1 1 conv35_relu24 conv36 -23330=4,3,22,22,96 0=96 1=1 5=1 6=55296\nEltwise add7 2 1 conv33_splitncnn_0 conv36 add7 -23330=4,3,22,22,96 0=1\nSplit splitncnn_7 1 2 add7 add7_splitncnn_0 add7_splitncnn_1 -23330=8,3,22,22,96,3,22,22,96\nConvolution conv37 1 1 add7_splitncnn_1 conv37_relu25 -23330=4,3,22,22,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise conv38 1 1 conv37_relu25 conv38_relu26 -23330=4,3,22,22,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution conv39 1 1 conv38_relu26 conv39 -23330=4,3,22,22,96 0=96 1=1 5=1 6=55296\nEltwise add8 2 1 add7_splitncnn_0 conv39 add8 -23330=4,3,22,22,96 0=1\nConvolution conv40 1 1 add8 conv40_relu27 -23330=4,3,22,22,576 0=576 1=1 5=1 6=55296 9=1\nSplit splitncnn_8 1 2 conv40_relu27 conv40_relu27_splitncnn_0 conv40_relu27_splitncnn_1 -23330=8,3,22,22,576,3,22,22,576\nConvolutionDepthWise conv41 1 1 conv40_relu27_splitncnn_1 conv41_relu28 -23330=4,3,11,11,576 0=576 1=3 3=2 4=1 5=1 6=5184 7=576 9=1\nConvolution conv42 1 1 conv41_relu28 conv42 -23330=4,3,11,11,160 0=160 1=1 5=1 6=92160\nSplit splitncnn_9 1 2 conv42 conv42_splitncnn_0 conv42_splitncnn_1 -23330=8,3,11,11,160,3,11,11,160\nConvolution conv43 1 1 conv42_splitncnn_1 conv43_relu29 -23330=4,3,11,11,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv44 1 1 conv43_relu29 conv44_relu30 -23330=4,3,11,11,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv45 1 1 conv44_relu30 conv45 -23330=4,3,11,11,160 0=160 1=1 5=1 6=153600\nEltwise add9 2 1 conv42_splitncnn_0 conv45 add9 -23330=4,3,11,11,160 0=1\nSplit splitncnn_10 1 2 add9 add9_splitncnn_0 add9_splitncnn_1 -23330=8,3,11,11,160,3,11,11,160\nConvolution conv46 1 1 add9_splitncnn_1 conv46_relu31 -23330=4,3,11,11,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv47 1 1 conv46_relu31 conv47_relu32 -23330=4,3,11,11,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv48 1 1 conv47_relu32 conv48 -23330=4,3,11,11,160 0=160 1=1 5=1 6=153600\nEltwise add10 2 1 add9_splitncnn_0 conv48 add10 -23330=4,3,11,11,160 0=1\nConvolution conv49 1 1 add10 conv49_relu33 -23330=4,3,11,11,960 0=960 1=1 5=1 6=153600 9=1\nConvolutionDepthWise conv50 1 1 conv49_relu33 conv50_relu34 -23330=4,3,11,11,960 0=960 1=3 4=1 5=1 6=8640 7=960 9=1\nConvolution conv51 1 1 conv50_relu34 conv51 -23330=4,3,11,11,320 0=320 1=1 5=1 6=307200\nConvolution conv52 1 1 conv51 conv52_relu35 -23330=4,3,11,11,1280 0=1280 1=1 5=1 6=409600 9=1\nConvolutionDepthWise yolo/conv1/dw 1 1 conv52_relu35 yolo/conv1/dw_yolo/conv1/dw/relu -23330=4,3,11,11,1280 0=1280 1=3 4=1 5=1 6=11520 7=1280 9=1\nConvolution yolo/conv1 1 1 yolo/conv1/dw_yolo/conv1/dw/relu yolo/conv1_yolo/conv1/relu -23330=4,3,11,11,576 0=576 1=1 5=1 6=737280 9=1\nSplit splitncnn_11 1 2 yolo/conv1_yolo/conv1/relu yolo/conv1_yolo/conv1/relu_splitncnn_0 yolo/conv1_yolo/conv1/relu_splitncnn_1 -23330=8,3,11,11,576,3,11,11,576\nDeconvolutionDepthWise upsample 1 1 yolo/conv1_yolo/conv1/relu_splitncnn_1 upsample -23330=4,3,21,21,576 0=576 1=1 3=2 6=576 7=576\nPooling maxpool 1 1 upsample maxpool -23330=4,3,22,22,576 1=2 3=1\nConvolutionDepthWise yolo/conv2/dw 1 1 conv40_relu27_splitncnn_0 yolo/conv2/dw_yolo/conv2/dw/relu -23330=4,3,22,22,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution yolo/conv2 1 1 yolo/conv2/dw_yolo/conv2/dw/relu yolo/conv2_yolo/conv2/relu -23330=4,3,22,22,576 0=576 1=1 5=1 6=331776 9=1\nEltwise yolo/conv2/sum 2 1 maxpool yolo/conv2_yolo/conv2/relu yolo/conv2/sum -23330=4,3,22,22,576 0=1\nConvolutionDepthWise yolo/conv3/dw 1 1 yolo/conv2/sum yolo/conv3/dw_yolo/conv3/dw/relu -23330=4,3,22,22,576 0=576 1=3 4=1 5=1 6=5184 7=576 9=1\nConvolution yolo/conv3 1 1 yolo/conv3/dw_yolo/conv3/dw/relu yolo/conv3_yolo/conv3/relu -23330=4,3,22,22,576 0=576 1=1 5=1 6=331776 9=1\nConvolution yolo/conv4 1 1 yolo/conv1_yolo/conv1/relu_splitncnn_0 yolo/conv4 -23330=4,3,11,11,75 0=75 1=1 5=1 6=43200\nConvolution yolo/conv5 1 1 yolo/conv3_yolo/conv3/relu yolo/conv5 -23330=4,3,22,22,75 0=75 1=1 5=1 6=43200\nYolov3DetectionOutput detection_out 2 1 yolo/conv4 yolo/conv5 output 1=3 2=3.000000e-01 -23304=12,2.000000e+01,3.700000e+01,4.900000e+01,9.400000e+01,7.300000e+01,2.010000e+02,1.430000e+02,2.650000e+02,1.530000e+02,1.210000e+02,2.800000e+02,2.790000e+02 -23305=6,1077936128,1082130432,1084227584,0,1065353216,1073741824 -23306=2,3.200000e+01,1.600000e+01\n" }, { "path": "benchmark/nanodet_m.param", "content": "7767517\n179 204\nInput input.1 0 1 input.1 -23330=4,3,320,320,3 0=320 1=320 2=3\nConvolution Conv_0 1 1 input.1 424 -23330=4,3,160,160,24 0=24 1=3 3=2 4=1 5=1 6=648 9=2 -23310=1,1.000000e-01\nPooling MaxPool_2 1 1 424 425 -23330=4,3,80,80,24 1=3 2=2 3=1 5=1\nSplit splitncnn_0 1 2 425 425_splitncnn_0 425_splitncnn_1 -23330=8,3,80,80,24,3,80,80,24\nConvolutionDepthWise Conv_3 1 1 425_splitncnn_1 943 -23330=4,3,40,40,24 0=24 1=3 3=2 4=1 5=1 6=216 7=24\nConvolution Conv_4 1 1 943 430 -23330=4,3,40,40,58 0=58 1=1 5=1 6=1392 9=2 -23310=1,1.000000e-01\nConvolution Conv_6 1 1 425_splitncnn_0 433 -23330=4,3,80,80,58 0=58 1=1 5=1 6=1392 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_8 1 1 433 952 -23330=4,3,40,40,58 0=58 1=3 3=2 4=1 5=1 6=522 7=58\nConvolution Conv_9 1 1 952 438 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConcat Concat_11 2 1 430 438 439 -23330=4,3,40,40,116\nShuffleChannel Reshape_16 1 1 439 444 -23330=4,3,40,40,116 0=2\nSplit splitncnn_1 1 2 444 444_splitncnn_0 444_splitncnn_1 -23330=8,3,40,40,116,3,40,40,116\nCrop Slice_27 1 1 444_splitncnn_1 455 -23330=4,3,40,40,58 -23309=1,0 -23310=1,58 -23311=1,0\nCrop Slice_30 1 1 444_splitncnn_0 458 -23330=4,3,40,40,58 -23309=1,58 -23310=1,116 -23311=1,0\nConvolution Conv_31 1 1 458 461 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_33 1 1 461 961 -23330=4,3,40,40,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution Conv_34 1 1 961 466 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConcat Concat_36 2 1 455 466 467 -23330=4,3,40,40,116\nShuffleChannel Reshape_41 1 1 467 472 -23330=4,3,40,40,116 0=2\nSplit splitncnn_2 1 2 472 472_splitncnn_0 472_splitncnn_1 -23330=8,3,40,40,116,3,40,40,116\nCrop Slice_52 1 1 472_splitncnn_1 483 -23330=4,3,40,40,58 -23309=1,0 -23310=1,58 -23311=1,0\nCrop Slice_55 1 1 472_splitncnn_0 486 -23330=4,3,40,40,58 -23309=1,58 -23310=1,116 -23311=1,0\nConvolution Conv_56 1 1 486 489 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_58 1 1 489 970 -23330=4,3,40,40,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution Conv_59 1 1 970 494 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConcat Concat_61 2 1 483 494 495 -23330=4,3,40,40,116\nShuffleChannel Reshape_66 1 1 495 500 -23330=4,3,40,40,116 0=2\nSplit splitncnn_3 1 2 500 500_splitncnn_0 500_splitncnn_1 -23330=8,3,40,40,116,3,40,40,116\nCrop Slice_77 1 1 500_splitncnn_1 511 -23330=4,3,40,40,58 -23309=1,0 -23310=1,58 -23311=1,0\nCrop Slice_80 1 1 500_splitncnn_0 514 -23330=4,3,40,40,58 -23309=1,58 -23310=1,116 -23311=1,0\nConvolution Conv_81 1 1 514 517 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_83 1 1 517 979 -23330=4,3,40,40,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution Conv_84 1 1 979 522 -23330=4,3,40,40,58 0=58 1=1 5=1 6=3364 9=2 -23310=1,1.000000e-01\nConcat Concat_86 2 1 511 522 523 -23330=4,3,40,40,116\nShuffleChannel Reshape_91 1 1 523 528 -23330=4,3,40,40,116 0=2\nSplit splitncnn_4 1 3 528 528_splitncnn_0 528_splitncnn_1 528_splitncnn_2 -23330=12,3,40,40,116,3,40,40,116,3,40,40,116\nConvolutionDepthWise Conv_92 1 1 528_splitncnn_2 985 -23330=4,3,20,20,116 0=116 1=3 3=2 4=1 5=1 6=1044 7=116\nConvolution Conv_93 1 1 985 533 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolution Conv_95 1 1 528_splitncnn_1 536 -23330=4,3,40,40,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_97 1 1 536 994 -23330=4,3,20,20,116 0=116 1=3 3=2 4=1 5=1 6=1044 7=116\nConvolution Conv_98 1 1 994 541 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_100 2 1 533 541 542 -23330=4,3,20,20,232\nShuffleChannel Reshape_105 1 1 542 547 -23330=4,3,20,20,232 0=2\nSplit splitncnn_5 1 2 547 547_splitncnn_0 547_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_116 1 1 547_splitncnn_1 558 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_119 1 1 547_splitncnn_0 561 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_120 1 1 561 564 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_122 1 1 564 1003 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_123 1 1 1003 569 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_125 2 1 558 569 570 -23330=4,3,20,20,232\nShuffleChannel Reshape_130 1 1 570 575 -23330=4,3,20,20,232 0=2\nSplit splitncnn_6 1 2 575 575_splitncnn_0 575_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_141 1 1 575_splitncnn_1 586 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_144 1 1 575_splitncnn_0 589 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_145 1 1 589 592 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_147 1 1 592 1012 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_148 1 1 1012 597 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_150 2 1 586 597 598 -23330=4,3,20,20,232\nShuffleChannel Reshape_155 1 1 598 603 -23330=4,3,20,20,232 0=2\nSplit splitncnn_7 1 2 603 603_splitncnn_0 603_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_166 1 1 603_splitncnn_1 614 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_169 1 1 603_splitncnn_0 617 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_170 1 1 617 620 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_172 1 1 620 1021 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_173 1 1 1021 625 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_175 2 1 614 625 626 -23330=4,3,20,20,232\nShuffleChannel Reshape_180 1 1 626 631 -23330=4,3,20,20,232 0=2\nSplit splitncnn_8 1 2 631 631_splitncnn_0 631_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_191 1 1 631_splitncnn_1 642 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_194 1 1 631_splitncnn_0 645 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_195 1 1 645 648 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_197 1 1 648 1030 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_198 1 1 1030 653 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_200 2 1 642 653 654 -23330=4,3,20,20,232\nShuffleChannel Reshape_205 1 1 654 659 -23330=4,3,20,20,232 0=2\nSplit splitncnn_9 1 2 659 659_splitncnn_0 659_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_216 1 1 659_splitncnn_1 670 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_219 1 1 659_splitncnn_0 673 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_220 1 1 673 676 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_222 1 1 676 1039 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_223 1 1 1039 681 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_225 2 1 670 681 682 -23330=4,3,20,20,232\nShuffleChannel Reshape_230 1 1 682 687 -23330=4,3,20,20,232 0=2\nSplit splitncnn_10 1 2 687 687_splitncnn_0 687_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_241 1 1 687_splitncnn_1 698 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_244 1 1 687_splitncnn_0 701 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_245 1 1 701 704 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_247 1 1 704 1048 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_248 1 1 1048 709 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_250 2 1 698 709 710 -23330=4,3,20,20,232\nShuffleChannel Reshape_255 1 1 710 715 -23330=4,3,20,20,232 0=2\nSplit splitncnn_11 1 2 715 715_splitncnn_0 715_splitncnn_1 -23330=8,3,20,20,232,3,20,20,232\nCrop Slice_266 1 1 715_splitncnn_1 726 -23330=4,3,20,20,116 -23309=1,0 -23310=1,116 -23311=1,0\nCrop Slice_269 1 1 715_splitncnn_0 729 -23330=4,3,20,20,116 -23309=1,116 -23310=1,232 -23311=1,0\nConvolution Conv_270 1 1 729 732 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_272 1 1 732 1057 -23330=4,3,20,20,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution Conv_273 1 1 1057 737 -23330=4,3,20,20,116 0=116 1=1 5=1 6=13456 9=2 -23310=1,1.000000e-01\nConcat Concat_275 2 1 726 737 738 -23330=4,3,20,20,232\nShuffleChannel Reshape_280 1 1 738 743 -23330=4,3,20,20,232 0=2\nSplit splitncnn_12 1 3 743 743_splitncnn_0 743_splitncnn_1 743_splitncnn_2 -23330=12,3,20,20,232,3,20,20,232,3,20,20,232\nConvolutionDepthWise Conv_281 1 1 743_splitncnn_2 1063 -23330=4,3,10,10,232 0=232 1=3 3=2 4=1 5=1 6=2088 7=232\nConvolution Conv_282 1 1 1063 748 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConvolution Conv_284 1 1 743_splitncnn_1 751 -23330=4,3,20,20,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_286 1 1 751 1072 -23330=4,3,10,10,232 0=232 1=3 3=2 4=1 5=1 6=2088 7=232\nConvolution Conv_287 1 1 1072 756 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConcat Concat_289 2 1 748 756 757 -23330=4,3,10,10,464\nShuffleChannel Reshape_294 1 1 757 762 -23330=4,3,10,10,464 0=2\nSplit splitncnn_13 1 2 762 762_splitncnn_0 762_splitncnn_1 -23330=8,3,10,10,464,3,10,10,464\nCrop Slice_305 1 1 762_splitncnn_1 773 -23330=4,3,10,10,232 -23309=1,0 -23310=1,232 -23311=1,0\nCrop Slice_308 1 1 762_splitncnn_0 776 -23330=4,3,10,10,232 -23309=1,232 -23310=1,464 -23311=1,0\nConvolution Conv_309 1 1 776 779 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_311 1 1 779 1081 -23330=4,3,10,10,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution Conv_312 1 1 1081 784 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConcat Concat_314 2 1 773 784 785 -23330=4,3,10,10,464\nShuffleChannel Reshape_319 1 1 785 790 -23330=4,3,10,10,464 0=2\nSplit splitncnn_14 1 2 790 790_splitncnn_0 790_splitncnn_1 -23330=8,3,10,10,464,3,10,10,464\nCrop Slice_330 1 1 790_splitncnn_1 801 -23330=4,3,10,10,232 -23309=1,0 -23310=1,232 -23311=1,0\nCrop Slice_333 1 1 790_splitncnn_0 804 -23330=4,3,10,10,232 -23309=1,232 -23310=1,464 -23311=1,0\nConvolution Conv_334 1 1 804 807 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_336 1 1 807 1090 -23330=4,3,10,10,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution Conv_337 1 1 1090 812 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConcat Concat_339 2 1 801 812 813 -23330=4,3,10,10,464\nShuffleChannel Reshape_344 1 1 813 818 -23330=4,3,10,10,464 0=2\nSplit splitncnn_15 1 2 818 818_splitncnn_0 818_splitncnn_1 -23330=8,3,10,10,464,3,10,10,464\nCrop Slice_355 1 1 818_splitncnn_1 829 -23330=4,3,10,10,232 -23309=1,0 -23310=1,232 -23311=1,0\nCrop Slice_358 1 1 818_splitncnn_0 832 -23330=4,3,10,10,232 -23309=1,232 -23310=1,464 -23311=1,0\nConvolution Conv_359 1 1 832 835 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_361 1 1 835 1099 -23330=4,3,10,10,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution Conv_362 1 1 1099 840 -23330=4,3,10,10,232 0=232 1=1 5=1 6=53824 9=2 -23310=1,1.000000e-01\nConcat Concat_364 2 1 829 840 841 -23330=4,3,10,10,464\nShuffleChannel Reshape_369 1 1 841 846 -23330=4,3,10,10,464 0=2\nConvolution Conv_370 1 1 528_splitncnn_0 847 -23330=4,3,40,40,96 0=96 1=1 5=1 6=11136\nConvolution Conv_371 1 1 743_splitncnn_0 848 -23330=4,3,20,20,96 0=96 1=1 5=1 6=22272\nConvolution Conv_372 1 1 846 849 -23330=4,3,10,10,96 0=96 1=1 5=1 6=44544\nSplit splitncnn_16 1 2 849 849_splitncnn_0 849_splitncnn_1 -23330=8,3,10,10,96,3,10,10,96\nInterp Resize_374 1 1 849_splitncnn_1 854 -23330=4,3,20,20,96 0=2 1=2.000000e+00 2=2.000000e+00\nBinaryOp Add_375 2 1 848 854 855 -23330=4,3,20,20,96\nSplit splitncnn_17 1 2 855 855_splitncnn_0 855_splitncnn_1 -23330=8,3,20,20,96,3,20,20,96\nInterp Resize_377 1 1 855_splitncnn_1 860 -23330=4,3,40,40,96 0=2 1=2.000000e+00 2=2.000000e+00\nBinaryOp Add_378 2 1 847 860 861 -23330=4,3,40,40,96\nSplit splitncnn_18 1 2 861 861_splitncnn_0 861_splitncnn_1 -23330=8,3,40,40,96,3,40,40,96\nInterp Resize_380 1 1 861_splitncnn_1 866 -23330=4,3,20,20,96 0=2 1=5.000000e-01 2=5.000000e-01\nBinaryOp Add_381 2 1 855_splitncnn_0 866 867 -23330=4,3,20,20,96\nSplit splitncnn_19 1 2 867 867_splitncnn_0 867_splitncnn_1 -23330=8,3,20,20,96,3,20,20,96\nInterp Resize_383 1 1 867_splitncnn_1 872 -23330=4,3,10,10,96 0=2 1=5.000000e-01 2=5.000000e-01\nBinaryOp Add_384 2 1 849_splitncnn_0 872 873 -23330=4,3,10,10,96\nConvolutionDepthWise Conv_385 1 1 861_splitncnn_0 876 -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_387 1 1 876 879 -23330=4,3,40,40,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_389 1 1 879 882 -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_391 1 1 882 885 -23330=4,3,40,40,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolution Conv_393 1 1 885 886 -23330=4,3,40,40,112 0=112 1=1 5=1 6=10752\nSlice Split_394 1 2 886 887 888 -23330=8,3,40,40,80,3,40,40,32 -23300=2,80,-233\nSigmoid Sigmoid_395 1 1 887 889 -23330=4,3,40,40,80\nReshape Reshape_397 1 1 889 891 -23330=4,2,1600,80,1 0=-1 1=80\nPermute Transpose_398 1 1 891 cls_pred_stride_8 -23330=4,2,80,1600,1 0=1\nReshape Reshape_400 1 1 888 894 -23330=4,2,1600,32,1 0=-1 1=32\nPermute Transpose_401 1 1 894 dis_pred_stride_8 -23330=4,2,32,1600,1 0=1\nConvolutionDepthWise Conv_402 1 1 867_splitncnn_0 898 -23330=4,3,20,20,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_404 1 1 898 901 -23330=4,3,20,20,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_406 1 1 901 904 -23330=4,3,20,20,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_408 1 1 904 907 -23330=4,3,20,20,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolution Conv_410 1 1 907 908 -23330=4,3,20,20,112 0=112 1=1 5=1 6=10752\nSlice Split_411 1 2 908 909 910 -23330=8,3,20,20,80,3,20,20,32 -23300=2,80,-233\nSigmoid Sigmoid_412 1 1 909 911 -23330=4,3,20,20,80\nReshape Reshape_414 1 1 911 913 -23330=4,2,400,80,1 0=-1 1=80\nPermute Transpose_415 1 1 913 cls_pred_stride_16 -23330=4,2,80,400,1 0=1\nReshape Reshape_417 1 1 910 916 -23330=4,2,400,32,1 0=-1 1=32\nPermute Transpose_418 1 1 916 dis_pred_stride_16 -23330=4,2,32,400,1 0=1\nConvolutionDepthWise Conv_419 1 1 873 920 -23330=4,3,10,10,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_421 1 1 920 923 -23330=4,3,10,10,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise Conv_423 1 1 923 926 -23330=4,3,10,10,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution Conv_425 1 1 926 929 -23330=4,3,10,10,96 0=96 1=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConvolution Conv_427 1 1 929 930 -23330=4,3,10,10,112 0=112 1=1 5=1 6=10752\nSlice Split_428 1 2 930 931 932 -23330=8,3,10,10,80,3,10,10,32 -23300=2,80,-233\nSigmoid Sigmoid_429 1 1 931 933 -23330=4,3,10,10,80\nReshape Reshape_431 1 1 933 935 -23330=4,2,100,80,1 0=-1 1=80\nPermute Transpose_432 1 1 935 cls_pred_stride_32 -23330=4,2,80,100,1 0=1\nReshape Reshape_434 1 1 932 938 -23330=4,2,100,32,1 0=-1 1=32\nPermute Transpose_435 1 1 938 dis_pred_stride_32 -23330=4,2,32,100,1 0=1\nNoop Output 6 1 cls_pred_stride_8 cls_pred_stride_16 cls_pred_stride_32 dis_pred_stride_8 dis_pred_stride_16 dis_pred_stride_32 output\n" }, { "path": "benchmark/proxylessnasnet.param", "content": "7767517\n91 104\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution first-3x3-conv 1 1 data first-3x3-conv_relu -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nConvolutionDepthWise A0_dw 1 1 first-3x3-conv_relu A0_dw_relu -23330=4,3,112,112,32 0=32 1=3 4=1 5=1 6=288 7=32 9=1\nConvolution A0_linear 1 1 A0_dw_relu A0_linear_bn -23330=4,3,112,112,32 0=32 1=1 5=1 6=1024\nConvolution B0_expand 1 1 A0_linear_bn B0_expand_relu -23330=4,3,112,112,48 0=48 1=1 5=1 6=1536 9=1\nConvolutionDepthWise B0_dw 1 1 B0_expand_relu B0_dw_relu -23330=4,3,56,56,48 0=48 1=5 3=2 4=2 5=1 6=1200 7=48 9=1\nConvolution B0_linear 1 1 B0_dw_relu B0_linear_bn -23330=4,3,56,56,32 0=32 1=1 5=1 6=1536\nSplit splitncnn_0 1 2 B0_linear_bn B0_linear_bn_splitncnn_0 B0_linear_bn_splitncnn_1 -23330=8,3,56,56,32,3,56,56,32\nConvolution B1_expand 1 1 B0_linear_bn_splitncnn_1 B1_expand_relu -23330=4,3,56,56,96 0=96 1=1 5=1 6=3072 9=1\nConvolutionDepthWise B1_dw 1 1 B1_expand_relu B1_dw_relu -23330=4,3,56,56,96 0=96 1=3 4=1 5=1 6=864 7=96 9=1\nConvolution B1_linear 1 1 B1_dw_relu B1_linear_bn -23330=4,3,56,56,32 0=32 1=1 5=1 6=3072\nBinaryOp unknownncnn_0 2 1 B0_linear_bn_splitncnn_0 B1_linear_bn unknownncnn_0 -23330=4,3,56,56,32\nConvolution C0_expand 1 1 unknownncnn_0 C0_expand_relu -23330=4,3,56,56,96 0=96 1=1 5=1 6=3072 9=1\nConvolutionDepthWise C0_dw 1 1 C0_expand_relu C0_dw_relu -23330=4,3,28,28,96 0=96 1=7 3=2 4=3 5=1 6=4704 7=96 9=1\nConvolution C0_linear 1 1 C0_dw_relu C0_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=3840\nSplit splitncnn_1 1 2 C0_linear_bn C0_linear_bn_splitncnn_0 C0_linear_bn_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution C1_expand 1 1 C0_linear_bn_splitncnn_1 C1_expand_relu -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise C1_dw 1 1 C1_expand_relu C1_dw_relu -23330=4,3,28,28,120 0=120 1=3 4=1 5=1 6=1080 7=120 9=1\nConvolution C1_linear 1 1 C1_dw_relu C1_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp unknownncnn_1 2 1 C0_linear_bn_splitncnn_0 C1_linear_bn unknownncnn_1 -23330=4,3,28,28,40\nSplit splitncnn_2 1 2 unknownncnn_1 unknownncnn_1_splitncnn_0 unknownncnn_1_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution C2_expand 1 1 unknownncnn_1_splitncnn_1 C2_expand_relu -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise C2_dw 1 1 C2_expand_relu C2_dw_relu -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=1\nConvolution C2_linear 1 1 C2_dw_relu C2_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp unknownncnn_2 2 1 unknownncnn_1_splitncnn_0 C2_linear_bn unknownncnn_2 -23330=4,3,28,28,40\nSplit splitncnn_3 1 2 unknownncnn_2 unknownncnn_2_splitncnn_0 unknownncnn_2_splitncnn_1 -23330=8,3,28,28,40,3,28,28,40\nConvolution C3_expand 1 1 unknownncnn_2_splitncnn_1 C3_expand_relu -23330=4,3,28,28,120 0=120 1=1 5=1 6=4800 9=1\nConvolutionDepthWise C3_dw 1 1 C3_expand_relu C3_dw_relu -23330=4,3,28,28,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=1\nConvolution C3_linear 1 1 C3_dw_relu C3_linear_bn -23330=4,3,28,28,40 0=40 1=1 5=1 6=4800\nBinaryOp unknownncnn_3 2 1 unknownncnn_2_splitncnn_0 C3_linear_bn unknownncnn_3 -23330=4,3,28,28,40\nConvolution D0_expand 1 1 unknownncnn_3 D0_expand_relu -23330=4,3,28,28,240 0=240 1=1 5=1 6=9600 9=1\nConvolutionDepthWise D0_dw 1 1 D0_expand_relu D0_dw_relu -23330=4,3,14,14,240 0=240 1=7 3=2 4=3 5=1 6=11760 7=240 9=1\nConvolution D0_linear 1 1 D0_dw_relu D0_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nSplit splitncnn_4 1 2 D0_linear_bn D0_linear_bn_splitncnn_0 D0_linear_bn_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution D1_expand 1 1 D0_linear_bn_splitncnn_1 D1_expand_relu -23330=4,3,14,14,240 0=240 1=1 5=1 6=19200 9=1\nConvolutionDepthWise D1_dw 1 1 D1_expand_relu D1_dw_relu -23330=4,3,14,14,240 0=240 1=5 4=2 5=1 6=6000 7=240 9=1\nConvolution D1_linear 1 1 D1_dw_relu D1_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nBinaryOp unknownncnn_4 2 1 D0_linear_bn_splitncnn_0 D1_linear_bn unknownncnn_4 -23330=4,3,14,14,80\nSplit splitncnn_5 1 2 unknownncnn_4 unknownncnn_4_splitncnn_0 unknownncnn_4_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution D2_expand 1 1 unknownncnn_4_splitncnn_1 D2_expand_relu -23330=4,3,14,14,240 0=240 1=1 5=1 6=19200 9=1\nConvolutionDepthWise D2_dw 1 1 D2_expand_relu D2_dw_relu -23330=4,3,14,14,240 0=240 1=5 4=2 5=1 6=6000 7=240 9=1\nConvolution D2_linear 1 1 D2_dw_relu D2_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nBinaryOp unknownncnn_5 2 1 unknownncnn_4_splitncnn_0 D2_linear_bn unknownncnn_5 -23330=4,3,14,14,80\nSplit splitncnn_6 1 2 unknownncnn_5 unknownncnn_5_splitncnn_0 unknownncnn_5_splitncnn_1 -23330=8,3,14,14,80,3,14,14,80\nConvolution D3_expand 1 1 unknownncnn_5_splitncnn_1 D3_expand_relu -23330=4,3,14,14,240 0=240 1=1 5=1 6=19200 9=1\nConvolutionDepthWise D3_dw 1 1 D3_expand_relu D3_dw_relu -23330=4,3,14,14,240 0=240 1=5 4=2 5=1 6=6000 7=240 9=1\nConvolution D3_linear 1 1 D3_dw_relu D3_linear_bn -23330=4,3,14,14,80 0=80 1=1 5=1 6=19200\nBinaryOp unknownncnn_6 2 1 unknownncnn_5_splitncnn_0 D3_linear_bn unknownncnn_6 -23330=4,3,14,14,80\nConvolution E0_expand 1 1 unknownncnn_6 E0_expand_relu -23330=4,3,14,14,480 0=480 1=1 5=1 6=38400 9=1\nConvolutionDepthWise E0_dw 1 1 E0_expand_relu E0_dw_relu -23330=4,3,14,14,480 0=480 1=5 4=2 5=1 6=12000 7=480 9=1\nConvolution E0_linear 1 1 E0_dw_relu E0_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=46080\nSplit splitncnn_7 1 2 E0_linear_bn E0_linear_bn_splitncnn_0 E0_linear_bn_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution E1_expand 1 1 E0_linear_bn_splitncnn_1 E1_expand_relu -23330=4,3,14,14,288 0=288 1=1 5=1 6=27648 9=1\nConvolutionDepthWise E1_dw 1 1 E1_expand_relu E1_dw_relu -23330=4,3,14,14,288 0=288 1=5 4=2 5=1 6=7200 7=288 9=1\nConvolution E1_linear 1 1 E1_dw_relu E1_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=27648\nBinaryOp unknownncnn_7 2 1 E0_linear_bn_splitncnn_0 E1_linear_bn unknownncnn_7 -23330=4,3,14,14,96\nSplit splitncnn_8 1 2 unknownncnn_7 unknownncnn_7_splitncnn_0 unknownncnn_7_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution E2_expand 1 1 unknownncnn_7_splitncnn_1 E2_expand_relu -23330=4,3,14,14,288 0=288 1=1 5=1 6=27648 9=1\nConvolutionDepthWise E2_dw 1 1 E2_expand_relu E2_dw_relu -23330=4,3,14,14,288 0=288 1=5 4=2 5=1 6=7200 7=288 9=1\nConvolution E2_linear 1 1 E2_dw_relu E2_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=27648\nBinaryOp unknownncnn_8 2 1 unknownncnn_7_splitncnn_0 E2_linear_bn unknownncnn_8 -23330=4,3,14,14,96\nSplit splitncnn_9 1 2 unknownncnn_8 unknownncnn_8_splitncnn_0 unknownncnn_8_splitncnn_1 -23330=8,3,14,14,96,3,14,14,96\nConvolution E3_expand 1 1 unknownncnn_8_splitncnn_1 E3_expand_relu -23330=4,3,14,14,288 0=288 1=1 5=1 6=27648 9=1\nConvolutionDepthWise E3_dw 1 1 E3_expand_relu E3_dw_relu -23330=4,3,14,14,288 0=288 1=5 4=2 5=1 6=7200 7=288 9=1\nConvolution E3_linear 1 1 E3_dw_relu E3_linear_bn -23330=4,3,14,14,96 0=96 1=1 5=1 6=27648\nBinaryOp unknownncnn_9 2 1 unknownncnn_8_splitncnn_0 E3_linear_bn unknownncnn_9 -23330=4,3,14,14,96\nConvolution F0_expand 1 1 unknownncnn_9 F0_expand_relu -23330=4,3,14,14,576 0=576 1=1 5=1 6=55296 9=1\nConvolutionDepthWise F0_dw 1 1 F0_expand_relu F0_dw_relu -23330=4,3,7,7,576 0=576 1=7 3=2 4=3 5=1 6=28224 7=576 9=1\nConvolution F0_linear 1 1 F0_dw_relu F0_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=110592\nSplit splitncnn_10 1 2 F0_linear_bn F0_linear_bn_splitncnn_0 F0_linear_bn_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F1_expand 1 1 F0_linear_bn_splitncnn_1 F1_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise F1_dw 1 1 F1_expand_relu F1_dw_relu -23330=4,3,7,7,1152 0=1152 1=7 4=3 5=1 6=56448 7=1152 9=1\nConvolution F1_linear 1 1 F1_dw_relu F1_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=221184\nBinaryOp unknownncnn_10 2 1 F0_linear_bn_splitncnn_0 F1_linear_bn unknownncnn_10 -23330=4,3,7,7,192\nSplit splitncnn_11 1 2 unknownncnn_10 unknownncnn_10_splitncnn_0 unknownncnn_10_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F2_expand 1 1 unknownncnn_10_splitncnn_1 F2_expand_relu -23330=4,3,7,7,576 0=576 1=1 5=1 6=110592 9=1\nConvolutionDepthWise F2_dw 1 1 F2_expand_relu F2_dw_relu -23330=4,3,7,7,576 0=576 1=7 4=3 5=1 6=28224 7=576 9=1\nConvolution F2_linear 1 1 F2_dw_relu F2_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=110592\nBinaryOp unknownncnn_11 2 1 unknownncnn_10_splitncnn_0 F2_linear_bn unknownncnn_11 -23330=4,3,7,7,192\nSplit splitncnn_12 1 2 unknownncnn_11 unknownncnn_11_splitncnn_0 unknownncnn_11_splitncnn_1 -23330=8,3,7,7,192,3,7,7,192\nConvolution F3_expand 1 1 unknownncnn_11_splitncnn_1 F3_expand_relu -23330=4,3,7,7,576 0=576 1=1 5=1 6=110592 9=1\nConvolutionDepthWise F3_dw 1 1 F3_expand_relu F3_dw_relu -23330=4,3,7,7,576 0=576 1=7 4=3 5=1 6=28224 7=576 9=1\nConvolution F3_linear 1 1 F3_dw_relu F3_linear_bn -23330=4,3,7,7,192 0=192 1=1 5=1 6=110592\nBinaryOp unknownncnn_12 2 1 unknownncnn_11_splitncnn_0 F3_linear_bn unknownncnn_12 -23330=4,3,7,7,192\nConvolution G0_expand 1 1 unknownncnn_12 G0_expand_relu -23330=4,3,7,7,1152 0=1152 1=1 5=1 6=221184 9=1\nConvolutionDepthWise G0_dw 1 1 G0_expand_relu G0_dw_relu -23330=4,3,7,7,1152 0=1152 1=7 4=3 5=1 6=56448 7=1152 9=1\nConvolution G0_linear 1 1 G0_dw_relu G0_linear_bn -23330=4,3,7,7,320 0=320 1=1 5=1 6=368640\nConvolution last-1x1-conv 1 1 G0_linear_bn last-1x1-conv_relu -23330=4,3,7,7,1280 0=1280 1=1 5=1 6=409600 9=1\nPooling avgpool 1 1 last-1x1-conv_relu flatten -23330=4,1,1280,1,1 0=1 1=7 4=1 5=1\nInnerProduct fc 1 1 flatten fc -23330=4,1,1000,1,1 0=1000 1=1 2=1280000\nSoftmax prob 1 1 fc output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/regnety_400m.param", "content": "7767517\n185 217\nInput input.1 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution Conv_0 1 1 data 387 -23330=4,3,112,112,32 0=32 1=3 3=2 4=1 5=1 6=864 9=1\nSplit splitncnn_0 1 2 387 387_splitncnn_0 387_splitncnn_1 -23330=8,3,112,112,32,3,112,112,32\nConvolution Conv_3 1 1 387_splitncnn_1 389 -23330=4,3,56,56,48 0=48 1=1 3=2 5=1 6=1536\nConvolution Conv_5 1 1 387_splitncnn_0 392 -23330=4,3,112,112,48 0=48 1=1 5=1 6=1536 9=1\nConvolutionDepthWise Conv_8 1 1 392 395 -23330=4,3,56,56,48 0=48 1=3 3=2 4=1 5=1 6=3456 7=6 9=1\nSplit splitncnn_1 1 2 395 395_splitncnn_0 395_splitncnn_1 -23330=8,3,56,56,48,3,56,56,48\nPooling GlobalAveragePool_11 1 1 395_splitncnn_1 396 -23330=4,1,48,1,1 0=1 4=1\nInnerProduct Conv_12 1 1 396 398 -23330=4,1,8,1,1 0=8 1=1 2=384 9=1\nInnerProduct Conv_14 1 1 398 400 -23330=4,1,48,1,1 0=48 1=1 2=384 9=4\nBinaryOp Mul_16 2 1 395_splitncnn_0 400 401 -23330=4,3,56,56,48 0=2\nConvolution Conv_17 1 1 401 403 -23330=4,3,56,56,48 0=48 1=1 5=1 6=2304\nBinaryOp Add_19 2 1 389 403 404 -23330=4,3,56,56,48\nReLU Relu_20 1 1 404 405 -23330=4,3,56,56,48\nSplit splitncnn_2 1 2 405 405_splitncnn_0 405_splitncnn_1 -23330=8,3,56,56,48,3,56,56,48\nConvolution Conv_21 1 1 405_splitncnn_1 407 -23330=4,3,28,28,104 0=104 1=1 3=2 5=1 6=4992\nConvolution Conv_23 1 1 405_splitncnn_0 410 -23330=4,3,56,56,104 0=104 1=1 5=1 6=4992 9=1\nConvolutionDepthWise Conv_26 1 1 410 413 -23330=4,3,28,28,104 0=104 1=3 3=2 4=1 5=1 6=7488 7=13 9=1\nSplit splitncnn_3 1 2 413 413_splitncnn_0 413_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nPooling GlobalAveragePool_29 1 1 413_splitncnn_1 414 -23330=4,1,104,1,1 0=1 4=1\nInnerProduct Conv_30 1 1 414 416 -23330=4,1,12,1,1 0=12 1=1 2=1248 9=1\nInnerProduct Conv_32 1 1 416 418 -23330=4,1,104,1,1 0=104 1=1 2=1248 9=4\nBinaryOp Mul_34 2 1 413_splitncnn_0 418 419 -23330=4,3,28,28,104 0=2\nConvolution Conv_35 1 1 419 421 -23330=4,3,28,28,104 0=104 1=1 5=1 6=10816\nBinaryOp Add_37 2 1 407 421 422 -23330=4,3,28,28,104\nReLU Relu_38 1 1 422 423 -23330=4,3,28,28,104\nSplit splitncnn_4 1 2 423 423_splitncnn_0 423_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nConvolution Conv_39 1 1 423_splitncnn_1 426 -23330=4,3,28,28,104 0=104 1=1 5=1 6=10816 9=1\nConvolutionDepthWise Conv_42 1 1 426 429 -23330=4,3,28,28,104 0=104 1=3 4=1 5=1 6=7488 7=13 9=1\nSplit splitncnn_5 1 2 429 429_splitncnn_0 429_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nPooling GlobalAveragePool_45 1 1 429_splitncnn_1 430 -23330=4,1,104,1,1 0=1 4=1\nInnerProduct Conv_46 1 1 430 432 -23330=4,1,26,1,1 0=26 1=1 2=2704 9=1\nInnerProduct Conv_48 1 1 432 434 -23330=4,1,104,1,1 0=104 1=1 2=2704 9=4\nBinaryOp Mul_50 2 1 429_splitncnn_0 434 435 -23330=4,3,28,28,104 0=2\nConvolution Conv_51 1 1 435 437 -23330=4,3,28,28,104 0=104 1=1 5=1 6=10816\nBinaryOp Add_53 2 1 423_splitncnn_0 437 438 -23330=4,3,28,28,104\nReLU Relu_54 1 1 438 439 -23330=4,3,28,28,104\nSplit splitncnn_6 1 2 439 439_splitncnn_0 439_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nConvolution Conv_55 1 1 439_splitncnn_1 442 -23330=4,3,28,28,104 0=104 1=1 5=1 6=10816 9=1\nConvolutionDepthWise Conv_58 1 1 442 445 -23330=4,3,28,28,104 0=104 1=3 4=1 5=1 6=7488 7=13 9=1\nSplit splitncnn_7 1 2 445 445_splitncnn_0 445_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nPooling GlobalAveragePool_61 1 1 445_splitncnn_1 446 -23330=4,1,104,1,1 0=1 4=1\nInnerProduct Conv_62 1 1 446 448 -23330=4,1,26,1,1 0=26 1=1 2=2704 9=1\nInnerProduct Conv_64 1 1 448 450 -23330=4,1,104,1,1 0=104 1=1 2=2704 9=4\nBinaryOp Mul_66 2 1 445_splitncnn_0 450 451 -23330=4,3,28,28,104 0=2\nConvolution Conv_67 1 1 451 453 -23330=4,3,28,28,104 0=104 1=1 5=1 6=10816\nBinaryOp Add_69 2 1 439_splitncnn_0 453 454 -23330=4,3,28,28,104\nReLU Relu_70 1 1 454 455 -23330=4,3,28,28,104\nSplit splitncnn_8 1 2 455 455_splitncnn_0 455_splitncnn_1 -23330=8,3,28,28,104,3,28,28,104\nConvolution Conv_71 1 1 455_splitncnn_1 457 -23330=4,3,14,14,208 0=208 1=1 3=2 5=1 6=21632\nConvolution Conv_73 1 1 455_splitncnn_0 460 -23330=4,3,28,28,208 0=208 1=1 5=1 6=21632 9=1\nConvolutionDepthWise Conv_76 1 1 460 463 -23330=4,3,14,14,208 0=208 1=3 3=2 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_9 1 2 463 463_splitncnn_0 463_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_79 1 1 463_splitncnn_1 464 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_80 1 1 464 466 -23330=4,1,26,1,1 0=26 1=1 2=5408 9=1\nInnerProduct Conv_82 1 1 466 468 -23330=4,1,208,1,1 0=208 1=1 2=5408 9=4\nBinaryOp Mul_84 2 1 463_splitncnn_0 468 469 -23330=4,3,14,14,208 0=2\nConvolution Conv_85 1 1 469 471 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_87 2 1 457 471 472 -23330=4,3,14,14,208\nReLU Relu_88 1 1 472 473 -23330=4,3,14,14,208\nSplit splitncnn_10 1 2 473 473_splitncnn_0 473_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_89 1 1 473_splitncnn_1 476 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264 9=1\nConvolutionDepthWise Conv_92 1 1 476 479 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_11 1 2 479 479_splitncnn_0 479_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_95 1 1 479_splitncnn_1 480 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_96 1 1 480 482 -23330=4,1,52,1,1 0=52 1=1 2=10816 9=1\nInnerProduct Conv_98 1 1 482 484 -23330=4,1,208,1,1 0=208 1=1 2=10816 9=4\nBinaryOp Mul_100 2 1 479_splitncnn_0 484 485 -23330=4,3,14,14,208 0=2\nConvolution Conv_101 1 1 485 487 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_103 2 1 473_splitncnn_0 487 488 -23330=4,3,14,14,208\nReLU Relu_104 1 1 488 489 -23330=4,3,14,14,208\nSplit splitncnn_12 1 2 489 489_splitncnn_0 489_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_105 1 1 489_splitncnn_1 492 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264 9=1\nConvolutionDepthWise Conv_108 1 1 492 495 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_13 1 2 495 495_splitncnn_0 495_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_111 1 1 495_splitncnn_1 496 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_112 1 1 496 498 -23330=4,1,52,1,1 0=52 1=1 2=10816 9=1\nInnerProduct Conv_114 1 1 498 500 -23330=4,1,208,1,1 0=208 1=1 2=10816 9=4\nBinaryOp Mul_116 2 1 495_splitncnn_0 500 501 -23330=4,3,14,14,208 0=2\nConvolution Conv_117 1 1 501 503 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_119 2 1 489_splitncnn_0 503 504 -23330=4,3,14,14,208\nReLU Relu_120 1 1 504 505 -23330=4,3,14,14,208\nSplit splitncnn_14 1 2 505 505_splitncnn_0 505_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_121 1 1 505_splitncnn_1 508 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264 9=1\nConvolutionDepthWise Conv_124 1 1 508 511 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_15 1 2 511 511_splitncnn_0 511_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_127 1 1 511_splitncnn_1 512 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_128 1 1 512 514 -23330=4,1,52,1,1 0=52 1=1 2=10816 9=1\nInnerProduct Conv_130 1 1 514 516 -23330=4,1,208,1,1 0=208 1=1 2=10816 9=4\nBinaryOp Mul_132 2 1 511_splitncnn_0 516 517 -23330=4,3,14,14,208 0=2\nConvolution Conv_133 1 1 517 519 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_135 2 1 505_splitncnn_0 519 520 -23330=4,3,14,14,208\nReLU Relu_136 1 1 520 521 -23330=4,3,14,14,208\nSplit splitncnn_16 1 2 521 521_splitncnn_0 521_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_137 1 1 521_splitncnn_1 524 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264 9=1\nConvolutionDepthWise Conv_140 1 1 524 527 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_17 1 2 527 527_splitncnn_0 527_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_143 1 1 527_splitncnn_1 528 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_144 1 1 528 530 -23330=4,1,52,1,1 0=52 1=1 2=10816 9=1\nInnerProduct Conv_146 1 1 530 532 -23330=4,1,208,1,1 0=208 1=1 2=10816 9=4\nBinaryOp Mul_148 2 1 527_splitncnn_0 532 533 -23330=4,3,14,14,208 0=2\nConvolution Conv_149 1 1 533 535 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_151 2 1 521_splitncnn_0 535 536 -23330=4,3,14,14,208\nReLU Relu_152 1 1 536 537 -23330=4,3,14,14,208\nSplit splitncnn_18 1 2 537 537_splitncnn_0 537_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_153 1 1 537_splitncnn_1 540 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264 9=1\nConvolutionDepthWise Conv_156 1 1 540 543 -23330=4,3,14,14,208 0=208 1=3 4=1 5=1 6=14976 7=26 9=1\nSplit splitncnn_19 1 2 543 543_splitncnn_0 543_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nPooling GlobalAveragePool_159 1 1 543_splitncnn_1 544 -23330=4,1,208,1,1 0=1 4=1\nInnerProduct Conv_160 1 1 544 546 -23330=4,1,52,1,1 0=52 1=1 2=10816 9=1\nInnerProduct Conv_162 1 1 546 548 -23330=4,1,208,1,1 0=208 1=1 2=10816 9=4\nBinaryOp Mul_164 2 1 543_splitncnn_0 548 549 -23330=4,3,14,14,208 0=2\nConvolution Conv_165 1 1 549 551 -23330=4,3,14,14,208 0=208 1=1 5=1 6=43264\nBinaryOp Add_167 2 1 537_splitncnn_0 551 552 -23330=4,3,14,14,208\nReLU Relu_168 1 1 552 553 -23330=4,3,14,14,208\nSplit splitncnn_20 1 2 553 553_splitncnn_0 553_splitncnn_1 -23330=8,3,14,14,208,3,14,14,208\nConvolution Conv_169 1 1 553_splitncnn_1 555 -23330=4,3,7,7,440 0=440 1=1 3=2 5=1 6=91520\nConvolution Conv_171 1 1 553_splitncnn_0 558 -23330=4,3,14,14,440 0=440 1=1 5=1 6=91520 9=1\nConvolutionDepthWise Conv_174 1 1 558 561 -23330=4,3,7,7,440 0=440 1=3 3=2 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_21 1 2 561 561_splitncnn_0 561_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_177 1 1 561_splitncnn_1 562 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_178 1 1 562 564 -23330=4,1,52,1,1 0=52 1=1 2=22880 9=1\nInnerProduct Conv_180 1 1 564 566 -23330=4,1,440,1,1 0=440 1=1 2=22880 9=4\nBinaryOp Mul_182 2 1 561_splitncnn_0 566 567 -23330=4,3,7,7,440 0=2\nConvolution Conv_183 1 1 567 569 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_185 2 1 555 569 570 -23330=4,3,7,7,440\nReLU Relu_186 1 1 570 571 -23330=4,3,7,7,440\nSplit splitncnn_22 1 2 571 571_splitncnn_0 571_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nConvolution Conv_187 1 1 571_splitncnn_1 574 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600 9=1\nConvolutionDepthWise Conv_190 1 1 574 577 -23330=4,3,7,7,440 0=440 1=3 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_23 1 2 577 577_splitncnn_0 577_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_193 1 1 577_splitncnn_1 578 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_194 1 1 578 580 -23330=4,1,110,1,1 0=110 1=1 2=48400 9=1\nInnerProduct Conv_196 1 1 580 582 -23330=4,1,440,1,1 0=440 1=1 2=48400 9=4\nBinaryOp Mul_198 2 1 577_splitncnn_0 582 583 -23330=4,3,7,7,440 0=2\nConvolution Conv_199 1 1 583 585 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_201 2 1 571_splitncnn_0 585 586 -23330=4,3,7,7,440\nReLU Relu_202 1 1 586 587 -23330=4,3,7,7,440\nSplit splitncnn_24 1 2 587 587_splitncnn_0 587_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nConvolution Conv_203 1 1 587_splitncnn_1 590 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600 9=1\nConvolutionDepthWise Conv_206 1 1 590 593 -23330=4,3,7,7,440 0=440 1=3 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_25 1 2 593 593_splitncnn_0 593_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_209 1 1 593_splitncnn_1 594 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_210 1 1 594 596 -23330=4,1,110,1,1 0=110 1=1 2=48400 9=1\nInnerProduct Conv_212 1 1 596 598 -23330=4,1,440,1,1 0=440 1=1 2=48400 9=4\nBinaryOp Mul_214 2 1 593_splitncnn_0 598 599 -23330=4,3,7,7,440 0=2\nConvolution Conv_215 1 1 599 601 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_217 2 1 587_splitncnn_0 601 602 -23330=4,3,7,7,440\nReLU Relu_218 1 1 602 603 -23330=4,3,7,7,440\nSplit splitncnn_26 1 2 603 603_splitncnn_0 603_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nConvolution Conv_219 1 1 603_splitncnn_1 606 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600 9=1\nConvolutionDepthWise Conv_222 1 1 606 609 -23330=4,3,7,7,440 0=440 1=3 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_27 1 2 609 609_splitncnn_0 609_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_225 1 1 609_splitncnn_1 610 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_226 1 1 610 612 -23330=4,1,110,1,1 0=110 1=1 2=48400 9=1\nInnerProduct Conv_228 1 1 612 614 -23330=4,1,440,1,1 0=440 1=1 2=48400 9=4\nBinaryOp Mul_230 2 1 609_splitncnn_0 614 615 -23330=4,3,7,7,440 0=2\nConvolution Conv_231 1 1 615 617 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_233 2 1 603_splitncnn_0 617 618 -23330=4,3,7,7,440\nReLU Relu_234 1 1 618 619 -23330=4,3,7,7,440\nSplit splitncnn_28 1 2 619 619_splitncnn_0 619_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nConvolution Conv_235 1 1 619_splitncnn_1 622 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600 9=1\nConvolutionDepthWise Conv_238 1 1 622 625 -23330=4,3,7,7,440 0=440 1=3 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_29 1 2 625 625_splitncnn_0 625_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_241 1 1 625_splitncnn_1 626 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_242 1 1 626 628 -23330=4,1,110,1,1 0=110 1=1 2=48400 9=1\nInnerProduct Conv_244 1 1 628 630 -23330=4,1,440,1,1 0=440 1=1 2=48400 9=4\nBinaryOp Mul_246 2 1 625_splitncnn_0 630 631 -23330=4,3,7,7,440 0=2\nConvolution Conv_247 1 1 631 633 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_249 2 1 619_splitncnn_0 633 634 -23330=4,3,7,7,440\nReLU Relu_250 1 1 634 635 -23330=4,3,7,7,440\nSplit splitncnn_30 1 2 635 635_splitncnn_0 635_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nConvolution Conv_251 1 1 635_splitncnn_1 638 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600 9=1\nConvolutionDepthWise Conv_254 1 1 638 641 -23330=4,3,7,7,440 0=440 1=3 4=1 5=1 6=31680 7=55 9=1\nSplit splitncnn_31 1 2 641 641_splitncnn_0 641_splitncnn_1 -23330=8,3,7,7,440,3,7,7,440\nPooling GlobalAveragePool_257 1 1 641_splitncnn_1 642 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Conv_258 1 1 642 644 -23330=4,1,110,1,1 0=110 1=1 2=48400 9=1\nInnerProduct Conv_260 1 1 644 646 -23330=4,1,440,1,1 0=440 1=1 2=48400 9=4\nBinaryOp Mul_262 2 1 641_splitncnn_0 646 647 -23330=4,3,7,7,440 0=2\nConvolution Conv_263 1 1 647 649 -23330=4,3,7,7,440 0=440 1=1 5=1 6=193600\nBinaryOp Add_265 2 1 635_splitncnn_0 649 650 -23330=4,3,7,7,440\nReLU Relu_266 1 1 650 651 -23330=4,3,7,7,440\nPooling GlobalAveragePool_267 1 1 651 660 -23330=4,1,440,1,1 0=1 4=1\nInnerProduct Gemm_274 1 1 660 661 -23330=4,1,1000,1,1 0=1000 1=1 2=440000\nSoftmax prob 1 1 661 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/resnet18.param", "content": "7767517\n50 58\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu -23330=4,3,112,112,64 0=64 1=7 3=2 4=3 5=1 6=9408 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 -23330=4,3,56,56,64 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1 -23330=8,3,56,56,64,3,56,56,64\nConvolution res2a_branch1 1 1 pool1_splitncnn_1 res2a_branch1_scale2a_branch1 -23330=4,3,56,56,64 0=64 1=1 5=1 6=4096\nConvolution res2a_branch2a 1 1 pool1_splitncnn_0 res2a_branch2a_res2a_branch2a_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1\nConvolution res2a_branch2b 1 1 res2a_branch2a_res2a_branch2a_relu res2a_branch2b_scale2a_branch2b -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864\nEltwise res2a 2 1 res2a_branch1_scale2a_branch1 res2a_branch2b_scale2a_branch2b res2a -23330=4,3,56,56,64 0=1\nReLU res2a_relu 1 1 res2a res2a_res2a_relu -23330=4,3,56,56,64\nSplit splitncnn_1 1 2 res2a_res2a_relu res2a_res2a_relu_splitncnn_0 res2a_res2a_relu_splitncnn_1 -23330=8,3,56,56,64,3,56,56,64\nConvolution res2b_branch2a 1 1 res2a_res2a_relu_splitncnn_1 res2b_branch2a_res2b_branch2a_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1\nConvolution res2b_branch2b 1 1 res2b_branch2a_res2b_branch2a_relu res2b_branch2b_scale2b_branch2b -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864\nEltwise res2b 2 1 res2a_res2a_relu_splitncnn_0 res2b_branch2b_scale2b_branch2b res2b -23330=4,3,56,56,64 0=1\nReLU res2b_relu 1 1 res2b res2b_res2b_relu -23330=4,3,56,56,64\nSplit splitncnn_2 1 2 res2b_res2b_relu res2b_res2b_relu_splitncnn_0 res2b_res2b_relu_splitncnn_1 -23330=8,3,56,56,64,3,56,56,64\nConvolution res3a_branch1 1 1 res2b_res2b_relu_splitncnn_1 res3a_branch1_scale3a_branch1 -23330=4,3,28,28,128 0=128 1=1 3=2 5=1 6=8192\nConvolution res3a_branch2a 1 1 res2b_res2b_relu_splitncnn_0 res3a_branch2a_res3a_branch2a_relu -23330=4,3,28,28,128 0=128 1=3 3=2 4=1 5=1 6=73728 9=1\nConvolution res3a_branch2b 1 1 res3a_branch2a_res3a_branch2a_relu res3a_branch2b_scale3a_branch2b -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456\nEltwise res3a 2 1 res3a_branch1_scale3a_branch1 res3a_branch2b_scale3a_branch2b res3a -23330=4,3,28,28,128 0=1\nReLU res3a_relu 1 1 res3a res3a_res3a_relu -23330=4,3,28,28,128\nSplit splitncnn_3 1 2 res3a_res3a_relu res3a_res3a_relu_splitncnn_0 res3a_res3a_relu_splitncnn_1 -23330=8,3,28,28,128,3,28,28,128\nConvolution res3b_branch2a 1 1 res3a_res3a_relu_splitncnn_1 res3b_branch2a_res3b_branch2a_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1\nConvolution res3b_branch2b 1 1 res3b_branch2a_res3b_branch2a_relu res3b_branch2b_scale3b_branch2b -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456\nEltwise res3b 2 1 res3a_res3a_relu_splitncnn_0 res3b_branch2b_scale3b_branch2b res3b -23330=4,3,28,28,128 0=1\nReLU res3b_relu 1 1 res3b res3b_res3b_relu -23330=4,3,28,28,128\nSplit splitncnn_4 1 2 res3b_res3b_relu res3b_res3b_relu_splitncnn_0 res3b_res3b_relu_splitncnn_1 -23330=8,3,28,28,128,3,28,28,128\nConvolution res4a_branch1 1 1 res3b_res3b_relu_splitncnn_1 res4a_branch1_scale4a_branch1 -23330=4,3,14,14,256 0=256 1=1 3=2 5=1 6=32768\nConvolution res4a_branch2a 1 1 res3b_res3b_relu_splitncnn_0 res4a_branch2a_res4a_branch2a_relu -23330=4,3,14,14,256 0=256 1=3 3=2 4=1 5=1 6=294912 9=1\nConvolution res4a_branch2b 1 1 res4a_branch2a_res4a_branch2a_relu res4a_branch2b_scale4a_branch2b -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824\nEltwise res4a 2 1 res4a_branch1_scale4a_branch1 res4a_branch2b_scale4a_branch2b res4a -23330=4,3,14,14,256 0=1\nReLU res4a_relu 1 1 res4a res4a_res4a_relu -23330=4,3,14,14,256\nSplit splitncnn_5 1 2 res4a_res4a_relu res4a_res4a_relu_splitncnn_0 res4a_res4a_relu_splitncnn_1 -23330=8,3,14,14,256,3,14,14,256\nConvolution res4b_branch2a 1 1 res4a_res4a_relu_splitncnn_1 res4b_branch2a_res4b_branch2a_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4b_branch2b 1 1 res4b_branch2a_res4b_branch2a_relu res4b_branch2b_scale4b_branch2b -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824\nEltwise res4b 2 1 res4a_res4a_relu_splitncnn_0 res4b_branch2b_scale4b_branch2b res4b -23330=4,3,14,14,256 0=1\nReLU res4b_relu 1 1 res4b res4b_res4b_relu -23330=4,3,14,14,256\nSplit splitncnn_6 1 2 res4b_res4b_relu res4b_res4b_relu_splitncnn_0 res4b_res4b_relu_splitncnn_1 -23330=8,3,14,14,256,3,14,14,256\nConvolution res5a_branch1 1 1 res4b_res4b_relu_splitncnn_1 res5a_branch1_scale5a_branch1 -23330=4,3,7,7,512 0=512 1=1 3=2 5=1 6=131072\nConvolution res5a_branch2a 1 1 res4b_res4b_relu_splitncnn_0 res5a_branch2a_res5a_branch2a_relu -23330=4,3,7,7,512 0=512 1=3 3=2 4=1 5=1 6=1179648 9=1\nConvolution res5a_branch2b 1 1 res5a_branch2a_res5a_branch2a_relu res5a_branch2b_scale5a_branch2b -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296\nEltwise res5a 2 1 res5a_branch1_scale5a_branch1 res5a_branch2b_scale5a_branch2b res5a -23330=4,3,7,7,512 0=1\nReLU res5a_relu 1 1 res5a res5a_res5a_relu -23330=4,3,7,7,512\nSplit splitncnn_7 1 2 res5a_res5a_relu res5a_res5a_relu_splitncnn_0 res5a_res5a_relu_splitncnn_1 -23330=8,3,7,7,512,3,7,7,512\nConvolution res5b_branch2a 1 1 res5a_res5a_relu_splitncnn_1 res5b_branch2a_res5b_branch2a_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution res5b_branch2b 1 1 res5b_branch2a_res5b_branch2a_relu res5b_branch2b_scale5b_branch2b -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296\nEltwise res5b 2 1 res5a_res5a_relu_splitncnn_0 res5b_branch2b_scale5b_branch2b res5b -23330=4,3,7,7,512 0=1\nReLU res5b_relu 1 1 res5b res5b_res5b_relu -23330=4,3,7,7,512\nPooling pool5 1 1 res5b_res5b_relu pool5 -23330=4,3,1,1,512 0=1 1=7\nInnerProduct fc1000 1 1 pool5 fc1000 -23330=4,1,1000,1,1 0=1000 1=1 2=512000\nSoftmax prob 1 1 fc1000 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/resnet18_int8.param", "content": "7767517\n50 58\nInput data 0 1 data 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu 0=64 1=7 3=2 4=3 5=1 6=9408 8=2 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1\nConvolution res2a_branch1 1 1 pool1_splitncnn_1 res2a_branch1_scale2a_branch1 0=64 1=1 5=1 6=4096 8=2\nConvolution res2a_branch2a 1 1 pool1_splitncnn_0 res2a_branch2a_res2a_branch2a_relu 0=64 1=3 4=1 5=1 6=36864 8=102 9=1\nConvolution res2a_branch2b 1 1 res2a_branch2a_res2a_branch2a_relu res2a_branch2b_scale2a_branch2b 0=64 1=3 4=1 5=1 6=36864 8=2\nEltwise res2a 2 1 res2a_branch1_scale2a_branch1 res2a_branch2b_scale2a_branch2b res2a 0=1\nReLU res2a_relu 1 1 res2a res2a_res2a_relu\nSplit splitncnn_1 1 2 res2a_res2a_relu res2a_res2a_relu_splitncnn_0 res2a_res2a_relu_splitncnn_1\nConvolution res2b_branch2a 1 1 res2a_res2a_relu_splitncnn_1 res2b_branch2a_res2b_branch2a_relu 0=64 1=3 4=1 5=1 6=36864 8=102 9=1\nConvolution res2b_branch2b 1 1 res2b_branch2a_res2b_branch2a_relu res2b_branch2b_scale2b_branch2b 0=64 1=3 4=1 5=1 6=36864 8=2\nEltwise res2b 2 1 res2a_res2a_relu_splitncnn_0 res2b_branch2b_scale2b_branch2b res2b 0=1\nReLU res2b_relu 1 1 res2b res2b_res2b_relu\nSplit splitncnn_2 1 2 res2b_res2b_relu res2b_res2b_relu_splitncnn_0 res2b_res2b_relu_splitncnn_1\nConvolution res3a_branch1 1 1 res2b_res2b_relu_splitncnn_1 res3a_branch1_scale3a_branch1 0=128 1=1 3=2 5=1 6=8192 8=2\nConvolution res3a_branch2a 1 1 res2b_res2b_relu_splitncnn_0 res3a_branch2a_res3a_branch2a_relu 0=128 1=3 3=2 4=1 5=1 6=73728 8=102 9=1\nConvolution res3a_branch2b 1 1 res3a_branch2a_res3a_branch2a_relu res3a_branch2b_scale3a_branch2b 0=128 1=3 4=1 5=1 6=147456 8=2\nEltwise res3a 2 1 res3a_branch1_scale3a_branch1 res3a_branch2b_scale3a_branch2b res3a 0=1\nReLU res3a_relu 1 1 res3a res3a_res3a_relu\nSplit splitncnn_3 1 2 res3a_res3a_relu res3a_res3a_relu_splitncnn_0 res3a_res3a_relu_splitncnn_1\nConvolution res3b_branch2a 1 1 res3a_res3a_relu_splitncnn_1 res3b_branch2a_res3b_branch2a_relu 0=128 1=3 4=1 5=1 6=147456 8=102 9=1\nConvolution res3b_branch2b 1 1 res3b_branch2a_res3b_branch2a_relu res3b_branch2b_scale3b_branch2b 0=128 1=3 4=1 5=1 6=147456 8=2\nEltwise res3b 2 1 res3a_res3a_relu_splitncnn_0 res3b_branch2b_scale3b_branch2b res3b 0=1\nReLU res3b_relu 1 1 res3b res3b_res3b_relu\nSplit splitncnn_4 1 2 res3b_res3b_relu res3b_res3b_relu_splitncnn_0 res3b_res3b_relu_splitncnn_1\nConvolution res4a_branch1 1 1 res3b_res3b_relu_splitncnn_1 res4a_branch1_scale4a_branch1 0=256 1=1 3=2 5=1 6=32768 8=2\nConvolution res4a_branch2a 1 1 res3b_res3b_relu_splitncnn_0 res4a_branch2a_res4a_branch2a_relu 0=256 1=3 3=2 4=1 5=1 6=294912 8=102 9=1\nConvolution res4a_branch2b 1 1 res4a_branch2a_res4a_branch2a_relu res4a_branch2b_scale4a_branch2b 0=256 1=3 4=1 5=1 6=589824 8=2\nEltwise res4a 2 1 res4a_branch1_scale4a_branch1 res4a_branch2b_scale4a_branch2b res4a 0=1\nReLU res4a_relu 1 1 res4a res4a_res4a_relu\nSplit splitncnn_5 1 2 res4a_res4a_relu res4a_res4a_relu_splitncnn_0 res4a_res4a_relu_splitncnn_1\nConvolution res4b_branch2a 1 1 res4a_res4a_relu_splitncnn_1 res4b_branch2a_res4b_branch2a_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4b_branch2b 1 1 res4b_branch2a_res4b_branch2a_relu res4b_branch2b_scale4b_branch2b 0=256 1=3 4=1 5=1 6=589824 8=2\nEltwise res4b 2 1 res4a_res4a_relu_splitncnn_0 res4b_branch2b_scale4b_branch2b res4b 0=1\nReLU res4b_relu 1 1 res4b res4b_res4b_relu\nSplit splitncnn_6 1 2 res4b_res4b_relu res4b_res4b_relu_splitncnn_0 res4b_res4b_relu_splitncnn_1\nConvolution res5a_branch1 1 1 res4b_res4b_relu_splitncnn_1 res5a_branch1_scale5a_branch1 0=512 1=1 3=2 5=1 6=131072 8=2\nConvolution res5a_branch2a 1 1 res4b_res4b_relu_splitncnn_0 res5a_branch2a_res5a_branch2a_relu 0=512 1=3 3=2 4=1 5=1 6=1179648 8=102 9=1\nConvolution res5a_branch2b 1 1 res5a_branch2a_res5a_branch2a_relu res5a_branch2b_scale5a_branch2b 0=512 1=3 4=1 5=1 6=2359296 8=2\nEltwise res5a 2 1 res5a_branch1_scale5a_branch1 res5a_branch2b_scale5a_branch2b res5a 0=1\nReLU res5a_relu 1 1 res5a res5a_res5a_relu\nSplit splitncnn_7 1 2 res5a_res5a_relu res5a_res5a_relu_splitncnn_0 res5a_res5a_relu_splitncnn_1\nConvolution res5b_branch2a 1 1 res5a_res5a_relu_splitncnn_1 res5b_branch2a_res5b_branch2a_relu 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution res5b_branch2b 1 1 res5b_branch2a_res5b_branch2a_relu res5b_branch2b_scale5b_branch2b 0=512 1=3 4=1 5=1 6=2359296 8=2\nEltwise res5b 2 1 res5a_res5a_relu_splitncnn_0 res5b_branch2b_scale5b_branch2b res5b 0=1\nReLU res5b_relu 1 1 res5b res5b_res5b_relu\nPooling pool5 1 1 res5b_res5b_relu pool5 0=1 1=7\nInnerProduct fc1000 1 1 pool5 fc1000 0=1000 1=1 2=512000\nSoftmax prob 1 1 fc1000 output\n" }, { "path": "benchmark/resnet50.param", "content": "7767517\n106 122\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu -23330=4,3,112,112,64 0=64 1=7 3=2 4=3 5=1 6=9408 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 -23330=4,3,56,56,64 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1 -23330=8,3,56,56,64,3,56,56,64\nConvolution res2a_branch1 1 1 pool1_splitncnn_1 res2a_branch1_scale2a_branch1 -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384\nConvolution res2a_branch2a 1 1 pool1_splitncnn_0 res2a_branch2a_res2a_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=4096 9=1\nConvolution res2a_branch2b 1 1 res2a_branch2a_res2a_branch2a_relu res2a_branch2b_res2a_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1\nConvolution res2a_branch2c 1 1 res2a_branch2b_res2a_branch2b_relu res2a_branch2c_scale2a_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384\nEltwise res2a 2 1 res2a_branch1_scale2a_branch1 res2a_branch2c_scale2a_branch2c res2a -23330=4,3,56,56,256 0=1\nReLU res2a_relu 1 1 res2a res2a_res2a_relu -23330=4,3,56,56,256\nSplit splitncnn_1 1 2 res2a_res2a_relu res2a_res2a_relu_splitncnn_0 res2a_res2a_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256\nConvolution res2b_branch2a 1 1 res2a_res2a_relu_splitncnn_1 res2b_branch2a_res2b_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=16384 9=1\nConvolution res2b_branch2b 1 1 res2b_branch2a_res2b_branch2a_relu res2b_branch2b_res2b_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1\nConvolution res2b_branch2c 1 1 res2b_branch2b_res2b_branch2b_relu res2b_branch2c_scale2b_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384\nEltwise res2b 2 1 res2a_res2a_relu_splitncnn_0 res2b_branch2c_scale2b_branch2c res2b -23330=4,3,56,56,256 0=1\nReLU res2b_relu 1 1 res2b res2b_res2b_relu -23330=4,3,56,56,256\nSplit splitncnn_2 1 2 res2b_res2b_relu res2b_res2b_relu_splitncnn_0 res2b_res2b_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256\nConvolution res2c_branch2a 1 1 res2b_res2b_relu_splitncnn_1 res2c_branch2a_res2c_branch2a_relu -23330=4,3,56,56,64 0=64 1=1 5=1 6=16384 9=1\nConvolution res2c_branch2b 1 1 res2c_branch2a_res2c_branch2a_relu res2c_branch2b_res2c_branch2b_relu -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=36864 9=1\nConvolution res2c_branch2c 1 1 res2c_branch2b_res2c_branch2b_relu res2c_branch2c_scale2c_branch2c -23330=4,3,56,56,256 0=256 1=1 5=1 6=16384\nEltwise res2c 2 1 res2b_res2b_relu_splitncnn_0 res2c_branch2c_scale2c_branch2c res2c -23330=4,3,56,56,256 0=1\nReLU res2c_relu 1 1 res2c res2c_res2c_relu -23330=4,3,56,56,256\nSplit splitncnn_3 1 2 res2c_res2c_relu res2c_res2c_relu_splitncnn_0 res2c_res2c_relu_splitncnn_1 -23330=8,3,56,56,256,3,56,56,256\nConvolution res3a_branch1 1 1 res2c_res2c_relu_splitncnn_1 res3a_branch1_scale3a_branch1 -23330=4,3,28,28,512 0=512 1=1 3=2 5=1 6=131072\nConvolution res3a_branch2a 1 1 res2c_res2c_relu_splitncnn_0 res3a_branch2a_res3a_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 3=2 5=1 6=32768 9=1\nConvolution res3a_branch2b 1 1 res3a_branch2a_res3a_branch2a_relu res3a_branch2b_res3a_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1\nConvolution res3a_branch2c 1 1 res3a_branch2b_res3a_branch2b_relu res3a_branch2c_scale3a_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536\nEltwise res3a 2 1 res3a_branch1_scale3a_branch1 res3a_branch2c_scale3a_branch2c res3a -23330=4,3,28,28,512 0=1\nReLU res3a_relu 1 1 res3a res3a_res3a_relu -23330=4,3,28,28,512\nSplit splitncnn_4 1 2 res3a_res3a_relu res3a_res3a_relu_splitncnn_0 res3a_res3a_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512\nConvolution res3b_branch2a 1 1 res3a_res3a_relu_splitncnn_1 res3b_branch2a_res3b_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1\nConvolution res3b_branch2b 1 1 res3b_branch2a_res3b_branch2a_relu res3b_branch2b_res3b_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1\nConvolution res3b_branch2c 1 1 res3b_branch2b_res3b_branch2b_relu res3b_branch2c_scale3b_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536\nEltwise res3b 2 1 res3a_res3a_relu_splitncnn_0 res3b_branch2c_scale3b_branch2c res3b -23330=4,3,28,28,512 0=1\nReLU res3b_relu 1 1 res3b res3b_res3b_relu -23330=4,3,28,28,512\nSplit splitncnn_5 1 2 res3b_res3b_relu res3b_res3b_relu_splitncnn_0 res3b_res3b_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512\nConvolution res3c_branch2a 1 1 res3b_res3b_relu_splitncnn_1 res3c_branch2a_res3c_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1\nConvolution res3c_branch2b 1 1 res3c_branch2a_res3c_branch2a_relu res3c_branch2b_res3c_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1\nConvolution res3c_branch2c 1 1 res3c_branch2b_res3c_branch2b_relu res3c_branch2c_scale3c_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536\nEltwise res3c 2 1 res3b_res3b_relu_splitncnn_0 res3c_branch2c_scale3c_branch2c res3c -23330=4,3,28,28,512 0=1\nReLU res3c_relu 1 1 res3c res3c_res3c_relu -23330=4,3,28,28,512\nSplit splitncnn_6 1 2 res3c_res3c_relu res3c_res3c_relu_splitncnn_0 res3c_res3c_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512\nConvolution res3d_branch2a 1 1 res3c_res3c_relu_splitncnn_1 res3d_branch2a_res3d_branch2a_relu -23330=4,3,28,28,128 0=128 1=1 5=1 6=65536 9=1\nConvolution res3d_branch2b 1 1 res3d_branch2a_res3d_branch2a_relu res3d_branch2b_res3d_branch2b_relu -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=147456 9=1\nConvolution res3d_branch2c 1 1 res3d_branch2b_res3d_branch2b_relu res3d_branch2c_scale3d_branch2c -23330=4,3,28,28,512 0=512 1=1 5=1 6=65536\nEltwise res3d 2 1 res3c_res3c_relu_splitncnn_0 res3d_branch2c_scale3d_branch2c res3d -23330=4,3,28,28,512 0=1\nReLU res3d_relu 1 1 res3d res3d_res3d_relu -23330=4,3,28,28,512\nSplit splitncnn_7 1 2 res3d_res3d_relu res3d_res3d_relu_splitncnn_0 res3d_res3d_relu_splitncnn_1 -23330=8,3,28,28,512,3,28,28,512\nConvolution res4a_branch1 1 1 res3d_res3d_relu_splitncnn_1 res4a_branch1_scale4a_branch1 -23330=4,3,14,14,1024 0=1024 1=1 3=2 5=1 6=524288\nConvolution res4a_branch2a 1 1 res3d_res3d_relu_splitncnn_0 res4a_branch2a_res4a_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 3=2 5=1 6=131072 9=1\nConvolution res4a_branch2b 1 1 res4a_branch2a_res4a_branch2a_relu res4a_branch2b_res4a_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4a_branch2c 1 1 res4a_branch2b_res4a_branch2b_relu res4a_branch2c_scale4a_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4a 2 1 res4a_branch1_scale4a_branch1 res4a_branch2c_scale4a_branch2c res4a -23330=4,3,14,14,1024 0=1\nReLU res4a_relu 1 1 res4a res4a_res4a_relu -23330=4,3,14,14,1024\nSplit splitncnn_8 1 2 res4a_res4a_relu res4a_res4a_relu_splitncnn_0 res4a_res4a_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res4b_branch2a 1 1 res4a_res4a_relu_splitncnn_1 res4b_branch2a_res4b_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1\nConvolution res4b_branch2b 1 1 res4b_branch2a_res4b_branch2a_relu res4b_branch2b_res4b_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4b_branch2c 1 1 res4b_branch2b_res4b_branch2b_relu res4b_branch2c_scale4b_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4b 2 1 res4a_res4a_relu_splitncnn_0 res4b_branch2c_scale4b_branch2c res4b -23330=4,3,14,14,1024 0=1\nReLU res4b_relu 1 1 res4b res4b_res4b_relu -23330=4,3,14,14,1024\nSplit splitncnn_9 1 2 res4b_res4b_relu res4b_res4b_relu_splitncnn_0 res4b_res4b_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res4c_branch2a 1 1 res4b_res4b_relu_splitncnn_1 res4c_branch2a_res4c_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1\nConvolution res4c_branch2b 1 1 res4c_branch2a_res4c_branch2a_relu res4c_branch2b_res4c_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4c_branch2c 1 1 res4c_branch2b_res4c_branch2b_relu res4c_branch2c_scale4c_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4c 2 1 res4b_res4b_relu_splitncnn_0 res4c_branch2c_scale4c_branch2c res4c -23330=4,3,14,14,1024 0=1\nReLU res4c_relu 1 1 res4c res4c_res4c_relu -23330=4,3,14,14,1024\nSplit splitncnn_10 1 2 res4c_res4c_relu res4c_res4c_relu_splitncnn_0 res4c_res4c_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res4d_branch2a 1 1 res4c_res4c_relu_splitncnn_1 res4d_branch2a_res4d_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1\nConvolution res4d_branch2b 1 1 res4d_branch2a_res4d_branch2a_relu res4d_branch2b_res4d_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4d_branch2c 1 1 res4d_branch2b_res4d_branch2b_relu res4d_branch2c_scale4d_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4d 2 1 res4c_res4c_relu_splitncnn_0 res4d_branch2c_scale4d_branch2c res4d -23330=4,3,14,14,1024 0=1\nReLU res4d_relu 1 1 res4d res4d_res4d_relu -23330=4,3,14,14,1024\nSplit splitncnn_11 1 2 res4d_res4d_relu res4d_res4d_relu_splitncnn_0 res4d_res4d_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res4e_branch2a 1 1 res4d_res4d_relu_splitncnn_1 res4e_branch2a_res4e_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1\nConvolution res4e_branch2b 1 1 res4e_branch2a_res4e_branch2a_relu res4e_branch2b_res4e_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4e_branch2c 1 1 res4e_branch2b_res4e_branch2b_relu res4e_branch2c_scale4e_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4e 2 1 res4d_res4d_relu_splitncnn_0 res4e_branch2c_scale4e_branch2c res4e -23330=4,3,14,14,1024 0=1\nReLU res4e_relu 1 1 res4e res4e_res4e_relu -23330=4,3,14,14,1024\nSplit splitncnn_12 1 2 res4e_res4e_relu res4e_res4e_relu_splitncnn_0 res4e_res4e_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res4f_branch2a 1 1 res4e_res4e_relu_splitncnn_1 res4f_branch2a_res4f_branch2a_relu -23330=4,3,14,14,256 0=256 1=1 5=1 6=262144 9=1\nConvolution res4f_branch2b 1 1 res4f_branch2a_res4f_branch2a_relu res4f_branch2b_res4f_branch2b_relu -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution res4f_branch2c 1 1 res4f_branch2b_res4f_branch2b_relu res4f_branch2c_scale4f_branch2c -23330=4,3,14,14,1024 0=1024 1=1 5=1 6=262144\nEltwise res4f 2 1 res4e_res4e_relu_splitncnn_0 res4f_branch2c_scale4f_branch2c res4f -23330=4,3,14,14,1024 0=1\nReLU res4f_relu 1 1 res4f res4f_res4f_relu -23330=4,3,14,14,1024\nSplit splitncnn_13 1 2 res4f_res4f_relu res4f_res4f_relu_splitncnn_0 res4f_res4f_relu_splitncnn_1 -23330=8,3,14,14,1024,3,14,14,1024\nConvolution res5a_branch1 1 1 res4f_res4f_relu_splitncnn_1 res5a_branch1_scale5a_branch1 -23330=4,3,7,7,2048 0=2048 1=1 3=2 5=1 6=2097152\nConvolution res5a_branch2a 1 1 res4f_res4f_relu_splitncnn_0 res5a_branch2a_res5a_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 3=2 5=1 6=524288 9=1\nConvolution res5a_branch2b 1 1 res5a_branch2a_res5a_branch2a_relu res5a_branch2b_res5a_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution res5a_branch2c 1 1 res5a_branch2b_res5a_branch2b_relu res5a_branch2c_scale5a_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576\nEltwise res5a 2 1 res5a_branch1_scale5a_branch1 res5a_branch2c_scale5a_branch2c res5a -23330=4,3,7,7,2048 0=1\nReLU res5a_relu 1 1 res5a res5a_res5a_relu -23330=4,3,7,7,2048\nSplit splitncnn_14 1 2 res5a_res5a_relu res5a_res5a_relu_splitncnn_0 res5a_res5a_relu_splitncnn_1 -23330=8,3,7,7,2048,3,7,7,2048\nConvolution res5b_branch2a 1 1 res5a_res5a_relu_splitncnn_1 res5b_branch2a_res5b_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 5=1 6=1048576 9=1\nConvolution res5b_branch2b 1 1 res5b_branch2a_res5b_branch2a_relu res5b_branch2b_res5b_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution res5b_branch2c 1 1 res5b_branch2b_res5b_branch2b_relu res5b_branch2c_scale5b_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576\nEltwise res5b 2 1 res5a_res5a_relu_splitncnn_0 res5b_branch2c_scale5b_branch2c res5b -23330=4,3,7,7,2048 0=1\nReLU res5b_relu 1 1 res5b res5b_res5b_relu -23330=4,3,7,7,2048\nSplit splitncnn_15 1 2 res5b_res5b_relu res5b_res5b_relu_splitncnn_0 res5b_res5b_relu_splitncnn_1 -23330=8,3,7,7,2048,3,7,7,2048\nConvolution res5c_branch2a 1 1 res5b_res5b_relu_splitncnn_1 res5c_branch2a_res5c_branch2a_relu -23330=4,3,7,7,512 0=512 1=1 5=1 6=1048576 9=1\nConvolution res5c_branch2b 1 1 res5c_branch2a_res5c_branch2a_relu res5c_branch2b_res5c_branch2b_relu -23330=4,3,7,7,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution res5c_branch2c 1 1 res5c_branch2b_res5c_branch2b_relu res5c_branch2c_scale5c_branch2c -23330=4,3,7,7,2048 0=2048 1=1 5=1 6=1048576\nEltwise res5c 2 1 res5b_res5b_relu_splitncnn_0 res5c_branch2c_scale5c_branch2c res5c -23330=4,3,7,7,2048 0=1\nReLU res5c_relu 1 1 res5c res5c_res5c_relu -23330=4,3,7,7,2048\nPooling pool5 1 1 res5c_res5c_relu pool5 -23330=4,3,1,1,2048 0=1 1=7\nInnerProduct fc1000 1 1 pool5 fc1000 -23330=4,1,1000,1,1 0=1000 1=1 2=2048000\nSoftmax prob 1 1 fc1000 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/resnet50_int8.param", "content": "7767517\n106 122\nInput data 0 1 data 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu 0=64 1=7 3=2 4=3 5=1 6=9408 8=2 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1\nConvolution res2a_branch1 1 1 pool1_splitncnn_1 res2a_branch1_scale2a_branch1 0=256 1=1 5=1 6=16384 8=2\nConvolution res2a_branch2a 1 1 pool1_splitncnn_0 res2a_branch2a_res2a_branch2a_relu 0=64 1=1 5=1 6=4096 8=102 9=1\nConvolution res2a_branch2b 1 1 res2a_branch2a_res2a_branch2a_relu res2a_branch2b_res2a_branch2b_relu 0=64 1=3 4=1 5=1 6=36864 8=102 9=1\nConvolution res2a_branch2c 1 1 res2a_branch2b_res2a_branch2b_relu res2a_branch2c_scale2a_branch2c 0=256 1=1 5=1 6=16384 8=2\nEltwise res2a 2 1 res2a_branch1_scale2a_branch1 res2a_branch2c_scale2a_branch2c res2a 0=1\nReLU res2a_relu 1 1 res2a res2a_res2a_relu\nSplit splitncnn_1 1 2 res2a_res2a_relu res2a_res2a_relu_splitncnn_0 res2a_res2a_relu_splitncnn_1\nConvolution res2b_branch2a 1 1 res2a_res2a_relu_splitncnn_1 res2b_branch2a_res2b_branch2a_relu 0=64 1=1 5=1 6=16384 8=102 9=1\nConvolution res2b_branch2b 1 1 res2b_branch2a_res2b_branch2a_relu res2b_branch2b_res2b_branch2b_relu 0=64 1=3 4=1 5=1 6=36864 8=102 9=1\nConvolution res2b_branch2c 1 1 res2b_branch2b_res2b_branch2b_relu res2b_branch2c_scale2b_branch2c 0=256 1=1 5=1 6=16384 8=2\nEltwise res2b 2 1 res2a_res2a_relu_splitncnn_0 res2b_branch2c_scale2b_branch2c res2b 0=1\nReLU res2b_relu 1 1 res2b res2b_res2b_relu\nSplit splitncnn_2 1 2 res2b_res2b_relu res2b_res2b_relu_splitncnn_0 res2b_res2b_relu_splitncnn_1\nConvolution res2c_branch2a 1 1 res2b_res2b_relu_splitncnn_1 res2c_branch2a_res2c_branch2a_relu 0=64 1=1 5=1 6=16384 8=102 9=1\nConvolution res2c_branch2b 1 1 res2c_branch2a_res2c_branch2a_relu res2c_branch2b_res2c_branch2b_relu 0=64 1=3 4=1 5=1 6=36864 8=102 9=1\nConvolution res2c_branch2c 1 1 res2c_branch2b_res2c_branch2b_relu res2c_branch2c_scale2c_branch2c 0=256 1=1 5=1 6=16384 8=2\nEltwise res2c 2 1 res2b_res2b_relu_splitncnn_0 res2c_branch2c_scale2c_branch2c res2c 0=1\nReLU res2c_relu 1 1 res2c res2c_res2c_relu\nSplit splitncnn_3 1 2 res2c_res2c_relu res2c_res2c_relu_splitncnn_0 res2c_res2c_relu_splitncnn_1\nConvolution res3a_branch1 1 1 res2c_res2c_relu_splitncnn_1 res3a_branch1_scale3a_branch1 0=512 1=1 3=2 5=1 6=131072 8=2\nConvolution res3a_branch2a 1 1 res2c_res2c_relu_splitncnn_0 res3a_branch2a_res3a_branch2a_relu 0=128 1=1 3=2 5=1 6=32768 8=102 9=1\nConvolution res3a_branch2b 1 1 res3a_branch2a_res3a_branch2a_relu res3a_branch2b_res3a_branch2b_relu 0=128 1=3 4=1 5=1 6=147456 8=102 9=1\nConvolution res3a_branch2c 1 1 res3a_branch2b_res3a_branch2b_relu res3a_branch2c_scale3a_branch2c 0=512 1=1 5=1 6=65536 8=2\nEltwise res3a 2 1 res3a_branch1_scale3a_branch1 res3a_branch2c_scale3a_branch2c res3a 0=1\nReLU res3a_relu 1 1 res3a res3a_res3a_relu\nSplit splitncnn_4 1 2 res3a_res3a_relu res3a_res3a_relu_splitncnn_0 res3a_res3a_relu_splitncnn_1\nConvolution res3b_branch2a 1 1 res3a_res3a_relu_splitncnn_1 res3b_branch2a_res3b_branch2a_relu 0=128 1=1 5=1 6=65536 8=102 9=1\nConvolution res3b_branch2b 1 1 res3b_branch2a_res3b_branch2a_relu res3b_branch2b_res3b_branch2b_relu 0=128 1=3 4=1 5=1 6=147456 8=102 9=1\nConvolution res3b_branch2c 1 1 res3b_branch2b_res3b_branch2b_relu res3b_branch2c_scale3b_branch2c 0=512 1=1 5=1 6=65536 8=2\nEltwise res3b 2 1 res3a_res3a_relu_splitncnn_0 res3b_branch2c_scale3b_branch2c res3b 0=1\nReLU res3b_relu 1 1 res3b res3b_res3b_relu\nSplit splitncnn_5 1 2 res3b_res3b_relu res3b_res3b_relu_splitncnn_0 res3b_res3b_relu_splitncnn_1\nConvolution res3c_branch2a 1 1 res3b_res3b_relu_splitncnn_1 res3c_branch2a_res3c_branch2a_relu 0=128 1=1 5=1 6=65536 8=102 9=1\nConvolution res3c_branch2b 1 1 res3c_branch2a_res3c_branch2a_relu res3c_branch2b_res3c_branch2b_relu 0=128 1=3 4=1 5=1 6=147456 8=102 9=1\nConvolution res3c_branch2c 1 1 res3c_branch2b_res3c_branch2b_relu res3c_branch2c_scale3c_branch2c 0=512 1=1 5=1 6=65536 8=2\nEltwise res3c 2 1 res3b_res3b_relu_splitncnn_0 res3c_branch2c_scale3c_branch2c res3c 0=1\nReLU res3c_relu 1 1 res3c res3c_res3c_relu\nSplit splitncnn_6 1 2 res3c_res3c_relu res3c_res3c_relu_splitncnn_0 res3c_res3c_relu_splitncnn_1\nConvolution res3d_branch2a 1 1 res3c_res3c_relu_splitncnn_1 res3d_branch2a_res3d_branch2a_relu 0=128 1=1 5=1 6=65536 8=102 9=1\nConvolution res3d_branch2b 1 1 res3d_branch2a_res3d_branch2a_relu res3d_branch2b_res3d_branch2b_relu 0=128 1=3 4=1 5=1 6=147456 8=102 9=1\nConvolution res3d_branch2c 1 1 res3d_branch2b_res3d_branch2b_relu res3d_branch2c_scale3d_branch2c 0=512 1=1 5=1 6=65536 8=2\nEltwise res3d 2 1 res3c_res3c_relu_splitncnn_0 res3d_branch2c_scale3d_branch2c res3d 0=1\nReLU res3d_relu 1 1 res3d res3d_res3d_relu\nSplit splitncnn_7 1 2 res3d_res3d_relu res3d_res3d_relu_splitncnn_0 res3d_res3d_relu_splitncnn_1\nConvolution res4a_branch1 1 1 res3d_res3d_relu_splitncnn_1 res4a_branch1_scale4a_branch1 0=1024 1=1 3=2 5=1 6=524288 8=2\nConvolution res4a_branch2a 1 1 res3d_res3d_relu_splitncnn_0 res4a_branch2a_res4a_branch2a_relu 0=256 1=1 3=2 5=1 6=131072 8=102 9=1\nConvolution res4a_branch2b 1 1 res4a_branch2a_res4a_branch2a_relu res4a_branch2b_res4a_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4a_branch2c 1 1 res4a_branch2b_res4a_branch2b_relu res4a_branch2c_scale4a_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4a 2 1 res4a_branch1_scale4a_branch1 res4a_branch2c_scale4a_branch2c res4a 0=1\nReLU res4a_relu 1 1 res4a res4a_res4a_relu\nSplit splitncnn_8 1 2 res4a_res4a_relu res4a_res4a_relu_splitncnn_0 res4a_res4a_relu_splitncnn_1\nConvolution res4b_branch2a 1 1 res4a_res4a_relu_splitncnn_1 res4b_branch2a_res4b_branch2a_relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution res4b_branch2b 1 1 res4b_branch2a_res4b_branch2a_relu res4b_branch2b_res4b_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4b_branch2c 1 1 res4b_branch2b_res4b_branch2b_relu res4b_branch2c_scale4b_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4b 2 1 res4a_res4a_relu_splitncnn_0 res4b_branch2c_scale4b_branch2c res4b 0=1\nReLU res4b_relu 1 1 res4b res4b_res4b_relu\nSplit splitncnn_9 1 2 res4b_res4b_relu res4b_res4b_relu_splitncnn_0 res4b_res4b_relu_splitncnn_1\nConvolution res4c_branch2a 1 1 res4b_res4b_relu_splitncnn_1 res4c_branch2a_res4c_branch2a_relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution res4c_branch2b 1 1 res4c_branch2a_res4c_branch2a_relu res4c_branch2b_res4c_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4c_branch2c 1 1 res4c_branch2b_res4c_branch2b_relu res4c_branch2c_scale4c_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4c 2 1 res4b_res4b_relu_splitncnn_0 res4c_branch2c_scale4c_branch2c res4c 0=1\nReLU res4c_relu 1 1 res4c res4c_res4c_relu\nSplit splitncnn_10 1 2 res4c_res4c_relu res4c_res4c_relu_splitncnn_0 res4c_res4c_relu_splitncnn_1\nConvolution res4d_branch2a 1 1 res4c_res4c_relu_splitncnn_1 res4d_branch2a_res4d_branch2a_relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution res4d_branch2b 1 1 res4d_branch2a_res4d_branch2a_relu res4d_branch2b_res4d_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4d_branch2c 1 1 res4d_branch2b_res4d_branch2b_relu res4d_branch2c_scale4d_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4d 2 1 res4c_res4c_relu_splitncnn_0 res4d_branch2c_scale4d_branch2c res4d 0=1\nReLU res4d_relu 1 1 res4d res4d_res4d_relu\nSplit splitncnn_11 1 2 res4d_res4d_relu res4d_res4d_relu_splitncnn_0 res4d_res4d_relu_splitncnn_1\nConvolution res4e_branch2a 1 1 res4d_res4d_relu_splitncnn_1 res4e_branch2a_res4e_branch2a_relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution res4e_branch2b 1 1 res4e_branch2a_res4e_branch2a_relu res4e_branch2b_res4e_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4e_branch2c 1 1 res4e_branch2b_res4e_branch2b_relu res4e_branch2c_scale4e_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4e 2 1 res4d_res4d_relu_splitncnn_0 res4e_branch2c_scale4e_branch2c res4e 0=1\nReLU res4e_relu 1 1 res4e res4e_res4e_relu\nSplit splitncnn_12 1 2 res4e_res4e_relu res4e_res4e_relu_splitncnn_0 res4e_res4e_relu_splitncnn_1\nConvolution res4f_branch2a 1 1 res4e_res4e_relu_splitncnn_1 res4f_branch2a_res4f_branch2a_relu 0=256 1=1 5=1 6=262144 8=102 9=1\nConvolution res4f_branch2b 1 1 res4f_branch2a_res4f_branch2a_relu res4f_branch2b_res4f_branch2b_relu 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution res4f_branch2c 1 1 res4f_branch2b_res4f_branch2b_relu res4f_branch2c_scale4f_branch2c 0=1024 1=1 5=1 6=262144 8=2\nEltwise res4f 2 1 res4e_res4e_relu_splitncnn_0 res4f_branch2c_scale4f_branch2c res4f 0=1\nReLU res4f_relu 1 1 res4f res4f_res4f_relu\nSplit splitncnn_13 1 2 res4f_res4f_relu res4f_res4f_relu_splitncnn_0 res4f_res4f_relu_splitncnn_1\nConvolution res5a_branch1 1 1 res4f_res4f_relu_splitncnn_1 res5a_branch1_scale5a_branch1 0=2048 1=1 3=2 5=1 6=2097152 8=2\nConvolution res5a_branch2a 1 1 res4f_res4f_relu_splitncnn_0 res5a_branch2a_res5a_branch2a_relu 0=512 1=1 3=2 5=1 6=524288 8=102 9=1\nConvolution res5a_branch2b 1 1 res5a_branch2a_res5a_branch2a_relu res5a_branch2b_res5a_branch2b_relu 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution res5a_branch2c 1 1 res5a_branch2b_res5a_branch2b_relu res5a_branch2c_scale5a_branch2c 0=2048 1=1 5=1 6=1048576 8=2\nEltwise res5a 2 1 res5a_branch1_scale5a_branch1 res5a_branch2c_scale5a_branch2c res5a 0=1\nReLU res5a_relu 1 1 res5a res5a_res5a_relu\nSplit splitncnn_14 1 2 res5a_res5a_relu res5a_res5a_relu_splitncnn_0 res5a_res5a_relu_splitncnn_1\nConvolution res5b_branch2a 1 1 res5a_res5a_relu_splitncnn_1 res5b_branch2a_res5b_branch2a_relu 0=512 1=1 5=1 6=1048576 8=102 9=1\nConvolution res5b_branch2b 1 1 res5b_branch2a_res5b_branch2a_relu res5b_branch2b_res5b_branch2b_relu 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution res5b_branch2c 1 1 res5b_branch2b_res5b_branch2b_relu res5b_branch2c_scale5b_branch2c 0=2048 1=1 5=1 6=1048576 8=2\nEltwise res5b 2 1 res5a_res5a_relu_splitncnn_0 res5b_branch2c_scale5b_branch2c res5b 0=1\nReLU res5b_relu 1 1 res5b res5b_res5b_relu\nSplit splitncnn_15 1 2 res5b_res5b_relu res5b_res5b_relu_splitncnn_0 res5b_res5b_relu_splitncnn_1\nConvolution res5c_branch2a 1 1 res5b_res5b_relu_splitncnn_1 res5c_branch2a_res5c_branch2a_relu 0=512 1=1 5=1 6=1048576 8=102 9=1\nConvolution res5c_branch2b 1 1 res5c_branch2a_res5c_branch2a_relu res5c_branch2b_res5c_branch2b_relu 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution res5c_branch2c 1 1 res5c_branch2b_res5c_branch2b_relu res5c_branch2c_scale5c_branch2c 0=2048 1=1 5=1 6=1048576 8=2\nEltwise res5c 2 1 res5b_res5b_relu_splitncnn_0 res5c_branch2c_scale5c_branch2c res5c 0=1\nReLU res5c_relu 1 1 res5c res5c_res5c_relu\nPooling pool5 1 1 res5c_res5c_relu pool5 0=1 1=7\nInnerProduct fc1000 1 1 pool5 fc1000 0=1000 1=1 2=2048000\nSoftmax prob 1 1 fc1000 output\n" }, { "path": "benchmark/shufflenet.param", "content": "7767517\n120 136\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu -23330=4,3,112,112,24 0=24 1=3 3=2 4=1 5=1 6=648 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 -23330=4,3,56,56,24 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nPooling resx1_match_conv 1 1 pool1_splitncnn_1 resx1_match_conv -23330=4,3,28,28,24 0=1 1=3 2=2\nConvolution resx1_conv1 1 1 pool1_splitncnn_0 resx1_conv1_resx1_conv1_relu -23330=4,3,56,56,54 0=54 1=1 5=1 6=1296 9=1\nConvolutionDepthWise resx1_conv2 1 1 resx1_conv1_resx1_conv1_relu resx1_conv2_resx1_conv2_scale -23330=4,3,28,28,54 0=54 1=3 3=2 4=1 5=1 6=486 7=54\nConvolutionDepthWise resx1_conv3 1 1 resx1_conv2_resx1_conv2_scale resx1_conv3_resx1_conv3_scale -23330=4,3,28,28,216 0=216 1=1 5=1 6=3888 7=3\nConcat resx1_concat 2 1 resx1_match_conv resx1_conv3_resx1_conv3_scale resx1_concat -23330=4,3,28,28,240\nReLU resx1_concat_relu 1 1 resx1_concat resx1_concat_resx1_concat_relu -23330=4,3,28,28,240\nSplit splitncnn_1 1 2 resx1_concat_resx1_concat_relu resx1_concat_resx1_concat_relu_splitncnn_0 resx1_concat_resx1_concat_relu_splitncnn_1 -23330=8,3,28,28,240,3,28,28,240\nConvolutionDepthWise resx2_conv1 1 1 resx1_concat_resx1_concat_relu_splitncnn_1 resx2_conv1_resx2_conv1_relu -23330=4,3,28,28,60 0=60 1=1 5=1 6=4800 7=3 9=1\nShuffleChannel shuffle2 1 1 resx2_conv1_resx2_conv1_relu shuffle2 -23330=4,3,28,28,60 0=3\nConvolutionDepthWise resx2_conv2 1 1 shuffle2 resx2_conv2_resx2_conv2_scale -23330=4,3,28,28,60 0=60 1=3 4=1 5=1 6=540 7=60\nConvolutionDepthWise resx2_conv3 1 1 resx2_conv2_resx2_conv2_scale resx2_conv3_resx2_conv3_scale -23330=4,3,28,28,240 0=240 1=1 5=1 6=4800 7=3\nEltwise resx2_elewise 2 1 resx1_concat_resx1_concat_relu_splitncnn_0 resx2_conv3_resx2_conv3_scale resx2_elewise -23330=4,3,28,28,240 0=1\nReLU resx2_elewise_relu 1 1 resx2_elewise resx2_elewise_resx2_elewise_relu -23330=4,3,28,28,240\nSplit splitncnn_2 1 2 resx2_elewise_resx2_elewise_relu resx2_elewise_resx2_elewise_relu_splitncnn_0 resx2_elewise_resx2_elewise_relu_splitncnn_1 -23330=8,3,28,28,240,3,28,28,240\nConvolutionDepthWise resx3_conv1 1 1 resx2_elewise_resx2_elewise_relu_splitncnn_1 resx3_conv1_resx3_conv1_relu -23330=4,3,28,28,60 0=60 1=1 5=1 6=4800 7=3 9=1\nShuffleChannel shuffle3 1 1 resx3_conv1_resx3_conv1_relu shuffle3 -23330=4,3,28,28,60 0=3\nConvolutionDepthWise resx3_conv2 1 1 shuffle3 resx3_conv2_resx3_conv2_scale -23330=4,3,28,28,60 0=60 1=3 4=1 5=1 6=540 7=60\nConvolutionDepthWise resx3_conv3 1 1 resx3_conv2_resx3_conv2_scale resx3_conv3_resx3_conv3_scale -23330=4,3,28,28,240 0=240 1=1 5=1 6=4800 7=3\nEltwise resx3_elewise 2 1 resx2_elewise_resx2_elewise_relu_splitncnn_0 resx3_conv3_resx3_conv3_scale resx3_elewise -23330=4,3,28,28,240 0=1\nReLU resx3_elewise_relu 1 1 resx3_elewise resx3_elewise_resx3_elewise_relu -23330=4,3,28,28,240\nSplit splitncnn_3 1 2 resx3_elewise_resx3_elewise_relu resx3_elewise_resx3_elewise_relu_splitncnn_0 resx3_elewise_resx3_elewise_relu_splitncnn_1 -23330=8,3,28,28,240,3,28,28,240\nConvolutionDepthWise resx4_conv1 1 1 resx3_elewise_resx3_elewise_relu_splitncnn_1 resx4_conv1_resx4_conv1_relu -23330=4,3,28,28,60 0=60 1=1 5=1 6=4800 7=3 9=1\nShuffleChannel shuffle4 1 1 resx4_conv1_resx4_conv1_relu shuffle4 -23330=4,3,28,28,60 0=3\nConvolutionDepthWise resx4_conv2 1 1 shuffle4 resx4_conv2_resx4_conv2_scale -23330=4,3,28,28,60 0=60 1=3 4=1 5=1 6=540 7=60\nConvolutionDepthWise resx4_conv3 1 1 resx4_conv2_resx4_conv2_scale resx4_conv3_resx4_conv3_scale -23330=4,3,28,28,240 0=240 1=1 5=1 6=4800 7=3\nEltwise resx4_elewise 2 1 resx3_elewise_resx3_elewise_relu_splitncnn_0 resx4_conv3_resx4_conv3_scale resx4_elewise -23330=4,3,28,28,240 0=1\nReLU resx4_elewise_relu 1 1 resx4_elewise resx4_elewise_resx4_elewise_relu -23330=4,3,28,28,240\nSplit splitncnn_4 1 2 resx4_elewise_resx4_elewise_relu resx4_elewise_resx4_elewise_relu_splitncnn_0 resx4_elewise_resx4_elewise_relu_splitncnn_1 -23330=8,3,28,28,240,3,28,28,240\nPooling resx5_match_conv 1 1 resx4_elewise_resx4_elewise_relu_splitncnn_1 resx5_match_conv -23330=4,3,14,14,240 0=1 1=3 2=2\nConvolutionDepthWise resx5_conv1 1 1 resx4_elewise_resx4_elewise_relu_splitncnn_0 resx5_conv1_resx5_conv1_relu -23330=4,3,28,28,60 0=60 1=1 5=1 6=4800 7=3 9=1\nShuffleChannel shuffle5 1 1 resx5_conv1_resx5_conv1_relu shuffle5 -23330=4,3,28,28,60 0=3\nConvolutionDepthWise resx5_conv2 1 1 shuffle5 resx5_conv2_resx5_conv2_scale -23330=4,3,14,14,60 0=60 1=3 3=2 4=1 5=1 6=540 7=60\nConvolutionDepthWise resx5_conv3 1 1 resx5_conv2_resx5_conv2_scale resx5_conv3_resx5_conv3_scale -23330=4,3,14,14,240 0=240 1=1 5=1 6=4800 7=3\nConcat resx5_concat 2 1 resx5_match_conv resx5_conv3_resx5_conv3_scale resx5_concat -23330=4,3,14,14,480\nReLU resx5_concat_relu 1 1 resx5_concat resx5_concat_resx5_concat_relu -23330=4,3,14,14,480\nSplit splitncnn_5 1 2 resx5_concat_resx5_concat_relu resx5_concat_resx5_concat_relu_splitncnn_0 resx5_concat_resx5_concat_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx6_conv1 1 1 resx5_concat_resx5_concat_relu_splitncnn_1 resx6_conv1_resx6_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle6 1 1 resx6_conv1_resx6_conv1_relu shuffle6 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx6_conv2 1 1 shuffle6 resx6_conv2_resx6_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx6_conv3 1 1 resx6_conv2_resx6_conv2_scale resx6_conv3_resx6_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx6_elewise 2 1 resx5_concat_resx5_concat_relu_splitncnn_0 resx6_conv3_resx6_conv3_scale resx6_elewise -23330=4,3,14,14,480 0=1\nReLU resx6_elewise_relu 1 1 resx6_elewise resx6_elewise_resx6_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_6 1 2 resx6_elewise_resx6_elewise_relu resx6_elewise_resx6_elewise_relu_splitncnn_0 resx6_elewise_resx6_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx7_conv1 1 1 resx6_elewise_resx6_elewise_relu_splitncnn_1 resx7_conv1_resx7_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle7 1 1 resx7_conv1_resx7_conv1_relu shuffle7 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx7_conv2 1 1 shuffle7 resx7_conv2_resx7_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx7_conv3 1 1 resx7_conv2_resx7_conv2_scale resx7_conv3_resx7_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx7_elewise 2 1 resx6_elewise_resx6_elewise_relu_splitncnn_0 resx7_conv3_resx7_conv3_scale resx7_elewise -23330=4,3,14,14,480 0=1\nReLU resx7_elewise_relu 1 1 resx7_elewise resx7_elewise_resx7_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_7 1 2 resx7_elewise_resx7_elewise_relu resx7_elewise_resx7_elewise_relu_splitncnn_0 resx7_elewise_resx7_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx8_conv1 1 1 resx7_elewise_resx7_elewise_relu_splitncnn_1 resx8_conv1_resx8_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle8 1 1 resx8_conv1_resx8_conv1_relu shuffle8 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx8_conv2 1 1 shuffle8 resx8_conv2_resx8_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx8_conv3 1 1 resx8_conv2_resx8_conv2_scale resx8_conv3_resx8_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx8_elewise 2 1 resx7_elewise_resx7_elewise_relu_splitncnn_0 resx8_conv3_resx8_conv3_scale resx8_elewise -23330=4,3,14,14,480 0=1\nReLU resx8_elewise_relu 1 1 resx8_elewise resx8_elewise_resx8_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_8 1 2 resx8_elewise_resx8_elewise_relu resx8_elewise_resx8_elewise_relu_splitncnn_0 resx8_elewise_resx8_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx9_conv1 1 1 resx8_elewise_resx8_elewise_relu_splitncnn_1 resx9_conv1_resx9_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle9 1 1 resx9_conv1_resx9_conv1_relu shuffle9 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx9_conv2 1 1 shuffle9 resx9_conv2_resx9_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx9_conv3 1 1 resx9_conv2_resx9_conv2_scale resx9_conv3_resx9_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx9_elewise 2 1 resx8_elewise_resx8_elewise_relu_splitncnn_0 resx9_conv3_resx9_conv3_scale resx9_elewise -23330=4,3,14,14,480 0=1\nReLU resx9_elewise_relu 1 1 resx9_elewise resx9_elewise_resx9_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_9 1 2 resx9_elewise_resx9_elewise_relu resx9_elewise_resx9_elewise_relu_splitncnn_0 resx9_elewise_resx9_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx10_conv1 1 1 resx9_elewise_resx9_elewise_relu_splitncnn_1 resx10_conv1_resx10_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle10 1 1 resx10_conv1_resx10_conv1_relu shuffle10 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx10_conv2 1 1 shuffle10 resx10_conv2_resx10_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx10_conv3 1 1 resx10_conv2_resx10_conv2_scale resx10_conv3_resx10_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx10_elewise 2 1 resx9_elewise_resx9_elewise_relu_splitncnn_0 resx10_conv3_resx10_conv3_scale resx10_elewise -23330=4,3,14,14,480 0=1\nReLU resx10_elewise_relu 1 1 resx10_elewise resx10_elewise_resx10_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_10 1 2 resx10_elewise_resx10_elewise_relu resx10_elewise_resx10_elewise_relu_splitncnn_0 resx10_elewise_resx10_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx11_conv1 1 1 resx10_elewise_resx10_elewise_relu_splitncnn_1 resx11_conv1_resx11_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle11 1 1 resx11_conv1_resx11_conv1_relu shuffle11 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx11_conv2 1 1 shuffle11 resx11_conv2_resx11_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx11_conv3 1 1 resx11_conv2_resx11_conv2_scale resx11_conv3_resx11_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx11_elewise 2 1 resx10_elewise_resx10_elewise_relu_splitncnn_0 resx11_conv3_resx11_conv3_scale resx11_elewise -23330=4,3,14,14,480 0=1\nReLU resx11_elewise_relu 1 1 resx11_elewise resx11_elewise_resx11_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_11 1 2 resx11_elewise_resx11_elewise_relu resx11_elewise_resx11_elewise_relu_splitncnn_0 resx11_elewise_resx11_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nConvolutionDepthWise resx12_conv1 1 1 resx11_elewise_resx11_elewise_relu_splitncnn_1 resx12_conv1_resx12_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle12 1 1 resx12_conv1_resx12_conv1_relu shuffle12 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx12_conv2 1 1 shuffle12 resx12_conv2_resx12_conv2_scale -23330=4,3,14,14,120 0=120 1=3 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx12_conv3 1 1 resx12_conv2_resx12_conv2_scale resx12_conv3_resx12_conv3_scale -23330=4,3,14,14,480 0=480 1=1 5=1 6=19200 7=3\nEltwise resx12_elewise 2 1 resx11_elewise_resx11_elewise_relu_splitncnn_0 resx12_conv3_resx12_conv3_scale resx12_elewise -23330=4,3,14,14,480 0=1\nReLU resx12_elewise_relu 1 1 resx12_elewise resx12_elewise_resx12_elewise_relu -23330=4,3,14,14,480\nSplit splitncnn_12 1 2 resx12_elewise_resx12_elewise_relu resx12_elewise_resx12_elewise_relu_splitncnn_0 resx12_elewise_resx12_elewise_relu_splitncnn_1 -23330=8,3,14,14,480,3,14,14,480\nPooling resx13_match_conv 1 1 resx12_elewise_resx12_elewise_relu_splitncnn_1 resx13_match_conv -23330=4,3,7,7,480 0=1 1=3 2=2\nConvolutionDepthWise resx13_conv1 1 1 resx12_elewise_resx12_elewise_relu_splitncnn_0 resx13_conv1_resx13_conv1_relu -23330=4,3,14,14,120 0=120 1=1 5=1 6=19200 7=3 9=1\nShuffleChannel shuffle13 1 1 resx13_conv1_resx13_conv1_relu shuffle13 -23330=4,3,14,14,120 0=3\nConvolutionDepthWise resx13_conv2 1 1 shuffle13 resx13_conv2_resx13_conv2_scale -23330=4,3,7,7,120 0=120 1=3 3=2 4=1 5=1 6=1080 7=120\nConvolutionDepthWise resx13_conv3 1 1 resx13_conv2_resx13_conv2_scale resx13_conv3_resx13_conv3_scale -23330=4,3,7,7,480 0=480 1=1 5=1 6=19200 7=3\nConcat resx13_concat 2 1 resx13_match_conv resx13_conv3_resx13_conv3_scale resx13_concat -23330=4,3,7,7,960\nReLU resx13_concat_relu 1 1 resx13_concat resx13_concat_resx13_concat_relu -23330=4,3,7,7,960\nSplit splitncnn_13 1 2 resx13_concat_resx13_concat_relu resx13_concat_resx13_concat_relu_splitncnn_0 resx13_concat_resx13_concat_relu_splitncnn_1 -23330=8,3,7,7,960,3,7,7,960\nConvolutionDepthWise resx14_conv1 1 1 resx13_concat_resx13_concat_relu_splitncnn_1 resx14_conv1_resx14_conv1_relu -23330=4,3,7,7,240 0=240 1=1 5=1 6=76800 7=3 9=1\nShuffleChannel shuffle14 1 1 resx14_conv1_resx14_conv1_relu shuffle14 -23330=4,3,7,7,240 0=3\nConvolutionDepthWise resx14_conv2 1 1 shuffle14 resx14_conv2_resx14_conv2_scale -23330=4,3,7,7,240 0=240 1=3 4=1 5=1 6=2160 7=240\nConvolutionDepthWise resx14_conv3 1 1 resx14_conv2_resx14_conv2_scale resx14_conv3_resx14_conv3_scale -23330=4,3,7,7,960 0=960 1=1 5=1 6=76800 7=3\nEltwise resx14_elewise 2 1 resx13_concat_resx13_concat_relu_splitncnn_0 resx14_conv3_resx14_conv3_scale resx14_elewise -23330=4,3,7,7,960 0=1\nReLU resx14_elewise_relu 1 1 resx14_elewise resx14_elewise_resx14_elewise_relu -23330=4,3,7,7,960\nSplit splitncnn_14 1 2 resx14_elewise_resx14_elewise_relu resx14_elewise_resx14_elewise_relu_splitncnn_0 resx14_elewise_resx14_elewise_relu_splitncnn_1 -23330=8,3,7,7,960,3,7,7,960\nConvolutionDepthWise resx15_conv1 1 1 resx14_elewise_resx14_elewise_relu_splitncnn_1 resx15_conv1_resx15_conv1_relu -23330=4,3,7,7,240 0=240 1=1 5=1 6=76800 7=3 9=1\nShuffleChannel shuffle15 1 1 resx15_conv1_resx15_conv1_relu shuffle15 -23330=4,3,7,7,240 0=3\nConvolutionDepthWise resx15_conv2 1 1 shuffle15 resx15_conv2_resx15_conv2_scale -23330=4,3,7,7,240 0=240 1=3 4=1 5=1 6=2160 7=240\nConvolutionDepthWise resx15_conv3 1 1 resx15_conv2_resx15_conv2_scale resx15_conv3_resx15_conv3_scale -23330=4,3,7,7,960 0=960 1=1 5=1 6=76800 7=3\nEltwise resx15_elewise 2 1 resx14_elewise_resx14_elewise_relu_splitncnn_0 resx15_conv3_resx15_conv3_scale resx15_elewise -23330=4,3,7,7,960 0=1\nReLU resx15_elewise_relu 1 1 resx15_elewise resx15_elewise_resx15_elewise_relu -23330=4,3,7,7,960\nSplit splitncnn_15 1 2 resx15_elewise_resx15_elewise_relu resx15_elewise_resx15_elewise_relu_splitncnn_0 resx15_elewise_resx15_elewise_relu_splitncnn_1 -23330=8,3,7,7,960,3,7,7,960\nConvolutionDepthWise resx16_conv1 1 1 resx15_elewise_resx15_elewise_relu_splitncnn_1 resx16_conv1_resx16_conv1_relu -23330=4,3,7,7,240 0=240 1=1 5=1 6=76800 7=3 9=1\nShuffleChannel shuffle16 1 1 resx16_conv1_resx16_conv1_relu shuffle16 -23330=4,3,7,7,240 0=3\nConvolutionDepthWise resx16_conv2 1 1 shuffle16 resx16_conv2_resx16_conv2_scale -23330=4,3,7,7,240 0=240 1=3 4=1 5=1 6=2160 7=240\nConvolutionDepthWise resx16_conv3 1 1 resx16_conv2_resx16_conv2_scale resx16_conv3_resx16_conv3_scale -23330=4,3,7,7,960 0=960 1=1 5=1 6=76800 7=3\nEltwise resx16_elewise 2 1 resx15_elewise_resx15_elewise_relu_splitncnn_0 resx16_conv3_resx16_conv3_scale resx16_elewise -23330=4,3,7,7,960 0=1\nReLU resx16_elewise_relu 1 1 resx16_elewise resx16_elewise_resx16_elewise_relu -23330=4,3,7,7,960\nPooling pool_ave 1 1 resx16_elewise_resx16_elewise_relu pool_ave -23330=4,1,960,1,1 0=1 4=1\nInnerProduct fc1000 1 1 pool_ave fc1000 -23330=4,1,1000,1,1 0=1000 1=1 2=960000\nSoftmax prob 1 1 fc1000 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/shufflenet_v2.param", "content": "7767517\n109 125\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1 1 1 data conv1_conv1_relu -23330=4,3,112,112,24 0=24 1=3 3=2 4=1 5=1 6=648 9=1\nPooling pool1 1 1 conv1_conv1_relu pool1 -23330=4,3,56,56,24 1=3 2=2\nSplit splitncnn_0 1 2 pool1 pool1_splitncnn_0 pool1_splitncnn_1 -23330=8,3,56,56,24,3,56,56,24\nConvolutionDepthWise branch1_1_conv1 1 1 pool1_splitncnn_1 branch1_1_conv1_branch1_1_conv1_scale -23330=4,3,28,28,24 0=24 1=3 3=2 4=1 5=1 6=216 7=24\nConvolution branch1_1_conv2 1 1 branch1_1_conv1_branch1_1_conv1_scale branch1_1_conv2_branch1_1_conv2_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=1392 9=1\nConvolution branch1_2_conv1 1 1 pool1_splitncnn_0 branch1_2_conv1_branch1_2_conv1_relu -23330=4,3,56,56,58 0=58 1=1 5=1 6=1392 9=1\nConvolutionDepthWise branch1_2_conv2 1 1 branch1_2_conv1_branch1_2_conv1_relu branch1_2_conv2_branch1_2_conv2_scale -23330=4,3,28,28,58 0=58 1=3 3=2 4=1 5=1 6=522 7=58\nConvolution branch1_2_conv3 1 1 branch1_2_conv2_branch1_2_conv2_scale branch1_2_conv3_branch1_2_conv3_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConcat concat1 2 1 branch1_1_conv2_branch1_1_conv2_relu branch1_2_conv3_branch1_2_conv3_relu concat1 -23330=4,3,28,28,116\nShuffleChannel shuffle1 1 1 concat1 shuffle1 -23330=4,3,28,28,116 0=2\nSlice slice2 1 2 shuffle1 branch2_1 branch2_2 -23330=8,3,28,28,58,3,28,28,58 -23300=2,58,-233\nConvolution branch2_2_conv1 1 1 branch2_2 branch2_2_conv1_branch2_2_conv1_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConvolutionDepthWise branch2_2_conv2 1 1 branch2_2_conv1_branch2_2_conv1_relu branch2_2_conv2_branch2_2_conv2_scale -23330=4,3,28,28,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution branch2_2_conv3 1 1 branch2_2_conv2_branch2_2_conv2_scale branch2_2_conv3_branch2_2_conv3_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConcat concat2 2 1 branch2_1 branch2_2_conv3_branch2_2_conv3_relu concat2 -23330=4,3,28,28,116\nShuffleChannel shuffle2 1 1 concat2 shuffle2 -23330=4,3,28,28,116 0=2\nSlice slice3 1 2 shuffle2 branch3_1 branch3_2 -23330=8,3,28,28,58,3,28,28,58 -23300=2,58,-233\nConvolution branch3_2_conv1 1 1 branch3_2 branch3_2_conv1_branch3_2_conv1_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConvolutionDepthWise branch3_2_conv2 1 1 branch3_2_conv1_branch3_2_conv1_relu branch3_2_conv2_branch3_2_conv2_scale -23330=4,3,28,28,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution branch3_2_conv3 1 1 branch3_2_conv2_branch3_2_conv2_scale branch3_2_conv3_branch3_2_conv3_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConcat concat3 2 1 branch3_1 branch3_2_conv3_branch3_2_conv3_relu concat3 -23330=4,3,28,28,116\nShuffleChannel shuffle3 1 1 concat3 shuffle3 -23330=4,3,28,28,116 0=2\nSlice slice4 1 2 shuffle3 branch4_1 branch4_2 -23330=8,3,28,28,58,3,28,28,58 -23300=2,58,-233\nConvolution branch4_2_conv1 1 1 branch4_2 branch4_2_conv1_branch4_2_conv1_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConvolutionDepthWise branch4_2_conv2 1 1 branch4_2_conv1_branch4_2_conv1_relu branch4_2_conv2_branch4_2_conv2_scale -23330=4,3,28,28,58 0=58 1=3 4=1 5=1 6=522 7=58\nConvolution branch4_2_conv3 1 1 branch4_2_conv2_branch4_2_conv2_scale branch4_2_conv3_branch4_2_conv3_relu -23330=4,3,28,28,58 0=58 1=1 5=1 6=3364 9=1\nConcat concat4 2 1 branch4_1 branch4_2_conv3_branch4_2_conv3_relu concat4 -23330=4,3,28,28,116\nShuffleChannel shuffle4 1 1 concat4 shuffle4 -23330=4,3,28,28,116 0=2\nSplit splitncnn_1 1 2 shuffle4 shuffle4_splitncnn_0 shuffle4_splitncnn_1 -23330=8,3,28,28,116,3,28,28,116\nConvolutionDepthWise branch5_1_conv1 1 1 shuffle4_splitncnn_1 branch5_1_conv1_branch5_1_conv1_scale -23330=4,3,14,14,116 0=116 1=3 3=2 4=1 5=1 6=1044 7=116\nConvolution branch5_1_conv2 1 1 branch5_1_conv1_branch5_1_conv1_scale branch5_1_conv2_branch5_1_conv2_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolution branch5_2_conv1 1 1 shuffle4_splitncnn_0 branch5_2_conv1_branch5_2_conv1_relu -23330=4,3,28,28,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch5_2_conv2 1 1 branch5_2_conv1_branch5_2_conv1_relu branch5_2_conv2_branch5_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 3=2 4=1 5=1 6=1044 7=116\nConvolution branch5_2_conv3 1 1 branch5_2_conv2_branch5_2_conv2_scale branch5_2_conv3_branch5_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat5 2 1 branch5_1_conv2_branch5_1_conv2_relu branch5_2_conv3_branch5_2_conv3_relu concat5 -23330=4,3,14,14,232\nShuffleChannel shuffle5 1 1 concat5 shuffle5 -23330=4,3,14,14,232 0=2\nSlice slice6 1 2 shuffle5 branch6_1 branch6_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch6_2_conv1 1 1 branch6_2 branch6_2_conv1_branch6_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch6_2_conv2 1 1 branch6_2_conv1_branch6_2_conv1_relu branch6_2_conv2_branch6_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch6_2_conv3 1 1 branch6_2_conv2_branch6_2_conv2_scale branch6_2_conv3_branch6_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat6 2 1 branch6_1 branch6_2_conv3_branch6_2_conv3_relu concat6 -23330=4,3,14,14,232\nShuffleChannel shuffle6 1 1 concat6 shuffle6 -23330=4,3,14,14,232 0=2\nSlice slice7 1 2 shuffle6 branch7_1 branch7_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch7_2_conv1 1 1 branch7_2 branch7_2_conv1_branch7_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch7_2_conv2 1 1 branch7_2_conv1_branch7_2_conv1_relu branch7_2_conv2_branch7_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch7_2_conv3 1 1 branch7_2_conv2_branch7_2_conv2_scale branch7_2_conv3_branch7_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat7 2 1 branch7_1 branch7_2_conv3_branch7_2_conv3_relu concat7 -23330=4,3,14,14,232\nShuffleChannel shuffle7 1 1 concat7 shuffle7 -23330=4,3,14,14,232 0=2\nSlice slice8 1 2 shuffle7 branch8_1 branch8_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch8_2_conv1 1 1 branch8_2 branch8_2_conv1_branch8_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch8_2_conv2 1 1 branch8_2_conv1_branch8_2_conv1_relu branch8_2_conv2_branch8_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch8_2_conv3 1 1 branch8_2_conv2_branch8_2_conv2_scale branch8_2_conv3_branch8_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat8 2 1 branch8_1 branch8_2_conv3_branch8_2_conv3_relu concat8 -23330=4,3,14,14,232\nShuffleChannel shuffle8 1 1 concat8 shuffle8 -23330=4,3,14,14,232 0=2\nSlice slice9 1 2 shuffle8 branch9_1 branch9_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch9_2_conv1 1 1 branch9_2 branch9_2_conv1_branch9_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch9_2_conv2 1 1 branch9_2_conv1_branch9_2_conv1_relu branch9_2_conv2_branch9_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch9_2_conv3 1 1 branch9_2_conv2_branch9_2_conv2_scale branch9_2_conv3_branch9_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat9 2 1 branch9_1 branch9_2_conv3_branch9_2_conv3_relu concat9 -23330=4,3,14,14,232\nShuffleChannel shuffle9 1 1 concat9 shuffle9 -23330=4,3,14,14,232 0=2\nSlice slice10 1 2 shuffle9 branch10_1 branch10_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch10_2_conv1 1 1 branch10_2 branch10_2_conv1_branch10_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch10_2_conv2 1 1 branch10_2_conv1_branch10_2_conv1_relu branch10_2_conv2_branch10_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch10_2_conv3 1 1 branch10_2_conv2_branch10_2_conv2_scale branch10_2_conv3_branch10_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat10 2 1 branch10_1 branch10_2_conv3_branch10_2_conv3_relu concat10 -23330=4,3,14,14,232\nShuffleChannel shuffle10 1 1 concat10 shuffle10 -23330=4,3,14,14,232 0=2\nSlice slice11 1 2 shuffle10 branch11_1 branch11_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch11_2_conv1 1 1 branch11_2 branch11_2_conv1_branch11_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch11_2_conv2 1 1 branch11_2_conv1_branch11_2_conv1_relu branch11_2_conv2_branch11_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch11_2_conv3 1 1 branch11_2_conv2_branch11_2_conv2_scale branch11_2_conv3_branch11_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat11 2 1 branch11_1 branch11_2_conv3_branch11_2_conv3_relu concat11 -23330=4,3,14,14,232\nShuffleChannel shuffle11 1 1 concat11 shuffle11 -23330=4,3,14,14,232 0=2\nSlice slice12 1 2 shuffle11 branch12_1 branch12_2 -23330=8,3,14,14,116,3,14,14,116 -23300=2,116,-233\nConvolution branch12_2_conv1 1 1 branch12_2 branch12_2_conv1_branch12_2_conv1_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConvolutionDepthWise branch12_2_conv2 1 1 branch12_2_conv1_branch12_2_conv1_relu branch12_2_conv2_branch12_2_conv2_scale -23330=4,3,14,14,116 0=116 1=3 4=1 5=1 6=1044 7=116\nConvolution branch12_2_conv3 1 1 branch12_2_conv2_branch12_2_conv2_scale branch12_2_conv3_branch12_2_conv3_relu -23330=4,3,14,14,116 0=116 1=1 5=1 6=13456 9=1\nConcat concat12 2 1 branch12_1 branch12_2_conv3_branch12_2_conv3_relu concat12 -23330=4,3,14,14,232\nShuffleChannel shuffle12 1 1 concat12 shuffle12 -23330=4,3,14,14,232 0=2\nSplit splitncnn_2 1 2 shuffle12 shuffle12_splitncnn_0 shuffle12_splitncnn_1 -23330=8,3,14,14,232,3,14,14,232\nConvolutionDepthWise branch13_1_conv1 1 1 shuffle12_splitncnn_1 branch13_1_conv1_branch13_1_conv1_scale -23330=4,3,7,7,232 0=232 1=3 3=2 4=1 5=1 6=2088 7=232\nConvolution branch13_1_conv2 1 1 branch13_1_conv1_branch13_1_conv1_scale branch13_1_conv2_branch13_1_conv2_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConvolution branch13_2_conv1 1 1 shuffle12_splitncnn_0 branch13_2_conv1_branch13_2_conv1_relu -23330=4,3,14,14,232 0=232 1=1 5=1 6=53824 9=1\nConvolutionDepthWise branch13_2_conv2 1 1 branch13_2_conv1_branch13_2_conv1_relu branch13_2_conv2_branch13_2_conv2_scale -23330=4,3,7,7,232 0=232 1=3 3=2 4=1 5=1 6=2088 7=232\nConvolution branch13_2_conv3 1 1 branch13_2_conv2_branch13_2_conv2_scale branch13_2_conv3_branch13_2_conv3_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConcat concat13 2 1 branch13_1_conv2_branch13_1_conv2_relu branch13_2_conv3_branch13_2_conv3_relu concat13 -23330=4,3,7,7,464\nShuffleChannel shuffle13 1 1 concat13 shuffle13 -23330=4,3,7,7,464 0=2\nSlice slice14 1 2 shuffle13 branch14_1 branch14_2 -23330=8,3,7,7,232,3,7,7,232 -23300=2,232,-233\nConvolution branch14_2_conv1 1 1 branch14_2 branch14_2_conv1_branch14_2_conv1_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConvolutionDepthWise branch14_2_conv2 1 1 branch14_2_conv1_branch14_2_conv1_relu branch14_2_conv2_branch14_2_conv2_scale -23330=4,3,7,7,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution branch14_2_conv3 1 1 branch14_2_conv2_branch14_2_conv2_scale branch14_2_conv3_branch14_2_conv3_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConcat concat14 2 1 branch14_1 branch14_2_conv3_branch14_2_conv3_relu concat14 -23330=4,3,7,7,464\nShuffleChannel shuffle14 1 1 concat14 shuffle14 -23330=4,3,7,7,464 0=2\nSlice slice15 1 2 shuffle14 branch15_1 branch15_2 -23330=8,3,7,7,232,3,7,7,232 -23300=2,232,-233\nConvolution branch15_2_conv1 1 1 branch15_2 branch15_2_conv1_branch15_2_conv1_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConvolutionDepthWise branch15_2_conv2 1 1 branch15_2_conv1_branch15_2_conv1_relu branch15_2_conv2_branch15_2_conv2_scale -23330=4,3,7,7,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution branch15_2_conv3 1 1 branch15_2_conv2_branch15_2_conv2_scale branch15_2_conv3_branch15_2_conv3_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConcat concat15 2 1 branch15_1 branch15_2_conv3_branch15_2_conv3_relu concat15 -23330=4,3,7,7,464\nShuffleChannel shuffle15 1 1 concat15 shuffle15 -23330=4,3,7,7,464 0=2\nSlice slice16 1 2 shuffle15 branch16_1 branch16_2 -23330=8,3,7,7,232,3,7,7,232 -23300=2,232,-233\nConvolution branch16_2_conv1 1 1 branch16_2 branch16_2_conv1_branch16_2_conv1_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConvolutionDepthWise branch16_2_conv2 1 1 branch16_2_conv1_branch16_2_conv1_relu branch16_2_conv2_branch16_2_conv2_scale -23330=4,3,7,7,232 0=232 1=3 4=1 5=1 6=2088 7=232\nConvolution branch16_2_conv3 1 1 branch16_2_conv2_branch16_2_conv2_scale branch16_2_conv3_branch16_2_conv3_relu -23330=4,3,7,7,232 0=232 1=1 5=1 6=53824 9=1\nConcat concat16 2 1 branch16_1 branch16_2_conv3_branch16_2_conv3_relu concat16 -23330=4,3,7,7,464\nShuffleChannel shuffle16 1 1 concat16 shuffle16 -23330=4,3,7,7,464 0=2\nConvolution conv5 1 1 shuffle16 conv5_conv5_relu -23330=4,3,7,7,1024 0=1024 1=1 5=1 6=475136 9=1\nPooling pool_ave 1 1 conv5_conv5_relu pool_ave -23330=4,1,1024,1,1 0=1 4=1\nInnerProduct fc1000 1 1 pool_ave fc1000 -23330=4,1,1000,1,1 0=1000 1=1 2=1024000\nSoftmax prob 1 1 fc1000 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/squeezenet.param", "content": "7767517\n48 56\nInput data 0 1 data -23330=4,3,227,227,3 0=227 1=227 2=3\nConvolution conv1 1 1 data conv1_relu_conv1 -23330=4,3,113,113,64 0=64 1=3 3=2 5=1 6=1728 9=1\nPooling pool1 1 1 conv1_relu_conv1 pool1 -23330=4,3,56,56,64 1=3 2=2\nConvolution fire2/squeeze1x1 1 1 pool1 fire2/squeeze1x1_fire2/relu_squeeze1x1 -23330=4,3,56,56,16 0=16 1=1 5=1 6=1024 9=1\nSplit splitncnn_0 1 2 fire2/squeeze1x1_fire2/relu_squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 -23330=8,3,56,56,16,3,56,56,16\nConvolution fire2/expand1x1 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 fire2/expand1x1_fire2/relu_expand1x1 -23330=4,3,56,56,64 0=64 1=1 5=1 6=1024 9=1\nConvolution fire2/expand3x3 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/expand3x3_fire2/relu_expand3x3 -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=9216 9=1\nConcat fire2/concat 2 1 fire2/expand1x1_fire2/relu_expand1x1 fire2/expand3x3_fire2/relu_expand3x3 fire2/concat -23330=4,3,56,56,128\nConvolution fire3/squeeze1x1 1 1 fire2/concat fire3/squeeze1x1_fire3/relu_squeeze1x1 -23330=4,3,56,56,16 0=16 1=1 5=1 6=2048 9=1\nSplit splitncnn_1 1 2 fire3/squeeze1x1_fire3/relu_squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 -23330=8,3,56,56,16,3,56,56,16\nConvolution fire3/expand1x1 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 fire3/expand1x1_fire3/relu_expand1x1 -23330=4,3,56,56,64 0=64 1=1 5=1 6=1024 9=1\nConvolution fire3/expand3x3 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/expand3x3_fire3/relu_expand3x3 -23330=4,3,56,56,64 0=64 1=3 4=1 5=1 6=9216 9=1\nConcat fire3/concat 2 1 fire3/expand1x1_fire3/relu_expand1x1 fire3/expand3x3_fire3/relu_expand3x3 fire3/concat -23330=4,3,56,56,128\nPooling pool3 1 1 fire3/concat pool3 -23330=4,3,28,28,128 1=3 2=2\nConvolution fire4/squeeze1x1 1 1 pool3 fire4/squeeze1x1_fire4/relu_squeeze1x1 -23330=4,3,28,28,32 0=32 1=1 5=1 6=4096 9=1\nSplit splitncnn_2 1 2 fire4/squeeze1x1_fire4/relu_squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 -23330=8,3,28,28,32,3,28,28,32\nConvolution fire4/expand1x1 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 fire4/expand1x1_fire4/relu_expand1x1 -23330=4,3,28,28,128 0=128 1=1 5=1 6=4096 9=1\nConvolution fire4/expand3x3 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/expand3x3_fire4/relu_expand3x3 -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=36864 9=1\nConcat fire4/concat 2 1 fire4/expand1x1_fire4/relu_expand1x1 fire4/expand3x3_fire4/relu_expand3x3 fire4/concat -23330=4,3,28,28,256\nConvolution fire5/squeeze1x1 1 1 fire4/concat fire5/squeeze1x1_fire5/relu_squeeze1x1 -23330=4,3,28,28,32 0=32 1=1 5=1 6=8192 9=1\nSplit splitncnn_3 1 2 fire5/squeeze1x1_fire5/relu_squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 -23330=8,3,28,28,32,3,28,28,32\nConvolution fire5/expand1x1 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 fire5/expand1x1_fire5/relu_expand1x1 -23330=4,3,28,28,128 0=128 1=1 5=1 6=4096 9=1\nConvolution fire5/expand3x3 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/expand3x3_fire5/relu_expand3x3 -23330=4,3,28,28,128 0=128 1=3 4=1 5=1 6=36864 9=1\nConcat fire5/concat 2 1 fire5/expand1x1_fire5/relu_expand1x1 fire5/expand3x3_fire5/relu_expand3x3 fire5/concat -23330=4,3,28,28,256\nPooling pool5 1 1 fire5/concat pool5 -23330=4,3,14,14,256 1=3 2=2\nConvolution fire6/squeeze1x1 1 1 pool5 fire6/squeeze1x1_fire6/relu_squeeze1x1 -23330=4,3,14,14,48 0=48 1=1 5=1 6=12288 9=1\nSplit splitncnn_4 1 2 fire6/squeeze1x1_fire6/relu_squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 -23330=8,3,14,14,48,3,14,14,48\nConvolution fire6/expand1x1 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 fire6/expand1x1_fire6/relu_expand1x1 -23330=4,3,14,14,192 0=192 1=1 5=1 6=9216 9=1\nConvolution fire6/expand3x3 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/expand3x3_fire6/relu_expand3x3 -23330=4,3,14,14,192 0=192 1=3 4=1 5=1 6=82944 9=1\nConcat fire6/concat 2 1 fire6/expand1x1_fire6/relu_expand1x1 fire6/expand3x3_fire6/relu_expand3x3 fire6/concat -23330=4,3,14,14,384\nConvolution fire7/squeeze1x1 1 1 fire6/concat fire7/squeeze1x1_fire7/relu_squeeze1x1 -23330=4,3,14,14,48 0=48 1=1 5=1 6=18432 9=1\nSplit splitncnn_5 1 2 fire7/squeeze1x1_fire7/relu_squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 -23330=8,3,14,14,48,3,14,14,48\nConvolution fire7/expand1x1 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 fire7/expand1x1_fire7/relu_expand1x1 -23330=4,3,14,14,192 0=192 1=1 5=1 6=9216 9=1\nConvolution fire7/expand3x3 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/expand3x3_fire7/relu_expand3x3 -23330=4,3,14,14,192 0=192 1=3 4=1 5=1 6=82944 9=1\nConcat fire7/concat 2 1 fire7/expand1x1_fire7/relu_expand1x1 fire7/expand3x3_fire7/relu_expand3x3 fire7/concat -23330=4,3,14,14,384\nConvolution fire8/squeeze1x1 1 1 fire7/concat fire8/squeeze1x1_fire8/relu_squeeze1x1 -23330=4,3,14,14,64 0=64 1=1 5=1 6=24576 9=1\nSplit splitncnn_6 1 2 fire8/squeeze1x1_fire8/relu_squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 -23330=8,3,14,14,64,3,14,14,64\nConvolution fire8/expand1x1 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 fire8/expand1x1_fire8/relu_expand1x1 -23330=4,3,14,14,256 0=256 1=1 5=1 6=16384 9=1\nConvolution fire8/expand3x3 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/expand3x3_fire8/relu_expand3x3 -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=147456 9=1\nConcat fire8/concat 2 1 fire8/expand1x1_fire8/relu_expand1x1 fire8/expand3x3_fire8/relu_expand3x3 fire8/concat -23330=4,3,14,14,512\nConvolution fire9/squeeze1x1 1 1 fire8/concat fire9/squeeze1x1_fire9/relu_squeeze1x1 -23330=4,3,14,14,64 0=64 1=1 5=1 6=32768 9=1\nSplit splitncnn_7 1 2 fire9/squeeze1x1_fire9/relu_squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 -23330=8,3,14,14,64,3,14,14,64\nConvolution fire9/expand1x1 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 fire9/expand1x1_fire9/relu_expand1x1 -23330=4,3,14,14,256 0=256 1=1 5=1 6=16384 9=1\nConvolution fire9/expand3x3 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/expand3x3_fire9/relu_expand3x3 -23330=4,3,14,14,256 0=256 1=3 4=1 5=1 6=147456 9=1\nConcat fire9/concat 2 1 fire9/expand1x1_fire9/relu_expand1x1 fire9/expand3x3_fire9/relu_expand3x3 fire9/concat_drop9 -23330=4,3,14,14,512\nConvolution conv10 1 1 fire9/concat_drop9 conv10_relu_conv10 -23330=4,3,16,16,1000 0=1000 1=1 4=1 5=1 6=512000 9=1\nPooling pool10 1 1 conv10_relu_conv10 pool10 -23330=4,1,1000,1,1 0=1 4=1\nSoftmax prob 1 1 pool10 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/squeezenet_int8.param", "content": "7767517\n48 56\nInput data 0 1 data 0=227 1=227 2=3\nConvolution conv1 1 1 data conv1_relu_conv1 0=64 1=3 3=2 5=1 6=1728 8=2 9=1\nPooling pool1 1 1 conv1_relu_conv1 pool1 1=3 2=2\nConvolution fire2/squeeze1x1 1 1 pool1 fire2/squeeze1x1_fire2/relu_squeeze1x1 0=16 1=1 5=1 6=1024 8=102 9=1\nSplit splitncnn_0 1 2 fire2/squeeze1x1_fire2/relu_squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1\nConvolution fire2/expand1x1 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 fire2/expand1x1_fire2/relu_expand1x1 0=64 1=1 5=1 6=1024 8=2 9=1\nConvolution fire2/expand3x3 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/expand3x3_fire2/relu_expand3x3 0=64 1=3 4=1 5=1 6=9216 8=2 9=1\nConcat fire2/concat 2 1 fire2/expand1x1_fire2/relu_expand1x1 fire2/expand3x3_fire2/relu_expand3x3 fire2/concat\nConvolution fire3/squeeze1x1 1 1 fire2/concat fire3/squeeze1x1_fire3/relu_squeeze1x1 0=16 1=1 5=1 6=2048 8=102 9=1\nSplit splitncnn_1 1 2 fire3/squeeze1x1_fire3/relu_squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1\nConvolution fire3/expand1x1 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 fire3/expand1x1_fire3/relu_expand1x1 0=64 1=1 5=1 6=1024 8=2 9=1\nConvolution fire3/expand3x3 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/expand3x3_fire3/relu_expand3x3 0=64 1=3 4=1 5=1 6=9216 8=2 9=1\nConcat fire3/concat 2 1 fire3/expand1x1_fire3/relu_expand1x1 fire3/expand3x3_fire3/relu_expand3x3 fire3/concat\nPooling pool3 1 1 fire3/concat pool3 1=3 2=2\nConvolution fire4/squeeze1x1 1 1 pool3 fire4/squeeze1x1_fire4/relu_squeeze1x1 0=32 1=1 5=1 6=4096 8=102 9=1\nSplit splitncnn_2 1 2 fire4/squeeze1x1_fire4/relu_squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1\nConvolution fire4/expand1x1 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 fire4/expand1x1_fire4/relu_expand1x1 0=128 1=1 5=1 6=4096 8=2 9=1\nConvolution fire4/expand3x3 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/expand3x3_fire4/relu_expand3x3 0=128 1=3 4=1 5=1 6=36864 8=2 9=1\nConcat fire4/concat 2 1 fire4/expand1x1_fire4/relu_expand1x1 fire4/expand3x3_fire4/relu_expand3x3 fire4/concat\nConvolution fire5/squeeze1x1 1 1 fire4/concat fire5/squeeze1x1_fire5/relu_squeeze1x1 0=32 1=1 5=1 6=8192 8=102 9=1\nSplit splitncnn_3 1 2 fire5/squeeze1x1_fire5/relu_squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1\nConvolution fire5/expand1x1 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 fire5/expand1x1_fire5/relu_expand1x1 0=128 1=1 5=1 6=4096 8=2 9=1\nConvolution fire5/expand3x3 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/expand3x3_fire5/relu_expand3x3 0=128 1=3 4=1 5=1 6=36864 8=2 9=1\nConcat fire5/concat 2 1 fire5/expand1x1_fire5/relu_expand1x1 fire5/expand3x3_fire5/relu_expand3x3 fire5/concat\nPooling pool5 1 1 fire5/concat pool5 1=3 2=2\nConvolution fire6/squeeze1x1 1 1 pool5 fire6/squeeze1x1_fire6/relu_squeeze1x1 0=48 1=1 5=1 6=12288 8=102 9=1\nSplit splitncnn_4 1 2 fire6/squeeze1x1_fire6/relu_squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1\nConvolution fire6/expand1x1 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 fire6/expand1x1_fire6/relu_expand1x1 0=192 1=1 5=1 6=9216 8=2 9=1\nConvolution fire6/expand3x3 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/expand3x3_fire6/relu_expand3x3 0=192 1=3 4=1 5=1 6=82944 8=2 9=1\nConcat fire6/concat 2 1 fire6/expand1x1_fire6/relu_expand1x1 fire6/expand3x3_fire6/relu_expand3x3 fire6/concat\nConvolution fire7/squeeze1x1 1 1 fire6/concat fire7/squeeze1x1_fire7/relu_squeeze1x1 0=48 1=1 5=1 6=18432 8=102 9=1\nSplit splitncnn_5 1 2 fire7/squeeze1x1_fire7/relu_squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1\nConvolution fire7/expand1x1 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 fire7/expand1x1_fire7/relu_expand1x1 0=192 1=1 5=1 6=9216 8=2 9=1\nConvolution fire7/expand3x3 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/expand3x3_fire7/relu_expand3x3 0=192 1=3 4=1 5=1 6=82944 8=2 9=1\nConcat fire7/concat 2 1 fire7/expand1x1_fire7/relu_expand1x1 fire7/expand3x3_fire7/relu_expand3x3 fire7/concat\nConvolution fire8/squeeze1x1 1 1 fire7/concat fire8/squeeze1x1_fire8/relu_squeeze1x1 0=64 1=1 5=1 6=24576 8=102 9=1\nSplit splitncnn_6 1 2 fire8/squeeze1x1_fire8/relu_squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1\nConvolution fire8/expand1x1 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 fire8/expand1x1_fire8/relu_expand1x1 0=256 1=1 5=1 6=16384 8=2 9=1\nConvolution fire8/expand3x3 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/expand3x3_fire8/relu_expand3x3 0=256 1=3 4=1 5=1 6=147456 8=2 9=1\nConcat fire8/concat 2 1 fire8/expand1x1_fire8/relu_expand1x1 fire8/expand3x3_fire8/relu_expand3x3 fire8/concat\nConvolution fire9/squeeze1x1 1 1 fire8/concat fire9/squeeze1x1_fire9/relu_squeeze1x1 0=64 1=1 5=1 6=32768 8=102 9=1\nSplit splitncnn_7 1 2 fire9/squeeze1x1_fire9/relu_squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1\nConvolution fire9/expand1x1 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 fire9/expand1x1_fire9/relu_expand1x1 0=256 1=1 5=1 6=16384 8=2 9=1\nConvolution fire9/expand3x3 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/expand3x3_fire9/relu_expand3x3 0=256 1=3 4=1 5=1 6=147456 8=2 9=1\nConcat fire9/concat 2 1 fire9/expand1x1_fire9/relu_expand1x1 fire9/expand3x3_fire9/relu_expand3x3 fire9/concat_drop9\nConvolution conv10 1 1 fire9/concat_drop9 conv10_relu_conv10 0=1000 1=1 4=1 5=1 6=512000 8=2 9=1\nPooling pool10 1 1 conv10_relu_conv10 pool10 0=1 4=1\nSoftmax prob 1 1 pool10 output\n" }, { "path": "benchmark/squeezenet_ssd.param", "content": "7767517\n119 152\nInput data 0 1 data -23330=4,3,300,300,3 0=300 1=300 2=3\nSplit splitncnn_0 1 7 data data_splitncnn_0 data_splitncnn_1 data_splitncnn_2 data_splitncnn_3 data_splitncnn_4 data_splitncnn_5 data_splitncnn_6 -23330=28,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3,3,300,300,3\nConvolution conv1 1 1 data_splitncnn_6 conv1_relu_conv1 -23330=4,3,149,149,64 0=64 1=3 3=2 5=1 6=1728 9=1\nPooling pool1 1 1 conv1_relu_conv1 pool1 -23330=4,3,74,74,64 1=3 2=2\nConvolution fire2/squeeze1x1 1 1 pool1 fire2/squeeze1x1_fire2/relu_squeeze1x1 -23330=4,3,74,74,16 0=16 1=1 5=1 6=1024 9=1\nSplit splitncnn_1 1 2 fire2/squeeze1x1_fire2/relu_squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 -23330=8,3,74,74,16,3,74,74,16\nConvolution fire2/expand1x1 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 fire2/expand1x1_fire2/relu_expand1x1 -23330=4,3,74,74,64 0=64 1=1 5=1 6=1024 9=1\nConvolution fire2/expand3x3 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/expand3x3_fire2/relu_expand3x3 -23330=4,3,74,74,64 0=64 1=3 4=1 5=1 6=9216 9=1\nConcat fire2/concat 2 1 fire2/expand1x1_fire2/relu_expand1x1 fire2/expand3x3_fire2/relu_expand3x3 fire2/concat -23330=4,3,74,74,128\nConvolution fire3/squeeze1x1 1 1 fire2/concat fire3/squeeze1x1_fire3/relu_squeeze1x1 -23330=4,3,74,74,16 0=16 1=1 5=1 6=2048 9=1\nSplit splitncnn_2 1 2 fire3/squeeze1x1_fire3/relu_squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 -23330=8,3,74,74,16,3,74,74,16\nConvolution fire3/expand1x1 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 fire3/expand1x1_fire3/relu_expand1x1 -23330=4,3,74,74,64 0=64 1=1 5=1 6=1024 9=1\nConvolution fire3/expand3x3 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/expand3x3_fire3/relu_expand3x3 -23330=4,3,74,74,64 0=64 1=3 4=1 5=1 6=9216 9=1\nConcat fire3/concat 2 1 fire3/expand1x1_fire3/relu_expand1x1 fire3/expand3x3_fire3/relu_expand3x3 fire3/concat -23330=4,3,74,74,128\nPooling pool3 1 1 fire3/concat pool3 -23330=4,3,37,37,128 1=3 2=2\nConvolution fire4/squeeze1x1 1 1 pool3 fire4/squeeze1x1_fire4/relu_squeeze1x1 -23330=4,3,37,37,32 0=32 1=1 5=1 6=4096 9=1\nSplit splitncnn_3 1 2 fire4/squeeze1x1_fire4/relu_squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 -23330=8,3,37,37,32,3,37,37,32\nConvolution fire4/expand1x1 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 fire4/expand1x1_fire4/relu_expand1x1 -23330=4,3,37,37,128 0=128 1=1 5=1 6=4096 9=1\nConvolution fire4/expand3x3 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/expand3x3_fire4/relu_expand3x3 -23330=4,3,37,37,128 0=128 1=3 4=1 5=1 6=36864 9=1\nConcat fire4/concat 2 1 fire4/expand1x1_fire4/relu_expand1x1 fire4/expand3x3_fire4/relu_expand3x3 fire4/concat -23330=4,3,37,37,256\nConvolution fire5/squeeze1x1 1 1 fire4/concat fire5/squeeze1x1_fire5/relu_squeeze1x1 -23330=4,3,37,37,32 0=32 1=1 5=1 6=8192 9=1\nSplit splitncnn_4 1 2 fire5/squeeze1x1_fire5/relu_squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 -23330=8,3,37,37,32,3,37,37,32\nConvolution fire5/expand1x1 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 fire5/expand1x1_fire5/relu_expand1x1 -23330=4,3,37,37,128 0=128 1=1 5=1 6=4096 9=1\nConvolution fire5/expand3x3 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/expand3x3_fire5/relu_expand3x3 -23330=4,3,37,37,128 0=128 1=3 4=1 5=1 6=36864 9=1\nConcat fire5/concat 2 1 fire5/expand1x1_fire5/relu_expand1x1 fire5/expand3x3_fire5/relu_expand3x3 fire5/concat -23330=4,3,37,37,256\nSplit splitncnn_5 1 2 fire5/concat fire5/concat_splitncnn_0 fire5/concat_splitncnn_1 -23330=8,3,37,37,256,3,37,37,256\nPooling pool5 1 1 fire5/concat_splitncnn_1 pool5 -23330=4,3,18,18,256 1=3 2=2\nConvolution fire6/squeeze1x1 1 1 pool5 fire6/squeeze1x1_fire6/relu_squeeze1x1 -23330=4,3,18,18,48 0=48 1=1 5=1 6=12288 9=1\nSplit splitncnn_6 1 2 fire6/squeeze1x1_fire6/relu_squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 -23330=8,3,18,18,48,3,18,18,48\nConvolution fire6/expand1x1 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 fire6/expand1x1_fire6/relu_expand1x1 -23330=4,3,18,18,192 0=192 1=1 5=1 6=9216 9=1\nConvolution fire6/expand3x3 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/expand3x3_fire6/relu_expand3x3 -23330=4,3,18,18,192 0=192 1=3 4=1 5=1 6=82944 9=1\nConcat fire6/concat 2 1 fire6/expand1x1_fire6/relu_expand1x1 fire6/expand3x3_fire6/relu_expand3x3 fire6/concat -23330=4,3,18,18,384\nConvolution fire7/squeeze1x1 1 1 fire6/concat fire7/squeeze1x1_fire7/relu_squeeze1x1 -23330=4,3,18,18,48 0=48 1=1 5=1 6=18432 9=1\nSplit splitncnn_7 1 2 fire7/squeeze1x1_fire7/relu_squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 -23330=8,3,18,18,48,3,18,18,48\nConvolution fire7/expand1x1 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 fire7/expand1x1_fire7/relu_expand1x1 -23330=4,3,18,18,192 0=192 1=1 5=1 6=9216 9=1\nConvolution fire7/expand3x3 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/expand3x3_fire7/relu_expand3x3 -23330=4,3,18,18,192 0=192 1=3 4=1 5=1 6=82944 9=1\nConcat fire7/concat 2 1 fire7/expand1x1_fire7/relu_expand1x1 fire7/expand3x3_fire7/relu_expand3x3 fire7/concat -23330=4,3,18,18,384\nConvolution fire8/squeeze1x1 1 1 fire7/concat fire8/squeeze1x1_fire8/relu_squeeze1x1 -23330=4,3,18,18,64 0=64 1=1 5=1 6=24576 9=1\nSplit splitncnn_8 1 2 fire8/squeeze1x1_fire8/relu_squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 -23330=8,3,18,18,64,3,18,18,64\nConvolution fire8/expand1x1 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 fire8/expand1x1_fire8/relu_expand1x1 -23330=4,3,18,18,256 0=256 1=1 5=1 6=16384 9=1\nConvolution fire8/expand3x3 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/expand3x3_fire8/relu_expand3x3 -23330=4,3,18,18,256 0=256 1=3 4=1 5=1 6=147456 9=1\nConcat fire8/concat 2 1 fire8/expand1x1_fire8/relu_expand1x1 fire8/expand3x3_fire8/relu_expand3x3 fire8/concat -23330=4,3,18,18,512\nConvolution fire9/squeeze1x1 1 1 fire8/concat fire9/squeeze1x1_fire9/relu_squeeze1x1 -23330=4,3,18,18,64 0=64 1=1 5=1 6=32768 9=1\nSplit splitncnn_9 1 2 fire9/squeeze1x1_fire9/relu_squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 -23330=8,3,18,18,64,3,18,18,64\nConvolution fire9/expand1x1 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 fire9/expand1x1_fire9/relu_expand1x1 -23330=4,3,18,18,256 0=256 1=1 5=1 6=16384 9=1\nConvolution fire9/expand3x3 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/expand3x3_fire9/relu_expand3x3 -23330=4,3,18,18,256 0=256 1=3 4=1 5=1 6=147456 9=1\nConcat fire9/concat 2 1 fire9/expand1x1_fire9/relu_expand1x1 fire9/expand3x3_fire9/relu_expand3x3 fire9/concat -23330=4,3,18,18,512\nSplit splitncnn_10 1 4 fire9/concat fire9/concat_splitncnn_0 fire9/concat_splitncnn_1 fire9/concat_splitncnn_2 fire9/concat_splitncnn_3 -23330=16,3,18,18,512,3,18,18,512,3,18,18,512,3,18,18,512\nPooling pool9 1 1 fire9/concat_splitncnn_3 pool9 -23330=4,3,9,9,512 1=3 2=2\nConvolution fire10/squeeze1x1 1 1 pool9 fire10/squeeze1x1_fire10/relu_squeeze1x1 -23330=4,3,9,9,96 0=96 1=1 5=1 6=49152 9=1\nSplit splitncnn_11 1 2 fire10/squeeze1x1_fire10/relu_squeeze1x1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_0 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_1 -23330=8,3,9,9,96,3,9,9,96\nConvolution fire10/expand1x1 1 1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_1 fire10/expand1x1_fire10/relu_expand1x1 -23330=4,3,9,9,384 0=384 1=1 5=1 6=36864 9=1\nConvolution fire10/expand3x3 1 1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_0 fire10/expand3x3_fire10/relu_expand3x3 -23330=4,3,9,9,384 0=384 1=3 4=1 5=1 6=331776 9=1\nConcat fire10/concat 2 1 fire10/expand1x1_fire10/relu_expand1x1 fire10/expand3x3_fire10/relu_expand3x3 fire10/concat -23330=4,3,9,9,768\nSplit splitncnn_12 1 4 fire10/concat fire10/concat_splitncnn_0 fire10/concat_splitncnn_1 fire10/concat_splitncnn_2 fire10/concat_splitncnn_3 -23330=16,3,9,9,768,3,9,9,768,3,9,9,768,3,9,9,768\nPooling pool10 1 1 fire10/concat_splitncnn_3 pool10 -23330=4,3,4,4,768 1=3 2=2\nConvolution fire11/squeeze1x1 1 1 pool10 fire11/squeeze1x1_fire11/relu_squeeze1x1 -23330=4,3,4,4,96 0=96 1=1 5=1 6=73728 9=1\nSplit splitncnn_13 1 2 fire11/squeeze1x1_fire11/relu_squeeze1x1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_0 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_1 -23330=8,3,4,4,96,3,4,4,96\nConvolution fire11/expand1x1 1 1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_1 fire11/expand1x1_fire11/relu_expand1x1 -23330=4,3,4,4,384 0=384 1=1 5=1 6=36864 9=1\nConvolution fire11/expand3x3 1 1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_0 fire11/expand3x3_fire11/relu_expand3x3 -23330=4,3,4,4,384 0=384 1=3 4=1 5=1 6=331776 9=1\nConcat fire11/concat 2 1 fire11/expand1x1_fire11/relu_expand1x1 fire11/expand3x3_fire11/relu_expand3x3 fire11/concat -23330=4,3,4,4,768\nSplit splitncnn_14 1 4 fire11/concat fire11/concat_splitncnn_0 fire11/concat_splitncnn_1 fire11/concat_splitncnn_2 fire11/concat_splitncnn_3 -23330=16,3,4,4,768,3,4,4,768,3,4,4,768,3,4,4,768\nConvolution conv12_1 1 1 fire11/concat_splitncnn_3 conv12_1_conv12_1/relu -23330=4,3,4,4,128 0=128 1=1 5=1 6=98304 9=1\nConvolution conv12_2 1 1 conv12_1_conv12_1/relu conv12_2_conv12_2/relu -23330=4,3,2,2,256 0=256 1=3 3=2 4=1 5=1 6=294912 9=1\nSplit splitncnn_15 1 4 conv12_2_conv12_2/relu conv12_2_conv12_2/relu_splitncnn_0 conv12_2_conv12_2/relu_splitncnn_1 conv12_2_conv12_2/relu_splitncnn_2 conv12_2_conv12_2/relu_splitncnn_3 -23330=16,3,2,2,256,3,2,2,256,3,2,2,256,3,2,2,256\nConvolution conv13_1 1 1 conv12_2_conv12_2/relu_splitncnn_3 conv13_1_conv13_1/relu -23330=4,3,2,2,64 0=64 1=1 5=1 6=16384 9=1\nConvolution conv13_2 1 1 conv13_1_conv13_1/relu conv13_2_conv13_2/relu -23330=4,3,1,1,128 0=128 1=3 3=2 4=1 5=1 6=73728 9=1\nSplit splitncnn_16 1 3 conv13_2_conv13_2/relu conv13_2_conv13_2/relu_splitncnn_0 conv13_2_conv13_2/relu_splitncnn_1 conv13_2_conv13_2/relu_splitncnn_2 -23330=12,3,1,1,128,3,1,1,128,3,1,1,128\nBatchNorm fire5/bn 1 1 fire5/concat_splitncnn_0 fire5/normal_fire5/scale -23330=4,3,37,37,256 0=256\nSplit splitncnn_17 1 3 fire5/normal_fire5/scale fire5/normal_fire5/scale_splitncnn_0 fire5/normal_fire5/scale_splitncnn_1 fire5/normal_fire5/scale_splitncnn_2 -23330=12,3,37,37,256,3,37,37,256,3,37,37,256\nConvolution fire5_mbox_loc 1 1 fire5/normal_fire5/scale_splitncnn_2 fire5_mbox_loc -23330=4,3,37,37,16 0=16 1=3 4=1 5=1 6=36864\nPermute fire5_mbox_loc_perm 1 1 fire5_mbox_loc fire5_mbox_loc_perm -23330=4,3,16,37,37 0=3\nFlatten fire5_mbox_loc_flat 1 1 fire5_mbox_loc_perm fire5_mbox_loc_flat -23330=4,1,21904,1,1\nConvolution fire5_mbox_conf 1 1 fire5/normal_fire5/scale_splitncnn_1 fire5_mbox_conf -23330=4,3,37,37,84 0=84 1=3 4=1 5=1 6=193536\nPermute fire5_mbox_conf_perm 1 1 fire5_mbox_conf fire5_mbox_conf_perm -23330=4,3,84,37,37 0=3\nFlatten fire5_mbox_conf_flat 1 1 fire5_mbox_conf_perm fire5_mbox_conf_flat -23330=4,1,114996,1,1\nPriorBox fire5_mbox_priorbox 2 1 fire5/normal_fire5/scale_splitncnn_0 data_splitncnn_5 fire5_mbox_priorbox -23330=4,2,21904,2,1 -23300=1,2.100000e+01 -23301=1,4.500000e+01 -23302=1,2.000000e+00 9=-233 10=-233 11=8.000000e+00 12=8.000000e+00 13=5.000000e-01\nConvolution fire9_mbox_loc 1 1 fire9/concat_splitncnn_2 fire9_mbox_loc -23330=4,3,18,18,24 0=24 1=3 4=1 5=1 6=110592\nPermute fire9_mbox_loc_perm 1 1 fire9_mbox_loc fire9_mbox_loc_perm -23330=4,3,24,18,18 0=3\nFlatten fire9_mbox_loc_flat 1 1 fire9_mbox_loc_perm fire9_mbox_loc_flat -23330=4,1,7776,1,1\nConvolution fire9_mbox_conf 1 1 fire9/concat_splitncnn_1 fire9_mbox_conf -23330=4,3,18,18,126 0=126 1=3 4=1 5=1 6=580608\nPermute fire9_mbox_conf_perm 1 1 fire9_mbox_conf fire9_mbox_conf_perm -23330=4,3,126,18,18 0=3\nFlatten fire9_mbox_conf_flat 1 1 fire9_mbox_conf_perm fire9_mbox_conf_flat -23330=4,1,40824,1,1\nPriorBox fire9_mbox_priorbox 2 1 fire9/concat_splitncnn_0 data_splitncnn_4 fire9_mbox_priorbox -23330=4,2,7776,2,1 -23300=1,4.500000e+01 -23301=1,9.900000e+01 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 11=1.600000e+01 12=1.600000e+01 13=5.000000e-01\nConvolution fire10_mbox_loc 1 1 fire10/concat_splitncnn_2 fire10_mbox_loc -23330=4,3,9,9,24 0=24 1=3 4=1 5=1 6=165888\nPermute fire10_mbox_loc_perm 1 1 fire10_mbox_loc fire10_mbox_loc_perm -23330=4,3,24,9,9 0=3\nFlatten fire10_mbox_loc_flat 1 1 fire10_mbox_loc_perm fire10_mbox_loc_flat -23330=4,1,1944,1,1\nConvolution fire10_mbox_conf 1 1 fire10/concat_splitncnn_1 fire10_mbox_conf -23330=4,3,9,9,126 0=126 1=3 4=1 5=1 6=870912\nPermute fire10_mbox_conf_perm 1 1 fire10_mbox_conf fire10_mbox_conf_perm -23330=4,3,126,9,9 0=3\nFlatten fire10_mbox_conf_flat 1 1 fire10_mbox_conf_perm fire10_mbox_conf_flat -23330=4,1,10206,1,1\nPriorBox fire10_mbox_priorbox 2 1 fire10/concat_splitncnn_0 data_splitncnn_3 fire10_mbox_priorbox -23330=4,2,1944,2,1 -23300=1,9.900000e+01 -23301=1,1.530000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 11=3.200000e+01 12=3.200000e+01 13=5.000000e-01\nConvolution fire11_mbox_loc 1 1 fire11/concat_splitncnn_2 fire11_mbox_loc -23330=4,3,4,4,24 0=24 1=3 4=1 5=1 6=165888\nPermute fire11_mbox_loc_perm 1 1 fire11_mbox_loc fire11_mbox_loc_perm -23330=4,3,24,4,4 0=3\nFlatten fire11_mbox_loc_flat 1 1 fire11_mbox_loc_perm fire11_mbox_loc_flat -23330=4,1,384,1,1\nConvolution fire11_mbox_conf 1 1 fire11/concat_splitncnn_1 fire11_mbox_conf -23330=4,3,4,4,126 0=126 1=3 4=1 5=1 6=870912\nPermute fire11_mbox_conf_perm 1 1 fire11_mbox_conf fire11_mbox_conf_perm -23330=4,3,126,4,4 0=3\nFlatten fire11_mbox_conf_flat 1 1 fire11_mbox_conf_perm fire11_mbox_conf_flat -23330=4,1,2016,1,1\nPriorBox fire11_mbox_priorbox 2 1 fire11/concat_splitncnn_0 data_splitncnn_2 fire11_mbox_priorbox -23330=4,2,384,2,1 -23300=1,1.530000e+02 -23301=1,2.070000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 11=6.400000e+01 12=6.400000e+01 13=5.000000e-01\nConvolution conv12_2_mbox_loc 1 1 conv12_2_conv12_2/relu_splitncnn_2 conv12_2_mbox_loc -23330=4,3,2,2,24 0=24 1=3 4=1 5=1 6=55296\nPermute conv12_2_mbox_loc_perm 1 1 conv12_2_mbox_loc conv12_2_mbox_loc_perm -23330=4,3,24,2,2 0=3\nFlatten conv12_2_mbox_loc_flat 1 1 conv12_2_mbox_loc_perm conv12_2_mbox_loc_flat -23330=4,1,96,1,1\nConvolution conv12_2_mbox_conf 1 1 conv12_2_conv12_2/relu_splitncnn_1 conv12_2_mbox_conf -23330=4,3,2,2,126 0=126 1=3 4=1 5=1 6=290304\nPermute conv12_2_mbox_conf_perm 1 1 conv12_2_mbox_conf conv12_2_mbox_conf_perm -23330=4,3,126,2,2 0=3\nFlatten conv12_2_mbox_conf_flat 1 1 conv12_2_mbox_conf_perm conv12_2_mbox_conf_flat -23330=4,1,504,1,1\nPriorBox conv12_2_mbox_priorbox 2 1 conv12_2_conv12_2/relu_splitncnn_0 data_splitncnn_1 conv12_2_mbox_priorbox -23330=4,2,96,2,1 -23300=1,2.070000e+02 -23301=1,2.610000e+02 -23302=2,2.000000e+00,3.000000e+00 9=-233 10=-233 11=1.000000e+02 12=1.000000e+02 13=5.000000e-01\nConvolution conv13_2_mbox_loc 1 1 conv13_2_conv13_2/relu_splitncnn_2 conv13_2_mbox_loc -23330=4,3,1,1,16 0=16 1=3 4=1 5=1 6=18432\nPermute conv13_2_mbox_loc_perm 1 1 conv13_2_mbox_loc conv13_2_mbox_loc_perm -23330=4,3,16,1,1 0=3\nFlatten conv13_2_mbox_loc_flat 1 1 conv13_2_mbox_loc_perm conv13_2_mbox_loc_flat -23330=4,1,16,1,1\nConvolution conv13_2_mbox_conf 1 1 conv13_2_conv13_2/relu_splitncnn_1 conv13_2_mbox_conf -23330=4,3,1,1,84 0=84 1=3 4=1 5=1 6=96768\nPermute conv13_2_mbox_conf_perm 1 1 conv13_2_mbox_conf conv13_2_mbox_conf_perm -23330=4,3,84,1,1 0=3\nFlatten conv13_2_mbox_conf_flat 1 1 conv13_2_mbox_conf_perm conv13_2_mbox_conf_flat -23330=4,1,84,1,1\nPriorBox conv13_2_mbox_priorbox 2 1 conv13_2_conv13_2/relu_splitncnn_0 data_splitncnn_0 conv13_2_mbox_priorbox -23330=4,2,16,2,1 -23300=1,2.610000e+02 -23301=1,3.150000e+02 -23302=1,2.000000e+00 9=-233 10=-233 11=3.000000e+02 12=3.000000e+02 13=5.000000e-01\nConcat mbox_loc 6 1 fire5_mbox_loc_flat fire9_mbox_loc_flat fire10_mbox_loc_flat fire11_mbox_loc_flat conv12_2_mbox_loc_flat conv13_2_mbox_loc_flat mbox_loc -23330=4,1,32120,1,1\nConcat mbox_conf 6 1 fire5_mbox_conf_flat fire9_mbox_conf_flat fire10_mbox_conf_flat fire11_mbox_conf_flat conv12_2_mbox_conf_flat conv13_2_mbox_conf_flat mbox_conf -23330=4,1,168630,1,1\nConcat mbox_priorbox 6 1 fire5_mbox_priorbox fire9_mbox_priorbox fire10_mbox_priorbox fire11_mbox_priorbox conv12_2_mbox_priorbox conv13_2_mbox_priorbox mbox_priorbox -23330=4,2,32120,2,1 0=1\nReshape mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape -23330=4,2,21,8030,1 0=21 1=-1\nSoftmax mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax -23330=4,2,21,8030,1 0=1 1=1\nFlatten mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten -23330=4,1,168630,1,1\nDetectionOutput detection_out 3 1 mbox_loc mbox_conf_flatten mbox_priorbox output 0=21 1=4.500000e-01 2=100 4=2.500000e-01\n" }, { "path": "benchmark/squeezenet_ssd_int8.param", "content": "7767517\n119 152\nInput data 0 1 data 0=300 1=300 2=3\nSplit splitncnn_0 1 7 data data_splitncnn_0 data_splitncnn_1 data_splitncnn_2 data_splitncnn_3 data_splitncnn_4 data_splitncnn_5 data_splitncnn_6\nConvolution conv1 1 1 data_splitncnn_6 conv1_relu_conv1 0=64 1=3 3=2 5=1 6=1728 8=2 9=1\nPooling pool1 1 1 conv1_relu_conv1 pool1 1=3 2=2\nConvolution fire2/squeeze1x1 1 1 pool1 fire2/squeeze1x1_fire2/relu_squeeze1x1 0=16 1=1 5=1 6=1024 8=102 9=1\nSplit splitncnn_1 1 2 fire2/squeeze1x1_fire2/relu_squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1\nConvolution fire2/expand1x1 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 fire2/expand1x1_fire2/relu_expand1x1 0=64 1=1 5=1 6=1024 8=2 9=1\nConvolution fire2/expand3x3 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/expand3x3_fire2/relu_expand3x3 0=64 1=3 4=1 5=1 6=9216 8=2 9=1\nConcat fire2/concat 2 1 fire2/expand1x1_fire2/relu_expand1x1 fire2/expand3x3_fire2/relu_expand3x3 fire2/concat\nConvolution fire3/squeeze1x1 1 1 fire2/concat fire3/squeeze1x1_fire3/relu_squeeze1x1 0=16 1=1 5=1 6=2048 8=102 9=1\nSplit splitncnn_2 1 2 fire3/squeeze1x1_fire3/relu_squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1\nConvolution fire3/expand1x1 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 fire3/expand1x1_fire3/relu_expand1x1 0=64 1=1 5=1 6=1024 8=2 9=1\nConvolution fire3/expand3x3 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/expand3x3_fire3/relu_expand3x3 0=64 1=3 4=1 5=1 6=9216 8=2 9=1\nConcat fire3/concat 2 1 fire3/expand1x1_fire3/relu_expand1x1 fire3/expand3x3_fire3/relu_expand3x3 fire3/concat\nPooling pool3 1 1 fire3/concat pool3 1=3 2=2\nConvolution fire4/squeeze1x1 1 1 pool3 fire4/squeeze1x1_fire4/relu_squeeze1x1 0=32 1=1 5=1 6=4096 8=102 9=1\nSplit splitncnn_3 1 2 fire4/squeeze1x1_fire4/relu_squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1\nConvolution fire4/expand1x1 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 fire4/expand1x1_fire4/relu_expand1x1 0=128 1=1 5=1 6=4096 8=2 9=1\nConvolution fire4/expand3x3 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/expand3x3_fire4/relu_expand3x3 0=128 1=3 4=1 5=1 6=36864 8=2 9=1\nConcat fire4/concat 2 1 fire4/expand1x1_fire4/relu_expand1x1 fire4/expand3x3_fire4/relu_expand3x3 fire4/concat\nConvolution fire5/squeeze1x1 1 1 fire4/concat fire5/squeeze1x1_fire5/relu_squeeze1x1 0=32 1=1 5=1 6=8192 8=102 9=1\nSplit splitncnn_4 1 2 fire5/squeeze1x1_fire5/relu_squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1\nConvolution fire5/expand1x1 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 fire5/expand1x1_fire5/relu_expand1x1 0=128 1=1 5=1 6=4096 8=2 9=1\nConvolution fire5/expand3x3 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/expand3x3_fire5/relu_expand3x3 0=128 1=3 4=1 5=1 6=36864 8=2 9=1\nConcat fire5/concat 2 1 fire5/expand1x1_fire5/relu_expand1x1 fire5/expand3x3_fire5/relu_expand3x3 fire5/concat\nSplit splitncnn_5 1 2 fire5/concat fire5/concat_splitncnn_0 fire5/concat_splitncnn_1\nPooling pool5 1 1 fire5/concat_splitncnn_1 pool5 1=3 2=2\nConvolution fire6/squeeze1x1 1 1 pool5 fire6/squeeze1x1_fire6/relu_squeeze1x1 0=48 1=1 5=1 6=12288 8=102 9=1\nSplit splitncnn_6 1 2 fire6/squeeze1x1_fire6/relu_squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1\nConvolution fire6/expand1x1 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 fire6/expand1x1_fire6/relu_expand1x1 0=192 1=1 5=1 6=9216 8=2 9=1\nConvolution fire6/expand3x3 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/expand3x3_fire6/relu_expand3x3 0=192 1=3 4=1 5=1 6=82944 8=2 9=1\nConcat fire6/concat 2 1 fire6/expand1x1_fire6/relu_expand1x1 fire6/expand3x3_fire6/relu_expand3x3 fire6/concat\nConvolution fire7/squeeze1x1 1 1 fire6/concat fire7/squeeze1x1_fire7/relu_squeeze1x1 0=48 1=1 5=1 6=18432 8=102 9=1\nSplit splitncnn_7 1 2 fire7/squeeze1x1_fire7/relu_squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1\nConvolution fire7/expand1x1 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 fire7/expand1x1_fire7/relu_expand1x1 0=192 1=1 5=1 6=9216 8=2 9=1\nConvolution fire7/expand3x3 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/expand3x3_fire7/relu_expand3x3 0=192 1=3 4=1 5=1 6=82944 8=2 9=1\nConcat fire7/concat 2 1 fire7/expand1x1_fire7/relu_expand1x1 fire7/expand3x3_fire7/relu_expand3x3 fire7/concat\nConvolution fire8/squeeze1x1 1 1 fire7/concat fire8/squeeze1x1_fire8/relu_squeeze1x1 0=64 1=1 5=1 6=24576 8=102 9=1\nSplit splitncnn_8 1 2 fire8/squeeze1x1_fire8/relu_squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1\nConvolution fire8/expand1x1 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 fire8/expand1x1_fire8/relu_expand1x1 0=256 1=1 5=1 6=16384 8=2 9=1\nConvolution fire8/expand3x3 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/expand3x3_fire8/relu_expand3x3 0=256 1=3 4=1 5=1 6=147456 8=2 9=1\nConcat fire8/concat 2 1 fire8/expand1x1_fire8/relu_expand1x1 fire8/expand3x3_fire8/relu_expand3x3 fire8/concat\nConvolution fire9/squeeze1x1 1 1 fire8/concat fire9/squeeze1x1_fire9/relu_squeeze1x1 0=64 1=1 5=1 6=32768 8=102 9=1\nSplit splitncnn_9 1 2 fire9/squeeze1x1_fire9/relu_squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1\nConvolution fire9/expand1x1 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 fire9/expand1x1_fire9/relu_expand1x1 0=256 1=1 5=1 6=16384 8=2 9=1\nConvolution fire9/expand3x3 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/expand3x3_fire9/relu_expand3x3 0=256 1=3 4=1 5=1 6=147456 8=2 9=1\nConcat fire9/concat 2 1 fire9/expand1x1_fire9/relu_expand1x1 fire9/expand3x3_fire9/relu_expand3x3 fire9/concat\nSplit splitncnn_10 1 4 fire9/concat fire9/concat_splitncnn_0 fire9/concat_splitncnn_1 fire9/concat_splitncnn_2 fire9/concat_splitncnn_3\nPooling pool9 1 1 fire9/concat_splitncnn_3 pool9 1=3 2=2\nConvolution fire10/squeeze1x1 1 1 pool9 fire10/squeeze1x1_fire10/relu_squeeze1x1 0=96 1=1 5=1 6=49152 8=102 9=1\nSplit splitncnn_11 1 2 fire10/squeeze1x1_fire10/relu_squeeze1x1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_0 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_1\nConvolution fire10/expand1x1 1 1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_1 fire10/expand1x1_fire10/relu_expand1x1 0=384 1=1 5=1 6=36864 8=2 9=1\nConvolution fire10/expand3x3 1 1 fire10/squeeze1x1_fire10/relu_squeeze1x1_splitncnn_0 fire10/expand3x3_fire10/relu_expand3x3 0=384 1=3 4=1 5=1 6=331776 8=2 9=1\nConcat fire10/concat 2 1 fire10/expand1x1_fire10/relu_expand1x1 fire10/expand3x3_fire10/relu_expand3x3 fire10/concat\nSplit splitncnn_12 1 4 fire10/concat fire10/concat_splitncnn_0 fire10/concat_splitncnn_1 fire10/concat_splitncnn_2 fire10/concat_splitncnn_3\nPooling pool10 1 1 fire10/concat_splitncnn_3 pool10 1=3 2=2\nConvolution fire11/squeeze1x1 1 1 pool10 fire11/squeeze1x1_fire11/relu_squeeze1x1 0=96 1=1 5=1 6=73728 8=102 9=1\nSplit splitncnn_13 1 2 fire11/squeeze1x1_fire11/relu_squeeze1x1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_0 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_1\nConvolution fire11/expand1x1 1 1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_1 fire11/expand1x1_fire11/relu_expand1x1 0=384 1=1 5=1 6=36864 8=2 9=1\nConvolution fire11/expand3x3 1 1 fire11/squeeze1x1_fire11/relu_squeeze1x1_splitncnn_0 fire11/expand3x3_fire11/relu_expand3x3 0=384 1=3 4=1 5=1 6=331776 8=2 9=1\nConcat fire11/concat 2 1 fire11/expand1x1_fire11/relu_expand1x1 fire11/expand3x3_fire11/relu_expand3x3 fire11/concat\nSplit splitncnn_14 1 4 fire11/concat fire11/concat_splitncnn_0 fire11/concat_splitncnn_1 fire11/concat_splitncnn_2 fire11/concat_splitncnn_3\nConvolution conv12_1 1 1 fire11/concat_splitncnn_3 conv12_1_conv12_1/relu 0=128 1=1 5=1 6=98304 8=102 9=1\nConvolution conv12_2 1 1 conv12_1_conv12_1/relu conv12_2_conv12_2/relu 0=256 1=3 3=2 4=1 5=1 6=294912 8=2 9=1\nSplit splitncnn_15 1 4 conv12_2_conv12_2/relu conv12_2_conv12_2/relu_splitncnn_0 conv12_2_conv12_2/relu_splitncnn_1 conv12_2_conv12_2/relu_splitncnn_2 conv12_2_conv12_2/relu_splitncnn_3\nConvolution conv13_1 1 1 conv12_2_conv12_2/relu_splitncnn_3 conv13_1_conv13_1/relu 0=64 1=1 5=1 6=16384 8=102 9=1\nConvolution conv13_2 1 1 conv13_1_conv13_1/relu conv13_2_conv13_2/relu 0=128 1=3 3=2 4=1 5=1 6=73728 8=2 9=1\nSplit splitncnn_16 1 3 conv13_2_conv13_2/relu conv13_2_conv13_2/relu_splitncnn_0 conv13_2_conv13_2/relu_splitncnn_1 conv13_2_conv13_2/relu_splitncnn_2\nBatchNorm fire5/bn 1 1 fire5/concat_splitncnn_0 fire5/normal_fire5/scale 0=256\nSplit splitncnn_17 1 3 fire5/normal_fire5/scale fire5/normal_fire5/scale_splitncnn_0 fire5/normal_fire5/scale_splitncnn_1 fire5/normal_fire5/scale_splitncnn_2\nConvolution fire5_mbox_loc 1 1 fire5/normal_fire5/scale_splitncnn_2 fire5_mbox_loc 0=16 1=3 4=1 5=1 6=36864 8=2\nPermute fire5_mbox_loc_perm 1 1 fire5_mbox_loc fire5_mbox_loc_perm 0=3\nFlatten fire5_mbox_loc_flat 1 1 fire5_mbox_loc_perm fire5_mbox_loc_flat\nConvolution fire5_mbox_conf 1 1 fire5/normal_fire5/scale_splitncnn_1 fire5_mbox_conf 0=84 1=3 4=1 5=1 6=193536 8=2\nPermute fire5_mbox_conf_perm 1 1 fire5_mbox_conf fire5_mbox_conf_perm 0=3\nFlatten fire5_mbox_conf_flat 1 1 fire5_mbox_conf_perm fire5_mbox_conf_flat\nPriorBox fire5_mbox_priorbox 2 1 fire5/normal_fire5/scale_splitncnn_0 data_splitncnn_5 fire5_mbox_priorbox -23300=1,21.000000 -23301=1,45.000000 -23302=1,2.000000 9=-233 10=-233 11=8.000000 12=8.000000 13=0.500000\nConvolution fire9_mbox_loc 1 1 fire9/concat_splitncnn_2 fire9_mbox_loc 0=24 1=3 4=1 5=1 6=110592 8=2\nPermute fire9_mbox_loc_perm 1 1 fire9_mbox_loc fire9_mbox_loc_perm 0=3\nFlatten fire9_mbox_loc_flat 1 1 fire9_mbox_loc_perm fire9_mbox_loc_flat\nConvolution fire9_mbox_conf 1 1 fire9/concat_splitncnn_1 fire9_mbox_conf 0=126 1=3 4=1 5=1 6=580608 8=2\nPermute fire9_mbox_conf_perm 1 1 fire9_mbox_conf fire9_mbox_conf_perm 0=3\nFlatten fire9_mbox_conf_flat 1 1 fire9_mbox_conf_perm fire9_mbox_conf_flat\nPriorBox fire9_mbox_priorbox 2 1 fire9/concat_splitncnn_0 data_splitncnn_4 fire9_mbox_priorbox -23300=1,45.000000 -23301=1,99.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 11=16.000000 12=16.000000 13=0.500000\nConvolution fire10_mbox_loc 1 1 fire10/concat_splitncnn_2 fire10_mbox_loc 0=24 1=3 4=1 5=1 6=165888 8=2\nPermute fire10_mbox_loc_perm 1 1 fire10_mbox_loc fire10_mbox_loc_perm 0=3\nFlatten fire10_mbox_loc_flat 1 1 fire10_mbox_loc_perm fire10_mbox_loc_flat\nConvolution fire10_mbox_conf 1 1 fire10/concat_splitncnn_1 fire10_mbox_conf 0=126 1=3 4=1 5=1 6=870912 8=2\nPermute fire10_mbox_conf_perm 1 1 fire10_mbox_conf fire10_mbox_conf_perm 0=3\nFlatten fire10_mbox_conf_flat 1 1 fire10_mbox_conf_perm fire10_mbox_conf_flat\nPriorBox fire10_mbox_priorbox 2 1 fire10/concat_splitncnn_0 data_splitncnn_3 fire10_mbox_priorbox -23300=1,99.000000 -23301=1,153.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 11=32.000000 12=32.000000 13=0.500000\nConvolution fire11_mbox_loc 1 1 fire11/concat_splitncnn_2 fire11_mbox_loc 0=24 1=3 4=1 5=1 6=165888 8=2\nPermute fire11_mbox_loc_perm 1 1 fire11_mbox_loc fire11_mbox_loc_perm 0=3\nFlatten fire11_mbox_loc_flat 1 1 fire11_mbox_loc_perm fire11_mbox_loc_flat\nConvolution fire11_mbox_conf 1 1 fire11/concat_splitncnn_1 fire11_mbox_conf 0=126 1=3 4=1 5=1 6=870912 8=2\nPermute fire11_mbox_conf_perm 1 1 fire11_mbox_conf fire11_mbox_conf_perm 0=3\nFlatten fire11_mbox_conf_flat 1 1 fire11_mbox_conf_perm fire11_mbox_conf_flat\nPriorBox fire11_mbox_priorbox 2 1 fire11/concat_splitncnn_0 data_splitncnn_2 fire11_mbox_priorbox -23300=1,153.000000 -23301=1,207.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 11=64.000000 12=64.000000 13=0.500000\nConvolution conv12_2_mbox_loc 1 1 conv12_2_conv12_2/relu_splitncnn_2 conv12_2_mbox_loc 0=24 1=3 4=1 5=1 6=55296 8=2\nPermute conv12_2_mbox_loc_perm 1 1 conv12_2_mbox_loc conv12_2_mbox_loc_perm 0=3\nFlatten conv12_2_mbox_loc_flat 1 1 conv12_2_mbox_loc_perm conv12_2_mbox_loc_flat\nConvolution conv12_2_mbox_conf 1 1 conv12_2_conv12_2/relu_splitncnn_1 conv12_2_mbox_conf 0=126 1=3 4=1 5=1 6=290304 8=2\nPermute conv12_2_mbox_conf_perm 1 1 conv12_2_mbox_conf conv12_2_mbox_conf_perm 0=3\nFlatten conv12_2_mbox_conf_flat 1 1 conv12_2_mbox_conf_perm conv12_2_mbox_conf_flat\nPriorBox conv12_2_mbox_priorbox 2 1 conv12_2_conv12_2/relu_splitncnn_0 data_splitncnn_1 conv12_2_mbox_priorbox -23300=1,207.000000 -23301=1,261.000000 -23302=2,2.000000,3.000000 9=-233 10=-233 11=100.000000 12=100.000000 13=0.500000\nConvolution conv13_2_mbox_loc 1 1 conv13_2_conv13_2/relu_splitncnn_2 conv13_2_mbox_loc 0=16 1=3 4=1 5=1 6=18432 8=2\nPermute conv13_2_mbox_loc_perm 1 1 conv13_2_mbox_loc conv13_2_mbox_loc_perm 0=3\nFlatten conv13_2_mbox_loc_flat 1 1 conv13_2_mbox_loc_perm conv13_2_mbox_loc_flat\nConvolution conv13_2_mbox_conf 1 1 conv13_2_conv13_2/relu_splitncnn_1 conv13_2_mbox_conf 0=84 1=3 4=1 5=1 6=96768 8=2\nPermute conv13_2_mbox_conf_perm 1 1 conv13_2_mbox_conf conv13_2_mbox_conf_perm 0=3\nFlatten conv13_2_mbox_conf_flat 1 1 conv13_2_mbox_conf_perm conv13_2_mbox_conf_flat\nPriorBox conv13_2_mbox_priorbox 2 1 conv13_2_conv13_2/relu_splitncnn_0 data_splitncnn_0 conv13_2_mbox_priorbox -23300=1,261.000000 -23301=1,315.000000 -23302=1,2.000000 9=-233 10=-233 11=300.000000 12=300.000000 13=0.500000\nConcat mbox_loc 6 1 fire5_mbox_loc_flat fire9_mbox_loc_flat fire10_mbox_loc_flat fire11_mbox_loc_flat conv12_2_mbox_loc_flat conv13_2_mbox_loc_flat mbox_loc\nConcat mbox_conf 6 1 fire5_mbox_conf_flat fire9_mbox_conf_flat fire10_mbox_conf_flat fire11_mbox_conf_flat conv12_2_mbox_conf_flat conv13_2_mbox_conf_flat mbox_conf\nConcat mbox_priorbox 6 1 fire5_mbox_priorbox fire9_mbox_priorbox fire10_mbox_priorbox fire11_mbox_priorbox conv12_2_mbox_priorbox conv13_2_mbox_priorbox mbox_priorbox 0=1\nReshape mbox_conf_reshape 1 1 mbox_conf mbox_conf_reshape 0=21 1=-1\nSoftmax mbox_conf_softmax 1 1 mbox_conf_reshape mbox_conf_softmax 0=1 1=1\nFlatten mbox_conf_flatten 1 1 mbox_conf_softmax mbox_conf_flatten\nDetectionOutput detection_out 3 1 mbox_loc mbox_conf_flatten mbox_priorbox output 0=21 1=0.450000 2=100 4=0.250000\n" }, { "path": "benchmark/vgg16.param", "content": "7767517\n23 23\nInput data 0 1 data -23330=4,3,224,224,3 0=224 1=224 2=3\nConvolution conv1_1 1 1 data conv1_1_relu1_1 -23330=4,3,224,224,64 0=64 1=3 4=1 5=1 6=1728 9=1\nConvolution conv1_2 1 1 conv1_1_relu1_1 conv1_2_relu1_2 -23330=4,3,224,224,64 0=64 1=3 4=1 5=1 6=36864 9=1\nPooling pool1 1 1 conv1_2_relu1_2 pool1 -23330=4,3,112,112,64 1=2 2=2\nConvolution conv2_1 1 1 pool1 conv2_1_relu2_1 -23330=4,3,112,112,128 0=128 1=3 4=1 5=1 6=73728 9=1\nConvolution conv2_2 1 1 conv2_1_relu2_1 conv2_2_relu2_2 -23330=4,3,112,112,128 0=128 1=3 4=1 5=1 6=147456 9=1\nPooling pool2 1 1 conv2_2_relu2_2 pool2 -23330=4,3,56,56,128 1=2 2=2\nConvolution conv3_1 1 1 pool2 conv3_1_relu3_1 -23330=4,3,56,56,256 0=256 1=3 4=1 5=1 6=294912 9=1\nConvolution conv3_2 1 1 conv3_1_relu3_1 conv3_2_relu3_2 -23330=4,3,56,56,256 0=256 1=3 4=1 5=1 6=589824 9=1\nConvolution conv3_3 1 1 conv3_2_relu3_2 conv3_3_relu3_3 -23330=4,3,56,56,256 0=256 1=3 4=1 5=1 6=589824 9=1\nPooling pool3 1 1 conv3_3_relu3_3 pool3 -23330=4,3,28,28,256 1=2 2=2\nConvolution conv4_1 1 1 pool3 conv4_1_relu4_1 -23330=4,3,28,28,512 0=512 1=3 4=1 5=1 6=1179648 9=1\nConvolution conv4_2 1 1 conv4_1_relu4_1 conv4_2_relu4_2 -23330=4,3,28,28,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution conv4_3 1 1 conv4_2_relu4_2 conv4_3_relu4_3 -23330=4,3,28,28,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nPooling pool4 1 1 conv4_3_relu4_3 pool4 -23330=4,3,14,14,512 1=2 2=2\nConvolution conv5_1 1 1 pool4 conv5_1_relu5_1 -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution conv5_2 1 1 conv5_1_relu5_1 conv5_2_relu5_2 -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nConvolution conv5_3 1 1 conv5_2_relu5_2 conv5_3_relu5_3 -23330=4,3,14,14,512 0=512 1=3 4=1 5=1 6=2359296 9=1\nPooling pool5 1 1 conv5_3_relu5_3 pool5 -23330=4,3,7,7,512 1=2 2=2\nInnerProduct fc6 1 1 pool5 fc6_drop6 -23330=4,1,4096,1,1 0=4096 1=1 2=102760448 9=1\nInnerProduct fc7 1 1 fc6_drop6 fc7_drop7 -23330=4,1,4096,1,1 0=4096 1=1 2=16777216 9=1\nInnerProduct fc8 1 1 fc7_drop7 fc8 -23330=4,1,1000,1,1 0=1000 1=1 2=4096000\nSoftmax prob 1 1 fc8 output -23330=4,1,1000,1,1\n" }, { "path": "benchmark/vgg16_int8.param", "content": "7767517\n23 23\nInput data 0 1 data 0=224 1=224 2=3\nConvolution conv1_1 1 1 data conv1_1_relu1_1 0=64 1=3 4=1 5=1 6=1728 8=102 9=1\nConvolution conv1_2 1 1 conv1_1_relu1_1 conv1_2_relu1_2 0=64 1=3 4=1 5=1 6=36864 8=2 9=1\nPooling pool1 1 1 conv1_2_relu1_2 pool1 1=2 2=2\nConvolution conv2_1 1 1 pool1 conv2_1_relu2_1 0=128 1=3 4=1 5=1 6=73728 8=102 9=1\nConvolution conv2_2 1 1 conv2_1_relu2_1 conv2_2_relu2_2 0=128 1=3 4=1 5=1 6=147456 8=2 9=1\nPooling pool2 1 1 conv2_2_relu2_2 pool2 1=2 2=2\nConvolution conv3_1 1 1 pool2 conv3_1_relu3_1 0=256 1=3 4=1 5=1 6=294912 8=102 9=1\nConvolution conv3_2 1 1 conv3_1_relu3_1 conv3_2_relu3_2 0=256 1=3 4=1 5=1 6=589824 8=102 9=1\nConvolution conv3_3 1 1 conv3_2_relu3_2 conv3_3_relu3_3 0=256 1=3 4=1 5=1 6=589824 8=2 9=1\nPooling pool3 1 1 conv3_3_relu3_3 pool3 1=2 2=2\nConvolution conv4_1 1 1 pool3 conv4_1_relu4_1 0=512 1=3 4=1 5=1 6=1179648 8=102 9=1\nConvolution conv4_2 1 1 conv4_1_relu4_1 conv4_2_relu4_2 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution conv4_3 1 1 conv4_2_relu4_2 conv4_3_relu4_3 0=512 1=3 4=1 5=1 6=2359296 8=2 9=1\nPooling pool4 1 1 conv4_3_relu4_3 pool4 1=2 2=2\nConvolution conv5_1 1 1 pool4 conv5_1_relu5_1 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution conv5_2 1 1 conv5_1_relu5_1 conv5_2_relu5_2 0=512 1=3 4=1 5=1 6=2359296 8=102 9=1\nConvolution conv5_3 1 1 conv5_2_relu5_2 conv5_3_relu5_3 0=512 1=3 4=1 5=1 6=2359296 8=2 9=1\nPooling pool5 1 1 conv5_3_relu5_3 pool5 1=2 2=2\nInnerProduct fc6 1 1 pool5 fc6_drop6 0=4096 1=1 2=102760448 8=2 9=1\nInnerProduct fc7 1 1 fc6_drop6 fc7_drop7 0=4096 1=1 2=16777216 8=2 9=1\nInnerProduct fc8 1 1 fc7_drop7 fc8 0=1000 1=1 2=4096000 8=2\nSoftmax prob 1 1 fc8 output\n" }, { "path": "benchmark/vision_transformer.param", "content": "7767517\n144 192\nInput input 0 1 input\nMemoryData backbone.cls_token 0 1 backbone.cls_token 0=768 1=1\nMemoryData backbone.pos_embed 0 1 backbone.pos_embed 0=768 1=145\nConvolution Conv_0 1 1 input onnx::Shape_153 0=768 1=32 11=32 2=1 12=1 3=32 13=32 4=0 14=0 15=0 16=0 5=1 6=2359296\nReshape Reshape_8 1 1 onnx::Shape_153 onnx::Transpose_161 0=-1 1=768\nPermute Transpose_9 1 1 onnx::Transpose_161 onnx::Concat_162 0=1\nConcat Concat_10 2 1 backbone.cls_token onnx::Concat_162 onnx::Add_163 0=0\nBinaryOp Add_11 2 1 onnx::Add_163 backbone.pos_embed input.1 0=0\nSplit splitncnn_0 1 2 input.1 input.1_splitncnn_0 input.1_splitncnn_1\nLayerNorm LayerNorm_12 1 1 input.1_splitncnn_1 qkv_input 0=768 1=1.000000e-06 2=1\nSplit splitncnn_1 1 3 qkv_input qkv_input_splitncnn_0 qkv_input_splitncnn_1 qkv_input_splitncnn_2\nMultiHeadAttention MultiHeadAttention_21 3 1 qkv_input_splitncnn_2 qkv_input_splitncnn_1 qkv_input_splitncnn_0 onnx::Add_174 0=768 1=12 2=589824\nBinaryOp Add_22 2 1 input.1_splitncnn_0 onnx::Add_174 input.4 0=0\nSplit splitncnn_2 1 2 input.4 input.4_splitncnn_0 input.4_splitncnn_1\nLayerNorm LayerNorm_23 1 1 input.4_splitncnn_1 mmdeploy::Gemm_176 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_24 1 1 mmdeploy::Gemm_176 mmdeploy::Gelu_177 0=3072 1=1 2=2359296\nGELU Gelu_25 1 1 mmdeploy::Gelu_177 input.8 0=1\nInnerProduct Gemm_26 1 1 input.8 input.12 0=768 1=1 2=2359296\nBinaryOp Add_27 2 1 input.4_splitncnn_0 input.12 input.16 0=0\nSplit splitncnn_3 1 2 input.16 input.16_splitncnn_0 input.16_splitncnn_1\nLayerNorm LayerNorm_28 1 1 input.16_splitncnn_1 qkv_input.3 0=768 1=1.000000e-06 2=1\nSplit splitncnn_4 1 3 qkv_input.3 qkv_input.3_splitncnn_0 qkv_input.3_splitncnn_1 qkv_input.3_splitncnn_2\nMultiHeadAttention MultiHeadAttention_37 3 1 qkv_input.3_splitncnn_2 qkv_input.3_splitncnn_1 qkv_input.3_splitncnn_0 onnx::Add_190 0=768 1=12 2=589824\nBinaryOp Add_38 2 1 input.16_splitncnn_0 onnx::Add_190 input.20 0=0\nSplit splitncnn_5 1 2 input.20 input.20_splitncnn_0 input.20_splitncnn_1\nLayerNorm LayerNorm_39 1 1 input.20_splitncnn_1 mmdeploy::Gemm_192 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_40 1 1 mmdeploy::Gemm_192 mmdeploy::Gelu_193 0=3072 1=1 2=2359296\nGELU Gelu_41 1 1 mmdeploy::Gelu_193 input.24 0=1\nInnerProduct Gemm_42 1 1 input.24 input.28 0=768 1=1 2=2359296\nBinaryOp Add_43 2 1 input.20_splitncnn_0 input.28 input.32 0=0\nSplit splitncnn_6 1 2 input.32 input.32_splitncnn_0 input.32_splitncnn_1\nLayerNorm LayerNorm_44 1 1 input.32_splitncnn_1 qkv_input.7 0=768 1=1.000000e-06 2=1\nSplit splitncnn_7 1 3 qkv_input.7 qkv_input.7_splitncnn_0 qkv_input.7_splitncnn_1 qkv_input.7_splitncnn_2\nMultiHeadAttention MultiHeadAttention_53 3 1 qkv_input.7_splitncnn_2 qkv_input.7_splitncnn_1 qkv_input.7_splitncnn_0 onnx::Add_206 0=768 1=12 2=589824\nBinaryOp Add_54 2 1 input.32_splitncnn_0 onnx::Add_206 input.36 0=0\nSplit splitncnn_8 1 2 input.36 input.36_splitncnn_0 input.36_splitncnn_1\nLayerNorm LayerNorm_55 1 1 input.36_splitncnn_1 mmdeploy::Gemm_208 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_56 1 1 mmdeploy::Gemm_208 mmdeploy::Gelu_209 0=3072 1=1 2=2359296\nGELU Gelu_57 1 1 mmdeploy::Gelu_209 input.40 0=1\nInnerProduct Gemm_58 1 1 input.40 input.44 0=768 1=1 2=2359296\nBinaryOp Add_59 2 1 input.36_splitncnn_0 input.44 input.48 0=0\nSplit splitncnn_9 1 2 input.48 input.48_splitncnn_0 input.48_splitncnn_1\nLayerNorm LayerNorm_60 1 1 input.48_splitncnn_1 qkv_input.11 0=768 1=1.000000e-06 2=1\nSplit splitncnn_10 1 3 qkv_input.11 qkv_input.11_splitncnn_0 qkv_input.11_splitncnn_1 qkv_input.11_splitncnn_2\nMultiHeadAttention MultiHeadAttention_69 3 1 qkv_input.11_splitncnn_2 qkv_input.11_splitncnn_1 qkv_input.11_splitncnn_0 onnx::Add_222 0=768 1=12 2=589824\nBinaryOp Add_70 2 1 input.48_splitncnn_0 onnx::Add_222 input.52 0=0\nSplit splitncnn_11 1 2 input.52 input.52_splitncnn_0 input.52_splitncnn_1\nLayerNorm LayerNorm_71 1 1 input.52_splitncnn_1 mmdeploy::Gemm_224 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_72 1 1 mmdeploy::Gemm_224 mmdeploy::Gelu_225 0=3072 1=1 2=2359296\nGELU Gelu_73 1 1 mmdeploy::Gelu_225 input.56 0=1\nInnerProduct Gemm_74 1 1 input.56 input.60 0=768 1=1 2=2359296\nBinaryOp Add_75 2 1 input.52_splitncnn_0 input.60 input.64 0=0\nSplit splitncnn_12 1 2 input.64 input.64_splitncnn_0 input.64_splitncnn_1\nLayerNorm LayerNorm_76 1 1 input.64_splitncnn_1 qkv_input.15 0=768 1=1.000000e-06 2=1\nSplit splitncnn_13 1 3 qkv_input.15 qkv_input.15_splitncnn_0 qkv_input.15_splitncnn_1 qkv_input.15_splitncnn_2\nMultiHeadAttention MultiHeadAttention_85 3 1 qkv_input.15_splitncnn_2 qkv_input.15_splitncnn_1 qkv_input.15_splitncnn_0 onnx::Add_238 0=768 1=12 2=589824\nBinaryOp Add_86 2 1 input.64_splitncnn_0 onnx::Add_238 input.68 0=0\nSplit splitncnn_14 1 2 input.68 input.68_splitncnn_0 input.68_splitncnn_1\nLayerNorm LayerNorm_87 1 1 input.68_splitncnn_1 mmdeploy::Gemm_240 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_88 1 1 mmdeploy::Gemm_240 mmdeploy::Gelu_241 0=3072 1=1 2=2359296\nGELU Gelu_89 1 1 mmdeploy::Gelu_241 input.72 0=1\nInnerProduct Gemm_90 1 1 input.72 input.76 0=768 1=1 2=2359296\nBinaryOp Add_91 2 1 input.68_splitncnn_0 input.76 input.80 0=0\nSplit splitncnn_15 1 2 input.80 input.80_splitncnn_0 input.80_splitncnn_1\nLayerNorm LayerNorm_92 1 1 input.80_splitncnn_1 qkv_input.19 0=768 1=1.000000e-06 2=1\nSplit splitncnn_16 1 3 qkv_input.19 qkv_input.19_splitncnn_0 qkv_input.19_splitncnn_1 qkv_input.19_splitncnn_2\nMultiHeadAttention MultiHeadAttention_101 3 1 qkv_input.19_splitncnn_2 qkv_input.19_splitncnn_1 qkv_input.19_splitncnn_0 onnx::Add_254 0=768 1=12 2=589824\nBinaryOp Add_102 2 1 input.80_splitncnn_0 onnx::Add_254 input.84 0=0\nSplit splitncnn_17 1 2 input.84 input.84_splitncnn_0 input.84_splitncnn_1\nLayerNorm LayerNorm_103 1 1 input.84_splitncnn_1 mmdeploy::Gemm_256 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_104 1 1 mmdeploy::Gemm_256 mmdeploy::Gelu_257 0=3072 1=1 2=2359296\nGELU Gelu_105 1 1 mmdeploy::Gelu_257 input.88 0=1\nInnerProduct Gemm_106 1 1 input.88 input.92 0=768 1=1 2=2359296\nBinaryOp Add_107 2 1 input.84_splitncnn_0 input.92 input.96 0=0\nSplit splitncnn_18 1 2 input.96 input.96_splitncnn_0 input.96_splitncnn_1\nLayerNorm LayerNorm_108 1 1 input.96_splitncnn_1 qkv_input.23 0=768 1=1.000000e-06 2=1\nSplit splitncnn_19 1 3 qkv_input.23 qkv_input.23_splitncnn_0 qkv_input.23_splitncnn_1 qkv_input.23_splitncnn_2\nMultiHeadAttention MultiHeadAttention_117 3 1 qkv_input.23_splitncnn_2 qkv_input.23_splitncnn_1 qkv_input.23_splitncnn_0 onnx::Add_270 0=768 1=12 2=589824\nBinaryOp Add_118 2 1 input.96_splitncnn_0 onnx::Add_270 input.100 0=0\nSplit splitncnn_20 1 2 input.100 input.100_splitncnn_0 input.100_splitncnn_1\nLayerNorm LayerNorm_119 1 1 input.100_splitncnn_1 mmdeploy::Gemm_272 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_120 1 1 mmdeploy::Gemm_272 mmdeploy::Gelu_273 0=3072 1=1 2=2359296\nGELU Gelu_121 1 1 mmdeploy::Gelu_273 input.104 0=1\nInnerProduct Gemm_122 1 1 input.104 input.108 0=768 1=1 2=2359296\nBinaryOp Add_123 2 1 input.100_splitncnn_0 input.108 input.112 0=0\nSplit splitncnn_21 1 2 input.112 input.112_splitncnn_0 input.112_splitncnn_1\nLayerNorm LayerNorm_124 1 1 input.112_splitncnn_1 qkv_input.27 0=768 1=1.000000e-06 2=1\nSplit splitncnn_22 1 3 qkv_input.27 qkv_input.27_splitncnn_0 qkv_input.27_splitncnn_1 qkv_input.27_splitncnn_2\nMultiHeadAttention MultiHeadAttention_133 3 1 qkv_input.27_splitncnn_2 qkv_input.27_splitncnn_1 qkv_input.27_splitncnn_0 onnx::Add_286 0=768 1=12 2=589824\nBinaryOp Add_134 2 1 input.112_splitncnn_0 onnx::Add_286 input.116 0=0\nSplit splitncnn_23 1 2 input.116 input.116_splitncnn_0 input.116_splitncnn_1\nLayerNorm LayerNorm_135 1 1 input.116_splitncnn_1 mmdeploy::Gemm_288 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_136 1 1 mmdeploy::Gemm_288 mmdeploy::Gelu_289 0=3072 1=1 2=2359296\nGELU Gelu_137 1 1 mmdeploy::Gelu_289 input.120 0=1\nInnerProduct Gemm_138 1 1 input.120 input.124 0=768 1=1 2=2359296\nBinaryOp Add_139 2 1 input.116_splitncnn_0 input.124 input.128 0=0\nSplit splitncnn_24 1 2 input.128 input.128_splitncnn_0 input.128_splitncnn_1\nLayerNorm LayerNorm_140 1 1 input.128_splitncnn_1 qkv_input.31 0=768 1=1.000000e-06 2=1\nSplit splitncnn_25 1 3 qkv_input.31 qkv_input.31_splitncnn_0 qkv_input.31_splitncnn_1 qkv_input.31_splitncnn_2\nMultiHeadAttention MultiHeadAttention_149 3 1 qkv_input.31_splitncnn_2 qkv_input.31_splitncnn_1 qkv_input.31_splitncnn_0 onnx::Add_302 0=768 1=12 2=589824\nBinaryOp Add_150 2 1 input.128_splitncnn_0 onnx::Add_302 input.132 0=0\nSplit splitncnn_26 1 2 input.132 input.132_splitncnn_0 input.132_splitncnn_1\nLayerNorm LayerNorm_151 1 1 input.132_splitncnn_1 mmdeploy::Gemm_304 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_152 1 1 mmdeploy::Gemm_304 mmdeploy::Gelu_305 0=3072 1=1 2=2359296\nGELU Gelu_153 1 1 mmdeploy::Gelu_305 input.136 0=1\nInnerProduct Gemm_154 1 1 input.136 input.140 0=768 1=1 2=2359296\nBinaryOp Add_155 2 1 input.132_splitncnn_0 input.140 input.144 0=0\nSplit splitncnn_27 1 2 input.144 input.144_splitncnn_0 input.144_splitncnn_1\nLayerNorm LayerNorm_156 1 1 input.144_splitncnn_1 qkv_input.35 0=768 1=1.000000e-06 2=1\nSplit splitncnn_28 1 3 qkv_input.35 qkv_input.35_splitncnn_0 qkv_input.35_splitncnn_1 qkv_input.35_splitncnn_2\nMultiHeadAttention MultiHeadAttention_165 3 1 qkv_input.35_splitncnn_2 qkv_input.35_splitncnn_1 qkv_input.35_splitncnn_0 onnx::Add_318 0=768 1=12 2=589824\nBinaryOp Add_166 2 1 input.144_splitncnn_0 onnx::Add_318 input.148 0=0\nSplit splitncnn_29 1 2 input.148 input.148_splitncnn_0 input.148_splitncnn_1\nLayerNorm LayerNorm_167 1 1 input.148_splitncnn_1 mmdeploy::Gemm_320 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_168 1 1 mmdeploy::Gemm_320 mmdeploy::Gelu_321 0=3072 1=1 2=2359296\nGELU Gelu_169 1 1 mmdeploy::Gelu_321 input.152 0=1\nInnerProduct Gemm_170 1 1 input.152 input.156 0=768 1=1 2=2359296\nBinaryOp Add_171 2 1 input.148_splitncnn_0 input.156 input.160 0=0\nSplit splitncnn_30 1 2 input.160 input.160_splitncnn_0 input.160_splitncnn_1\nLayerNorm LayerNorm_172 1 1 input.160_splitncnn_1 qkv_input.39 0=768 1=1.000000e-06 2=1\nSplit splitncnn_31 1 3 qkv_input.39 qkv_input.39_splitncnn_0 qkv_input.39_splitncnn_1 qkv_input.39_splitncnn_2\nMultiHeadAttention MultiHeadAttention_181 3 1 qkv_input.39_splitncnn_2 qkv_input.39_splitncnn_1 qkv_input.39_splitncnn_0 onnx::Add_334 0=768 1=12 2=589824\nBinaryOp Add_182 2 1 input.160_splitncnn_0 onnx::Add_334 input.164 0=0\nSplit splitncnn_32 1 2 input.164 input.164_splitncnn_0 input.164_splitncnn_1\nLayerNorm LayerNorm_183 1 1 input.164_splitncnn_1 mmdeploy::Gemm_336 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_184 1 1 mmdeploy::Gemm_336 mmdeploy::Gelu_337 0=3072 1=1 2=2359296\nGELU Gelu_185 1 1 mmdeploy::Gelu_337 input.168 0=1\nInnerProduct Gemm_186 1 1 input.168 input.172 0=768 1=1 2=2359296\nBinaryOp Add_187 2 1 input.164_splitncnn_0 input.172 input.176 0=0\nSplit splitncnn_33 1 2 input.176 input.176_splitncnn_0 input.176_splitncnn_1\nLayerNorm LayerNorm_188 1 1 input.176_splitncnn_1 qkv_input.43 0=768 1=1.000000e-06 2=1\nSplit splitncnn_34 1 3 qkv_input.43 qkv_input.43_splitncnn_0 qkv_input.43_splitncnn_1 qkv_input.43_splitncnn_2\nMultiHeadAttention MultiHeadAttention_197 3 1 qkv_input.43_splitncnn_2 qkv_input.43_splitncnn_1 qkv_input.43_splitncnn_0 onnx::Add_350 0=768 1=12 2=589824\nBinaryOp Add_198 2 1 input.176_splitncnn_0 onnx::Add_350 input.180 0=0\nSplit splitncnn_35 1 2 input.180 input.180_splitncnn_0 input.180_splitncnn_1\nLayerNorm LayerNorm_199 1 1 input.180_splitncnn_1 mmdeploy::Gemm_352 0=768 1=1.000000e-06 2=1\nInnerProduct Gemm_200 1 1 mmdeploy::Gemm_352 mmdeploy::Gelu_353 0=3072 1=1 2=2359296\nGELU Gelu_201 1 1 mmdeploy::Gelu_353 input.184 0=1\nInnerProduct Gemm_202 1 1 input.184 input.188 0=768 1=1 2=2359296\nBinaryOp Add_203 2 1 input.180_splitncnn_0 input.188 input.192 0=0\nLayerNorm LayerNorm_204 1 1 input.192 onnx::Gather_357 0=768 1=1.000000e-06 2=1\nCrop Gather_206 1 1 onnx::Gather_357 mmdeploy::Gemm_359 -23309=1,0 -23310=1,1 -23311=1,0\nInnerProduct Gemm_207 1 1 mmdeploy::Gemm_359 cls_score 0=1000 1=1 2=768000\nSoftmax Softmax_208 1 1 cls_score output 0=0 1=1\n" }, { "path": "benchmark/yolo-fastest-1.1.param", "content": "7767517\n131 154\nInput data 0 1 data -23330=4,3,320,320,3 0=320 1=320 2=3\nConvolution 0_22 1 1 data 0_22_bn_leaky -23330=4,3,160,160,8 0=8 1=3 3=2 4=1 5=1 6=216 9=2 -23310=1,1.000000e-01\nConvolution 1_31 1 1 0_22_bn_leaky 1_31_bn_leaky -23330=4,3,160,160,8 0=8 1=1 5=1 6=64 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 2_39 1 1 1_31_bn_leaky 2_39_bn_leaky -23330=4,3,160,160,8 0=8 1=3 4=1 5=1 6=72 7=8 9=2 -23310=1,1.000000e-01\nConvolution 3_48 1 1 2_39_bn_leaky 3_48_bn -23330=4,3,160,160,4 0=4 1=1 5=1 6=32\nSplit 3_48_bn_split 1 2 3_48_bn 3_48_bn_split_0 3_48_bn_split_1 -23330=8,3,160,160,4,3,160,160,4\nConvolution 4_57 1 1 3_48_bn_split_0 4_57_bn_leaky -23330=4,3,160,160,8 0=8 1=1 5=1 6=32 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 5_65 1 1 4_57_bn_leaky 5_65_bn_leaky -23330=4,3,160,160,8 0=8 1=3 4=1 5=1 6=72 7=8 9=2 -23310=1,1.000000e-01\nConvolution 6_74 1 1 5_65_bn_leaky 6_74_bn -23330=4,3,160,160,4 0=4 1=1 5=1 6=32\nEltwise 8_86 2 1 6_74_bn 3_48_bn_split_1 8_86 -23330=4,3,160,160,4 0=1\nConvolution 9_90 1 1 8_86 9_90_bn_leaky -23330=4,3,160,160,24 0=24 1=1 5=1 6=96 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 10_98 1 1 9_90_bn_leaky 10_98_bn_leaky -23330=4,3,80,80,24 0=24 1=3 3=2 4=1 5=1 6=216 7=24 9=2 -23310=1,1.000000e-01\nConvolution 11_107 1 1 10_98_bn_leaky 11_107_bn -23330=4,3,80,80,8 0=8 1=1 5=1 6=192\nSplit 11_107_bn_split 1 2 11_107_bn 11_107_bn_split_0 11_107_bn_split_1 -23330=8,3,80,80,8,3,80,80,8\nConvolution 12_116 1 1 11_107_bn_split_0 12_116_bn_leaky -23330=4,3,80,80,32 0=32 1=1 5=1 6=256 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 13_124 1 1 12_116_bn_leaky 13_124_bn_leaky -23330=4,3,80,80,32 0=32 1=3 4=1 5=1 6=288 7=32 9=2 -23310=1,1.000000e-01\nConvolution 14_133 1 1 13_124_bn_leaky 14_133_bn -23330=4,3,80,80,8 0=8 1=1 5=1 6=256\nEltwise 16_145 2 1 14_133_bn 11_107_bn_split_1 16_145 -23330=4,3,80,80,8 0=1\nSplit 16_145_split 1 2 16_145 16_145_split_0 16_145_split_1 -23330=8,3,80,80,8,3,80,80,8\nConvolution 17_149 1 1 16_145_split_0 17_149_bn_leaky -23330=4,3,80,80,32 0=32 1=1 5=1 6=256 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 18_157 1 1 17_149_bn_leaky 18_157_bn_leaky -23330=4,3,80,80,32 0=32 1=3 4=1 5=1 6=288 7=32 9=2 -23310=1,1.000000e-01\nConvolution 19_166 1 1 18_157_bn_leaky 19_166_bn -23330=4,3,80,80,8 0=8 1=1 5=1 6=256\nEltwise 21_179 2 1 19_166_bn 16_145_split_1 21_179 -23330=4,3,80,80,8 0=1\nConvolution 22_183 1 1 21_179 22_183_bn_leaky -23330=4,3,80,80,32 0=32 1=1 5=1 6=256 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 23_191 1 1 22_183_bn_leaky 23_191_bn_leaky -23330=4,3,40,40,32 0=32 1=3 3=2 4=1 5=1 6=288 7=32 9=2 -23310=1,1.000000e-01\nConvolution 24_200 1 1 23_191_bn_leaky 24_200_bn -23330=4,3,40,40,8 0=8 1=1 5=1 6=256\nSplit 24_200_bn_split 1 2 24_200_bn 24_200_bn_split_0 24_200_bn_split_1 -23330=8,3,40,40,8,3,40,40,8\nConvolution 25_209 1 1 24_200_bn_split_0 25_209_bn_leaky -23330=4,3,40,40,48 0=48 1=1 5=1 6=384 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 26_217 1 1 25_209_bn_leaky 26_217_bn_leaky -23330=4,3,40,40,48 0=48 1=3 4=1 5=1 6=432 7=48 9=2 -23310=1,1.000000e-01\nConvolution 27_226 1 1 26_217_bn_leaky 27_226_bn -23330=4,3,40,40,8 0=8 1=1 5=1 6=384\nEltwise 29_238 2 1 27_226_bn 24_200_bn_split_1 29_238 -23330=4,3,40,40,8 0=1\nSplit 29_238_split 1 2 29_238 29_238_split_0 29_238_split_1 -23330=8,3,40,40,8,3,40,40,8\nConvolution 30_242 1 1 29_238_split_0 30_242_bn_leaky -23330=4,3,40,40,48 0=48 1=1 5=1 6=384 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 31_250 1 1 30_242_bn_leaky 31_250_bn_leaky -23330=4,3,40,40,48 0=48 1=3 4=1 5=1 6=432 7=48 9=2 -23310=1,1.000000e-01\nConvolution 32_259 1 1 31_250_bn_leaky 32_259_bn -23330=4,3,40,40,8 0=8 1=1 5=1 6=384\nEltwise 34_273 2 1 32_259_bn 29_238_split_1 34_273 -23330=4,3,40,40,8 0=1\nConvolution 35_277 1 1 34_273 35_277_bn_leaky -23330=4,3,40,40,48 0=48 1=1 5=1 6=384 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 36_285 1 1 35_277_bn_leaky 36_285_bn_leaky -23330=4,3,40,40,48 0=48 1=3 4=1 5=1 6=432 7=48 9=2 -23310=1,1.000000e-01\nConvolution 37_294 1 1 36_285_bn_leaky 37_294_bn -23330=4,3,40,40,16 0=16 1=1 5=1 6=768\nSplit 37_294_bn_split 1 2 37_294_bn 37_294_bn_split_0 37_294_bn_split_1 -23330=8,3,40,40,16,3,40,40,16\nConvolution 38_303 1 1 37_294_bn_split_0 38_303_bn_leaky -23330=4,3,40,40,96 0=96 1=1 5=1 6=1536 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 39_311 1 1 38_303_bn_leaky 39_311_bn_leaky -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution 40_320 1 1 39_311_bn_leaky 40_320_bn -23330=4,3,40,40,16 0=16 1=1 5=1 6=1536\nEltwise 42_332 2 1 40_320_bn 37_294_bn_split_1 42_332 -23330=4,3,40,40,16 0=1\nSplit 42_332_split 1 2 42_332 42_332_split_0 42_332_split_1 -23330=8,3,40,40,16,3,40,40,16\nConvolution 43_336 1 1 42_332_split_0 43_336_bn_leaky -23330=4,3,40,40,96 0=96 1=1 5=1 6=1536 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 44_344 1 1 43_336_bn_leaky 44_344_bn_leaky -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution 45_353 1 1 44_344_bn_leaky 45_353_bn -23330=4,3,40,40,16 0=16 1=1 5=1 6=1536\nEltwise 47_365 2 1 45_353_bn 42_332_split_1 47_365 -23330=4,3,40,40,16 0=1\nSplit 47_365_split 1 2 47_365 47_365_split_0 47_365_split_1 -23330=8,3,40,40,16,3,40,40,16\nConvolution 48_369 1 1 47_365_split_0 48_369_bn_leaky -23330=4,3,40,40,96 0=96 1=1 5=1 6=1536 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 49_377 1 1 48_369_bn_leaky 49_377_bn_leaky -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution 50_386 1 1 49_377_bn_leaky 50_386_bn -23330=4,3,40,40,16 0=16 1=1 5=1 6=1536\nEltwise 52_399 2 1 50_386_bn 47_365_split_1 52_399 -23330=4,3,40,40,16 0=1\nSplit 52_399_split 1 2 52_399 52_399_split_0 52_399_split_1 -23330=8,3,40,40,16,3,40,40,16\nConvolution 53_403 1 1 52_399_split_0 53_403_bn_leaky -23330=4,3,40,40,96 0=96 1=1 5=1 6=1536 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 54_411 1 1 53_403_bn_leaky 54_411_bn_leaky -23330=4,3,40,40,96 0=96 1=3 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution 55_420 1 1 54_411_bn_leaky 55_420_bn -23330=4,3,40,40,16 0=16 1=1 5=1 6=1536\nEltwise 57_433 2 1 55_420_bn 52_399_split_1 57_433 -23330=4,3,40,40,16 0=1\nConvolution 58_437 1 1 57_433 58_437_bn_leaky -23330=4,3,40,40,96 0=96 1=1 5=1 6=1536 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 59_445 1 1 58_437_bn_leaky 59_445_bn_leaky -23330=4,3,20,20,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96 9=2 -23310=1,1.000000e-01\nConvolution 60_454 1 1 59_445_bn_leaky 60_454_bn -23330=4,3,20,20,24 0=24 1=1 5=1 6=2304\nSplit 60_454_bn_split 1 2 60_454_bn 60_454_bn_split_0 60_454_bn_split_1 -23330=8,3,20,20,24,3,20,20,24\nConvolution 61_463 1 1 60_454_bn_split_0 61_463_bn_leaky -23330=4,3,20,20,136 0=136 1=1 5=1 6=3264 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 62_471 1 1 61_463_bn_leaky 62_471_bn_leaky -23330=4,3,20,20,136 0=136 1=3 4=1 5=1 6=1224 7=136 9=2 -23310=1,1.000000e-01\nConvolution 63_480 1 1 62_471_bn_leaky 63_480_bn -23330=4,3,20,20,24 0=24 1=1 5=1 6=3264\nEltwise 65_492 2 1 63_480_bn 60_454_bn_split_1 65_492 -23330=4,3,20,20,24 0=1\nSplit 65_492_split 1 2 65_492 65_492_split_0 65_492_split_1 -23330=8,3,20,20,24,3,20,20,24\nConvolution 66_496 1 1 65_492_split_0 66_496_bn_leaky -23330=4,3,20,20,136 0=136 1=1 5=1 6=3264 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 67_504 1 1 66_496_bn_leaky 67_504_bn_leaky -23330=4,3,20,20,136 0=136 1=3 4=1 5=1 6=1224 7=136 9=2 -23310=1,1.000000e-01\nConvolution 68_513 1 1 67_504_bn_leaky 68_513_bn -23330=4,3,20,20,24 0=24 1=1 5=1 6=3264\nEltwise 70_526 2 1 68_513_bn 65_492_split_1 70_526 -23330=4,3,20,20,24 0=1\nSplit 70_526_split 1 2 70_526 70_526_split_0 70_526_split_1 -23330=8,3,20,20,24,3,20,20,24\nConvolution 71_530 1 1 70_526_split_0 71_530_bn_leaky -23330=4,3,20,20,136 0=136 1=1 5=1 6=3264 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 72_538 1 1 71_530_bn_leaky 72_538_bn_leaky -23330=4,3,20,20,136 0=136 1=3 4=1 5=1 6=1224 7=136 9=2 -23310=1,1.000000e-01\nConvolution 73_547 1 1 72_538_bn_leaky 73_547_bn -23330=4,3,20,20,24 0=24 1=1 5=1 6=3264\nEltwise 75_559 2 1 73_547_bn 70_526_split_1 75_559 -23330=4,3,20,20,24 0=1\nSplit 75_559_split 1 2 75_559 75_559_split_0 75_559_split_1 -23330=8,3,20,20,24,3,20,20,24\nConvolution 76_563 1 1 75_559_split_0 76_563_bn_leaky -23330=4,3,20,20,136 0=136 1=1 5=1 6=3264 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 77_571 1 1 76_563_bn_leaky 77_571_bn_leaky -23330=4,3,20,20,136 0=136 1=3 4=1 5=1 6=1224 7=136 9=2 -23310=1,1.000000e-01\nConvolution 78_580 1 1 77_571_bn_leaky 78_580_bn -23330=4,3,20,20,24 0=24 1=1 5=1 6=3264\nEltwise 80_593 2 1 78_580_bn 75_559_split_1 80_593 -23330=4,3,20,20,24 0=1\nSplit 80_593_split 1 2 80_593 80_593_split_0 80_593_split_1 -23330=8,3,20,20,24,3,20,20,24\nConvolution 81_597 1 1 80_593_split_0 81_597_bn_leaky -23330=4,3,20,20,136 0=136 1=1 5=1 6=3264 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 82_605 1 1 81_597_bn_leaky 82_605_bn_leaky -23330=4,3,10,10,136 0=136 1=3 3=2 4=1 5=1 6=1224 7=136 9=2 -23310=1,1.000000e-01\nConvolution 83_615 1 1 82_605_bn_leaky 83_615_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=6528\nSplit 83_615_bn_split 1 2 83_615_bn 83_615_bn_split_0 83_615_bn_split_1 -23330=8,3,10,10,48,3,10,10,48\nConvolution 84_624 1 1 83_615_bn_split_0 84_624_bn_leaky -23330=4,3,10,10,224 0=224 1=1 5=1 6=10752 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 85_632 1 1 84_624_bn_leaky 85_632_bn_leaky -23330=4,3,10,10,224 0=224 1=3 4=1 5=1 6=2016 7=224 9=2 -23310=1,1.000000e-01\nConvolution 86_641 1 1 85_632_bn_leaky 86_641_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=10752\nEltwise 88_653 2 1 86_641_bn 83_615_bn_split_1 88_653 -23330=4,3,10,10,48 0=1\nSplit 88_653_split 1 2 88_653 88_653_split_0 88_653_split_1 -23330=8,3,10,10,48,3,10,10,48\nConvolution 89_657 1 1 88_653_split_0 89_657_bn_leaky -23330=4,3,10,10,224 0=224 1=1 5=1 6=10752 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 90_665 1 1 89_657_bn_leaky 90_665_bn_leaky -23330=4,3,10,10,224 0=224 1=3 4=1 5=1 6=2016 7=224 9=2 -23310=1,1.000000e-01\nConvolution 91_674 1 1 90_665_bn_leaky 91_674_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=10752\nEltwise 93_686 2 1 91_674_bn 88_653_split_1 93_686 -23330=4,3,10,10,48 0=1\nSplit 93_686_split 1 2 93_686 93_686_split_0 93_686_split_1 -23330=8,3,10,10,48,3,10,10,48\nConvolution 94_690 1 1 93_686_split_0 94_690_bn_leaky -23330=4,3,10,10,224 0=224 1=1 5=1 6=10752 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 95_698 1 1 94_690_bn_leaky 95_698_bn_leaky -23330=4,3,10,10,224 0=224 1=3 4=1 5=1 6=2016 7=224 9=2 -23310=1,1.000000e-01\nConvolution 96_707 1 1 95_698_bn_leaky 96_707_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=10752\nEltwise 98_719 2 1 96_707_bn 93_686_split_1 98_719 -23330=4,3,10,10,48 0=1\nSplit 98_719_split 1 2 98_719 98_719_split_0 98_719_split_1 -23330=8,3,10,10,48,3,10,10,48\nConvolution 99_723 1 1 98_719_split_0 99_723_bn_leaky -23330=4,3,10,10,224 0=224 1=1 5=1 6=10752 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 100_731 1 1 99_723_bn_leaky 100_731_bn_leaky -23330=4,3,10,10,224 0=224 1=3 4=1 5=1 6=2016 7=224 9=2 -23310=1,1.000000e-01\nConvolution 101_740 1 1 100_731_bn_leaky 101_740_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=10752\nEltwise 103_752 2 1 101_740_bn 98_719_split_1 103_752 -23330=4,3,10,10,48 0=1\nSplit 103_752_split 1 2 103_752 103_752_split_0 103_752_split_1 -23330=8,3,10,10,48,3,10,10,48\nConvolution 104_756 1 1 103_752_split_0 104_756_bn_leaky -23330=4,3,10,10,224 0=224 1=1 5=1 6=10752 9=2 -23310=1,1.000000e-01\nConvolutionDepthWise 105_764 1 1 104_756_bn_leaky 105_764_bn_leaky -23330=4,3,10,10,224 0=224 1=3 4=1 5=1 6=2016 7=224 9=2 -23310=1,1.000000e-01\nConvolution 106_773 1 1 105_764_bn_leaky 106_773_bn -23330=4,3,10,10,48 0=48 1=1 5=1 6=10752\nEltwise 108_784 2 1 106_773_bn 103_752_split_1 108_784 -23330=4,3,10,10,48 0=1\nSplit 108_784_split 1 4 108_784 108_784_split_0 108_784_split_1 108_784_split_2 108_784_split_3 -23330=16,3,10,10,48,3,10,10,48,3,10,10,48,3,10,10,48\nPooling 109_788 1 1 108_784_split_0 109_788 -23330=4,3,10,10,48 1=3 3=1 5=1\nPooling 111_795 1 1 108_784_split_1 111_795 -23330=4,3,10,10,48 1=5 3=2 5=1\nPooling 113_802 1 1 108_784_split_2 113_802 -23330=4,3,10,10,48 1=9 3=4 5=1\nConcat 114_806 4 1 113_802 111_795 109_788 108_784_split_3 114_806 -23330=4,3,10,10,192\nConvolution 115_811 1 1 114_806 115_811_bn_leaky -23330=4,3,10,10,96 0=96 1=1 5=1 6=18432 9=2 -23310=1,1.000000e-01\nSplit 115_811_bn_leaky_split 1 2 115_811_bn_leaky 115_811_bn_leaky_split_0 115_811_bn_leaky_split_1 -23330=8,3,10,10,96,3,10,10,96\nConvolutionDepthWise 116_819 1 1 115_811_bn_leaky_split_0 116_819_bn_leaky -23330=4,3,10,10,96 0=96 1=5 4=2 5=1 6=2400 7=96 9=2 -23310=1,1.000000e-01\nConvolution 117_828 1 1 116_819_bn_leaky 117_828_bn -23330=4,3,10,10,96 0=96 1=1 5=1 6=9216\nConvolutionDepthWise 118_836 1 1 117_828_bn 118_836_bn_leaky -23330=4,3,10,10,96 0=96 1=5 4=2 5=1 6=2400 7=96 9=2 -23310=1,1.000000e-01\nConvolution 119_845 1 1 118_836_bn_leaky 119_845_bn -23330=4,3,10,10,96 0=96 1=1 5=1 6=9216\nConvolution 120_854 1 1 119_845_bn 120_854 -23330=4,3,10,10,255 0=255 1=1 5=1 6=24480\nInterp 123_882 1 1 115_811_bn_leaky_split_1 123_882 -23330=4,3,20,20,96 0=1 1=2.000000e+00 2=2.000000e+00\nConcat 124_885 2 1 123_882 80_593_split_1 124_885 -23330=4,3,20,20,120\nConvolutionDepthWise 125_888 1 1 124_885 125_888_bn_leaky -23330=4,3,20,20,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=2 -23310=1,1.000000e-01\nConvolution 126_897 1 1 125_888_bn_leaky 126_897_bn -23330=4,3,20,20,120 0=120 1=1 5=1 6=14400\nConvolutionDepthWise 127_905 1 1 126_897_bn 127_905_bn_leaky -23330=4,3,20,20,120 0=120 1=5 4=2 5=1 6=3000 7=120 9=2 -23310=1,1.000000e-01\nConvolution 128_914 1 1 127_905_bn_leaky 128_914_bn -23330=4,3,20,20,120 0=120 1=1 5=1 6=14400\nConvolution 129_922 1 1 128_914_bn 129_922 -23330=4,3,20,20,255 0=255 1=1 5=1 6=30600\nYolov3DetectionOutput detection_out 2 1 120_854 129_922 output -23330=4,2,6,1431,1 0=80 1=3 2=5.500000e-01 -23304=12,1.200000e+01,1.800000e+01,3.700000e+01,4.900000e+01,5.200000e+01,1.320000e+02,1.150000e+02,7.300000e+01,1.190000e+02,1.990000e+02,2.420000e+02,2.380000e+02 -23305=6,1077936128,1082130432,1084227584,0,1065353216,1073741824 -23306=2,3.200000e+01,1.600000e+01\n" }, { "path": "benchmark/yolo-fastestv2.param", "content": "7767517\n144 166\nInput input.1 0 1 input.1 -23330=4,3,352,352,3 0=352 1=352 2=3\nConvolution Conv_0 1 1 input.1 447 -23330=4,3,176,176,24 0=24 1=3 3=2 4=1 5=1 6=648 9=1\nPooling MaxPool_2 1 1 447 448 -23330=4,3,88,88,24 1=3 2=2 3=1 5=1\nSplit splitncnn_0 1 2 448 448_splitncnn_0 448_splitncnn_1 -23330=8,3,88,88,24,3,88,88,24\nConvolutionDepthWise Conv_3 1 1 448_splitncnn_1 800 -23330=4,3,44,44,24 0=24 1=3 3=2 4=1 5=1 6=216 7=24\nConvolution Conv_4 1 1 800 453 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConvolution Conv_6 1 1 448_splitncnn_0 456 -23330=4,3,88,88,24 0=24 1=1 5=1 6=576 9=1\nConvolutionDepthWise Conv_8 1 1 456 809 -23330=4,3,44,44,24 0=24 1=3 3=2 4=1 5=1 6=216 7=24\nConvolution Conv_9 1 1 809 461 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConcat Concat_11 2 1 453 461 462 -23330=4,3,44,44,48\nShuffleChannel Reshape_16 1 1 462 467 -23330=4,3,44,44,48 0=2 1=1\nSlice Gather_20 1 2 467 469 471 -23330=8,3,44,44,24,3,44,44,24 -23300=2,-233,-233\nConvolution Conv_21 1 1 471 474 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConvolutionDepthWise Conv_23 1 1 474 818 -23330=4,3,44,44,24 0=24 1=3 4=1 5=1 6=216 7=24\nConvolution Conv_24 1 1 818 479 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConcat Concat_26 2 1 469 479 480 -23330=4,3,44,44,48\nShuffleChannel Reshape_31 1 1 480 485 -23330=4,3,44,44,48 0=2 1=1\nSlice Gather_35 1 2 485 487 489 -23330=8,3,44,44,24,3,44,44,24 -23300=2,-233,-233\nConvolution Conv_36 1 1 489 492 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConvolutionDepthWise Conv_38 1 1 492 827 -23330=4,3,44,44,24 0=24 1=3 4=1 5=1 6=216 7=24\nConvolution Conv_39 1 1 827 497 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConcat Concat_41 2 1 487 497 498 -23330=4,3,44,44,48\nShuffleChannel Reshape_46 1 1 498 503 -23330=4,3,44,44,48 0=2 1=1\nSlice Gather_50 1 2 503 505 507 -23330=8,3,44,44,24,3,44,44,24 -23300=2,-233,-233\nConvolution Conv_51 1 1 507 510 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConvolutionDepthWise Conv_53 1 1 510 836 -23330=4,3,44,44,24 0=24 1=3 4=1 5=1 6=216 7=24\nConvolution Conv_54 1 1 836 515 -23330=4,3,44,44,24 0=24 1=1 5=1 6=576 9=1\nConcat Concat_56 2 1 505 515 516 -23330=4,3,44,44,48\nSplit splitncnn_1 1 2 516 516_splitncnn_0 516_splitncnn_1 -23330=8,3,44,44,48,3,44,44,48\nConvolutionDepthWise Conv_57 1 1 516_splitncnn_1 842 -23330=4,3,22,22,48 0=48 1=3 3=2 4=1 5=1 6=432 7=48\nConvolution Conv_58 1 1 842 521 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolution Conv_60 1 1 516_splitncnn_0 524 -23330=4,3,44,44,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_62 1 1 524 851 -23330=4,3,22,22,48 0=48 1=3 3=2 4=1 5=1 6=432 7=48\nConvolution Conv_63 1 1 851 529 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_65 2 1 521 529 530 -23330=4,3,22,22,96\nShuffleChannel Reshape_70 1 1 530 535 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_74 1 2 535 537 539 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_75 1 1 539 542 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_77 1 1 542 860 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_78 1 1 860 547 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_80 2 1 537 547 548 -23330=4,3,22,22,96\nShuffleChannel Reshape_85 1 1 548 553 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_89 1 2 553 555 557 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_90 1 1 557 560 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_92 1 1 560 869 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_93 1 1 869 565 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_95 2 1 555 565 566 -23330=4,3,22,22,96\nShuffleChannel Reshape_100 1 1 566 571 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_104 1 2 571 573 575 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_105 1 1 575 578 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_107 1 1 578 878 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_108 1 1 878 583 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_110 2 1 573 583 584 -23330=4,3,22,22,96\nShuffleChannel Reshape_115 1 1 584 589 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_119 1 2 589 591 593 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_120 1 1 593 596 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_122 1 1 596 887 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_123 1 1 887 601 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_125 2 1 591 601 602 -23330=4,3,22,22,96\nShuffleChannel Reshape_130 1 1 602 607 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_134 1 2 607 609 611 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_135 1 1 611 614 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_137 1 1 614 896 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_138 1 1 896 619 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_140 2 1 609 619 620 -23330=4,3,22,22,96\nShuffleChannel Reshape_145 1 1 620 625 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_149 1 2 625 627 629 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_150 1 1 629 632 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_152 1 1 632 905 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_153 1 1 905 637 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_155 2 1 627 637 638 -23330=4,3,22,22,96\nShuffleChannel Reshape_160 1 1 638 643 -23330=4,3,22,22,96 0=2 1=1\nSlice Gather_164 1 2 643 645 647 -23330=8,3,22,22,48,3,22,22,48 -23300=2,-233,-233\nConvolution Conv_165 1 1 647 650 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConvolutionDepthWise Conv_167 1 1 650 914 -23330=4,3,22,22,48 0=48 1=3 4=1 5=1 6=432 7=48\nConvolution Conv_168 1 1 914 655 -23330=4,3,22,22,48 0=48 1=1 5=1 6=2304 9=1\nConcat Concat_170 2 1 645 655 656 -23330=4,3,22,22,96\nSplit splitncnn_2 1 3 656 656_splitncnn_0 656_splitncnn_1 656_splitncnn_2 -23330=12,3,22,22,96,3,22,22,96,3,22,22,96\nConvolutionDepthWise Conv_171 1 1 656_splitncnn_2 920 -23330=4,3,11,11,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96\nConvolution Conv_172 1 1 920 661 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConvolution Conv_174 1 1 656_splitncnn_1 664 -23330=4,3,22,22,96 0=96 1=1 5=1 6=9216 9=1\nConvolutionDepthWise Conv_176 1 1 664 929 -23330=4,3,11,11,96 0=96 1=3 3=2 4=1 5=1 6=864 7=96\nConvolution Conv_177 1 1 929 669 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConcat Concat_179 2 1 661 669 670 -23330=4,3,11,11,192\nShuffleChannel Reshape_184 1 1 670 675 -23330=4,3,11,11,192 0=2 1=1\nSlice Gather_188 1 2 675 677 679 -23330=8,3,11,11,96,3,11,11,96 -23300=2,-233,-233\nConvolution Conv_189 1 1 679 682 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConvolutionDepthWise Conv_191 1 1 682 938 -23330=4,3,11,11,96 0=96 1=3 4=1 5=1 6=864 7=96\nConvolution Conv_192 1 1 938 687 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConcat Concat_194 2 1 677 687 688 -23330=4,3,11,11,192\nShuffleChannel Reshape_199 1 1 688 693 -23330=4,3,11,11,192 0=2 1=1\nSlice Gather_203 1 2 693 695 697 -23330=8,3,11,11,96,3,11,11,96 -23300=2,-233,-233\nConvolution Conv_204 1 1 697 700 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConvolutionDepthWise Conv_206 1 1 700 947 -23330=4,3,11,11,96 0=96 1=3 4=1 5=1 6=864 7=96\nConvolution Conv_207 1 1 947 705 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConcat Concat_209 2 1 695 705 706 -23330=4,3,11,11,192\nShuffleChannel Reshape_214 1 1 706 711 -23330=4,3,11,11,192 0=2 1=1\nSlice Gather_218 1 2 711 713 715 -23330=8,3,11,11,96,3,11,11,96 -23300=2,-233,-233\nConvolution Conv_219 1 1 715 718 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConvolutionDepthWise Conv_221 1 1 718 956 -23330=4,3,11,11,96 0=96 1=3 4=1 5=1 6=864 7=96\nConvolution Conv_222 1 1 956 723 -23330=4,3,11,11,96 0=96 1=1 5=1 6=9216 9=1\nConcat Concat_224 2 1 713 723 724 -23330=4,3,11,11,192\nSplit splitncnn_3 1 2 724 724_splitncnn_0 724_splitncnn_1 -23330=8,3,11,11,192,3,11,11,192\nConvolution Conv_225 1 1 724_splitncnn_1 727 -23330=4,3,11,11,72 0=72 1=1 5=1 6=13824 9=1\nSplit splitncnn_4 1 2 727 727_splitncnn_0 727_splitncnn_1 -23330=8,3,11,11,72,3,11,11,72\nConvolutionDepthWise Conv_227 1 1 727_splitncnn_1 730 -23330=4,3,11,11,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_229 1 1 730 968 -23330=4,3,11,11,72 0=72 1=1 5=1 6=5184\nConvolutionDepthWise Conv_230 1 1 968 735 -23330=4,3,11,11,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_232 1 1 735 974 -23330=4,3,11,11,72 0=72 1=1 5=1 6=5184\nSplit splitncnn_5 1 2 974 974_splitncnn_0 974_splitncnn_1 -23330=8,3,11,11,72,3,11,11,72\nConvolutionDepthWise Conv_233 1 1 727_splitncnn_0 740 -23330=4,3,11,11,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_235 1 1 740 980 -23330=4,3,11,11,72 0=72 1=1 5=1 6=5184\nConvolutionDepthWise Conv_236 1 1 980 745 -23330=4,3,11,11,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_238 1 1 745 986 -23330=4,3,11,11,72 0=72 1=1 5=1 6=5184\nInterp Resize_240 1 1 724_splitncnn_0 752 -23330=4,3,22,22,192 0=1 1=2.000000e+00 2=2.000000e+00\nConcat Concat_241 2 1 752 656_splitncnn_0 753 -23330=4,3,22,22,288\nConvolution Conv_242 1 1 753 756 -23330=4,3,22,22,72 0=72 1=1 5=1 6=20736 9=1\nSplit splitncnn_6 1 2 756 756_splitncnn_0 756_splitncnn_1 -23330=8,3,22,22,72,3,22,22,72\nConvolutionDepthWise Conv_244 1 1 756_splitncnn_1 759 -23330=4,3,22,22,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_246 1 1 759 995 -23330=4,3,22,22,72 0=72 1=1 5=1 6=5184\nConvolutionDepthWise Conv_247 1 1 995 764 -23330=4,3,22,22,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_249 1 1 764 1001 -23330=4,3,22,22,72 0=72 1=1 5=1 6=5184\nSplit splitncnn_7 1 2 1001 1001_splitncnn_0 1001_splitncnn_1 -23330=8,3,22,22,72,3,22,22,72\nConvolutionDepthWise Conv_250 1 1 756_splitncnn_0 769 -23330=4,3,22,22,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_252 1 1 769 1007 -23330=4,3,22,22,72 0=72 1=1 5=1 6=5184\nConvolutionDepthWise Conv_253 1 1 1007 774 -23330=4,3,22,22,72 0=72 1=5 4=2 5=1 6=1800 7=72 9=1\nConvolution Conv_255 1 1 774 1013 -23330=4,3,22,22,72 0=72 1=1 5=1 6=5184\nConvolution Conv_256 1 1 1013 783 -23330=4,3,22,22,12 0=12 1=1 5=1 6=864 9=4\nConvolution Conv_257 1 1 1001_splitncnn_1 784 -23330=4,3,22,22,3 0=3 1=1 5=1 6=216 9=4\nConvolution Conv_258 1 1 1001_splitncnn_0 779 -23330=4,3,22,22,80 0=80 1=1 5=1 6=5760\nConvolution Conv_259 1 1 986 788 -23330=4,3,11,11,12 0=12 1=1 5=1 6=864 9=4\nConvolution Conv_260 1 1 974_splitncnn_1 789 -23330=4,3,11,11,3 0=3 1=1 5=1 6=216 9=4\nConvolution Conv_261 1 1 974_splitncnn_0 782 -23330=4,3,11,11,80 0=80 1=1 5=1 6=5760\nPermute Transpose_264 1 1 779 785 -23330=4,3,80,22,22 0=5\nSoftmax Softmax_265 1 1 785 786 -23330=4,3,80,22,22 0=2 1=1\nPermute Transpose_266 1 1 786 787 -23330=4,3,22,22,80 0=5\nPermute Transpose_269 1 1 782 790 -23330=4,3,80,11,11 0=5\nSoftmax Softmax_270 1 1 790 791 -23330=4,3,80,11,11 0=2 1=1\nPermute Transpose_271 1 1 791 792 -23330=4,3,11,11,80 0=5\nConcat Concat_272 3 1 783 784 787 793 -23330=4,3,22,22,95\nPermute Transpose_273 1 1 793 794 -23330=4,3,95,22,22 0=3\nConcat Concat_274 3 1 788 789 792 795 -23330=4,3,11,11,95\nPermute Transpose_275 1 1 795 796 -23330=4,3,95,11,11 0=3\nNoop output 2 1 794 796 output\n" }, { "path": "benchmark/yolov4-tiny.param", "content": "7767517\n45 53\nInput data 0 1 data -23330=4,3,416,416,3 0=416 1=416 2=3\nConvolution 0_25 1 1 data 0_25_bn_leaky -23330=4,3,208,208,32 0=32 1=3 3=2 4=1 5=1 6=864 9=2 -23310=1,1.000000e-01\nConvolution 1_33 1 1 0_25_bn_leaky 1_33_bn_leaky -23330=4,3,104,104,64 0=64 1=3 3=2 4=1 5=1 6=18432 9=2 -23310=1,1.000000e-01\nConvolution 2_41 1 1 1_33_bn_leaky 2_41_bn_leaky -23330=4,3,104,104,64 0=64 1=3 4=1 5=1 6=36864 9=2 -23310=1,1.000000e-01\nSplit 2_41_bn_leaky_split 1 2 2_41_bn_leaky 2_41_bn_leaky_split_0 2_41_bn_leaky_split_1 -23330=8,3,104,104,64,3,104,104,64\nCrop 3_49 1 1 2_41_bn_leaky_split_0 3_49 -23330=4,3,104,104,32 2=32 3=104 4=104 5=32\nConvolution 4_54 1 1 3_49 4_54_bn_leaky -23330=4,3,104,104,32 0=32 1=3 4=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nSplit 4_54_bn_leaky_split 1 2 4_54_bn_leaky 4_54_bn_leaky_split_0 4_54_bn_leaky_split_1 -23330=8,3,104,104,32,3,104,104,32\nConvolution 5_62 1 1 4_54_bn_leaky_split_0 5_62_bn_leaky -23330=4,3,104,104,32 0=32 1=3 4=1 5=1 6=9216 9=2 -23310=1,1.000000e-01\nConcat 6_70 2 1 5_62_bn_leaky 4_54_bn_leaky_split_1 6_70 -23330=4,3,104,104,64\nConvolution 7_73 1 1 6_70 7_73_bn_leaky -23330=4,3,104,104,64 0=64 1=1 5=1 6=4096 9=2 -23310=1,1.000000e-01\nConcat 8_81 2 1 2_41_bn_leaky_split_1 7_73_bn_leaky 8_81 -23330=4,3,104,104,128\nPooling 9_84 1 1 8_81 9_84 -23330=4,3,52,52,128 1=2 2=2 14=1 15=1 5=1\nConvolution 10_88 1 1 9_84 10_88_bn_leaky -23330=4,3,52,52,128 0=128 1=3 4=1 5=1 6=147456 9=2 -23310=1,1.000000e-01\nSplit 10_88_bn_leaky_split 1 2 10_88_bn_leaky 10_88_bn_leaky_split_0 10_88_bn_leaky_split_1 -23330=8,3,52,52,128,3,52,52,128\nCrop 11_96 1 1 10_88_bn_leaky_split_0 11_96 -23330=4,3,52,52,64 2=64 3=52 4=52 5=64\nConvolution 12_101 1 1 11_96 12_101_bn_leaky -23330=4,3,52,52,64 0=64 1=3 4=1 5=1 6=36864 9=2 -23310=1,1.000000e-01\nSplit 12_101_bn_leaky_split 1 2 12_101_bn_leaky 12_101_bn_leaky_split_0 12_101_bn_leaky_split_1 -23330=8,3,52,52,64,3,52,52,64\nConvolution 13_109 1 1 12_101_bn_leaky_split_0 13_109_bn_leaky -23330=4,3,52,52,64 0=64 1=3 4=1 5=1 6=36864 9=2 -23310=1,1.000000e-01\nConcat 14_117 2 1 13_109_bn_leaky 12_101_bn_leaky_split_1 14_117 -23330=4,3,52,52,128\nConvolution 15_120 1 1 14_117 15_120_bn_leaky -23330=4,3,52,52,128 0=128 1=1 5=1 6=16384 9=2 -23310=1,1.000000e-01\nConcat 16_128 2 1 10_88_bn_leaky_split_1 15_120_bn_leaky 16_128 -23330=4,3,52,52,256\nPooling 17_131 1 1 16_128 17_131 -23330=4,3,26,26,256 1=2 2=2 14=1 15=1 5=1\nConvolution 18_135 1 1 17_131 18_135_bn_leaky -23330=4,3,26,26,256 0=256 1=3 4=1 5=1 6=589824 9=2 -23310=1,1.000000e-01\nSplit 18_135_bn_leaky_split 1 2 18_135_bn_leaky 18_135_bn_leaky_split_0 18_135_bn_leaky_split_1 -23330=8,3,26,26,256,3,26,26,256\nCrop 19_143 1 1 18_135_bn_leaky_split_0 19_143 -23330=4,3,26,26,128 2=128 3=26 4=26 5=128\nConvolution 20_148 1 1 19_143 20_148_bn_leaky -23330=4,3,26,26,128 0=128 1=3 4=1 5=1 6=147456 9=2 -23310=1,1.000000e-01\nSplit 20_148_bn_leaky_split 1 2 20_148_bn_leaky 20_148_bn_leaky_split_0 20_148_bn_leaky_split_1 -23330=8,3,26,26,128,3,26,26,128\nConvolution 21_156 1 1 20_148_bn_leaky_split_0 21_156_bn_leaky -23330=4,3,26,26,128 0=128 1=3 4=1 5=1 6=147456 9=2 -23310=1,1.000000e-01\nConcat 22_164 2 1 21_156_bn_leaky 20_148_bn_leaky_split_1 22_164 -23330=4,3,26,26,256\nConvolution 23_167 1 1 22_164 23_167_bn_leaky -23330=4,3,26,26,256 0=256 1=1 5=1 6=65536 9=2 -23310=1,1.000000e-01\nSplit 23_167_bn_leaky_split 1 2 23_167_bn_leaky 23_167_bn_leaky_split_0 23_167_bn_leaky_split_1 -23330=8,3,26,26,256,3,26,26,256\nConcat 24_175 2 1 18_135_bn_leaky_split_1 23_167_bn_leaky_split_0 24_175 -23330=4,3,26,26,512\nPooling 25_178 1 1 24_175 25_178 -23330=4,3,13,13,512 1=2 2=2 14=1 15=1 5=1\nConvolution 26_182 1 1 25_178 26_182_bn_leaky -23330=4,3,13,13,512 0=512 1=3 4=1 5=1 6=2359296 9=2 -23310=1,1.000000e-01\nConvolution 27_192 1 1 26_182_bn_leaky 27_192_bn_leaky -23330=4,3,13,13,256 0=256 1=1 5=1 6=131072 9=2 -23310=1,1.000000e-01\nSplit 27_192_bn_leaky_split 1 2 27_192_bn_leaky 27_192_bn_leaky_split_0 27_192_bn_leaky_split_1 -23330=8,3,13,13,256,3,13,13,256\nConvolution 28_200 1 1 27_192_bn_leaky_split_0 28_200_bn_leaky -23330=4,3,13,13,512 0=512 1=3 4=1 5=1 6=1179648 9=2 -23310=1,1.000000e-01\nConvolution 29_208 1 1 28_200_bn_leaky 29_208 -23330=4,3,13,13,255 0=255 1=1 5=1 6=130560\nConvolution 32_237 1 1 27_192_bn_leaky_split_1 32_237_bn_leaky -23330=4,3,13,13,128 0=128 1=1 5=1 6=32768 9=2 -23310=1,1.000000e-01\nInterp 33_245 1 1 32_237_bn_leaky 33_245 -23330=4,3,26,26,128 0=1 1=2.000000e+00 2=2.000000e+00\nConcat 34_248 2 1 33_245 23_167_bn_leaky_split_1 34_248 -23330=4,3,26,26,384\nConvolution 35_251 1 1 34_248 35_251_bn_leaky -23330=4,3,26,26,256 0=256 1=3 4=1 5=1 6=884736 9=2 -23310=1,1.000000e-01\nConvolution 36_259 1 1 35_251_bn_leaky 36_259 -23330=4,3,26,26,255 0=255 1=1 5=1 6=65280\nYolov3DetectionOutput detection_out 2 1 29_208 36_259 output -23330=4,2,6,1637,1 0=80 1=3 2=3.000001e-01 -23304=12,1.000000e+01,1.400000e+01,2.300000e+01,2.700000e+01,3.700000e+01,5.800000e+01,8.100000e+01,8.200000e+01,1.350000e+02,1.690000e+02,3.440000e+02,3.190000e+02 -23305=6,1077936128,1082130432,1084227584,1065353216,1073741824,1077936128 -23306=2,3.360000e+01,1.680000e+01\n" }, { "path": "build-android.cmd", "content": ":: Set android ndk root\n@ECHO OFF\n@SETLOCAL\n@SET ANDROID_NDK=\n\n:: Set ninja.exe\n:: @SET NINJA_EXE=\n\n:: android armv7\nmkdir build-android-armv7-vulkan\npushd build-android-armv7-vulkan\ncmake -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=%ANDROID_NDK%/build/cmake/android.toolchain.cmake -DCMAKE_MAKE_PROGRAM=\"%ANDROID_NDK%/prebuilt/windows-x86_64/bin/make.exe\" -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-19 -DNCNN_VULKAN=ON ..\ncmake --build . --parallel %NUMBER_OF_PROCESSORS%\ncmake --build . --target install\npopd\n\n:: android aarch64\nmkdir build-android-aarch64-vulkan\npushd build-android-aarch64-vulkan\ncmake -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=%ANDROID_NDK%/build/cmake/android.toolchain.cmake -DCMAKE_MAKE_PROGRAM=\"%ANDROID_NDK%/prebuilt/windows-x86_64/bin/make.exe\" -DANDROID_ABI=\"arm64-v8a\" -DANDROID_PLATFORM=android-21 -DNCNN_VULKAN=ON ..\ncmake --build . --parallel %NUMBER_OF_PROCESSORS%\ncmake --build . --target install\npopd\n\n:: android x86\nmkdir build-android-x86\npushd build-android-x86\ncmake -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=%ANDROID_NDK%/build/cmake/android.toolchain.cmake -DCMAKE_MAKE_PROGRAM=\"%ANDROID_NDK%/prebuilt/windows-x86_64/bin/make.exe\" -DANDROID_ABI=\"x86\" -DANDROID_PLATFORM=android-19 -DNCNN_VULKAN=ON ..\ncmake --build . --parallel %NUMBER_OF_PROCESSORS%\ncmake --build . --target install\npopd\n\n:: android x86_64\nmkdir build-android-x86_64\npushd build-android-x86_64\ncmake -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=%ANDROID_NDK%/build/cmake/android.toolchain.cmake -DCMAKE_MAKE_PROGRAM=\"%ANDROID_NDK%/prebuilt/windows-x86_64/bin/make.exe\" -DANDROID_ABI=\"x86_64\" -DANDROID_PLATFORM=android-21 -DNCNN_VULKAN=ON ..\ncmake --build . --parallel %NUMBER_OF_PROCESSORS%\ncmake --build . --target install\npopd\n\n:: android riscv64\nmkdir build-android-riscv64\npushd build-android-riscv64\ncmake -G \"Unix Makefiles\" -DCMAKE_TOOLCHAIN_FILE=%ANDROID_NDK%/build/cmake/android.toolchain.cmake -DCMAKE_MAKE_PROGRAM=\"%ANDROID_NDK%/prebuilt/windows-x86_64/bin/make.exe\" -DANDROID_ABI=\"riscv64\" -DANDROID_PLATFORM=android-35 -DNCNN_VULKAN=ON ..\ncmake --build . --parallel %NUMBER_OF_PROCESSORS%\ncmake --build . --target install\npopd\n\n@ENDLOCAL\n" }, { "path": "build.sh", "content": "#!/usr/bin/env bash\n\n##### android armv7 without neon\nmkdir -p build-android-armv7-without-neon\npushd build-android-armv7-without-neon\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=OFF -DANDROID_PLATFORM=android-19 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### android armv7\nmkdir -p build-android-armv7\npushd build-android-armv7\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON -DANDROID_PLATFORM=android-19 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### android aarch64\nmkdir -p build-android-aarch64\npushd build-android-aarch64\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"arm64-v8a\" -DANDROID_PLATFORM=android-21 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### android x86\nmkdir -p build-android-x86\npushd build-android-x86\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"x86\" -DANDROID_PLATFORM=android-19 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### android x86_64\nmkdir -p build-android-x86_64\npushd build-android-x86_64\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"x86_64\" -DANDROID_PLATFORM=android-21 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### android riscv64\nmkdir -p build-android-riscv64\npushd build-android-riscv64\ncmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=\"riscv64\" -DANDROID_PLATFORM=android-35 -DNCNN_VULKAN=ON ..\nmake -j4\nmake install\npopd\n\n##### linux of hisiv300 (forgot the chip name) toolchain with neon and openmp\nmkdir -p build-hisiv300-linux\npushd build-hisiv300-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv300.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of hisiv500 (Hi3516CV200 and Hi3519V101) toolchain with neon and openmp\nmkdir -p build-hisiv500-linux\npushd build-hisiv500-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv500.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of hisiv600 (Hi3559V100) toolchain with neon and no openmp (due to only one cpu, close openmp)\nmkdir -p build-hisiv600-linux\npushd build-hisiv600-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv600.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of himix100 (Hi3559a) toolchain with neon and openmp\nmkdir -p build-himix100-linux\npushd build-himix100-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix100.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of arm-linux-gnueabi toolchain\nmkdir -p build-arm-linux-gnueabi\npushd build-arm-linux-gnueabi\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of arm-linux-gnueabihf toolchain\nmkdir -p build-arm-linux-gnueabihf\npushd build-arm-linux-gnueabihf\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux of v831 toolchain with neon and openmp\nmkdir -p build-v831-linux\npushd build-v831-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/v831.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux for aarch64-linux-gnu toolchain\nmkdir -p build-aarch64-linux-gnu\npushd build-aarch64-linux-gnu\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### linux host system with gcc/g++\nmkdir -p build-host-gcc-linux\npushd build-host-gcc-linux\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/host.gcc.toolchain.cmake ..\nmake -j4\nmake install\npopd\n\n##### MacOS\nmkdir -p build-mac\npushd build-mac\ncmake -DNCNN_OPENMP=OFF \\\n -DNCNN_BENCHMARK=ON \\\n ..\nmake -j8\nmake install\npopd\n" }, { "path": "cmake/ncnnConfig.cmake.in", "content": "set(NCNN_VERSION @NCNN_VERSION@)\nset(NCNN_OPENMP @NCNN_OPENMP@)\nset(NCNN_THREADS @NCNN_THREADS@)\nset(NCNN_VULKAN @NCNN_VULKAN@)\nset(NCNN_SHARED_LIB @NCNN_SHARED_LIB@)\nset(NCNN_SYSTEM_GLSLANG @NCNN_SYSTEM_GLSLANG@)\nset(NCNN_SIMPLEVK @NCNN_SIMPLEVK@)\n\nif(NCNN_OPENMP)\n find_package(OpenMP)\nendif()\n\nif(NCNN_THREADS)\n set(CMAKE_THREAD_PREFER_PTHREAD TRUE)\n set(THREADS_PREFER_PTHREAD_FLAG TRUE)\n find_package(Threads REQUIRED)\nendif()\n\nif(NCNN_VULKAN)\n if(NOT NCNN_SIMPLEVK)\n find_package(Vulkan REQUIRED)\n endif()\n\n if(NOT NCNN_SHARED_LIB)\n if(NCNN_SYSTEM_GLSLANG)\n find_package(SPIRV-Tools QUIET)\n find_package(SPIRV-Tools-opt QUIET)\n find_package(glslang QUIET)\n if(NOT glslang_FOUND)\n set(GLSLANG_TARGET_DIR \"@GLSLANG_TARGET_DIR@\")\n include(${GLSLANG_TARGET_DIR}/OSDependentTargets.cmake)\n include(${GLSLANG_TARGET_DIR}/OGLCompilerTargets.cmake)\n if(EXISTS \"${GLSLANG_TARGET_DIR}/HLSLTargets.cmake\")\n # hlsl support can be optional\n include(\"${GLSLANG_TARGET_DIR}/HLSLTargets.cmake\")\n endif()\n include(${GLSLANG_TARGET_DIR}/glslangTargets.cmake)\n include(${GLSLANG_TARGET_DIR}/SPIRVTargets.cmake)\n endif()\n else()\n set(glslang_DIR \"${CMAKE_CURRENT_LIST_DIR}/../../../@CMAKE_INSTALL_LIBDIR@/cmake/glslang\")\n find_package(glslang QUIET)\n endif()\n endif()\nendif()\n\ninclude(${CMAKE_CURRENT_LIST_DIR}/ncnn.cmake)\n\nif(TARGET ncnn)\n set(ncnn_FOUND TRUE)\n if(NOT ncnn_FIND_QUIETLY)\n message(STATUS \"Found ncnn: ${NCNN_VERSION}\")\n endif()\nendif()\n" }, { "path": "cmake/ncnn_add_layer.cmake", "content": "\nmacro(ncnn_add_arch_opt_layer class NCNN_TARGET_ARCH_OPT NCNN_TARGET_ARCH_OPT_CFLAGS)\n set(NCNN_${NCNN_TARGET_ARCH}_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}.h)\n set(NCNN_${NCNN_TARGET_ARCH}_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}.cpp)\n\n if(WITH_LAYER_${name} AND EXISTS ${NCNN_${NCNN_TARGET_ARCH}_HEADER} AND EXISTS ${NCNN_${NCNN_TARGET_ARCH}_SOURCE})\n\n set(NCNN_${NCNN_TARGET_ARCH_OPT}_HEADER ${CMAKE_CURRENT_BINARY_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.h)\n set(NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE ${CMAKE_CURRENT_BINARY_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.cpp)\n\n add_custom_command(\n OUTPUT ${NCNN_${NCNN_TARGET_ARCH_OPT}_HEADER}\n COMMAND ${CMAKE_COMMAND} -DSRC=${NCNN_${NCNN_TARGET_ARCH}_HEADER} -DDST=${NCNN_${NCNN_TARGET_ARCH_OPT}_HEADER} -DCLASS=${class} -P \"${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_${NCNN_TARGET_ARCH_OPT}_source.cmake\"\n DEPENDS ${NCNN_${NCNN_TARGET_ARCH}_HEADER}\n COMMENT \"Generating source ${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.h\"\n VERBATIM\n )\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT}_HEADER} PROPERTIES GENERATED TRUE)\n\n add_custom_command(\n OUTPUT ${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE}\n COMMAND ${CMAKE_COMMAND} -DSRC=${NCNN_${NCNN_TARGET_ARCH}_SOURCE} -DDST=${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE} -DCLASS=${class} -P \"${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_${NCNN_TARGET_ARCH_OPT}_source.cmake\"\n DEPENDS ${NCNN_${NCNN_TARGET_ARCH}_SOURCE}\n COMMENT \"Generating source ${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.cpp\"\n VERBATIM\n )\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE} PROPERTIES GENERATED TRUE)\n\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE} PROPERTIES COMPILE_FLAGS ${NCNN_TARGET_ARCH_OPT_CFLAGS})\n\n list(APPEND ncnn_SRCS ${NCNN_${NCNN_TARGET_ARCH_OPT}_HEADER} ${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE})\n\n # generate layer_declaration and layer_registry file\n set(layer_declaration \"${layer_declaration}#include \\\"layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.h\\\"\\n\")\n set(layer_declaration \"${layer_declaration}namespace ncnn { DEFINE_LAYER_CREATOR(${class}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}) }\\n\")\n\n set(layer_registry_${NCNN_TARGET_ARCH_OPT} \"${layer_registry_${NCNN_TARGET_ARCH_OPT}}#if NCNN_STRING\\n{\\\"${class}\\\", ${class}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}_layer_creator},\\n#else\\n{${class}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}_layer_creator},\\n#endif\\n\")\n else()\n # no isa optimized version\n if(WITH_LAYER_${name})\n set(layer_registry_${NCNN_TARGET_ARCH_OPT} \"${layer_registry_${NCNN_TARGET_ARCH_OPT}}#if NCNN_STRING\\n{\\\"${class}\\\", ${class}_layer_creator},\\n#else\\n{${class}_layer_creator},\\n#endif\\n\")\n else()\n set(layer_registry_${NCNN_TARGET_ARCH_OPT} \"${layer_registry_${NCNN_TARGET_ARCH_OPT}}#if NCNN_STRING\\n{\\\"${class}\\\", 0},\\n#else\\n{0},\\n#endif\\n\")\n endif()\n endif()\nendmacro()\n\nmacro(ncnn_add_arch_opt_source class NCNN_TARGET_ARCH_OPT NCNN_TARGET_ARCH_OPT_CFLAGS)\n set(NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT}.cpp)\n\n if(WITH_LAYER_${name} AND EXISTS ${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE})\n if(NCNN_RUNTIME_CPU)\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE} PROPERTIES COMPILE_FLAGS ${NCNN_TARGET_ARCH_OPT_CFLAGS})\n endif()\n list(APPEND ncnn_SRCS ${NCNN_${NCNN_TARGET_ARCH_OPT}_SOURCE})\n endif()\nendmacro()\n\nmacro(ncnn_add_arch_opt_layer_source class NCNN_TARGET_ARCH_OPT_BASE NCNN_TARGET_ARCH_OPT NCNN_TARGET_ARCH_OPT_CFLAGS)\n set(NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT_BASE}.cpp)\n\n if(WITH_LAYER_${name} AND EXISTS ${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_SOURCE})\n\n set(NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE ${CMAKE_CURRENT_BINARY_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}.cpp)\n\n add_custom_command(\n OUTPUT ${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE}\n COMMAND ${CMAKE_COMMAND} -DSRC=${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_SOURCE} -DDST=${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE} -DCLASS=${class} -P \"${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_${NCNN_TARGET_ARCH_OPT}_source.cmake\"\n DEPENDS ${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_SOURCE}\n COMMENT \"Generating source ${name}_${NCNN_TARGET_ARCH}_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}.cpp\"\n VERBATIM\n )\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE} PROPERTIES GENERATED TRUE)\n\n if(NCNN_RUNTIME_CPU)\n set_source_files_properties(${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE} PROPERTIES COMPILE_FLAGS ${NCNN_TARGET_ARCH_OPT_CFLAGS})\n endif()\n list(APPEND ncnn_SRCS ${NCNN_${NCNN_TARGET_ARCH_OPT_BASE}_${NCNN_TARGET_ARCH_OPT}_SOURCE})\n endif()\nendmacro()\n\nmacro(ncnn_add_layer class)\n string(TOLOWER ${class} name)\n\n # WITH_LAYER_xxx option\n if(${ARGC} EQUAL 2)\n option(WITH_LAYER_${name} \"build with layer ${name}\" ${ARGV1})\n else()\n option(WITH_LAYER_${name} \"build with layer ${name}\" ON)\n endif()\n\n if(NCNN_CMAKE_VERBOSE)\n message(STATUS \"WITH_LAYER_${name} = ${WITH_LAYER_${name}}\")\n endif()\n\n if(WITH_LAYER_${name})\n list(APPEND ncnn_SRCS ${CMAKE_CURRENT_SOURCE_DIR}/layer/${name}.cpp)\n\n # look for arch specific implementation and append source\n # optimized implementation for armv7, aarch64 or x86\n set(LAYER_ARCH_SRC ${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}.cpp)\n if(EXISTS ${LAYER_ARCH_SRC})\n set(WITH_LAYER_${name}_${NCNN_TARGET_ARCH} 1)\n list(APPEND ncnn_SRCS ${LAYER_ARCH_SRC})\n endif()\n\n set(LAYER_VULKAN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/layer/vulkan/${name}_vulkan.cpp)\n if(NCNN_VULKAN AND EXISTS ${LAYER_VULKAN_SRC})\n set(WITH_LAYER_${name}_vulkan 1)\n list(APPEND ncnn_SRCS ${LAYER_VULKAN_SRC})\n endif()\n endif()\n\n # generate layer_declaration and layer_registry file\n if(WITH_LAYER_${name})\n set(layer_declaration \"${layer_declaration}#include \\\"layer/${name}.h\\\"\\n\")\n set(layer_declaration \"${layer_declaration}namespace ncnn { DEFINE_LAYER_CREATOR(${class}) }\\n\")\n\n source_group (\"sources\\\\\\\\layers\" FILES \"${CMAKE_CURRENT_SOURCE_DIR}/layer/${name}.cpp\")\n endif()\n\n if(WITH_LAYER_${name}_${NCNN_TARGET_ARCH})\n set(layer_declaration \"${layer_declaration}#include \\\"layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}.h\\\"\\n\")\n set(layer_declaration \"${layer_declaration}namespace ncnn { DEFINE_LAYER_CREATOR(${class}_${NCNN_TARGET_ARCH}) }\\n\")\n\n source_group (\"sources\\\\\\\\layers\\\\\\\\${NCNN_TARGET_ARCH}\" FILES \"${CMAKE_CURRENT_SOURCE_DIR}/layer/${NCNN_TARGET_ARCH}/${name}_${NCNN_TARGET_ARCH}.cpp\")\n endif()\n\n if(WITH_LAYER_${name}_vulkan)\n set(layer_declaration \"${layer_declaration}#include \\\"layer/vulkan/${name}_vulkan.h\\\"\\n\")\n set(layer_declaration \"${layer_declaration}namespace ncnn { DEFINE_LAYER_CREATOR(${class}_vulkan) }\\n\")\n\n file(GLOB NCNN_SHADER_SRCS \"layer/vulkan/shader/${name}.comp\")\n file(GLOB NCNN_SHADER_SUBSRCS \"layer/vulkan/shader/${name}_*.comp\")\n list(APPEND NCNN_SHADER_SRCS ${NCNN_SHADER_SUBSRCS})\n foreach(NCNN_SHADER_SRC ${NCNN_SHADER_SRCS})\n ncnn_add_shader(${NCNN_SHADER_SRC})\n endforeach()\n\n source_group (\"sources\\\\\\\\layers\\\\\\\\vulkan\" FILES \"${CMAKE_CURRENT_SOURCE_DIR}/layer/vulkan/${name}_vulkan.cpp\")\n endif()\n\n if(WITH_LAYER_${name})\n set(layer_registry \"${layer_registry}#if NCNN_STRING\\n{\\\"${class}\\\", ${class}_layer_creator},\\n#else\\n{${class}_layer_creator},\\n#endif\\n\")\n else()\n set(layer_registry \"${layer_registry}#if NCNN_STRING\\n{\\\"${class}\\\", 0},\\n#else\\n{0},\\n#endif\\n\")\n endif()\n\n if(WITH_LAYER_${name}_${NCNN_TARGET_ARCH})\n set(layer_registry_arch \"${layer_registry_arch}#if NCNN_STRING\\n{\\\"${class}\\\", ${class}_${NCNN_TARGET_ARCH}_layer_creator},\\n#else\\n{${class}_${NCNN_TARGET_ARCH}_layer_creator},\\n#endif\\n\")\n else()\n set(layer_registry_arch \"${layer_registry_arch}#if NCNN_STRING\\n{\\\"${class}\\\", 0},\\n#else\\n{0},\\n#endif\\n\")\n endif()\n\n if(WITH_LAYER_${name}_vulkan)\n set(layer_registry_vulkan \"${layer_registry_vulkan}#if NCNN_STRING\\n{\\\"${class}\\\", ${class}_vulkan_layer_creator},\\n#else\\n{${class}_vulkan_layer_creator},\\n#endif\\n\")\n else()\n set(layer_registry_vulkan \"${layer_registry_vulkan}#if NCNN_STRING\\n{\\\"${class}\\\", 0},\\n#else\\n{0},\\n#endif\\n\")\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"x86\")\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\")\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512)\n ncnn_add_arch_opt_layer(${class} avx512 \"/arch:AVX512 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_FMA)\n ncnn_add_arch_opt_layer(${class} fma \"/arch:AVX /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX)\n ncnn_add_arch_opt_layer(${class} avx \"/arch:AVX /D__SSSE3__ /D__SSE4_1__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512VNNI)\n ncnn_add_arch_opt_source(${class} avx512vnni \"/arch:AVX512 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512VNNI__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512BF16)\n ncnn_add_arch_opt_source(${class} avx512bf16 \"/arch:AVX512 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512BF16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512FP16)\n ncnn_add_arch_opt_source(${class} avx512fp16 \"/arch:AVX512 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512FP16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNI)\n ncnn_add_arch_opt_source(${class} avxvnni \"/arch:AVX2 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT8)\n ncnn_add_arch_opt_source(${class} avxvnniint8 \"/arch:AVX2 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__ /D__AVXVNNIINT8__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT16)\n ncnn_add_arch_opt_source(${class} avxvnniint16 \"/arch:AVX2 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__ /D__AVXVNNIINT16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXNECONVERT)\n ncnn_add_arch_opt_source(${class} avxneconvert \"/arch:AVX2 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXNECONVERT__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX2)\n ncnn_add_arch_opt_source(${class} avx2 \"/arch:AVX2 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_XOP)\n ncnn_add_arch_opt_source(${class} xop \"/arch:AVX /D__SSSE3__ /D__SSE4_1__ /D__XOP__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_F16C)\n ncnn_add_arch_opt_source(${class} f16c \"/arch:AVX /D__SSSE3__ /D__SSE4_1__ /D__F16C__\")\n endif()\n elseif(CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512)\n ncnn_add_arch_opt_layer(${class} avx512 \"/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_FMA)\n ncnn_add_arch_opt_layer(${class} fma \"/arch:AVX -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX)\n ncnn_add_arch_opt_layer(${class} avx \"/arch:AVX /D__SSSE3__ /D__SSE4_1__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512VNNI)\n ncnn_add_arch_opt_source(${class} avx512vnni \"/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512vnni /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512VNNI__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512BF16)\n ncnn_add_arch_opt_source(${class} avx512bf16 \"/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512bf16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512BF16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512FP16)\n ncnn_add_arch_opt_source(${class} avx512fp16 \"/arch:AVX512 -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512fp16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVX512FP16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNI)\n ncnn_add_arch_opt_source(${class} avxvnni \"/arch:AVX2 -mfma -mf16c -mavxvnni /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT8)\n ncnn_add_arch_opt_source(${class} avxvnniint8 \"/arch:AVX2 -mfma -mf16c -mavxvnni -mavxvnniint8 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__ /D__AVXVNNIINT8__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT16)\n ncnn_add_arch_opt_source(${class} avxvnniint16 \"/arch:AVX2 -mfma -mf16c -mavxvnni -mavxvnniint16 /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXVNNI__ /D__AVXVNNIINT16__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXNECONVERT)\n ncnn_add_arch_opt_source(${class} avxneconvert \"/arch:AVX2 -mfma -mf16c -mavxneconvert /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__ /D__AVXNECONVERT__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX2)\n ncnn_add_arch_opt_source(${class} avx2 \"/arch:AVX2 -mfma -mf16c /D__SSSE3__ /D__SSE4_1__ /D__FMA__ /D__F16C__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_XOP)\n ncnn_add_arch_opt_source(${class} xop \"/arch:AVX -mxop /D__SSSE3__ /D__SSE4_1__ /D__XOP__\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_F16C)\n ncnn_add_arch_opt_source(${class} f16c \"/arch:AVX -mf16c /D__SSSE3__ /D__SSE4_1__ /D__F16C__\")\n endif()\n else()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512)\n ncnn_add_arch_opt_layer(${class} avx512 \"-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_FMA)\n ncnn_add_arch_opt_layer(${class} fma \"-mavx -mfma -mf16c\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX)\n ncnn_add_arch_opt_layer(${class} avx \"-mavx\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512VNNI)\n ncnn_add_arch_opt_source(${class} avx512vnni \"-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512vnni\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512BF16)\n ncnn_add_arch_opt_source(${class} avx512bf16 \"-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512bf16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX512FP16)\n ncnn_add_arch_opt_source(${class} avx512fp16 \"-mavx512f -mavx512cd -mavx512bw -mavx512dq -mavx512vl -mfma -mf16c -mavx512fp16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNI)\n ncnn_add_arch_opt_source(${class} avxvnni \"-mavx2 -mfma -mf16c -mavxvnni\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT8)\n ncnn_add_arch_opt_source(${class} avxvnniint8 \"-mavx2 -mfma -mf16c -mavxvnni -mavxvnniint8\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNIINT16)\n ncnn_add_arch_opt_source(${class} avxvnniint16 \"-mavx2 -mfma -mf16c -mavxvnni -mavxvnniint16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVXNECONVERT)\n ncnn_add_arch_opt_source(${class} avxneconvert \"-mavx2 -mfma -mf16c -mavxneconvert\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_AVX2)\n ncnn_add_arch_opt_source(${class} avx2 \"-mavx2 -mfma -mf16c\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_XOP)\n ncnn_add_arch_opt_source(${class} xop \"-mavx -mxop\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_F16C)\n ncnn_add_arch_opt_source(${class} f16c \"-mavx -mf16c\")\n endif()\n endif()\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"arm\" AND (CMAKE_SIZEOF_VOID_P EQUAL 4 AND NOT NCNN_TARGET_ILP32))\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\" OR (CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\"))\n if(NCNN_VFPV4)\n ncnn_add_arch_opt_source(${class} vfpv4 \"/arch:VFPv4 /D__ARM_FP=0x0E\")\n endif()\n else()\n if(NCNN_VFPV4)\n if(NCNN_COMPILER_SUPPORT_ARM_VFPV4)\n ncnn_add_arch_opt_source(${class} vfpv4 \"-mfpu=neon-vfpv4\")\n elseif(NCNN_COMPILER_SUPPORT_ARM_VFPV4_FP16)\n ncnn_add_arch_opt_source(${class} vfpv4 \"-mfpu=neon-vfpv4 -mfp16-format=ieee\")\n endif()\n endif()\n endif()\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"arm\" AND (CMAKE_SIZEOF_VOID_P EQUAL 8 OR NCNN_TARGET_ILP32))\n if(CMAKE_CXX_COMPILER_ID MATCHES \"MSVC\")\n if(NCNN_VFPV4)\n ncnn_add_arch_opt_source(${class} vfpv4 \" \")\n endif()\n if(NCNN_ARM82)\n ncnn_add_arch_opt_source(${class} asimdhp \"/arch:armv8.2 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82DOT)\n ncnn_add_arch_opt_source(${class} asimddp \"/arch:armv8.2 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82FP16FML)\n ncnn_add_arch_opt_source(${class} asimdfhm \"/arch:armv8.2 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_FP16_FML\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84BF16)\n ncnn_add_arch_opt_source(${class} bf16 \"/arch:armv8.4 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_BF16_VECTOR_ARITHMETIC\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84I8MM)\n ncnn_add_arch_opt_source(${class} i8mm \"/arch:armv8.4 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_MATMUL_INT8\")\n endif()\n # TODO add support for sve family\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE2)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEBF16)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEI8MM)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEF32MM)\n endif()\n elseif(CMAKE_CXX_COMPILER_ID MATCHES \"Clang\" AND CMAKE_CXX_SIMULATE_ID MATCHES \"MSVC\" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES \"MSVC\")\n if(NCNN_VFPV4)\n ncnn_add_arch_opt_source(${class} vfpv4 \" \")\n endif()\n if(NCNN_ARM82)\n ncnn_add_arch_opt_source(${class} asimdhp \"/arch:armv8.2 -march=armv8.2-a+fp16 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82DOT)\n ncnn_add_arch_opt_source(${class} asimddp \"/arch:armv8.2 -march=armv8.2-a+fp16+dotprod /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82FP16FML)\n ncnn_add_arch_opt_source(${class} asimdfhm \"/arch:armv8.2 -march=armv8.2-a+fp16+fp16fml /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_FP16_FML\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84BF16)\n ncnn_add_arch_opt_source(${class} bf16 \"/arch:armv8.4 -march=armv8.4-a+fp16+dotprod+bf16 /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_BF16_VECTOR_ARITHMETIC\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84I8MM)\n ncnn_add_arch_opt_source(${class} i8mm \"/arch:armv8.4 -march=armv8.4-a+fp16+dotprod+i8mm /D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC /D__ARM_FEATURE_DOTPROD /D__ARM_FEATURE_FP16_FML /D__ARM_FEATURE_MATMUL_INT8\")\n endif()\n # TODO add support for sve family\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE2)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEBF16)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEI8MM)\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEF32MM)\n endif()\n else()\n if(NCNN_VFPV4)\n ncnn_add_arch_opt_source(${class} vfpv4 \" \")\n endif()\n if(NCNN_ARM82)\n ncnn_add_arch_opt_source(${class} asimdhp \"-march=armv8.2-a+fp16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82DOT)\n ncnn_add_arch_opt_source(${class} asimddp \"-march=armv8.2-a+fp16+dotprod\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM82FP16FML)\n # clang 9.0.9 shipped with android ndk-r21 is missing __ARM_FEATURE_FP16_FML macro for asimdfhm target\n ncnn_add_arch_opt_source(${class} asimdfhm \"-march=armv8.2-a+fp16+fp16fml -D__ARM_FEATURE_FP16_FML\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84BF16)\n ncnn_add_arch_opt_source(${class} bf16 \"-march=armv8.4-a+fp16+dotprod+bf16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM84I8MM)\n ncnn_add_arch_opt_source(${class} i8mm \"-march=armv8.4-a+fp16+dotprod+i8mm\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE)\n ncnn_add_arch_opt_source(${class} sve \"-march=armv8.6-a+fp16+dotprod+sve\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVE2)\n ncnn_add_arch_opt_source(${class} sve2 \"-march=armv8.6-a+fp16+dotprod+sve2\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEBF16)\n ncnn_add_arch_opt_source(${class} svebf16 \"-march=armv8.6-a+fp16+dotprod+sve+bf16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEI8MM)\n ncnn_add_arch_opt_source(${class} svei8mm \"-march=armv8.6-a+fp16+dotprod+sve+i8mm\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ARM86SVEF32MM)\n ncnn_add_arch_opt_source(${class} svef32mm \"-march=armv8.6-a+fp16+dotprod+sve+f32mm\")\n endif()\n endif()\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"mips\")\n if(NCNN_RUNTIME_CPU AND NCNN_MSA)\n ncnn_add_arch_opt_layer(${class} msa \"-mmsa\")\n endif()\n if(NCNN_MMI)\n ncnn_add_arch_opt_source(${class} mmi \"-mloongson-mmi\")\n endif()\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"loongarch\")\n if(NCNN_RUNTIME_CPU AND NCNN_LASX)\n ncnn_add_arch_opt_layer(${class} lasx \"-mlasx -mlsx\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_LSX)\n ncnn_add_arch_opt_layer(${class} lsx \"-mlsx\")\n endif()\n endif()\n\n if(NCNN_TARGET_ARCH STREQUAL \"riscv\")\n if(CMAKE_SIZEOF_VOID_P EQUAL 8)\n if(NCNN_RUNTIME_CPU AND NCNN_RVV)\n ncnn_add_arch_opt_layer(${class} rvv \"-march=rv64gcv\")\n endif()\n if(NCNN_ZFH)\n if(NOT NCNN_RUNTIME_CPU AND NCNN_ZVFH)\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv64gcv_zfh_zvfh -D__fp16=_Float16\")\n elseif(NOT NCNN_RUNTIME_CPU AND NCNN_XTHEADVECTOR)\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv64gc_zfh_xtheadvector -D__riscv_zvfh=1 -D__fp16=_Float16\")\n else()\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv64gc_zfh -D__fp16=_Float16\")\n endif()\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_XTHEADVECTOR)\n # linker complains the conflict of v and xtheadvector, so disable generating any riscv attributes\n ncnn_add_arch_opt_layer(${class} xtheadvector \"-march=rv64gc_xtheadvector -mno-riscv-attribute -Wa,-mno-arch-attr\")\n ncnn_add_arch_opt_layer_source(${class} zfh xtheadvector \"-march=rv64gc_zfh_xtheadvector -mno-riscv-attribute -Wa,-mno-arch-attr -D__fp16=_Float16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ZVFH)\n ncnn_add_arch_opt_layer_source(${class} zfh rvv \"-march=rv64gcv_zfh_zvfh -D__fp16=_Float16\")\n endif()\n elseif(CMAKE_SIZEOF_VOID_P EQUAL 4)\n if(NCNN_RUNTIME_CPU AND NCNN_RVV)\n ncnn_add_arch_opt_layer(${class} rvv \"-march=rv32gcv\")\n endif()\n if(NCNN_ZFH)\n if(NOT NCNN_RUNTIME_CPU AND NCNN_ZVFH)\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv32gcv_zfh_zvfh -D__fp16=_Float16\")\n elseif(NOT NCNN_RUNTIME_CPU AND NCNN_XTHEADVECTOR)\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv32gc_zfh_xtheadvector -D__riscv_zvfh=1 -D__fp16=_Float16\")\n else()\n ncnn_add_arch_opt_source(${class} zfh \"-march=rv32gc_zfh -D__fp16=_Float16\")\n endif()\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_XTHEADVECTOR)\n # linker complains the conflict of v and xtheadvector, so disable generating any riscv attributes\n ncnn_add_arch_opt_layer(${class} xtheadvector \"-march=rv32gc_xtheadvector -mno-riscv-attribute -Wa,-mno-arch-attr\")\n ncnn_add_arch_opt_layer_source(${class} zfh xtheadvector \"-march=rv32gc_zfh_xtheadvector -mno-riscv-attribute -Wa,-mno-arch-attr -D__fp16=_Float16\")\n endif()\n if(NCNN_RUNTIME_CPU AND NCNN_ZVFH)\n ncnn_add_arch_opt_layer_source(${class} zfh rvv \"-march=rv32gcv_zfh_zvfh -D__fp16=_Float16\")\n endif()\n endif()\n endif()\n\n # generate layer_type_enum file\n set(layer_type_enum \"${layer_type_enum}${class} = ${__LAYER_TYPE_ENUM_INDEX},\\n\")\n math(EXPR __LAYER_TYPE_ENUM_INDEX \"${__LAYER_TYPE_ENUM_INDEX}+1\")\nendmacro()\n" }, { "path": "cmake/ncnn_add_param.cmake", "content": "\nmacro(ncnn_add_param NCNN_PARAM_SRC)\n # Get the file name with extension\n get_filename_component(NCNN_PARAM_SRC_NAME_WE ${NCNN_PARAM_SRC} NAME)\n # Manually remove \".param\" since NAME_WE treats \".1.param\" as a multi-extension\n string(REPLACE \".param\" \"\" NCNN_PARAM_SRC_NAME_WE \"${NCNN_PARAM_SRC_NAME_WE}\")\n # Replace characters invalid in C identifiers ('.' and '-') with underscores\n string(REPLACE \".param\" \"\" NCNN_PARAM_SRC_NAME_WE \"${NCNN_PARAM_SRC_NAME_WE}\")\n # Replace characters invalid in C identifiers ('.' and '-') with underscores\n string(REPLACE \".\" \"_\" NCNN_PARAM_SRC_NAME_WE \"${NCNN_PARAM_SRC_NAME_WE}\")\n string(REPLACE \"-\" \"_\" NCNN_PARAM_SRC_NAME_WE \"${NCNN_PARAM_SRC_NAME_WE}\")\n # Check if the result is empty\n if (NOT NCNN_PARAM_SRC_NAME_WE)\n message(FATAL_ERROR \"Failed to extract valid filename from '${NCNN_PARAM_SRC}'\")\n endif()\n # Check if the extracted filename is a valid C identifier\n string(REGEX MATCH \"^[A-Za-z_][A-Za-z0-9_]*$\" is_valid \"${NCNN_PARAM_SRC_NAME_WE}\")\n if (NOT is_valid)\n message(FATAL_ERROR \"Extracted filename '${NCNN_PARAM_SRC_NAME_WE}' is not a valid C identifier\")\n endif()\n\n set(NCNN_PARAM_HEADER ${CMAKE_CURRENT_BINARY_DIR}/param/${NCNN_PARAM_SRC_NAME_WE}.hex.h)\n\n add_custom_command(\n OUTPUT ${NCNN_PARAM_HEADER}\n COMMAND ${CMAKE_COMMAND} -DPARAM_SRC=${NCNN_PARAM_SRC} -DPARAM_SRC_NAME_WE=${NCNN_PARAM_SRC_NAME_WE} -DPARAM_HEADER=${NCNN_PARAM_HEADER} -P \"${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_param_header.cmake\"\n DEPENDS ${NCNN_PARAM_SRC}\n COMMENT \"Preprocessing param source ${NCNN_PARAM_SRC_NAME_WE}.param\"\n VERBATIM\n )\n set_source_files_properties(${NCNN_PARAM_HEADER} PROPERTIES GENERATED TRUE)\n\n get_filename_component(NCNN_PARAM_HEADER_NAME ${NCNN_PARAM_HEADER} NAME)\n string(APPEND param_header_data \"#include \\\"param/${NCNN_PARAM_HEADER_NAME}\\\"\\n\")\n\n list(APPEND NCNN_PARAM_HEX_FILES ${NCNN_PARAM_HEADER})\nendmacro()\n" }, { "path": "cmake/ncnn_add_shader.cmake", "content": "\nmacro(ncnn_add_shader NCNN_SHADER_SRC)\n get_filename_component(NCNN_SHADER_SRC_NAME_WE ${NCNN_SHADER_SRC} NAME_WE)\n set(NCNN_SHADER_COMP_HEADER ${CMAKE_CURRENT_BINARY_DIR}/layer/vulkan/shader/${NCNN_SHADER_SRC_NAME_WE}.comp.hex.h)\n\n add_custom_command(\n OUTPUT ${NCNN_SHADER_COMP_HEADER}\n COMMAND ${CMAKE_COMMAND} -DSHADER_SRC=${NCNN_SHADER_SRC} -DSHADER_COMP_HEADER=${NCNN_SHADER_COMP_HEADER} -P \"${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_shader_comp_header.cmake\"\n DEPENDS ${NCNN_SHADER_SRC}\n COMMENT \"Preprocessing shader source ${NCNN_SHADER_SRC_NAME_WE}.comp\"\n VERBATIM\n )\n set_source_files_properties(${NCNN_SHADER_COMP_HEADER} PROPERTIES GENERATED TRUE)\n\n get_filename_component(NCNN_SHADER_COMP_HEADER_NAME ${NCNN_SHADER_COMP_HEADER} NAME)\n string(APPEND layer_shader_spv_data \"#include \\\"layer/vulkan/shader/${NCNN_SHADER_COMP_HEADER_NAME}\\\"\\n\")\n\n get_filename_component(NCNN_SHADER_SRC_NAME_WE ${NCNN_SHADER_SRC} NAME_WE)\n string(APPEND layer_shader_registry \"{${NCNN_SHADER_SRC_NAME_WE}_comp_data,sizeof(${NCNN_SHADER_SRC_NAME_WE}_comp_data)},\\n\")\n\n list(APPEND NCNN_SHADER_SPV_HEX_FILES ${NCNN_SHADER_COMP_HEADER})\n\n # generate layer_shader_type_enum file\n set(layer_shader_type_enum \"${layer_shader_type_enum}${NCNN_SHADER_SRC_NAME_WE} = ${__LAYER_SHADER_TYPE_ENUM_INDEX},\\n\")\n math(EXPR __LAYER_SHADER_TYPE_ENUM_INDEX \"${__LAYER_SHADER_TYPE_ENUM_INDEX}+1\")\nendmacro()\n\n" }, { "path": "cmake/ncnn_generate_avx512_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_X86_H\" \"LAYER_${CLASS_UPPER}_X86_AVX512_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_x86\" \"${CLASS}_x86_avx512\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_x86.h\\\"\" \"#include \\\"${CLASS_LOWER}_x86_avx512.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_avx_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_X86_H\" \"LAYER_${CLASS_UPPER}_X86_AVX_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_x86\" \"${CLASS}_x86_avx\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_x86.h\\\"\" \"#include \\\"${CLASS_LOWER}_x86_avx.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_fma_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_X86_H\" \"LAYER_${CLASS_UPPER}_X86_FMA_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_x86\" \"${CLASS}_x86_fma\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_x86.h\\\"\" \"#include \\\"${CLASS_LOWER}_x86_fma.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_lasx_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_LOONGARCH_H\" \"LAYER_${CLASS_UPPER}_LOONGARCH_LASX_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_loongarch\" \"${CLASS}_loongarch_lasx\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_loongarch.h\\\"\" \"#include \\\"${CLASS_LOWER}_loongarch_lasx.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_lsx_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_LOONGARCH_H\" \"LAYER_${CLASS_UPPER}_LOONGARCH_LSX_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_loongarch\" \"${CLASS}_loongarch_lsx\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_loongarch.h\\\"\" \"#include \\\"${CLASS_LOWER}_loongarch_lsx.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_msa_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_MIPS_H\" \"LAYER_${CLASS_UPPER}_MIPS_MSA_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_mips\" \"${CLASS}_mips_msa\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_mips.h\\\"\" \"#include \\\"${CLASS_LOWER}_mips_msa.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_param_header.cmake", "content": "\n# must define PARAM_HEADER PARAM_SRC PARAM_SRC_NAME_WE\n\nfile(READ ${PARAM_SRC} param_data)\n\n# remove whitespace\nstring(REGEX REPLACE \"\\n +\" \"\\n\" param_data ${param_data})\n\n# replace more spaces to one space\nstring(REGEX REPLACE \"[ \\t]+\" \" \" param_data \"${param_data}\")\n\n# remove empty line\nstring(REGEX REPLACE \"\\n[\\n]+\" \"\\n\" param_data \"${param_data}\")\n\n# text to hex\nfile(WRITE ${CMAKE_CURRENT_BINARY_DIR}/param/${PARAM_SRC_NAME_WE}.text2hex.txt \"${param_data}\")\nfile(READ ${CMAKE_CURRENT_BINARY_DIR}/param/${PARAM_SRC_NAME_WE}.text2hex.txt param_data_hex HEX)\nstring(REGEX REPLACE \"([0-9a-f][0-9a-f])\" \"0x\\\\1,\" param_data_hex ${param_data_hex})\nstring(FIND \"${param_data_hex}\" \",\" tail_comma REVERSE)\nstring(SUBSTRING \"${param_data_hex}\" 0 ${tail_comma} param_data_hex)\n\n# generate model param header file\nfile(WRITE ${PARAM_HEADER} \"static const char ${PARAM_SRC_NAME_WE}_param_data[] = {${param_data_hex},0x00};\\n\")\n" }, { "path": "cmake/ncnn_generate_rvv_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_RISCV_H\" \"LAYER_${CLASS_UPPER}_RISCV_RVV_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_riscv\" \"${CLASS}_riscv_rvv\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_riscv.h\\\"\" \"#include \\\"${CLASS_LOWER}_riscv_rvv.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/ncnn_generate_shader_comp_header.cmake", "content": "\n# must define SHADER_COMP_HEADER SHADER_SRC\n\nfile(READ ${SHADER_SRC} comp_data)\n\n# skip leading comment\nstring(FIND \"${comp_data}\" \"#version\" version_start)\nif(NOT ${version_start} EQUAL -1)\n string(SUBSTRING \"${comp_data}\" ${version_start} -1 comp_data)\nendif()\n\n# remove whitespace\nstring(REGEX REPLACE \"\\n +\" \"\\n\" comp_data \"${comp_data}\")\n\n# remove comments\nstring(REGEX REPLACE \"//[^\\n]*\" \"\" comp_data \"${comp_data}\")\n\n# replace more spaces to one space\nstring(REGEX REPLACE \"[ \\t]+\" \" \" comp_data \"${comp_data}\")\n\n# remove empty line\nstring(REGEX REPLACE \"\\n[\\n]+\" \"\\n\" comp_data \"${comp_data}\")\n\nget_filename_component(SHADER_SRC_NAME_WE ${SHADER_SRC} NAME_WE)\n\n# text to hex\nfile(WRITE ${CMAKE_CURRENT_BINARY_DIR}/layer/vulkan/shader/${SHADER_SRC_NAME_WE}.text2hex.txt \"${comp_data}\")\nfile(READ ${CMAKE_CURRENT_BINARY_DIR}/layer/vulkan/shader/${SHADER_SRC_NAME_WE}.text2hex.txt comp_data_hex HEX)\nstring(REGEX REPLACE \"([0-9a-f][0-9a-f])\" \"0x\\\\1,\" comp_data_hex ${comp_data_hex})\nstring(FIND \"${comp_data_hex}\" \",\" tail_comma REVERSE)\nstring(SUBSTRING \"${comp_data_hex}\" 0 ${tail_comma} comp_data_hex)\n\nfile(WRITE ${SHADER_COMP_HEADER} \"static const char ${SHADER_SRC_NAME_WE}_comp_data[] = {${comp_data_hex}};\\n\")\n" }, { "path": "cmake/ncnn_generate_xtheadvector_source.cmake", "content": "\n# must define SRC DST CLASS\n\nfile(READ ${SRC} source_data)\n\n# replace\nstring(TOUPPER ${CLASS} CLASS_UPPER)\nstring(TOLOWER ${CLASS} CLASS_LOWER)\n\nstring(REGEX REPLACE \"LAYER_${CLASS_UPPER}_RISCV_H\" \"LAYER_${CLASS_UPPER}_RISCV_XTHEADVECTOR_H\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"${CLASS}_riscv\" \"${CLASS}_riscv_xtheadvector\" source_data \"${source_data}\")\nstring(REGEX REPLACE \"#include \\\"${CLASS_LOWER}_riscv.h\\\"\" \"#include \\\"${CLASS_LOWER}_riscv_xtheadvector.h\\\"\" source_data \"${source_data}\")\n\nfile(WRITE ${DST} \"${source_data}\")\n" }, { "path": "cmake/run_test.cmake", "content": "\r\nexecute_process(COMMAND $ENV{TESTS_EXECUTABLE_LOADER} $ENV{TESTS_EXECUTABLE_LOADER_ARGUMENTS} ${TEST_EXECUTABLE} $ENV{TESTS_ARGUMENTS} RESULT_VARIABLE result)\r\nif(NOT \"${result}\" STREQUAL \"0\")\r\n message(FATAL_ERROR \"Test failed with return value '${result}'\")\r\nendif()\r\n" }, { "path": "codeformat.sh", "content": "#!/usr/bin/env bash\n\n# we run clang-format and astyle twice to get stable format output\n\nformat_code() {\n find src/ tools/ tests/ examples/ benchmark/ python/ -type f -name '*.c' -o -name '*.cpp' -o -name '*.cc' -o -name '*.h' | grep -v python/pybind11 | grep -v stb_image | grep -v ruapu | xargs -i clang-format -i {}\n astyle -n -r \"benchmark/*.h,*.cpp,*.cc\" \"tests/*.h,*.cpp,*.cc\" \"tools/*.h,*.cpp,*.cc\" \"examples/*.h,*.cpp,*.cc\"\n astyle -n -r \"src/*.h,*.cpp,*.cc\" --exclude=src/stb_image.h --exclude=src/stb_image_write.h --exclude=src/ruapu.h\n astyle -n -r \"python/*.h,*.cpp,*.cc\" --exclude=python/pybind11\n}\n\nformat_code || { echo 'Formatting failed' ; exit 1; } #first time execute\nformat_code || { echo 'Formatting failed' ; exit 1; } #second time execute\n" }, { "path": "docs/Home.md", "content": "### input data and extract output\n```cpp\n#include \n#include \n#include \"net.h\"\n\nint main()\n{\n cv::Mat img = cv::imread(\"image.ppm\", CV_LOAD_IMAGE_GRAYSCALE);\n int w = img.cols;\n int h = img.rows;\n\n // subtract 128, norm to -1 ~ 1\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(img.data, ncnn::Mat::PIXEL_GRAY, w, h, 60, 60);\n float mean[1] = { 128.f };\n float norm[1] = { 1/128.f };\n in.substract_mean_normalize(mean, norm);\n\n ncnn::Net net;\n net.load_param(\"model.param\");\n net.load_model(\"model.bin\");\n\n ncnn::Extractor ex = net.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat feat;\n ex.extract(\"output\", feat);\n\n return 0;\n}\n\n```\n\n### print Mat content\n```cpp\nvoid pretty_print(const ncnn::Mat& m)\n{\n for (int q=0; q normed_feats(m.c);\n\n for (int i=0; i(y);\n uchar* sp = normed_feats[i].ptr(y);\n for (int x=0; x /sys/class/power_supply/battery/charging_enabled\n\ncurrent(μA) = adb shell cat /sys/class/power_supply/battery/current_now (multiply -1 for 810 chip)\n\nvoltage(μV) = adb shell cat /sys/class/power_supply/battery/voltage_now\n\npower consumption(mW) = current / 1000 * voltage / 1000 / 1000\n\nperformance per watt(1000fps/W) = fps / power consumption * 1000\n\n\n***\n\nThe binary size after debug stripping\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/1.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/2.jpg)\n\n***\n\nsqueezenet\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/3.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/4.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/5.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/6.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/7.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/8.jpg)\n***\n\nmobilenet\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/9.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/10.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/11.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/12.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/13.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/14.jpg)\n***\n\nvgg16\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/15.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/16.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/17.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/18.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/19.jpg)\n\n![](https://github.com/nihui/ncnn-assets/raw/master/20180413/20.jpg)\n" }, { "path": "docs/benchmark/vulkan-conformance-test.md", "content": "\n|device|gpu|api version|driver version|squeezenet|mobilenetssd|yolov3|\n|---|---|---|---|---|---|---|\n|intel-i7-7700|Intel(R) HD Graphics 630 (Kaby Lake GT2)|1.1.90|18.3.4|y|y|y|\n|GTX-1060|GeForce GTX 1060 3GB|1.1.95|418.172.0|y|y|y|\n|AMD-Radeon R9 M290X|AMD RADV PITCAIRN (LLVM 7.0.1)|1.1.70|18.3.4|y|y|y|\n|iphone-5s|Apple A7 GPU|1.0.82|0.2.1825|y|y|y|\n|huawei-nexus6p|Adreno (TM) 430|1.0.49|35.601.2388|y|y|y\n|vivo-y1731ca|Adreno (TM) 505|1.0.61|37.845.1429|y|n|n|\n|vivo-y85a|Adreno (TM) 506|1.0.61|2.944.3349|y|n|n|\n|vivo-x9s|Adreno (TM) 510|1.0.61|42.917.1172|y|y|y|\n|meizu-15|Adreno (TM) 512|1.0.38|29.189.223|n|n|n|\n|chuizi-jianguo-pro2|Adreno (TM) 512|1.0.38|21.219.2615|n|n|n|\n|xiaomi-note3|Adreno (TM) 512|1.0.38|39.369.2305|n|n|n|\n|oppo-r11|Adreno (TM) 512|1.0.38|42.977.756|n|n|n|\n|xiaomi-6x|Adreno (TM) 512|1.0.61|14.322.3739|y|y|y|\n|oppo-r11s+|Adreno (TM) 512|1.0.61|35.1004.3936|y|y|y|\n|vivo-x20a|Adreno (TM) 512|1.0.61|43.10.3141|y|y|y|\n|vivo-v1816a|Adreno (TM) 512|1.0.61|43.10.3141|y|y|y|\n|vivo-z1|Adreno (TM) 512|1.0.61|43.10.3141|y|y|y|\n|xiaomi-redmi-note5|Adreno (TM) 512|1.0.61|63.219.2354|y|y|y|\n|google-pixel|Adreno (TM) 530|1.1.87|512.354.0|y|y|y|\n|nubia-z17|Adreno (TM) 540|1.0.38|1.28.32|n|n|n|\n|samsung-galaxys8+|Adreno (TM) 540|1.0.61|29.896.3583|y|y|y|\n|oneplus-5t|Adreno (TM) 540|1.0.61|18.1023.2233|y|y|y|\n|google-pixel2|Adreno (TM) 540|1.1.66|512.313.0|y|y|y|\n|essential-ph-1|Adreno (TM) 540|1.1.66|512.319.0|y|y|y|\n|vivo-x23|Adreno (TM) 615|1.0.66|33.870.3328|y|y|y|\n|vivo-v1813ba|Adreno (TM) 615|1.0.66|33.870.3328|y|y|y|\n|xiaomi-8se|Adreno (TM) 616|1.0.66|30.913.18|y|y|y|\n|vivo-nex-a|Adreno (TM) 616|1.0.66|33.870.3328|y|y|y|\n|xiaomi-mix2s|Adreno (TM) 630|1.0.61|4.91.2976|y|y|y|\n|heisha-SKR-A0|Adreno (TM) 630|1.0.61|36.173.3586|y|y|y|\n|heisha-SKR-A0|Adreno (TM) 630|1.0.66|47.448.1532|y|y|y|\n|oneplus-6|Adreno (TM) 630|1.1.66|512.324.0|y|y|y|\n|vivo-iQOO|Adreno (TM) 640|1.1.87|512.361.0|y|y|y|\n|meitu-m8s|Mali-T880|1.0.14|500.910.1017|n|n|n|\n|huawei-p10|Mali-G71|1.0.53|151.949.2145|n|n|n|\n|huawei-mate9|Mali-G71|1.0.53|151.949.2145|n|n|n|\n|oppo-a73|Mali-G71|1.0.47|575.795.1934|n|n|n|\n|vivo-y97|Mali-G72|1.0.58|240.537.3580|n|n|n|\n|huawei-mate10|Mali-G72|1.0.66|14.0.0|y|y|y|\n|huawei-v10|Mali-G72|1.0.66|14.0.0|y|y|y|\n|huawei-vce-al00|Mali-G72|1.0.66|14.0.0|y|y|y|\n|huawei-mate20|Mali-G76|1.0.66|14.0.0|y|y|y|\n|huawei-pct-al10|Mali-G76|1.0.66|14.0.0|y|y|y|" }, { "path": "docs/developer-guide/aarch64-mix-assembly-and-intrinsic.md", "content": "```c\n// v寄存器全部使用 %.4s\n// 128-bit vreg matches %.4s\n// a += b * c\nfloat32x4_t _a = vld1q_f32(a);\nfloat32x4_t _b = vld1q_f32(b);\nfloat32x4_t _c = vld1q_f32(c);\nasm volatile(\n \"fmla %0.4s, %2.4s, %3.4s\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// v寄存器使用低64位 %.2s\n// low 64-bit vreg matches %.2s\n// a += b * c\nfloat32x2_t _a = vld1_f32(a);\nfloat32x2_t _b = vld1_f32(b);\nfloat32x2_t _c = vld1_f32(c);\nasm volatile(\n \"fmla %0.2s, %2.2s, %3.2s\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// v寄存器单路使用 %.s[0] %.s[1] %.s[2] %.s[3]\n// 32-bit register matches %.s[0]\n// a += b * c[0]\n// a += b * c[1]\n// a += b * c[2]\n// a += b * c[3]\nfloat32x4_t _a = vld1_f32(a);\nfloat32x4_t _b = vld1_f32(b);\nfloat32x4_t _c = vld1_f32(c);\nasm volatile(\n \"fmla %0.4s, %2.4s, %3.s[0]\"\n \"fmla %0.4s, %2.4s, %3.s[1]\"\n \"fmla %0.4s, %2.4s, %3.s[2]\"\n \"fmla %0.4s, %2.4s, %3.s[3]\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n\n\nqwq\n" }, { "path": "docs/developer-guide/add-custom-layer.zh.md", "content": "# NCNN增加自定义层\n\n## 举例\n\n这里举个例子添加自定义层次 如Relu6，即 std::min(6.f, std::max(0.f, val))\n\n```\nInput input 0 1 input\nConvolution conv2d 1 1 input conv2d 0=32 1=1 2=1 3=1 4=0 5=0 6=768\nRelu6 relu6 1 1 conv2d relu6\nPooling maxpool 1 1 relu6 maxpool 0=0 1=3 2=2 3=-233 4=0\n```\n\n\n\n## 定义源码h文件：src/layer/relu6.h\n\n```CPP\n#ifndef LAYER_RELU6_H\n#define LAYER_RELU6_H\n\n#include \"layer.h\"\n\nnamespace ncnn {\n\nclass Relu6 : public Layer\n{\npublic:\n Relu6();\n\n virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;\n};\n\n} // namespace ncnn\n\n#endif // LAYER_RELU6_H\n```\n\n\n\n## 定义源码CPP文件：src/layer/relu6.cpp\n\n```CPP\n#include \"relu6.h\"\n\n#include \n\nnamespace ncnn {\n\nRelu6::Relu6()\n{\n one_blob_only = true;\n support_inplace = true;\n}\n\nint Relu6::forward_inplace(Mat& bottom_top_blob, const Option& opt) const\n{\n int w = bottom_top_blob.w;\n int h = bottom_top_blob.h;\n int channels = bottom_top_blob.c;\n int size = w * h;\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int q=0; q < channels; q++)\n {\n float* ptr = bottom_top_blob.channel(q);\n\n for (int i=0; i weights(0);\n\n int ret = test_layer(\"Relu6\", pd, weights, a);\n if (ret != 0)\n {\n fprintf(stderr, \"test_relu6 failed a.dims=%d a=(%d %d %d)\\n\", a.dims, a.w, a.h, a.c);\n }\n\n return ret;\n}\n\nstatic int test_relu6_0()\n{\n return 0\n || test_relu6(RandomMat(5, 7, 24))\n || test_relu6(RandomMat(7, 9, 12))\n || test_relu6(RandomMat(3, 5, 13));\n}\n\nstatic int test_relu6_1()\n{\n return 0\n || test_relu6(RandomMat(15, 24))\n || test_relu6(RandomMat(17, 12))\n || test_relu6(RandomMat(19, 15));\n}\n\nstatic int test_relu6_2()\n{\n return 0\n || test_relu6(RandomMat(128))\n || test_relu6(RandomMat(124))\n || test_relu6(RandomMat(127));\n}\n\nint main()\n{\n SRAND(7767517);\n\n return 0\n || test_relu6_0()\n || test_relu6_1()\n || test_relu6_2();\n}\n\n```\n\n\n\n## 修改tests/CMakeLists.txt 注册Relu6测试用例\n\n```CPP\nncnn_add_layer_test(LSTM)\nncnn_add_layer_test(Yolov3DetectionOutput)\nncnn_add_layer_test(Relu6)\n```\n\n\n\n## 编译\n\n```\n按原NCNN步骤编译\n```\n\n\n\n## 单元测试\n\n```\n./test_relu6\n```\n\n" }, { "path": "docs/developer-guide/arm-a53-a55-dual-issue.md", "content": "## natural assembly\n* no register dependency, no penalty\n```\nld1 {v0.4s}, [r0], #16\nfmla v10.4s, v16.4s, v24.s[0]\nfmla v11.4s, v16.4s, v24.s[1]\nfmla v12.4s, v16.4s, v24.s[2]\nfmla v13.4s, v16.4s, v24.s[3]\n```\n\n## A53\n* 128bit vector load cannot be dual issued with fmla, wait 2 cycles\n* 64bit vector load cannot be dual issued with fmla, wait 1 cycle\n* 64bit integer load can be dual issued with fmla, no penalty\n* pointer update can be dual issued with fmla, no penalty\n* 64bit vector load and 64bit vector insert can be dual issued, no penalty\n* any vector load cannot be issued on the 4th cycle of each fmla (enters the accumulator pipeline)\n\n### practical guide\n* use 64bit vector load only\n* issue vector load every three fmla\n* 1 cycle to load 64bit, dual issue with the previous interleaved 64bit insert\n* load the remaining 64bit into integer register, dual issue with fmla\n* update pointer, dual issue with fmla\n* insert 64bit into vector from integer register, dual issue with the next interleaved 64bit load\n* add nop every three fmla if no load, seems to be faster\n```\nldr d0, [r0] // 1 cycle, v0 first 64bit\nfmla\nldr x23, [r0, #8] // 0 cycle, v0 second 64bit to temp register\nfmla\nadd r0, r0, #16 // 0 cycle, update pointer\nfmla\nldr d1, [r0] // 1 cycle, v1 first 64bit\nins v0.d[1], x23 // 0 cycle, v0 second 64bit complete\nfmla\nldr x23, [r0, #8] // 0 cycle, v1 second 64bit to temp register\nfmla\nadd r0, r0, #16 // 0 cycle, update pointer\nfmla\nins v1.d[1], x23 // 1 cycle, v1 second 64bit complete\nnop\nfmla\nfmla\nfmla\nnop\nnop\nfmla\nfmla\nfmla\n```\n\n## A55\n* Limited by the number of neon register read and write ports, most neon instructions cannot be dual-issued.\n* neon instructions have different latencies\n* 128bit vector load cannot be issued with fmla, WAR wait 2 cycles\n* 64bit integer load can be dual issued with fmla, no penalty\n* pointer update can be dual issued with fmla, no penalty\n* 64bit vector insert can be dual issued with fmla, no penalty\n\n### practical guide\n* A55 supports 128bit load and 256bit write in one clock. Support dual emission of two 64bit vector loads or single emission of 128bit vector load\n* `ldr`, dual issue with fmla\n* load the remaining 64bit into integer register, dual issue with fmla\n* update pointer, dual issue with fmla\n* insert 64bit into vector from integer register, dual issue with fmla\n* interleaved load loose register dependency\n* nop trick is not needed\n* Loop unrolling fma reduces pipeline bubbles\n* Some data type conversion neon instructions can be dual issued, such as `fsvts`\n```\nldr d0, [r0] // 0 cycle, v0 first 64bit\nfmla\nldr x23, [r0, #8] // 0 cycle, v0 second 64bit to temp register\nfmla\nadd r0, r0, #16 // 0 cycle, update pointer\nfmla\nldr d1, [r0] // 0 cycle, v1 first 64bit\nfmla\nins v0.d[1], x23 // 0 cycle, v0 second 64bit complete\nfmla\nldr x23, [r0, #8] // 0 cycle, v1 second 64bit to temp register\nfmla\nadd r0, r0, #16 // 0 cycle, update pointer\nfmla\nins v1.d[1], x23 // 0 cycle, v1 second 64bit complete\nfmla\n```\n" }, { "path": "docs/developer-guide/armv7-mix-assembly-and-intrinsic.md", "content": "```c\n// d寄存器全部使用 %P\n// d reg matches %P\n// a += b * c\nfloat32x2_t _a = vld1_f32(a);\nfloat32x2_t _b = vld1_f32(b);\nfloat32x2_t _c = vld1_f32(c);\nasm volatile(\n \"vmla.f32 %P0, %P2, %P3\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// q寄存器全部使用 %q\n// q reg matches %q\n// a += b * c\nfloat32x4_t _a = vld1q_f32(a);\nfloat32x4_t _b = vld1q_f32(b);\nfloat32x4_t _c = vld1q_f32(c);\nasm volatile(\n \"vmla.f32 %q0, %q2, %q3\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// d寄存器单路使用 %P[0] %P[1]\n// 32bit d reg matches %P[0]\n// a += b * c[0]\n// a += b * c[1]\nfloat32x2_t _a = vld1_f32(a);\nfloat32x2_t _b = vld1_f32(b);\nfloat32x2_t _c = vld1_f32(c);\nasm volatile(\n \"vmla.f32 %P0, %P2, %P3[0]\"\n \"vmla.f32 %P0, %P2, %P3[1]\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// q寄存器单路使用 %e[0] %e[1] %f[0] %f[1]\n// 32-bit q reg matches %e[0]\n// a += b * c[0]\n// a += b * c[1]\n// a += b * c[2]\n// a += b * c[3]\nfloat32x4_t _a = vld1q_f32(a);\nfloat32x4_t _b = vld1q_f32(b);\nfloat32x4_t _c = vld1q_f32(c);\nasm volatile(\n \"vmla.f32 %q0, %q2, %e3[0]\"\n \"vmla.f32 %q0, %q2, %e3[1]\"\n \"vmla.f32 %q0, %q2, %f3[0]\"\n \"vmla.f32 %q0, %q2, %f3[1]\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// q寄存器拆分d寄存器使用 %e %f\n// use %e %f to split q reg into two d regs\n// a += b * c[0]c[1]\n// a += b * c[2]c[3]\nfloat32x2_t _a = vldq_f32(a);\nfloat32x2_t _b = vldq_f32(b);\nfloat32x4_t _c = vld1q_f32(c);\nasm volatile(\n \"vmla.f32 %P0, %P2, %e3\"\n \"vmla.f32 %P0, %P2, %f3\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// d寄存器声明绑定\n// specify concrete d reg which want to save\n// vmla.f32 d0, d2, d4\nregister float32x2_t _a asm(\"d0\") = vld1_f32(a);\nregister float32x2_t _b asm(\"d2\") = vld1_f32(b);\nregister float32x2_t _c asm(\"d4\") = vld1_f32(c);\n\nasm volatile(\n \"vmla.f32 %P0, %P2, %P3\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n```c\n// q寄存器声明绑定\n// bind q reg with data\n// vmla.f32 q0, q1, q2\nregister float32x4_t _a asm(\"q0\") = vld1q_f32(a);\nregister float32x4_t _b asm(\"q1\") = vld1q_f32(b);\nregister float32x4_t _c asm(\"q2\") = vld1q_f32(c);\n\nasm volatile(\n \"vmla.f32 %q0, %q2, %q3\"\n : \"=w\"(_a) // %0\n : \"0\"(_a),\n \"w\"(_b), // %2\n \"w\"(_c) // %3\n :\n);\n```\n\n如果不是因为编译器的bug，寄存器绑定是用不着的，然而。。。\n\nhttps://gcc.gnu.org/bugzilla/show_bug.cgi?id=41538\n\nqwq\n" }, { "path": "docs/developer-guide/binaryop-broadcasting.md", "content": "### broadcasting rule\n\nncnn BinaryOp accepts blobs with different shape\n\nC = BinaryOp(A, B)\n\nshape notation convention is [w], [w,h], [w,h,c], [w,h,d,c]\n\n* binaryop with scalar and scalar-like\n\n|A|B|C|\n|---|---|---|\n|[2]|scalar / [1]|[2]|\n|[2,3]|scalar / [1] / [1,1]|[2,3]|\n|[2,3,4]|scalar / [1] / [1,1] / [1,1,1]|[2,3,4]|\n|[2,3,4,5]|scalar / [1] / [1,1] / [1,1,1] / [1,1,1,1]|[2,3,4,5]|\n\n* no broadcast\n\n|A|B|C|\n|---|---|---|\n|[2]|[2]|[2]|\n|[2,3]|[2,3]|[2,3]|\n|[2,3,4]|[2,3,4]|[2,3,4]|\n|[2,3,4,5]|[2,3,4,5]|[2,3,4,5]|\n\n* explicit broadcast B\n\n|A|B|C|\n|---|---|---|\n|[2,3]|[1,3]|[2,3]|\n|[2,3]|[2,1]|[2,3]|\n|[2,3,4]|[1,3,4]|[2,3,4]|\n|[2,3,4]|[2,1,4]|[2,3,4]|\n|[2,3,4]|[2,3,1]|[2,3,4]|\n|[2,3,4]|[1,1,4]|[2,3,4]|\n|[2,3,4]|[1,3,1]|[2,3,4]|\n|[2,3,4]|[2,1,1]|[2,3,4]|\n|[2,3,4,5]|[1,3,4,5]|[2,3,4,5]|\n|[2,3,4,5]|[2,1,4,5]|[2,3,4,5]|\n|[2,3,4,5]|[2,3,1,5]|[2,3,4,5]|\n|[2,3,4,5]|[2,3,4,1]|[2,3,4,5]|\n|[2,3,4,5]|[1,1,4,5]|[2,3,4,5]|\n|[2,3,4,5]|[1,3,1,5]|[2,3,4,5]|\n|[2,3,4,5]|[1,3,4,1]|[2,3,4,5]|\n|[2,3,4,5]|[2,1,1,5]|[2,3,4,5]|\n|[2,3,4,5]|[2,1,4,1]|[2,3,4,5]|\n|[2,3,4,5]|[2,3,1,1]|[2,3,4,5]|\n|[2,3,4,5]|[1,1,1,5]|[2,3,4,5]|\n|[2,3,4,5]|[1,1,4,1]|[2,3,4,5]|\n|[2,3,4,5]|[1,3,1,1]|[2,3,4,5]|\n|[2,3,4,5]|[2,1,1,1]|[2,3,4,5]|\n\n* implicit broadcast B for inner axis\n\nIt broadcasts in the opposite direction of the numpy's implicit broadcasting behavior.\n\npnnx will insert reshape operator at the appropriate position to convert it to explicit broadcast automatically.\n\n|A|B|C|\n|---|---|---|\n|[2,3]|[3]|[2,3]|\n|[2,3,4]|[4]|[2,3,4]|\n|[2,3,4]|[3,4]|[2,3,4]|\n|[2,3,4,5]|[5]|[2,3,4,5]|\n|[2,3,4,5]|[4,5]|[2,3,4,5]|\n|[2,3,4,5]|[3,4,5]|[2,3,4,5]|\n\n* implicit broadcast B with 1 dimension rank for outer axis\n\nThis exists only for compatibility.\n\nWhen the size is the same, eg. [2,2] and [2], broadcast B for inner axis will be prioritized.\n\n|A|B|C|\n|---|---|---|\n|[2,3]|[2]|[2,3]|\n|[2,3,4]|[2]|[2,3,4]|\n|[2,3,4,5]|[2]|[2,3,4,5]|\n" }, { "path": "docs/developer-guide/build-ncnn-on-windows-xp.zh.md", "content": "# Build ncnn on Windows XP\n\n> **Contributors:** [@Sugar-Baby](https://github.com/Sugar-Baby) and [@AtomAlpaca](https://github.com/AtomAlpaca)\n\n## 0. 环境准备\n\n#### 0.1 虚拟机设置\n\n我使用的是[我的MSDN](https://www.imsdn.cn/)提供的[Windows XP SP3 x64版本](https://www.imsdn.cn/operating-systems/windows-xp/)。虚拟机使用Oracle VM VirtualBox，内存4GB，存储空间64GB（C盘16GB，D盘48GB）。\n\n**在虚拟机关机的情况下**，点击虚拟机管理器界面的\"设置\"-\"网络\"-\"高级\"，将控制芯片改为PCnet-FAST III，混杂模式设置为拒绝，勾选接入网线，点击\"OK\"保存。重启虚拟机就可以连接上网络了。\n\n点击虚拟机界面的\"设备\"-\"安装增强功能...\"，在虚拟机中进入\"我的电脑\"，刷新后出现\"VirtualBox Guest Additions (D: )\"，右键选择\"自动播放\"，完成安装后重启。\n\n点击虚拟机界面的\"设备\"-\"共享粘贴板\"，设置为\"双向\"。点击\"设备\"-\"共享文件夹\"-\"共享文件夹..\"，点击右侧加号，在\"共享文件夹路径\"中选择\"其他...\"，然后选择需要共享的主机文件夹。勾选\"自动挂载\"和\"固定分配\"，点击\"OK\"保存。在虚拟机中进入\"我的电脑\"，刷新后出现'VBoxSvr' 上的 <主机文件夹名称>，双击进入就可以双向传输文件了。\n\n#### 0.2 开发环境配置\n\n浏览器推荐[Mypal 68](https://www.mypal-browser.org/download.html)，注意要选择32位版本。Windows XP自带ZIP文件解压。安装后就可以访问互联网了。\n\n从Github下载[w64devkit](https://github.com/skeeto/w64devkit)，选择x86版本。这里下载的是一个自解压的7z文件，在虚拟机中解压即可。\n\n在\"开始\"-\"控制面板\"-\"切换到经典视图\"-\"系统\"-\"高级\"-\"环境变量\"-\"系统变量\"中，选择Path，点击\"编辑\"，在字符串末尾加入一个分号(;)，然后粘贴w64devkit下bin文件夹的目录。点击\"确定\"保存之后可以打开命令提示符输入例如c++的命令验证是否成功加入环境变量。\n\n由于年代过于久远，Git的官方release已经没有兼容Windows XP的版本了。最后一个兼容的版本(1.9.5)可以在[这里](https://www.xiazaiba.com/html/29352.html)下载。\n\n为了使用Git，需要安装[Win32 OpenSSL](https://slproweb.com/products/Win32OpenSSL.html)。选择Win32 OpenSSL Light版本。这个过程中会附带安装VC++ 2022运行时库。\n\n如果因为协议、代理等问题不能在虚拟机中使用Git，也可以下载ZIP版本后在虚拟机中解压。\n\n需要手动下载[CMake最后支持Windows XP的版本](https://github.com/Kitware/CMake/releases/download/v3.10.3/cmake-3.10.3-win32-x86.zip)。建议解压在C:\\Program Files下，并且需要设置系统变量，到CMake目录下的bin文件夹。具体可以参考上面w64devkit的方法。\n\n## 1. 编译\n\n### 1.1 使用 MinGW-w64\n\n运行\n\n```bash\ncd \nmkdir build\ncd build\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/windows-xp-mingw.toolchain.cmake -DNCNN_VULKAN=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_AVX=OFF -DCMAKE_BUILD_TYPE=Release -G \"MinGW Makefiles\" ..\nmake -j2\nmake install\n```\n\n由于平台性能的限制，Vulkan SDK 最低要求 Windows 7 SP1，XP 无法安装官方驱动和工具链，因此需要关闭Vulkan选项。同时需要使用简化版 OpenCV 替代库NCNN_SIMPLEOCV。\n\n### 1.2 使用 Clang\n\n需要先配置 MinGW-w64 环境，然后安装 Clang 6.0 或更高版本。\n\n```bash\ncd \nmkdir build\ncd build\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/windows-xp-clang.toolchain.cmake -DNCNN_SIMPLEOCV=ON -DNCNN_SIMPLEOMP=ON -DNCNN_AVX=OFF -DCMAKE_BUILD_TYPE=Release -G \"MinGW Makefiles\" ..\nmake -j2\nmake install\n```\n\n### 1.3 使用 Visual Studio (MSVC)\n\n需要安装支持 Windows XP 的 v141_xp 工具集：\n\n1. 打开 Visual Studio 安装程序（工具 → 获取工具和功能）\n2. 选择\"使用 C++ 的桌面开发\"\n3. 在摘要部分选择\"对 C++ 的 Windows XP 支持\"\n4. 点击修改\n\n```bash\ncd \nmkdir build\ncd build\ncmake -A WIN32 -G \"Visual Studio 17 2022\" -T v141_xp -DNCNN_SIMPLEOCV=ON -DNCNN_OPENMP=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_WITH_STATIC_CRT=ON -DCMAKE_TOOLCHAIN_FILE=../toolchains/windows-xp-msvc.toolchain.cmake ..\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\n## 2. 测试\n\n### 2.1 benchncnn\n\n将benchmark目录下的所有文件复制到build/benchmark目录下。在命令提示符中cd到build/benchmark， 然后运行\n\n```bash\nbenchncnn [测试的循环次数] [线程数] [节能模式]\n```\n\n其中，节能模式取值为0时关闭，为1时打开。\n\n### 2.2 examples\n\n从[这里](https://github.com/nihui/ncnn-assets/tree/master/models)可以下载到所有需要的param和bin文件。需要注意的是，ZF_faster_rcnn_final.bin开头的三个文件（.zip，.z01，.z02）最好先放在主机上解压出bin文件再传进虚拟机。\n\n把这些文件放在build/examples目录下。\n\n我写了一个bat脚本来批量测试这些模型：\n\n```batch\n@echo off\nsetlocal enabledelayedexpansion\n\nset EXAMPLES_DIR=\\BUILD\\EXAMPLES\nset IMAGE_PATH=\\IMAGES\\256-ncnn.png\nset LOG_FILE=test_results.log\n\necho NCNN Examples Test Results > %LOG_FILE%\necho ========================= >> %LOG_FILE%\necho Test started: %date% %time% >> %LOG_FILE%\necho. >> %LOG_FILE%\n\nfor %%f in (\"%EXAMPLES_DIR%\\*.exe\") do (\n set EXE_NAME=%%~nf\n set EXE_PATH=%%f\n echo Testing: !EXE_NAME! >> %LOG_FILE%\n echo -------------------------------- >> %LOG_FILE%\n\n !EXE_PATH! \"%IMAGE_PATH%\" >> %LOG_FILE% 2>&1\n\n if errorlevel 1 (\n echo [ERROR] !EXE_NAME! failed to run. >> %LOG_FILE%\n ) else (\n echo [SUCCESS] !EXE_NAME! completed. >> %LOG_FILE%\n )\n echo. >> %LOG_FILE%\n)\n\necho Test finished: %date% %time% >> %LOG_FILE%\necho Results saved to %LOG_FILE%\nendlocal\n```\n\n把这个bat脚本放在build/examples目录下，替换掉所有的``，双击运行。通过生成的test_results.log即可查看所有模型的结果。\n\n通过修改`set IMAGE_PATH=\\IMAGES\\256-ncnn.png`中的路径来更换需要测试的文件。" }, { "path": "docs/developer-guide/custom-allocator.md", "content": "Mat structure is now allocator-aware via an extra allocator parameter with default zero value.\n\nThe good-old ncnn::fastMalloc()/ncnn::fastFree() will be used for a null allocator.\n\nYou could pass a custom allocator to delegate all memory allocation and deallocation.\n\n```cpp\nclass Allocator\n{\npublic:\n virtual void* fastMalloc(size_t size) = 0;\n virtual void fastFree(void* ptr) = 0;\n};\n```\n\nncnn has already implemented two simple pooled Allocator class, with mutex lock or without it.\n\n```cpp\nncnn::PoolAllocator locked_mempool;\nncnn::UnlockedPoolAllocator unlocked_mempool;\n```\n\nthe two allocator types in ncnn\n\n* blob allocator\n\n used to allocate memory for all named blobs, which you could retrieve by Extractor::extract()\n* workspace allocator\n\n used to allocate memory for internal temporary use in layer implementation, such as the temp blob after padding in convolution\n\nby default, all Extractor instance use the two allocator in the default option\nYou can alter them by ncnn::set_default_option()\nor you can set them per Extractor by Extractor::set_blob_allocator()/Extractor::set_workspace_allocator()\n\nblob allocator is guaranteed to be called in-order in layer implementation during each Extractor lifecycle\nwhile workspace allocator may be called synchronously\n\nthe practical usage\n\n* one network, one-by-one inference\n\n shared unlocked blob allocator for all Extractor\n\n shared locked workspace allocator for all Extractor\n\n* one network, concurrent inference\n\n shared unlocked blob allocator for all Extractor in each thread\n\n shared locked workspace allocator for all Extractor among all threads\n\n* concurrent multiple networks, one-by-one inference for each network\n\n shared unlocked blob allocator for all Extractor of each network\n\n shared locked workspace allocator for all Extractor among all networks (for saving memory)\n\n* concurrent multiple networks, concurrent inference for each network\n\n shared unlocked blob allocator for all Extractor of each network in each thread\n\n shared locked workspace allocator for all Extractor among all networks (for saving memory)\n" }, { "path": "docs/developer-guide/element-packing.md", "content": "### what is packing and why\n\npacking is the form of storing multiple short-sized values as one long-sized value.\n\nelement packing is well mapped with the underlying simd register, which usually use one very wide register to store different types of values.\n\n|C|elemsize|elempack|\n|---|---|---|\n|double|8|1|\n|float|4|1|\n|int|4|1|\n|short|2|1|\n|signed char|1|1|\n\n|arm neon|elemsize|elempack|\n|---|---|---|\n|float64x2_t|16|2|\n|float32x4_t|16|4|\n|int32x4_t|16|4|\n|float16x4_t|8|4|\n|int8x8_t|8|8|\n\nThough the real count of values doubles when elempack is two, the wide-sized value is still treated as one value in the view of Mat structure. For example, we want to store 40 float values in Mat object, if elempack 1 is used, Mat width is then 40, while 10 if elempack 4 is used.\n\n|dims|w|h|c|cstep|elemsize|elempack|\n|---|---|---|---|---|---|---|\n|1|40|1|1|40|4|1|\n|1|10|1|1|10|16|4|\n\n### packing style convention\n\nIn practice, elempack 1, 4, 8 are the most common cases. It is possible to use any other packing style in theory.\n\nThe following table show the packing axis used in ncnn for different dimension.\n\n|dims|packing axis|shape before packing|shape after packing|\n|---|---|---|---|\n|1|w|w|w/elempack|\n|2|h|w, h|w, h/elempack|\n|3|c|w, h, c|w, h, c/elempack|\n\nIf the packing axis dim is not evenly divisible by elempack, zero padding may be used.\n\n```\noutw = (w + elempack - 1) / elempack;\n```\n\nThe following snippet shows the memory layout after elempack=4 on 3-dim Mat\n\n```\n// w=2 h=3 c=4 elempack=1\n0 1\n2 3\n4 5\n\n6 7\n8 9\n10 11\n\n12 13\n14 15\n16 17\n\n18 19\n20 21\n22 23\n\n// w=2 h=3 c=1 elempack=4\n(0,6,12,18) (1,7,13,19)\n(2,8,14,20) (3,9,15,21)\n(4,10,16,22) (5,11,17,23)\n```\n\n### how to convert elempack\n\nThere is a convenient wrapper function provided\n```\n// convert to elempack 4 if packing axis dim is evenly divisible by elempack\n// return the identity Mat otherwise\nncnn::Mat a;\nncnn::Mat a_packed;\nncnn::convert_packing(a, a_packed, 4);\nif (a_packed.elempack == 4)\n{\n // check if packing is successful\n}\n\n// convert to packing 1, aka unpacking, shall be always successful\nncnn::Mat b;\nncnn::Mat b_unpacked;\nncnn::convert_packing(b, b_unpacked, 1);\n```\n\n### handle general interleaved data\n\nHere is an example of using convert packing to convert RGB interleaved data to planar\n\n**NOTE:** The following code is just presented to explain what packing is and the conversion process. Do not use it in production due to its poor performance. Do use ncnn::Mat::from_pixels()\n\n```cpp\n// rgb_interleaved_u8 is RGB RGB RGB ...\n// rgb_interleaved_u8.w = w;\n// rgb_interleaved_u8.h = h;\n// rgb_interleaved_u8.c = 1;\n// rgb_interleaved_u8.elemsize = 3;\n// rgb_interleaved_u8.elempack = 3;\n\nncnn::Mat rgb_interleaved_u8(w, h, 1, 3, 3);\nncnn::Mat rgb_planar_u8;\n\nncnn::convert_packing(rgb_interleaved_u8, rgb_planar_u8, 1);\n\n// rgb_planar_u8 is now RRR ... GGG ... BBB ...\n// rgb_planar_u8.w = w;\n// rgb_planar_u8.h = h;\n// rgb_planar_u8.c = 3;\n// rgb_planar_u8.elemsize = 1;\n// rgb_planar_u8.elempack = 1;\n```\n" }, { "path": "docs/developer-guide/expression.md", "content": "### expression\n\nexpression is used in the reshape slice parameter to express the dynamic shape or subscript value based on the expression formula and input shape\n\nCompared with directly converting the expression calculation process into multiple operators, the motivation for using expression\n* No additional shape concat and other operators will be generated due to dynamic calculation, which greatly reduces the number of layers of the ncnn model and makes it easier to view the model structure and modify expression\n* Shape or subscript evaluations are usually single-digit operations, which are more suitable for direct completion on the CPU without layout conversion and kernel call overhead\n\nIn the param file, `Reshape` layer can contain 6=expression\n\nThe pnnx tool can automatically convert `pnnx.Expression` to the expr parameter of ncnn `Reshape`\n\n* Convert to 0w, 0h, 0d or 0c according to the input shape rank and `size(@0,1)`\n* Automatically remove the batch dimension according to the input batch index\n* Convert `pnnx.Expression` and `Tensor.reshape`/`Tensor.view` two operators are fused into ncnn `Reshape`\n* Automatically summarize the number of references, exclude duplicate references and sort the indexes of references\n* Convert the customary shape representation order, such as CHW to WHC\n\nExample pnnx.param where A and B are 3D tensors\n```\npnnx.Expression expr 2 1 A B shape expr=[add(size(@1,0),2),mul(size(@0,1),2),-1]\nTensor.reshape reshape 2 1 A shape out\n```\n\npnnx.py\n```python\nshape = [(B.size(0) + 2), (A.size(1) * 2), -1]\nout = A.reshape(*shape)\n```\n\nConverted to ncnn.param\n```\nReshape reshape 2 1 A B out 6=\"-1,*(0h,2),+(1c,2)\"\n```\n\n### syntax\n\nUse infix expression, format is `op(arg0,arg1,...)`, multiple operations can be nested, multiple sizes are separated by commas, and numbers can be integers or decimals\n\nAmong them, the commonly used `add` `sub` `mul` `div` `floor_div` are abbreviated as `+` `-` `*` `/` `//`, and other arithmetic operations use names, such as `sin` `ceil` `max`, etc.\n\n* `max(2,3)`\n* `floor(sin(3.14))`\n* `+(*(-2,1),10)` means (-2 * 1) + 10\n* `1,2,+(3,2)` list can represent output shape with 3-rank\n\nThe input shape can be referenced at runtime, format is `id(w|h|d|c)`, the maximum id is 9, which means that up to 10 inputs can be referenced\n\nAssuming that the Reshape layer has two input blobs, A and B, then\n\n* `0w,1h` means A.w, B.h\n* `*(+(0c,1c),2)` means (A.c + B.c) * 2\n\n### helper api\n\n```cpp\n#include \"expression.h\"\n\nint count_expression_blobs(const std::string& expr);\n\nint eval_list_expression(const std::string& expr, const std::vector& blobs, std::vector& outlist);\n```\n\n* `count_expression_blobs`\n\nPass expression to get the number of inputs it references, such as `0w,1h` returns 2\n\n* `eval_list_expression`\n\nEvaluate the result list according to expression and input blob calculate. If the calculation result is a floating point number, it will be automatically truncated to an integer.\n\n### supported operator\n\n|type|operators|\n|---|---|\n|float to int|`trunc` `ceil` `floor` `round`|\n|binary arithmetic|`+` `-` `*` `/` `//` `max` `min` `pow` `fmod` `remainder` `atan2` `logaddexp`|\n|unary arithmetic|`abs` `neg` `sign` `square` `sqrt` `rsqrt` `reciprocal` `exp` `log` `log10` `sin` `asin` `cos` `acos` `tan` `atan` `sinh` `asinh` `cosh` `acosh` `tanh` `atanh`|\n|integer bitwise|`and` `or` `xor` `lshift` `rshift`|\n" }, { "path": "docs/developer-guide/glsl-extension.md", "content": "# ncnn GLSL extension\n\n## rationale\nDifferent GPUs support different features, some support fp16 as buffer storage type, some support fp16 as operand variable, some old GPUs only support fp32\n\nWhen the GPU supports the `VK_KHR_16bit_storage` extension, in order to minimize the memory bandwidth consumption of the GPU, we will give priority to using fp16 as the storage type. Otherwise, we use `packHalf2x16` and `unpackHalf2x16` in GLSL 4.2 to compress 2 fp32 to uint, reducing read and write bandwidth.\n\nSimilarly, when the gpu supports the `VK_KHR_shader_float16_int8` extension, in order to speed up the calculation efficiency, we will give priority to using fp16 as the operation operand, which usually doubles the speed. Otherwise, we use fp32.\n\nTo ensure the widest compatibility, the following code for declaring descriptor binding and loading data will be written\n\n```c\n#if NCNN_fp16_storage // gpu supports 16bit storage\nlayout (binding = 0) buffer blob { f16vec4 blob_data[]; };\n#elif NCNN_fp16_packed // gpu supports GLSL 4.2\nlayout (binding = 0) buffer blob { uvec2 blob_data[]; };\n#else // gpu only supports fp32\nlayout (binding = 0) buffer blob { vec4 blob_data[]; };\n#endif\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n#if NCNN_fp16_storage && NCNN_fp16_arithmetic // gpu supports 16bit storage and shader float16\n f16vec4 x = blob_data[i];\n#elif NCNN_fp16_storage // gpu supports 16bit storage but no shader float16\n vec4 x = vec4(blob_data[i]);\n#elif NCNN_fp16_packed && NCNN_fp16_arithmetic // gpu supports GLSL 4.2 and shader float16\n f16vec4 x = f16vec4(unpackFloat2x16(blob_data[i].x), unpackFloat2x16(blob_data[i].y));\n#elif NCNN_fp16_packed // gpu supports GLSL 4.2\n vec4 x = vec4(unpackHalf2x16(blob_data[i].x), unpackHalf2x16(blob_data[i].y));\n#else // gpu only supports fp32\n vec4 x = blob_data[i];\n#endif\n}\n```\n\nAs you can see, just declaring the buffer type and reading a value consumes a lot of lines of code, which is a maintenance nightmare. Therefore, ncnn adds more flexible data types and auxiliary functions to reduce the size of the code and improve readability, and will automatically expand to the most efficient implementation according to the feature level supported by the GPU.\n\nThe above code, by using the ncnn glsl extension, can be simplified to\n\n```c\nlayout (binding = 0) buffer blob { sfpvec4 blob_data[]; };\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n afpvec4 x = buffer_ld4(blob_data, i);\n}\n```\n\nThe ncnn glsl extension provides the necessary data types for storage, computation, shared memory, and load, store, conversion functions for buffers and images. We also provide some buffer and image copy functions to prevent loss of precision when using fp16 as the intermediate data type, and to avoid unnecessary `unpackHalf2x16` and `packHalf2x16` pair.\n\n# entrypoint for compiling GLSL\n\nThe gpu.h header in the ncnn library exposes 3 APIs for compiling glsl code into spir-v binary, they support ncnn glsl extension, these 3 functions accept opt switch to control the expansion form of ncnn glsl extension. The first two accept raw glsl code strings, and the last one is used to create ncnn's built-in shader.\n\n```cpp\nnamespace ncnn {\n\n// online spirv compilation\nNCNN_EXPORT int compile_spirv_module(const char* comp_string, const Option& opt, std::vector& spirv);\nNCNN_EXPORT int compile_spirv_module(const char* comp_data, int comp_data_size, const Option& opt, std::vector& spirv);\nNCNN_EXPORT int compile_spirv_module(int shader_type_index, const Option& opt, std::vector& spirv);\n\n} // namespace ncnn\n```\n\n## compile ncnn extended GLSL code directly\n\nYou can write shader code with ncnn glsl extension, compiled to spir-v using ncnn functions. The compiled product is a standard-compliant spir-v binary, which can be directly used to create a pipeline object in the vulkan api\n\n```cpp\nstatic const char my_glsl_data[] = R\"(\n#version 450\n\nlayout (binding = 0) readonly buffer a_blob { sfpvec4 a_blob_data[]; };\nlayout (binding = 1) writeonly buffer b_blob { sfpvec4 b_blob_data[]; };\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n afpvec4 v = buffer_ld4(a_blob_data, i);\n\n v = v + 123;\n\n buffer_st4(b_blob_data, i, v);\n}\n)\";\n\nOption opt;\n // you can control the extension behavior\n // even if the gpu supports 16bit storage\nopt.use_fp16_storage = false;\n\nstd::vector spirv;\nncnn::compile_spirv_module(my_glsl_data, sizeof(my_glsl_data) - 1, opt, spirv);\n\n// To create pipeline object later\n// ncnn::Pipeline pipeline(vkdev);\n// pipeline.set_local_size_xyz(64, 1, 1);\n// pipeline.create(spirv.data(), spirv.size() * 4, specializations);\n```\n\n## ncnn built-in shader\n\nThe shader index inside ncnn is exposed in the `layer_shader_type.h` header and can be used if needed\n\n```cpp\n#include \"layer_shader_type.h\"\n\nint shader_type_index = LayerShaderType::convert_ycbcr;\n\nOption opt;\n\nstd::vector spirv;\nint retc = compile_spirv_module(shader_type_index, opt, spirv);\n```\n\n# data types\n\n## storage type\n\ndeclare buffer data layout in descriptor binding\n\n```c\nlayout (binding = 0) buffer top_blob { sfpvec4 top_blob_data[]; };\n```\n\n|storage type|fp32|fp16p|fp16s|bf16p|bf16s|\n|---|---|---|---|---|---|\n|sfp|float|uint|float16_t|uint|bfloat16_t|\n|sfpvec2|vec2|uint|f16vec2|uint|bf16vec2|\n|sfpvec4|vec4|uvec2|f16vec4|uvec2|bf16vec4|\n\n## arithmetic type\n\ndeclare local variable in glsl code\n\n```c\nvoid main()\n{\n afpvec4 v = a * b;\n}\n```\n\n|arithmetic type|fp32|fp16a|\n|---|---|---|\n|afp|float|float16_t|\n|afpvec2|vec2|f16vec2|\n|afpvec4|vec4|f16vec4|\n\n## local type\n\ndeclare variable in shared local memory\n\n```c\nshared lfp tmp_a[8][4][2];\n```\n\n|local type|fp32|fp16p / fp16s only|fp16s+fp16a|fp16s+fp16u|bf16p|bf16s|\n|---|---|---|---|---|---|---|\n|lfp|float|float|float|float16_t|float|bfloat16_t|\n|lfpvec4|vec4|uvec2|uint64_t|f16vec4|uvec2|bf16vec4|\n\n# buffer functions\n\n- load typed value from src[offset]\n\n```c\nafp buffer_ld1(sfp src, int offset);\nafpvec2 buffer_ld2(sfpvec2 src, int offset);\nafpvec4 buffer_ld4(sfpvec4 src, int offset);\n```\n\n- store typed value to dst[offset]\n\n```c\nvoid buffer_st1(sfp dst, int offset, afp v);\nvoid buffer_st2(sfpvec2 dst, int offset, afpvec2 v);\nvoid buffer_st4(sfpvec4 dst, int offset, afpvec4 v);\n```\n\n- copy typed value from src[src_offset] to dst[dst_offset]\n\n```c\nvoid buffer_cp1(sfp dst, int dst_offset, sfp src, int src_offset);\nvoid buffer_cp2(sfpvec2 dst, int dst_offset, sfpvec2 src, int src_offset);\nvoid buffer_cp4(sfpvec4 dst, int dst_offset, sfpvec4 src, int src_offset);\n```\n\n- copy and pack value from src[src_offsets[0],src_offsets[1],...] to dst[dst_offset]\n\n```c\nvoid buffer_cp1to4(sfpvec4 dst, int dst_offset, sfp src, ivec4 src_offsets);\n```\n\n- copy and unpack value from src[src_offset] to dst[dst_offsets[0],dst_offsets[1],...]\n\n```c\nvoid buffer_cp4to1(sfp dst, ivec4 dst_offsets, sfpvec4 src, int src_offset);\n```\n# local data conversion functions\n\n- storage buffer to local memory\n\n```c\nlfp buffer_sm1(sfp src, int offset);\nlfpvec4 buffer_sm4(sfpvec4 src, int offset);\n```\n\n- local memory to local variable\n\n```c\nafp lfp2afp(lfp v);\nafpvec4 lfp2afpvec4(lfpvec4 v);\n```\n\n- local variable to local memory\n\n```c\nlfp afp2lfp(afp v);\nlfpvec4 afp2lfpvec4(afpvec4 v);\n```\n\nNote: The common usage of local memory is to read from global memory first, store it in local memory, and then read local variables from local memory for subsequent use. Therefore, only storage type to local type and local type to arithmetic type conversion functions are provided here.\n\n# misc functions\n\n- prefer specialization constant over push constant\n\n```c\nT psc(T x)\n```\n\nDeclare the same variable in specialization constant AND push constant section, then `psc(x)` will become a compile-time constant when specialization constant given non-zero or be dynamic via push constant otherwise. This is often used for tensor shape specialization. We can usually resolve all shape information and make them be compile-time constants for more aggressive shader optimization.\n\n```c\nlayout (constant_id = 0) const int size = 0;\n\nlayout (push_constant) uniform parameter\n{\n int size;\n} p;\n\nvoid main()\n{\n const int s = psc(size);\n}\n```\n\n# platform macros\n\njudge if the current platform is moltenvk, for enabling some platform-specific workaround\n\n```c\n#if NCNN_moltenvk\n// enable workaround for moltenvk\n#endif\n```\n\nncnn adds additional macro definitions in the new version, which may conflict or confuse the existing glsl code. In order to obtain cross-version compatibility of ncnn, you can switch between the old and new codes according to the `ncnn_glsl_version` macro version.\n\n```c\n#if ncnn_glsl_version >= 1\n// use device macros introduced since version 1\n#endif\n```\n\nncnn additionally defines most of the vulkan device-related features as macros, which we can use to distinguish different platforms, device extensions, features, and properties\n\n### extension macros\n\nWhen the device supports an extension, `ncnn_` is defined as the extension version\n\n```c\nvoid main()\n{\n#if ncnn_VK_KHR_16bit_storage\n // here is the code for any device that supports VK_KHR_16bit_storage\n#endif\n\n#if ncnn_VK_KHR_sampler_ycbcr_conversion >= 10\n // here is the code for any device that supports VK_KHR_sampler_ycbcr_conversion and version >= 10\n#endif\n}\n```\n\n### device feature and property macros\n\nncnn will query device features and properties and then define them as macros.\n\nThe macro name is `ncnn_` or `ncnn_`\n\nThe `GL_EXT_shader_explicit_arithmetic_types_int64` extension will be automatically enabled without explicit code indication when the device supports `shaderInt64`\n\nThe `GL_EXT_shader_explicit_arithmetic_types_int16` extension will be automatically enabled without explicit code indication when the device supports `shaderInt16`\n\n```c\nvoid main()\n{\n#if ncnn_robustBufferAccess\n // here is the code for any device that supports robustBufferAccess feature\n#endif\n\n#if ncnn_vendorID == 4318\n // here is the vendor specific code, 4318 is nvidia graphics\n#endif\n\n#if ncnn_subgroupSize == 32\n // here is the code path optimized for subgroup_size == 32\n#endif\n\n // use macro definitions\n uint size; // dynamic value from some previous routines\n if (size < ncnn_subgroupSize)\n {\n#if ncnn_supportedOperations & 4\n // subgroup support arithmetic\n#endif\n\n#if ncnn_subgroup_arithmetic\n // shorthand style for checking subgroup arithmetic :P\n#endif\n }\n}\n```\n\n### validation layer macros\n\nncnn will define some additional convenient macros when the vulkan validation layer enabled\n\n* `ncnn_enable_validation_layer`\n* `NCNN_LOGE`\n\ncurrently, you have to modify the `ENABLE_VALIDATION_LAYER` definition at the beginning of `src/gpu.cpp` to `1` to enable these macros.\n\nThe `GL_EXT_debug_printf` extension will be enabled automatically without explicitly specifying it in your code.\n\n```c\nvoid main()\n{\n int gx = int(gl_GlobalInvocationID.x);\n\n#if ncnn_enable_validation_layer\n NCNN_LOGE(\"gx = %d\\n\", gx);\n#endif\n}\n```\n\nAt runtime, `NCNN_LOGE` will print out the value of `gx`\n\n### option macros\n\nenable glsl extension only if user enable some options\n\nThe `GL_EXT_shader_16bit_storage` extension will be automatically enabled without explicit code indication when the device supports 16-bit storage and the user turns on `opt.use_fp16_storage` or `opt.use_bf16_storage`\n\nThe `GL_EXT_shader_explicit_arithmetic_types_float16` extension will be automatically enabled without explicit code indication when the device supports 16-bit arithmetic and the user turns on `opt.use_fp16_arithmetic`\n\nThe `GL_EXT_shader_8bit_storage` extension will be automatically enabled without explicit code indication when the device supports 8-bit storage and the user turns on `opt.use_int8_storage`\n\nThe `GL_EXT_shader_explicit_arithmetic_types_int8` extension will be automatically enabled without explicit code indication when the device supports 8-bit arithmetic and the user turns on `opt.use_int8_arithmetic`\n\nThe `GL_EXT_bfloat16` extension will be automatically enabled without explicit code indication when the device supports bfloat16 storage and the user turns on `opt.use_bf16_storage`\n\n```c\nvoid main()\n{\n#if NCNN_fp16_storage\n // the user enable fp16 storage option and the device has fp16 storage support\n#endif\n\n#if NCNN_fp16_arithmetic\n // the user enable fp16 arithmetic option and the device has fp16 arithmetic support\n#endif\n}\n```\n\n|macro|defined by option|\n|---|---|\n|NCNN_fp16_packed|opt.use_fp16_packed|\n|NCNN_fp16_storage|opt.use_fp16_storage|\n|NCNN_fp16_arithmetic|opt.use_fp16_arithmetic|\n|NCNN_int8_packed|opt.use_int8_packed|\n|NCNN_int8_storage|opt.use_int8_storage|\n|NCNN_int8_arithmetic|opt.use_int8_arithmetic|\n|NCNN_bf16_packed|opt.use_bf16_packed|\n|NCNN_bf16_storage|opt.use_bf16_storage|\n|NCNN_shader_local_memory|opt.use_shader_local_memory|\n" }, { "path": "docs/developer-guide/glsl-extension.zh.md", "content": "# ncnn GLSL 扩展\n\n## 理由\n不同的 GPU 支持不同的功能，有的支持 fp16 作为缓冲存储类型，有的支持 fp16 作为操作数变量，有的老 GPU 只支持 fp32。\n\n当 GPU 支持 `VK_KHR_16bit_storage` 扩展时，为了尽量减少 GPU 的内存带宽消耗，我们会优先使用 fp16 作为存储类型。否则，我们使用 `packHalf2x16` 和 `unpackHalf2x16` 在 GLSL 4.2 中将 2 个 fp32 压缩为 uint，从而减少读写带宽。\n\n同样，当 GPU 支持 `VK_KHR_shader_float16_int8` 扩展时，为了加快计算效率，我们会优先使用 fp16 作为运算操作数，这通常会使速度翻倍。否则，我们使用 fp32。\n\n为了确保最广泛的兼容性，将编写以下用于声明描述符绑定和加载数据的代码\n\n```c\n#if NCNN_fp16_storage // GPU支持 16bit storage\nlayout (binding = 0) buffer blob { f16vec4 blob_data[]; };\n#elif NCNN_fp16_packed // GPU支持 GLSL 4.2\nlayout (binding = 0) buffer blob { uvec2 blob_data[]; };\n#else // GPU仅支持 fp32\nlayout (binding = 0) buffer blob { vec4 blob_data[]; };\n#endif\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n#if NCNN_fp16_storage && NCNN_fp16_arithmetic // GPU支持 16bit storage 和 shader float16\n f16vec4 x = blob_data[i];\n#elif NCNN_fp16_storage // GPU支持 16bit storage 但不包含 shader float16\n vec4 x = vec4(blob_data[i]);\n#elif NCNN_fp16_packed && NCNN_fp16_arithmetic // GPU支持 GLSL 4.2 和 shader float16\n f16vec4 x = f16vec4(unpackFloat2x16(blob_data[i].x), unpackFloat2x16(blob_data[i].y));\n#elif NCNN_fp16_packed // GPU支持 GLSL 4.2\n vec4 x = vec4(unpackHalf2x16(blob_data[i].x), unpackHalf2x16(blob_data[i].y));\n#else // GPU仅支持 fp32\n vec4 x = blob_data[i];\n#endif\n}\n```\n\n如您所见，仅声明缓冲区类型并读取值会消耗大量代码行，这是项目维护的噩梦。因此，ncnn 增加了更灵活的数据类型和辅助函数，以减小代码的大小并提高可读性，并且会根据 GPU 支持的功能级别自动扩展到最高效的实现。\n\n上面的代码，通过使用 ncnn GLSL 扩展，可以简化为\n\n```c\nlayout (binding = 0) buffer blob { sfpvec4 blob_data[]; };\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n afpvec4 x = buffer_ld4(blob_data, i);\n}\n```\n\nncnn GLSL 扩展为存储、计算、共享内存以及缓冲区和图像的加载、存储、转换函数提供了必要的数据类型。我们还提供了一些缓冲区和图像复制函数，以防止在使用 fp16 作为中间数据类型时丢失精度，并避免不必要的 `unpackHalf2x16` 和 `packHalf2x16` 配对。\n\n# 编译GLSL的入口点\n\nncnn库中的 gpu.h 头文件公开了3个用于将 GLSL 代码编译为 Spir-V 二进制的API函数，它们支持 ncnn GLSL 扩展，这3个函数接受 opt switch 来控制 ncnn GLSL 扩展形式。前两个函数接受原始 GLSL 代码字符串作为参数，最后一个函数用于创建 ncnn 的已存在的内置着色器。\n\n```cpp\nnamespace ncnn {\n\n// 在线 Spir-V 编译器\nNCNN_EXPORT int compile_spirv_module(const char* comp_string, const Option& opt, std::vector& spirv);\nNCNN_EXPORT int compile_spirv_module(const char* comp_data, int comp_data_size, const Option& opt, std::vector& spirv);\nNCNN_EXPORT int compile_spirv_module(int shader_type_index, const Option& opt, std::vector& spirv);\n\n} // namespace ncnn\n```\n\n## 直接编译ncnn扩展GLSL代码\n\n您可以使用 ncnn GLSL 扩展编写着色器代码，使用 ncnn 函数编译为 Spir-V。编译后的产品是符合标准的 Spir-V 二进制文件，可以直接用于在 Vulkan API 中创建流水线对象\n\n```cpp\nstatic const char my_glsl_data[] = R\"(\n#version 450\n\nlayout (binding = 0) readonly buffer a_blob { sfpvec4 a_blob_data[]; };\nlayout (binding = 1) writeonly buffer b_blob { sfpvec4 b_blob_data[]; };\n\nvoid main()\n{\n const int i = int(gl_GlobalInvocationID.x);\n\n afpvec4 v = buffer_ld4(a_blob_data, i);\n\n v = v + 123;\n\n buffer_st4(b_blob_data, i, v);\n}\n)\";\n\nOption opt;\n // 您可以控制Vulkan扩展行为\n // 当GPU支持16位存储的话\nopt.use_fp16_storage = false;\n\nstd::vector spirv;\nncnn::compile_spirv_module(my_glsl_data, sizeof(my_glsl_data) - 1, opt, spirv);\n\n// 稍后再创建管道对象\n// ncnn::Pipeline pipeline(vkdev);\n// pipeline.set_local_size_xyz(64, 1, 1);\n// pipeline.create(spirv.data(), spirv.size() * 4, specializations);\n```\n\n## ncnn内置着色器\n\nncnn内部的着色器索引在标头中公开，如果需要可以使用 `layer_shader_type.h`\n\n```cpp\n#include \"layer_shader_type.h\"\n\nint shader_type_index = LayerShaderType::convert_ycbcr;\n\nOption opt;\n\nstd::vector spirv;\nint retc = compile_spirv_module(shader_type_index, opt, spirv);\n```\n\n# 数据类型\n\n## 存储类型(storage type)\n\n在描述符绑定中声明缓冲区数据布局\n\n```c\nlayout (binding = 0) buffer top_blob { sfpvec4 top_blob_data[]; };\n```\n\n|存储类型|fp32|fp16p|fp16s|bf16p|bf16s|\n|---|---|---|---|---|---|\n|sfp|float|uint|float16_t|uint|bfloat16_t|\n|sfpvec2|vec2|uint|f16vec2|uint|bf16vec2|\n|sfpvec4|vec4|uvec2|f16vec4|uvec2|bf16vec4|\n\n## 算术类型(arithmetic type)\n\n在 GLSL 代码中声明局部变量\n\n```c\nvoid main()\n{\n afpvec4 v = a * b;\n}\n```\n\n|算术类型|fp32|fp16a|\n|---|---|---|\n|afp|float|float16_t|\n|afpvec2|vec2|f16vec2|\n|afpvec4|vec4|f16vec4|\n\n## 本地类型(local type)\n\n在共享本地内存中声明变量\n\n```c\nshared lfp tmp_a[8][4][2];\n```\n\n|本地类型|fp32|fp16p / fp16s only|fp16s+fp16a|fp16s+fp16u|bf16p|bf16s|\n|---|---|---|---|---|---|---|\n|lfp|float|float|float|float16_t|float|bfloat16_t|\n|lfpvec4|vec4|uvec2|uint64_t|f16vec4|uvec2|bf16vec4|\n\n# 缓冲区函数(buffer functions)\n\n- 从 src[offset] 加载已经确定类型的值\n\n```c\nafp buffer_ld1(sfp src, int offset);\nafpvec2 buffer_ld2(sfpvec2 src, int offset);\nafpvec4 buffer_ld4(sfpvec4 src, int offset);\n```\n\n- 将已确定类型的值存储到 dst[偏移量]\n\n```c\nvoid buffer_st1(sfp dst, int offset, afp v);\nvoid buffer_st2(sfpvec2 dst, int offset, afpvec2 v);\nvoid buffer_st4(sfpvec4 dst, int offset, afpvec4 v);\n```\n\n- 从已确定类型 src[src_offset] 的值拷贝到 dst[dst_offset]\n\n```c\nvoid buffer_cp1(sfp dst, int dst_offset, sfp src, int src_offset);\nvoid buffer_cp2(sfpvec2 dst, int dst_offset, sfpvec2 src, int src_offset);\nvoid buffer_cp4(sfpvec4 dst, int dst_offset, sfpvec4 src, int src_offset);\n```\n\n- 从 src[src_offsets[0],src_offsets[1],...] 的值拷贝并打包到 dst[dst_offset]\n\n```c\nvoid buffer_cp1to4(sfpvec4 dst, int dst_offset, sfp src, ivec4 src_offsets);\n```\n\n- 从 src[src_offset] 的值拷贝并解包到 dst[dst_offsets[0],dst_offsets[1],...]\n\n```c\nvoid buffer_cp4to1(sfp dst, ivec4 dst_offsets, sfpvec4 src, int src_offset);\n```\n\n# 本地数据转换函数\n\n- 存储缓冲区转换到本地内存\n\n```c\nlfp buffer_sm1(sfp src, int offset);\nlfpvec4 buffer_sm4(sfpvec4 src, int offset);\n```\n\n- 本地内存转换到局部变量\n\n```c\nafp lfp2afp(lfp v);\nafpvec4 lfp2afpvec4(lfpvec4 v);\n```\n\n- 局部变量转换到本地内存\n\n```c\nlfp afp2lfp(afp v);\nlfpvec4 afp2lfpvec4(afpvec4 v);\n```\n\n注意：本地内存的常见用法是先从全局内存中读取，存储在本地内存中，然后再从本地内存中读取局部变量以供后续使用。因此，此处仅提供存储类型到本地类型和本地类型到算术类型的转换函数。\n\n# 杂项函数\n\n- 更推荐使用专业化常量(specialization constants)，而不是推动常量(push constants)\n\n```c\nT psc(T x)\n```\n\n在 `专用常量` 和 `推送常量` 部分中声明相同的变量，然后在专用常量给定非零时 `psc(x)` 将成为编译时常量，否则将通过推送常量动态。这通常用于张量形状特化。我们通常可以解析所有形状信息，并使它们成为编译时常量，以实现让着色器得到更积极的优化。\n\n```c\nlayout (constant_id = 0) const int size = 0;\n\nlayout (push_constant) uniform parameter\n{\n int size;\n} p;\n\nvoid main()\n{\n const int s = psc(size);\n}\n```\n\n# 平台宏定义\n\n判断当前平台是否为 moltenvk，以启用对于某些特定于平台的解决方法\n\n```c\n#if NCNN_moltenvk\n// 启用moltenvk的解决方法\n#endif\n```\n\nncnn 在新版本中添加了额外的宏定义，可能与现在的 glsl 代码冲突或引起混淆。为了实现 ncnn 的跨版本兼容性，可以根据 `ncnn_glsl_version` 宏的版本号在新旧代码之间进行切换 。\n\n```c\n#if ncnn_glsl_version >= 1\n// 使用自版本 1 起引入的设备宏\n#endif\n```\n\nncnn 额外定义了大多数 vulcan 设备相关功能作为宏，我们可以用来区分不同的平台、设备扩展、功能和属性。\n\n### 扩展宏定义\n\n当设备支持某个扩展时，`ncnn_` 被定义为扩展版本\n\n```c\nvoid main()\n{\n#if ncnn_VK_KHR_16bit_storage\n // 支持 VK_KHR_16bit_storage 设备的代码\n#endif\n\n#if ncnn_VK_KHR_sampler_ycbcr_conversion >= 10\n // 支持 VK_KHR_sampler_ycbcr_conversion 且版本 >=10 的代码\n#endif\n}\n```\n\n### 设备特性和属性宏\n\nncnn 会查询设备特性和属性，然后将它们定义为宏。\n\n宏名称为 `ncnn_` 或 `ncnn_`\n\n当设备支持 `shaderInt64` 时，`GL_EXT_shader_explicit_arithmetic_types_int64` 扩展会自动启用，无需显式代码指示。\n\n当设备支持 `shaderInt16` 时，`GL_EXT_shader_explicit_arithmetic_types_int16` 扩展会自动启用，无需显式代码指示。\n\n```c\nvoid main()\n{\n#if ncnn_robustBufferAccess\n // 支持 robustBufferAccess 特性的设备代码\n#endif\n\n#if ncnn_vendorID == 4318\n // 供应商特定代码，4318 是 nvidia 显卡\n#endif\n\n#if ncnn_subgroupSize == 32\n // 为 subgroup_size == 32 优化的代码路径\n#endif\n\n // 使用宏定义\n uint size; // 来自先前例程的动态值\n if (size < ncnn_subgroupSize)\n {\n#if ncnn_supportedOperations & 4\n // subgroup 支持算术运算\n#endif\n\n#if ncnn_subgroup_arithmetic\n // 检查 subgroup 算术运算的简写形式\n#endif\n }\n}\n```\n\n### 验证层宏定义\n\n当启用 vulkan 验证层时，ncnn 会定义一些额外的便捷宏\n\n* `ncnn_enable_validation_layer`\n* `NCNN_LOGE`\n\n目前，你必须将 `src/gpu.cpp` 开头的 `ENABLE_VALIDATION_LAYER` 定义修改为 `1` 才能启用这些宏。\n\n`GL_EXT_debug_printf` 扩展会自动启用，无需在代码中显式指定。\n\n```c\nvoid main()\n{\n int gx = int(gl_GlobalInvocationID.x);\n\n#if ncnn_enable_validation_layer\n NCNN_LOGE(\"gx = %d\\n\", gx);\n#endif\n}\n```\n\n在运行时，`NCNN_LOGE` 将打印出 `gx` 的值\n\n### 选项宏\n\n仅当用户启用某些选项时才启用 GLSL 扩展\n\n`GL_EXT_shader_16bit_storage` 扩展会在设备支持 16 位存储且用户开启了 `opt.use_fp16_storage` 或 `opt.use_bf16_storage` 选项时，自动启用，无需显式代码指示。\n\n`GL_EXT_shader_explicit_arithmetic_types_float16` 扩展会在设备支持 16 位算术运算且用户开启了 `opt.use_fp16_arithmetic` 选项时，自动启用，无需显式代码指示。\n\n`GL_EXT_shader_8bit_storage` 扩展会在设备支持 8 位存储且用户开启了 `opt.use_int8_storage` 选项时，自动启用，无需显式代码指示。\n\n`GL_EXT_shader_explicit_arithmetic_types_int8` 扩展会在设备支持 8 位算术运算且用户开启了 `opt.use_int8_arithmetic` 选项时，自动启用，无需显式代码指示。\n\n`GL_EXT_bfloat16` 扩展会在设备支持 bfloat16 存储且用户开启了 `opt.use_bf16_storage` 选项时，自动启用，无需显式代码指示。\n\n```c\nvoid main()\n{\n#if NCNN_fp16_storage\n // 用户启用 fp16 存储选项，且设备支持 fp16 存储\n#endif\n\n#if NCNN_fp16_arithmetic\n // 用户启用 fp16 算术选项，且设备支持 fp16 算术运算\n#endif\n}\n```\n\n|宏定义|option中所定义的变量|\n|---|---|\n|NCNN_fp16_packed|opt.use_fp16_packed|\n|NCNN_fp16_storage|opt.use_fp16_storage|\n|NCNN_fp16_arithmetic|opt.use_fp16_arithmetic|\n|NCNN_int8_packed|opt.use_int8_packed|\n|NCNN_int8_storage|opt.use_int8_storage|\n|NCNN_int8_arithmetic|opt.use_int8_arithmetic|\n|NCNN_bf16_packed|opt.use_bf16_packed|\n|NCNN_bf16_storage|opt.use_bf16_storage|\n|NCNN_shader_local_memory|opt.use_shader_local_memory|\n" }, { "path": "docs/developer-guide/how-to-be-a-contributor.zh.md", "content": "### 如何提交代码\n\n#### 一、fork 分支\n在浏览器中打开 [ncnn](https://github.com/tencent/ncnn), `fork` 到自己的 repositories，例如\n```\nhttps://github.com/user/ncnn\n```\n\nclone 项目到本地，添加官方 remote 并 fetch:\n```\n$ git clone https://github.com/user/ncnn && cd ncnn\n$ git remote add tencent https://github.com/tencent/ncnn\n$ git fetch tencent\n```\n对于 `git clone` 下来的项目，它现在有两个 remote，分别是 origin 和 tencent：\n\n```\n$ git remote -v\norigin https://github.com/user/ncnn (fetch)\norigin https://github.com/user/ncnn (push)\ntencent https://github.com/Tencent/ncnn (fetch)\ntencent https://github.com/Tencent/ncnn (push)\n```\norigin 指向你 fork 的仓库地址；remote 即官方 repo。可以基于不同的 remote 创建和提交分支。\n\n例如切换到官方 master 分支，并基于此创建自己的分支（命名尽量言简意赅。一个分支只做一件事，方便 review 和 revert）\n```\n$ git checkout tencent/master\n$ git checkout -b add-conv-int8\n```\n\n或创建分支时指定基于官方 master 分支：\n```\n$ git checkout -b fix-typo-in-document tencent/master\n```\n\n> `git fetch` 是从远程获取最新代码到本地。如果是第二次 pr ncnn，直接从 `git fetch tencent` 开始即可，不需要 `git remote add tencent`，也不需要修改 `github.com/user/ncnn`。\n\n#### 二、代码习惯\n为了增加沟通效率，reviewer 一般要求 contributor 遵从以下规则\n\n* `if-else`和花括号`{`中间需要换行\n* 不能随意增删空行\n* tab 替换为 4 个空格\n* 为了保证平台兼容性，目前不使用`c++11`，`src`目录下尽量避免使用`template`\n* 若是新增功能或平台，`test`目录需有对应测试用例\n* 文档放到`doc`对应目录下，中文用`.zh.md`做后缀；英文直接用`.md`后缀\n\n开发完成后提交到自己的 repository\n```\n$ git commit -a\n$ git push origin add-conv-int8\n```\n推荐使用 [`commitizen`](https://pypi.org/project/commitizen/) 或 [`gitlint`](https://jorisroovers.com/gitlint/) 等工具格式化 commit message，方便事后检索海量提交记录\n\n#### 三、代码提交\n浏览器中打开 [ncnn pulls](https://github.com/Tencent/ncnn/pulls) ，此时应有此分支 pr 提示，点击 `Compare & pull request`\n\n* 标题**必须**是英文。未完成的分支应以 `WIP:` 开头，例如 `WIP: add conv int8`\n* 正文宜包含以下内容，中英不限\n * 内容概述和实现方式\n * 功能或性能测试\n * 测试结果\n\nCI 已集成了自动格式化，restyled-io 会在 pr 的同时生成 `Restyled add conv int8`，需要 merge 自动 restyled 的分支，例如\n```\n$ git fetch tencent\n$ git checkout add-conv-int8\n$ git merge tencent/restyled/pull-2078\n$ git push origin add-conv-int8\n```\n回到浏览器签署 CLA，所有 CI 测试通过后通知 reviewer merge 此分支。\n\n#### 四、彩蛋\n留下个人 qq 号会触发隐藏事件。" }, { "path": "docs/developer-guide/how-to-implement-custom-layer-step-by-step.md", "content": "# step1 create a new empty class\n```cpp\n// mylayer.h\n#include \"layer.h\"\nusing namespace ncnn;\n\n// a new layer type called MyLayer\nclass MyLayer : public Layer\n{\n};\n\n// mylayer.cpp\n#include \"mylayer.h\"\nDEFINE_LAYER_CREATOR(MyLayer)\n```\n\n# step2 declare layer parameters and weights\n```cpp\n// mylayer.h\n#include \"layer.h\"\nusing namespace ncnn;\n\nclass MyLayer : public Layer\n{\nprivate:\n int channels;// new code\n float gamma;// new code\n Mat weight;// new code\n};\n\n// mylayer.cpp\n#include \"mylayer.h\"\nDEFINE_LAYER_CREATOR(MyLayer)\n```\n\n# step3 implement load functions for parameters and weights\n```cpp\n// mylayer.h\n#include \"layer.h\"\nusing namespace ncnn;\n\nclass MyLayer : public Layer\n{\npublic:\n virtual int load_param(const ParamDict& pd);// new code\n virtual int load_model(const ModelBin& mb);// new code\n\nprivate:\n int channels;\n float eps;\n Mat gamma_data;\n};\n\n// mylayer.cpp\n#include \"mylayer.h\"\nDEFINE_LAYER_CREATOR(MyLayer)\n\n// new routine for loading parameters\nint MyLayer::load_param(const ParamDict& pd)\n{\n // details about the relations with param file\n // https://github.com/Tencent/ncnn/wiki/param-and-model-file-structure\n //\n channels = pd.get(0, 0);// parse 0= entry, default value 0\n eps = pd.get(1, 0.001f);// parse 1= entry, default value 0.001f\n\n return 0;// return zero if success\n}\n\n// new routine for loading weights\nint MyLayer::load_model(const ModelBin& mb)\n{\n // details about the relations with model file\n // https://github.com/Tencent/ncnn/wiki/param-and-model-file-structure\n //\n // read weights with length of channels * sizeof(float)\n // the second argument explains as follows\n // 0 judge the value type automatically, you may get float or float16 or uint8 etc\n // depends on the model storage and the supporting target hardware\n // 1 read float values anyway\n // 2 read float16 values anyway\n // 3 read uint8 values anyway\n gamma_data = mb.load(channels, 1);\n if (gamma_data.empty())\n return -100;// return non-zero on error, -100 indicates out-of-memory\n\n return 0;// return zero if success\n}\n```\n\n# step4 determine forward behavior\n```cpp\n// mylayer.h\n#include \"layer.h\"\nusing namespace ncnn;\n\nclass MyLayer : public Layer\n{\npublic:\n MyLayer();// new code\n virtual int load_param(const ParamDict& pd);\n virtual int load_model(const ModelBin& mb);\n\nprivate:\n int channels;\n float eps;\n Mat gamma_data;\n};\n\n// mylayer.cpp\n#include \"mylayer.h\"\nDEFINE_LAYER_CREATOR(MyLayer)\n\n// new routine for setting forward behavior\nMyLayer::MyLayer()\n{\n // one input and one output\n // typical one_blob_only type: Convolution, Pooling, ReLU, Softmax ...\n // typical non-one_blob_only type: Eltwise, Split, Concat, Slice ...\n one_blob_only = true;\n\n // do not change the blob size, modify data in-place\n // typical support_inplace type: ReLU, Sigmoid ...\n // typical non-support_inplace type: Convolution, Pooling ...\n support_inplace = true;\n}\n\nint MyLayer::load_param(const ParamDict& pd)\n{\n channels = pd.get(0, 0);\n eps = pd.get(1, 0.001f);\n\n // you could alter the behavior based on loaded parameter\n // if (eps == 0.001f)\n // {\n // one_blob_only = false;\n // support_inplace = false;\n // }\n\n return 0;\n}\n\nint MyLayer::load_model(const ModelBin& mb)\n{\n gamma_data = mb.load(channels, 1);\n if (gamma_data.empty())\n return -100;\n\n // you could alter the behavior based on loaded weight\n // if (gamma_data[0] == 0.f)\n // {\n // one_blob_only = false;\n // support_inplace = false;\n // }\n\n return 0;\n}\n```\n\n# step5 choose proper interface based on forward behavior\n```cpp\n// The base class Layer defines four interfaces for each forward behavior combination\n\n// 1\nvirtual int forward(const std::vector& bottom_blobs, std::vector& top_blobs, const Option& opt) const;\n\n// 2\nvirtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;\n\n// 3\nvirtual int forward_inplace(std::vector& bottom_top_blobs, const Option& opt) const;\n\n// 4\nvirtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;\n```\n**must** = layer must implement this function\n\n**optional** = layer may implement this function for optimal performance\n\nsometimes the graph inference path cannot call forward_inplace directly due to data sharing, in this situation the non-inplace forward routine will be used, which deep-copy the input blob and call inplace forward on it if the optional routine is not implemented. Thus, you could avoid this deep-copy by process input to output on-the-fly.\n\n|one_blob_only|support_inplace|1|2|3|4|\n|---|---|---|---|---|---|\n|false|false|must| | | |\n|false|true|optional| |must| |\n|true|false| |must| | |\n|true|true| |optional| |must|\n\n# step6 implement forward function\n```cpp\n// mylayer.h\n#include \"layer.h\"\nusing namespace ncnn;\n\nclass MyLayer : public Layer\n{\npublic:\n MyLayer();\n virtual int load_param(const ParamDict& pd);\n virtual int load_model(const ModelBin& mb);\n\n virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;// new code, optional\n virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;// new code\n\nprivate:\n int channels;\n float eps;\n Mat gamma_data;\n};\n\n// mylayer.cpp\n#include \"mylayer.h\"\nDEFINE_LAYER_CREATOR(MyLayer)\n\nMyLayer::MyLayer()\n{\n one_blob_only = true;\n support_inplace = true;\n}\n\nint MyLayer::load_param(const ParamDict& pd)\n{\n channels = pd.get(0, 0);\n eps = pd.get(1, 0.001f);\n\n return 0;\n}\n\nint MyLayer::load_model(const ModelBin& mb)\n{\n gamma_data = mb.load(channels, 1);\n if (gamma_data.empty())\n return -100;\n\n return 0;\n}\n\n// optional new routine for layer forward function, non-inplace version\nint MyLayer::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const\n{\n // check input dims, return non-zero on error\n if (bottom_blob.c != channels)\n return -1;\n\n // x = (x + eps) * gamma_per_channel\n\n int w = bottom_blob.w;\n int h = bottom_blob.h;\n size_t elemsize = bottom_blob.elemsize;\n int size = w * h;\n\n top_blob.create(w, h, channels, elemsize, opt.blob_allocator);\n if (top_blob.empty())\n return -100;// return non-zero on error, -100 indicates out-of-memory\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int q=0; q使用SSE优化之前，确保理解代码指针位置和移动原理，原生代码已经完成测试，输出结果正确。\n\n### 1.SSE数据类型\n\n​\tSSE数据类型形如：\n\n```c++\n__m\t\t\t //__m适用代表申请mm寄存器\n \t\t\t\t\t// bit 代表数据类型的字节长度，在SSE中为128 或 64\n \t\t\t\t\t// 默认type为单精度浮点（f32），其余为int 或double\n// 另外要注意所有SSE的类型除__m128和__m64外，随着版本更新有不同的类型，建议根据需要且确定硬件性能后选择合适的类型\n// 举例如下：\n__m128\t\t\t\t\t//4xf32 含有4个单精度浮点数；SSE\n__m64 \t\t\t\t//4xf32 含有2个单精度浮点数；SSE\n__m128i \t\t\t\t//8个int类型（8x16)\t\t ；SSE3\n__m128d\t\t\t\t\t//2个double类型(2x64)\n```\n\n### 2.SSE内联函数结构\n\n​\tSSE内联函数在线查询：[Intel® Intrinsics Guide](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#techs=SSE,SSE2,SSE3,SSSE3,SSE4_1,SSE4_2) 在此 单个指令的结构如下：\n\n- Synopsis ：摘要。描述指令的接口定义，需要引入的头文件，对应的指令，CPU必须支持的标志；\n- Description：描述该指令的行为；\n- Operation：逻辑层面描述指令行为；\n- Performance：在不同架构中所需要的延迟和执行所需要的时钟周期数（CPI）。\n\n​\t值得指出的的是此处默认使用小端存储，即左边为高位，右边为低位。\n\n​\t相似的内联函数有很多，在使用时候一定要注意Operation中的逻辑满足您的要求。\n\n​\t另外，在ncnn中，ncnn已经将部分SSE内联函数以NCNN内联的方式封装。在为NCNN添加SSE优化的算法的过程中，请首先考虑搜索“NCNNINLINE”宏封装的SSE函数。\n\n## 四：样例\n\n### 1.一个简单的样例：4x4矩阵乘法\n\n​\t矩阵乘法方面，已经有很多出色的成果。值得一读的比如[how-to-optimize-gemm](https://github.com/flame/how-to-optimize-gemm)，及 [以Arm Intrinsic优化矩阵乘法](https://github.com/tpoisonooo/how-to-optimize-gemm)。我建议感兴趣同学参考和学习这两份项目，来探究如何从0到1优化一份算法；\n\n​\t矩阵乘法原理很简单：\n\n​\t假设有A，B两个矩阵，如下：\n$$\nA_{[4][4]} = \n\\begin{bmatrix}\n\ta_0 & a_4 & a_8 & a_{12} \\\\\n\ta_1 & a_5 & a_9 & a_{13} \\\\\n\ta_2 & a_6 & a_{10} & a_{14} \\\\\n\ta_3 & a_7 & a_{11} & a_{15} \n\\end{bmatrix}\n~~\nB_{[4][4]} = \n\\begin{bmatrix}\n\tb_0 & b_4 & b_8 & b_{12} \\\\\n\tb_1 & b_5 & b_9 & b_{13} \\\\\n\tb_2 & b_6 & b_{10} & b_{14} \\\\\n\tb_3 & b_7 & b_{11} & b_{15} \n\\end{bmatrix}\n~~\nC_{[4][4]} = \n\\begin{bmatrix}\n\tc_0 & c_4 & c_8 & c_{12} \\\\\n\tc_1 & c_5 & c_9 & c_{13} \\\\\n\tc_2 & c_6 & c_{10} & c_{14} \\\\\n\tc_3 & c_7 & c_{11} & c_{15} \n\\end{bmatrix}\n$$\n​\t对于C 矩阵的第一列，我们有：\n$$\nc_0 = a_0b_0 + a_4b_1 + a_8b_2 + a_{12}b_3 \\\\\n\tc_1 = a_1b_0 + a_5b_1 + a_9b_2 + a_{13}b_3 \\\\\n\tc_2 = a_2b_0 + a_6b_1 + a_{10}b_2 + a_{14}b_3 \\\\\n\tc_3 = a_3b_0 + a_7b_1 + a_{11}b_2 + a_{15}b_3\n$$\n\n\n#### 1.编写测试代码和基准测量程序\n\n​\t在该样例中，测试代码很容易编写出来，我们只需要初始化4x4的二维数组，并返回指针即可。此时，可以不考虑泛用性，初始化为固定值即可。\n\n```c\n// <代码片段>\n...\nfloat A[16] = {0.0f};\t\t\t// 此处已经将输入和输出的矩阵默认展开成im2col 后的单行（inch = 1） 宽度为h*w = 16的矩阵\nfloat B[16] = {0.0f};\nfloat C[16] = {0.0f};\nmatrix_init_rand(A, 4, 4);\t\t// 随机初始化A数组\nmatrix_init_rand(B, 4, 4);\t\t// 随机初始化B数组\n```\n\n​\t编写验证正确性的测试代码。\n\n```c\n// <代码片段>\n...\nfloat T[16] = {...};\t\t\t// Target即为预测的C的结果数组，可用numpy或者纸笔计算\n...\nfloat error = 0.0001;\nbool CheckAuc(T, C, error);\t\t\n// 注意：float在计算机中不能完全表示，只能使用绝对误差的判别方法。gtest等测试框架的EXCEPT宏无法处理1.234e5这样结构的float数的对比。\n```\n\n​\t同样，编写计算耗时的基准测量代码，此处使用1000次操作所占的平均时间来作为基准。\n\n```c\n// <代码片段>\n...\nconst int loop = 1000;\nclock_gettime_(CLOCK_REALTIME, &time_start);\nfor(init i = 0; i < loop; i++)\n{\n\tmatirx_mult_native(C, A, B);\n}\nclock_gettime_(CLOCK_REALTIME, &time_end);\nclocks_c = (time_end.tv_sec - time_start.tv_sec) * 1000000 + (time_end.tv_sec - time_start.tv_sec) /1000;\n```\n\n#### 2.编写原生代码\n\n​\t编写原生代码，使得正确性测试能够通过。\n\n```c\n// <代码片段>\nstatic void matirx_mult_native(float *C, float *A, float *B)\n{\n for(int i_idx = 0; i_idx < 4; i_idx++)\n {\n for(int j_idx = 0; j_idx < 4; j_idx++)\n {\n for(int k_idx = 0; k_idx < 4; k_idx++)\n {\n C[4*j_idx + i_idx] += A[4*k_idx + i_idx] * B[4*j_idx + k_idx];\n }\n }\n }\n}\n```\n\n#### 3.优化原生代码\n\n​\t注意到上述代码中，先取c0 - c3 的计算作为样例考虑：\n$$\n\tc_0 = a_0b_0 + a_4b_1 + a_8b_2 + a_{12}b_3 \\\\\n\tc_1 = a_1b_0 + a_5b_1 + a_9b_2 + a_{13}b_3 \\\\\n\tc_2 = a_2b_0 + a_6b_1 + a_{10}b_2 + a_{14}b_3 \\\\\n\tc_3 = a_3b_0 + a_7b_1 + a_{11}b_2 + a_{15}b_3\n$$\n\n##### 1.装载寄存器\n\n- 考虑竖排a0-a1-a2-a3 为4个f32 数据，又因为SSE可以申请mm寄存器，单次保存128bit，那么不妨把a0-a4保存在寄存器中，\n\n- 对于b0-b3 则是，单次读取一个值，能够重复用4次，不妨考虑b0 重复4次，排满单个128bit的mm寄存器；\n\n- 同理把c0-c3也放入寄存器，从列方向上考虑，取名为_c0 \n\n ```c++\n _m128 _a0 = _mm_load_ps(a_ptr);\t\t\t//a0 -a1 -a2 -a3\n _m128 _a1 = _mm_load_ps(a_ptr + 4);\t\t//a4 -a5 -a6 -a7\n _m128 _a2 = _mm_load_ps(a_ptr + 8);\t\t//a8 -a9 -a10-a11\n _m128 _a3 = _mm_load_ps(a_ptr + 12);\t//a12-a13-a14-a15\n \n _m128 _b0 = _mm_load_ps1(b_ptr);\t\t// b0 - b0 - b0 - b0\n _m128 _b1 = _mm_load_ps1(b_ptr + 4);\t// b1 - b1 - b1 - b1\n _m128 _b2 = _mm_load_ps1(b_ptr + 8);\t// b2 - b2 - b2 - b2\n _m128 _b3 = _mm_load_ps1(b_ptr + 12);\t// b3 - b3 - b3 - b3\n ```\n\n##### 2.编写第一列的计算结果\n\n​\t对于_a0 -\\_a3 数据与\\_b0 数据相乘 ，有：\n\n```c++\n// 保存结果新建一个_c0 作为临时变量\n_m128 _c0 = _mm_set_ps1(0.0f);\n_c0 = _mm_mul_ps(_a0, _b0);\n_c0 = _mm_add_ps(_mm_mul_ps(_a1, _b1),_c0);\n_c0 = _mm_add_ps(_mm_mul_ps(_a2, _b2),_c0);\n_c0 = _mm_add_ps(_mm_mul_ps(_a3, _b3),_c0);\n// 把 _sum0存会以c指针开头的内存中，完美！\n_mm_store_ps(c_ptr, _c0);\n```\n\n##### 3.将单列输出扩展到所有列：\n\n​\t我们针对剩下的c中的c1 列也做相同的操作： 对于C1 列 有：\n$$\n\tc_4 = a_0b_4 + a_4b_5 + a_8b_6 + a_{12}b_7 \\\\\n\tc_5 = a_1b_4 + a_5b_5 + a_9b_6 + a_{13}b_7 \\\\\n\tc_6 = a_2b_4 + a_6b_5 + a_{10}b_6 + a_{14}b_7 \\\\\n\tc_7 = a_3b_4 + a_7b_5 + a_{11}b_6 + a_{15}b_7\n$$\n\n\n```c++\n// a 系列不变 b系列指针+1\n_m128 _b4 = _mm_load_ps1(b_ptr + 1);\t\t// b4 - b4 - b4 - b4\n_m128 _b5 = _mm_load_ps1(b_ptr + 4 + 1);\t// b5 - b5 - b5 - b5\n_m128 _b6 = _mm_load_ps1(b_ptr + 8 + 1);\t// b6 - b6 - b6 - b6\n_m128 _b7 = _mm_load_ps1(b_ptr + 12+ 1);\t// b7 - b7 - b7 - b7\n\n// 保存结果新建一个_c0 作为临时变量\n_m128 _c1 = _mm_set_ps1(0.0f);\n_c1 = _mm_mul_ps(_a0, _b4);\n_c1 = _mm_add_ps(_mm_mul_ps(_a1, _b5),_c1);\n_c1 = _mm_add_ps(_mm_mul_ps(_a2, _b6),_c1);\n_c1 = _mm_add_ps(_mm_mul_ps(_a3, _b7),_c1);\n// 把 _sum0存会以c指针开头的内存中，完美！\n_mm_store_ps(c_ptr, _c1);\n```\n\n​\t此时我们发现，对于C1列的操作与C0列及其相似，只不过是b_ptr的指针发生移动，不妨将其放到同一个循环中，有：\n\n```C++\n// a 系列不变\n_m128 _a0 = _mm_load_ps(a_ptr);\t\t\t//a0 -a1 -a2 -a3\n_m128 _a1 = _mm_load_ps(a_ptr + 4);\t\t//a4 -a5 -a6 -a7\n_m128 _a2 = _mm_load_ps(a_ptr + 8);\t\t//a8 -a9 -a10-a11\n_m128 _a3 = _mm_load_ps(a_ptr + 12);\t//a12-a13-a14-a15\n\nfor(int i = 0; i < 4; i++)\n{\n _m128 _b0 = _mm_load_ps1(b_ptr);\t\t// b0 - b0 - b0 - b0\n _m128 _b1 = _mm_load_ps1(b_ptr + 4);\t// b1 - b1 - b1 - b1\n _m128 _b2 = _mm_load_ps1(b_ptr + 8);\t// b2 - b2 - b2 - b2\n _m128 _b3 = _mm_load_ps1(b_ptr + 12);\t// b3 - b3 - b3 - b3\n \n _m128 _ci = _mm_set_ps1(0.0f);\n _ci = _mm_mul_ps(_a0, _b0);\n _ci = _mm_add_ps(_mm_mul_ps(_a1, _b1),_ci);\n _ci = _mm_add_ps(_mm_mul_ps(_a2, _b2),_ci);\n _ci = _mm_add_ps(_mm_mul_ps(_a3, _b3),_ci);\n \n _mm_store_ps(c_ptr, _ci);\n \n b_ptr += 1;\t\t\t\t// 移动b_ptr\n c_ptr += 4;\t\t\t\t// 移动保存内存的c_ptr\n}\n```\n\n### 2.NCNN中以SSE优化算子的注意事项\n\n#### 1.线程与openmp\n\n​\t以上计算Benchmark 和 SSE优化的方法大多集中在单个核心中，但是在实际使用ncnn中，ncnn使用Option opt 中提供的num_threads 给openmp赋值，以实现多线程并行化，同时运行在多个核心上。\n\n```c++\n#pragma omp parallel for num_threads(opt.num_threads)\n```\n\n​\t在优化成SSE代码的初期，可以考虑锁定为单线程，或者直接不用考虑线程的影响，仅对单核以SSE优化，保证单核的结果正确后，再加上opt的多线程进行结果测试。\n\n#### 2.展开循环\n\n​\t在实际ncnn实现的原生代码的算法中，循环是非常常见的。针对以SSE优化这类循环，遵循非常简单的原则：循环中，迭代器等于零时刻，整个输出的结果也是正确的。\n\n​\t那么，在我们使用SSE优化过程中，不妨以迭代器等于零的时刻，函数计算结果作为此时目标结果。在此基础上再利用SSE优化代码。与目标结果核对正确以后，再进一步去考虑迭代器等于1的情况（重复这个过程直到迭代器达到最大值）。在迭代器的每个元素下，SSE优化出的代码都与结果相等，那么我们可以说，该次优化是正确性，且完全覆盖了需执行代码。（一般来说不用考虑到最大值，根据数学归纳法，n有效，n+1有效，那么n的序列都是有效的）\n\n## 五：总结\n\n​\t本文描述SSE的使用及以4x4矩阵乘法的样例来优化SSE代码。\n\n​\t值得注意的是，SSE只是128bit数据宽度的指令集，但是也可以用来模拟256bit 和 512bit数据宽度，来实现以pack4拼接成pack8，甚至pack16的做法，只不过在输出结果管理上更加繁琐而已。感兴趣的同学可以尝试一下。\n\n## 六：引用\n\n1. [SSE指令扩展快查](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#techs=SSE,SSE2,SSE3,SSSE3,SSE4_1,SSE4_2)；\n2. 浮点性能基准计算-[浮点峰值那些事儿](https://zhuanlan.zhihu.com/p/28226956)\n3. 硬件性能基准测试计算样例：[M1芯片搞数据科学好使吗？5种基准测试给你答案](https://mp.weixin.qq.com/s/2N5cl_Z1MRF8dfbRo-sb4A)\n4. 讨论矩阵乘法如何优化的系列论文：[how-to-optimized-gemm](https://github.com/flame/how-to-optimize-gemm/wiki)\n5. 讨论以Arm Intrinsic 优化gemm的系列文章：[OpenBLAS gemm从零入门](https://zhuanlan.zhihu.com/p/65436463)\n" }, { "path": "docs/developer-guide/kvcache.md", "content": "# high-performance transformer inference with mha kv cache in ncnn\n\nThis document details the implementation and usage of the key-value (kv) cache for the `MultiHeadAttention` and `SDPA` layer in ncnn. This feature significantly accelerates autoregressive inference for Transformer-based models, such as large language models and other encoder-decoder architectures.\n\n## 1. what is kv cache?\n\n### the challenge of autoregressive inference\n\nTransformer models generate output token by token in a process called autoregressive decoding. In each step, the model takes the previously generated tokens as input to predict the next one. A core component of this is the self-attention mechanism, which computes query (q), key (k), and value (v) matrices based on the sequence generated so far.\n\nWithout optimization, the model must recompute the k and v matrices for all preceding tokens at every single step. For a sequence of length `N`, the computational cost for the self-attention mechanism is roughly proportional to `N^2`. As the sequence grows, this becomes a significant performance bottleneck.\n\n### the solution: kv cache\n\n**kv cache** is an optimization technique that stores the key and value tensors from previous decoding steps. When generating a new token, we only need to compute the k and v for the *current* token and append them to the cached values. The model then uses the full set of cached k and v tensors for the attention calculation.\n\n### key benefits\n\n- **dramatic speed-up:** It reduces the computational complexity of the self-attention mechanism from O(N^2) per step to approximately O(N). This drastically cuts down inference latency, especially for long sequences.\n- **reduced computation:** It eliminates redundant calculations, saving significant computational resources and energy.\n- **enables real-time applications:** The performance gain makes it feasible to deploy large Transformer models for interactive and real-time tasks.\n\n## 2. ncnn kv cache implementation\n\nncnn introduces kv cache support directly into its `MultiHeadAttention` and `SDPA` layer. The implementation is designed to be efficient and flexible, handling both the dynamic cache of self-attention and the static k/v of cross-attention found in encoder-decoder architectures.\n\n### self-attention vs. cross-attention cache logic\n\nThe caching strategy is fundamentally different for self-attention and cross-attention layers within a decoder.\n\n#### self-attention (dynamic cache)\n- **purpose:** Allows the decoder to attend to previously generated tokens in its own sequence (e.g., the text being generated).\n- **cache Logic:** The cache is **dynamic** and grows with each generated token. In step `t`, the k and v for token `t` are computed and appended to the cache from step `t-1`.\n- **ncnn implementation:** The `MultiHeadAttention` and `SDPA` layers for self-attention are modified to accept two additional inputs (`cache_k_in`, `cache_v_in`) and produce two corresponding outputs (`cache_k_out`, `cache_v_out`). The `7=1` parameter enables this dynamic caching behavior inside the layer.\n\n#### cross-attention (static k/v)\n- **purpose:** Allows the decoder to attend to the output of the encoder (e.g., attending to audio features in speech recognition, or an input sentence in translation).\n- **cache Logic:** The k and v matrices are derived from the encoder's output, which is computed only **once** per input sequence. Therefore, the k and v for cross-attention are **static** and do not change during the decoding process. They are \"cached\" in the sense that they are pre-computed and reused in every decoding step.\n- **ncnn implementation:** The `MultiHeadAttention` and `SDPA` layers for cross-attention are also configured with `7=1` and cache I/O blobs. However, the implementation correctly identifies cross-attention (where the query blob is different from the key/value blobs) and reuses the `cache_k_in` and `cache_v_in` directly, without performing concatenation. This allows the static encoder k/v to be passed efficiently through the network.\n\n## 3. ncnn kv cache memory layout\n\nThe memory layout of the kv cache is a critical design choice for performance. ncnn uses different layouts for `MultiHeadAttention` and `SDPA` to optimize for their respective calculation patterns.\n\n### `MultiHeadAttention` cache layout (Transposed)\n\nThe `MultiHeadAttention` layer uses a **transposed layout** for its cache blobs. The primary reason for this is to **ensure that data for each attention head is contiguous in memory, which significantly boosts gemm performance.**\n\n* **input blobs (q, k, v):** These typically have a shape where height represents the sequence length.\n * `ncnn::Mat` dimensions: `(w = embed_dim, h = seq_len)`\n\n* **cache blobs (`k_cache`, `v_cache`):** These are stored in a **transposed** format.\n * `ncnn::Mat` dimensions: `(w = seq_len, h = embed_dim)`\n\n**the rationale:**\n\n1. **slicing by Head:** During the attention calculation, the code slices the `k_cache` and `v_cache` matrices along their height to isolate the data for each head (e.g., using `row_range(head_index * embed_dim_per_head, embed_dim_per_head)`).\n2. **memory contiguity:** Because `ncnn::Mat` uses a row-major memory layout, this slicing operation on the transposed cache blob results in a sub-matrix where all the data for a single head is perfectly contiguous.\n3. **gemm efficiency:** Subsequent matrix multiplication operations (`q * k^T` and `Attention * v`) can then operate on these contiguous memory blocks. This maximizes CPU cache locality and the effectiveness of simd instructions, leading to a substantial increase in computational speed.\n\nIf a non-transposed layout were used, the data for each head would be strided in memory, causing frequent cache misses and dramatically slowing down the performance-critical gemm calculations. Therefore, this transposed layout is a deliberate and crucial optimization for computation.\n\n### `SDPA` cache layout (Standard)\n\nThe `SDPA` layer uses the **standard ncnn Mat layout**, where the sequence length is represented by the height.\n\n* **input blobs (q, k, v):** `(w = embed_dim, h = seq_len, c = num_heads)`\n* **cache blobs (`k_cache`, `v_cache`):** `(w = embed_dim, h = seq_len, c = num_heads)`\n\n**the rationale:**\n\nThe `SDPA` layer's internal implementation directly concatenates the cache blobs (`past_k`, `past_v`) with the current ones (`cur_k`, `cur_v`) along the height dimension (`seq_len`). This simpler approach avoids the need for a transposed layout while still being highly efficient, as the concatenation logic is handled inside the optimized C++ implementation.\n\n## 4. converting models to support kv cache\n\nTo enable kv cache, you must modify the model's `.param` file to add the necessary cache inputs and outputs to all `MultiHeadAttention` and `SDPA` layers in the decoder.\n\n### step 1: export a sequence-length-1 model\n\nFirst, export your model from its original framework (e.g., PyTorch) using a sequence length of 1 for the decoder. This creates a graph optimized for single-token generation, which is the core of the autoregressive decoding loop.\n\n### step 2: modify the .ncnn.param file\n\nAfter exporting, a script is needed to edit the generated `.ncnn.param` file to make it cache-aware.\n\n#### A. Adding kv cache to All MultiHeadAttention and SDPA Layers\n\nYou must add cache inputs/outputs to **every** `MultiHeadAttention` / `SDPA` layer in the decoder.\n\n- **change `input_count` and `output_count`:** Increase both by 2.\n- **add blob names:** Append new, unique blob names for `cache_k_in`, `cache_v_in`, `cache_k_out`, and `cache_v_out`.\n- **enable cache behavior:** Add the parameter `7=1`.\n\nHere is a robust Python function that automates this process:\n```python\ndef add_kv_cache_to_ncnn_param(filename):\n \"\"\"\n Modifies an ncnn.param file to add a kv cache mechanism to all\n MultiHeadAttention and SDPA layers and overwrites the original file.\n This handles both self-attention and cross-attention layers.\n \"\"\"\n import os\n\n if not os.path.exists(filename):\n print(f\"Error: The file '{filename}' was not found.\")\n return\n\n with open(filename, 'r', encoding='utf-8') as f:\n lines = f.readlines()\n\n header_line_index = 1 # line 2, after magic number\n header_parts = lines[header_line_index].strip().split()\n original_layer_count = int(header_parts[0])\n original_blob_count = int(header_parts[1])\n\n attention_indices = [i for i, line in enumerate(lines) if line.strip().startswith(\"MultiHeadAttention\") or line.strip().startswith(\"SDPA\")]\n attention_count = len(attention_indices)\n\n if attention_count == 0:\n print(\"No 'MultiHeadAttention' or 'SDPA' layers found. The file will not be modified.\")\n return\n\n # --- modify MultiHeadAttention and SDPA layers ---\n for i, line_index in enumerate(attention_indices):\n parts = lines[line_index].strip().split()\n layer_type, layer_name, input_count_str, output_count_str = parts[:4]\n input_count, output_count = int(input_count_str), int(output_count_str)\n\n blob_and_params = parts[4:]\n inputs = blob_and_params[:input_count]\n outputs = blob_and_params[input_count : input_count + output_count]\n params = blob_and_params[input_count + output_count:]\n\n # add cache I/O blobs and enable cache parameter\n inputs.extend([f\"cache_k_in_{i}\", f\"cache_v_in_{i}\"])\n outputs.extend([f\"cache_k_out_{i}\", f\"cache_v_out_{i}\"])\n params.append(\"7=1\")\n\n new_line_parts = [\n f\"{layer_type:<24}\", f\"{layer_name:<24}\",\n str(input_count + 2), str(output_count + 2),\n *inputs, *outputs, *params\n ]\n lines[line_index] = \" \".join(new_line_parts) + \"\\n\"\n\n # --- add a single input layer to provide all cache blobs ---\n new_layer_count = original_layer_count + 1\n # each mha needs 2 new *input* blobs and produces 2 new *output* blobs.\n # the total number of unique blobs increases by 4 for each mha.\n new_blob_count = original_blob_count + (attention_count * 4)\n lines[header_line_index] = f\"{new_layer_count} {new_blob_count}\\n\"\n\n # find where to insert the new input layer (after existing ones)\n insert_pos = header_line_index + 1\n while insert_pos < len(lines) and lines[insert_pos].strip().startswith(\"Input\"):\n insert_pos += 1\n\n cache_blob_names = [name for i in range(attention_count) for name in (f\"cache_k_in_{i}\", f\"cache_v_in_{i}\")]\n input_layer_line = (\n f\"{'Input':<24} {'kv_cache_in':<24} 0 {len(cache_blob_names)} \"\n f\"{' '.join(cache_blob_names)}\\n\"\n )\n lines.insert(insert_pos, input_layer_line)\n\n with open(filename, 'w', encoding='utf-8') as f:\n f.writelines(lines)\n\n print(f\"Successfully added kv cache to {attention_count} MultiHeadAttention / SDPA layers.\")\n\n# usage:\n# add_kv_cache_to_ncnn_param(\"your_model_decoder.ncnn.param\")\n```\n\n#### B. Supporting Dynamic Sequence Length in Gemm\nFeed-forward networks (`Gemm` layers) that process the output of attention blocks must support dynamic sequence lengths, as the cache grows. To achieve this, change the parameter `7=1` (constant input shape) to `7=0` (dynamic input shape) for the relevant `Gemm` layers.\n\n```python\ndef update_gemm_params(param_file_path):\n \"\"\"\n Finds all 'Gemm' layers and changes parameter '7=1' to '7=0'\n to support dynamic input shapes.\n \"\"\"\n import re\n with open(param_file_path, 'r') as f:\n lines = f.readlines()\n\n new_lines = []\n for line in lines:\n if line.strip().startswith('Gemm'):\n line = re.sub(r'(\\b7=)1\\b', r'\\g<1>0', line)\n new_lines.append(line)\n\n with open(param_file_path, 'w') as f:\n f.writelines(new_lines)\n print(f\"Updated Gemm layers in '{param_file_path}' to support dynamic inputs.\")\n\n# usage:\n# update_gemm_params(\"your_model_decoder.ncnn.param\")\n```\n\n## 5. implementing kv cache inference logic\n\nYour C++ inference code must manage the cache blobs across decoding steps.\n\n### step 1: identify cache blob indices\nAfter loading the network, identify the input and output blob indices for the cache. You can iterate through the mha layers and find the blobs you named in the conversion script.\n\n```cpp\n#include \"net.h\"\n#include \n#include \n\nstruct kvcache_info\n{\n std::vector input_indices;\n std::vector output_indices;\n};\n\nvoid find_mha_kvcache_blobs(const ncnn::Net& net, kvcache_info& info)\n{\n for (const ncnn::Layer* layer : net.layers())\n {\n // cache-enabled mha layer has 3 outputs (out, cache_k_out, cache_v_out) instead of 1\n if ((layer->typeindex == ncnn::LayerType::MultiHeadAttention || layer->typeindex == ncnn::LayerType::SDPA) && layer->tops.size() == 3)\n {\n // the script adds cache_k and cache_v as the last two inputs/outputs\n int input_count = layer->bottoms.size();\n int output_count = layer->tops.size();\n\n info.input_indices.push_back(layer->bottoms[input_count - 2]); // cache_k_in\n info.input_indices.push_back(layer->bottoms[input_count - 1]); // cache_v_in\n\n info.output_indices.push_back(layer->tops[output_count - 2]); // cache_k_out, i.e., tops[1]\n info.output_indices.push_back(layer->tops[output_count - 1]); // cache_v_out, i.e., tops[2]\n }\n }\n}\n```\n\n### step 2: prefill and decode loop\nThe inference process is split into two phases: \"prefill\" for the initial prompt and \"decode\" for subsequent single-token generation.\n\n- **prefill (`run_decoder_pre`):**\n - input: The entire initial sequence of token IDs\n - the kv cache is empty\n - run the decoder once\n - extract the output logits for the *last* token to predict the next token\n - extract the `out_cache_k` and `out_cache_v` blobs from all mha layers and store them\n\n- **decode (`run_decoder_step`):**\n - input: The single, most recently generated token ID\n - the kv cache blobs from the previous step are fed as input\n - run the decoder\n - extract the output logits to predict the next token\n - extract and store the updated kv cache blobs for the next step\n\nHere is a conceptual C++ implementation:\n\n```cpp\n// assume 'decoder_net' is loaded and 'kvcache_info' is populated.\n\n// --- prefill step (processes a sequence of tokens) ---\nvoid run_decoder_pre(const std::vector& tokens, const ncnn::Mat& encoder_states, std::vector& out_kv_cache)\n{\n ncnn::Extractor ex = decoder_net.create_extractor();\n\n ncnn::Mat input_embeds = prepare_input_embeds(tokens); // your embedding logic\n ex.input(\"in0\", input_embeds); // use your input blob name\n ex.input(\"encoder_out\", encoder_states); // use your encoder output blob name\n\n out_kv_cache.resize(kvcache_info.output_indices.size());\n for (size_t i = 0; i < kvcache_info.output_indices.size(); i++)\n {\n ex.extract(kvcache_info.output_indices[i], out_kv_cache[i]);\n }\n\n ncnn::Mat all_logits;\n ex.extract(\"out0\", all_logits); // Use your output blob name\n // ... process logits for the last token ...\n}\n\n// --- decode step (processes a single token) ---\nvoid run_decoder_step(int token, const ncnn::Mat& encoder_states, const std::vector& kv_cache, std::vector& out_kv_cache)\n{\n ncnn::Extractor ex = decoder_net.create_extractor();\n\n ncnn::Mat input_embeds = prepare_input_embeds({token});\n ex.input(\"in0\", input_embeds);\n ex.input(\"encoder_out\", encoder_states);\n\n // feed the existing cache\n for (size_t i = 0; i < kvcache_info.input_indices.size(); i++)\n {\n ex.input(kvcache_info.input_indices[i], kv_cache[i]);\n }\n\n // extract the updated cache\n out_kv_cache.resize(kvcache_info.output_indices.size());\n for (size_t i = 0; i < kvcache_info.output_indices.size(); i++)\n {\n ex.extract(kvcache_info.output_indices[i], out_kv_cache[i]);\n }\n\n ncnn::Mat logits;\n ex.extract(\"out0\", logits);\n // ... process logits to get the next token ...\n}\n\n// --- main inference loop ---\nvoid generate_sequence()\n{\n std::vector initial_tokens = { /* SOT and prompt tokens */ };\n ncnn::Mat encoder_states = run_encoder(); // compute encoder output once\n\n // 1. prefill stage\n std::vector kv_cache;\n run_decoder_pre(initial_tokens, encoder_states, kv_cache);\n int next_token = get_next_token_from_prefill_logits();\n\n // 2. autoregressive decoding loop\n while (next_token != EOT_TOKEN && sequence_length < MAX_LENGTH)\n {\n std::vector next_kv_cache;\n run_decoder_step(next_token, encoder_states, kv_cache, next_kv_cache);\n kv_cache = next_kv_cache; // update cache for the next iteration\n\n next_token = get_next_token_from_step_logits();\n // append next_token to your generated sequence\n }\n}\n```\nThis structured approach allows ncnn to perform highly efficient Transformer inference, correctly handling both dynamic self-attention and static cross-attention caches with an optimized memory layout.\n" }, { "path": "docs/developer-guide/layer-feat-mask.md", "content": "# layer feature mask\n\nEach ncnn layer allows a special parameter pair `31=X` to control specific bahavior.\n\nX is an unsigned integer with each bit contributing a feature mask.\n\nWe usually use it to configuring fine-graded behaviors for certain layers to maintain accuracy, reduce memory usage or optimize performance.\n\n|bit|value|mask|rationale|\n|---|---|---|---|\n|1<<0|1|no fp16 arithmetic|precision concern|\n|1<<1|2|no fp16 storage|precision concern|\n|1<<2|4|no bf16 storage|precision concern|\n|1<<3|8|no int8|debug dynamic quantized model|\n|1<<4|16|no vulkan|reduce overhead for cpu op - gpu split - cpu op|\n|1<<5|32|no sgemm|reduce some memory|\n|1<<6|64|no winograd|reduce some memory|\n|1<<7|128|no threading|force single thread|\n\nThese bits can be OR-combined into one value to control multiple behaviors simultaneously.\n\nFor example, `31=17` means disabling both vulkan and fp16 arithmetic.\n\n## disable fp16 for certain layer to fix overflow\n\n```ruby\n7767517\n3 3\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nConvolution conv1 1 1 conv0 conv1 0=128 1=3 6=36864 9=1\n```\n\nTypically, we use fp16 computation to improve inference speed.\nHowever, since the weight value of `conv1` is very large, fp16 accumulation may cause numerical overflow, so fp16 needs to be disabled individually for `conv1`, while other layers continue to use fp16 mode\n\nAdd `31=3` to disable fp16 storage and arithmetic.\n\n```ruby\n7767517\n3 3\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nConvolution conv1 1 1 conv0 conv1 0=128 1=3 6=36864 9=1 31=3\n```\n\n## disable vulkan for certain layer to improve performance\n\n```ruby\n7767517\n5 5\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nSomeCPULayer c0 1 1 conv0 c0 0=32\nReLU relu0 1 1 c0 relu0\nSomeCPULayer c1 1 1 relu0 c1 0=32\n```\n\nBetween the CPU layers, there is a simple calculation layer that supports vulkan. We can set `31=16` to force it to run on CPU. This can avoid the overhead of data upload, download and storage layout conversion between CPU and GPU. After all, CPU is fast enough for simple operations.\n\n```ruby\n7767517\n5 5\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nSomeCPULayer c0 1 1 conv0 c0 0=32\nReLU relu0 1 1 c0 relu0 31=16\nSomeCPULayer c1 1 1 relu0 c1 0=32\n```\n\n## disable winograd for certain layer to reduce memory usage\n\n```ruby\n7767517\n3 3\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nConvolution conv1 1 1 conv0 conv1 0=128 1=3 6=36864 9=1\n```\n\nThe winograd technology uses more memory for the purpose of improving convolution performance, but this is not always true. In some memory-constrained situations, or memory IO bottlenecks, we can disable the use of winograd on some layers in exchange for a smaller memory footprint. Add `31=64` to Convolution layer, which forces it to use implcit-gemm or tiled im2col-gemm implementation, reducing memory usage and sometimes improving vulkan performance.\n\n```ruby\n7767517\n3 3\nInput input 0 1 input0 0=22 1=22 2=32\nConvolution conv0 1 1 input0 conv0 0=32 1=1 6=1024 9=1\nConvolution conv1 1 1 conv0 conv1 0=128 1=3 6=36864 9=1 31=64\n```\n\n## disable threading for certain layer to improve performance\n\n```ruby\n7767517\n4 4\nInput input 0 1 input0 0=22 1=22 2=3\nConvolution conv0 1 1 input0 conv0 0=16 1=3 6=432\nHardSigmoid hs 1 1 conv0 hs0\nConvolution conv1 1 1 hs0 conv1 0=16 1=3 6=2304\n```\n\nThe overhead of multi-thread dispatch and merging is too large for small tensors. Add `31=128` to HardSigmoid layer, which forces it to execute in a single thread, reducing power consumption and improving performance.\n\n```ruby\n7767517\n4 4\nInput input 0 1 input0 0=22 1=22 2=3\nConvolution conv0 1 1 input0 conv0 0=16 1=3 6=432\nHardSigmoid hs 1 1 conv0 hs0 31=128\nConvolution conv1 1 1 hs0 conv1 0=16 1=3 6=2304\n```\n" }, { "path": "docs/developer-guide/layer-support-behavior.md", "content": "# Understanding `support_XYZ` Properties in ncnn's `Layer` Class\n\nThis document is for developers implementing new layers in `ncnn`. It explains the `support_XYZ` boolean properties in the `ncnn::Layer` base class. Correctly setting these properties declares the capabilities of your layer to the `ncnn` inference engine. This allows the engine to apply specific optimizations, such as enabling SIMD, half-precision floating-point computation, or Vulkan GPU acceleration, to achieve optimal performance and memory efficiency.\n\n## When to Set `support` Properties\n\nA layer can set its `support` properties in two ways:\n\n1. **Statically in the constructor**: If the layer's capabilities are fixed, the simplest way is to set them in its constructor.\n2. **Dynamically in `create_pipeline`**: If the layer's capabilities depend on parameters loaded from `load_param` or `load_model` (e.g., the data type of weights), you can set these properties dynamically within the `create_pipeline` method.\n\n---\n\n## Property Details\n\nHere is a detailed breakdown of each `support` property and what it means for your layer's implementation.\n\n### `one_blob_only`\n\n* **Purpose**: Declares that the layer accepts only one input `blob` and produces only one output `blob`.\n* **Requirements if `true`**: You must implement the single-input, single-output version of the `forward` method:\n ```cpp\n virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;\n ```\n* **Behavior**: When `true`, `ncnn` calls this overload. If `false` (default), the `std::vector` version of `forward` is called.\n\n### `support_inplace`\n\n* **Purpose**: Declares that the layer supports in-place computation, meaning the input and output can share the same memory. This significantly reduces memory overhead.\n* **Requirements if `true`**: You must implement the `forward_inplace` method. Depending on whether `one_blob_only` is also enabled, implement the corresponding version:\n ```cpp\n // If one_blob_only is true\n virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;\n\n // If one_blob_only is false\n virtual int forward_inplace(std::vector& bottom_top_blobs, const Option& opt) const;\n ```\n\n### `support_vulkan`\n\n* **Purpose**: Declares that the layer has a Vulkan implementation for GPU-accelerated inference.\n* **Requirements if `true`**:\n * Implement `forward` / `forward_inplace` methods that accept `VkMat` for input and output.\n * Implement `upload_model` to transfer weight data to the GPU.\n * Implement `create_pipeline` and `destroy_pipeline` to manage Vulkan `Pipeline` objects and other GPU resources.\n\n### `support_packing` (for CPU)\n\n* **Purpose**: Declares that the layer's **CPU implementation** can handle `Mat` data with a \"packing\" memory layout (i.e., `elempack > 1`). This is crucial for SIMD optimizations (e.g., processing 4 or 8 floats at once with NEON or AVX).\n* **Behavior if `true`**:\n * When the input `Mat` channel count is a multiple of the SIMD width, the `ncnn` engine ensures that the input `Mat` passed to `forward` / `forward_inplace` is packed (e.g., `elempack=4` or `elempack=8`).\n * Your implementation must correctly handle `Mat` data where `cstep` and `elempack` are not their default values.\n* **Behavior if `false`**:\n * The `ncnn` engine guarantees that the input `Mat` passed to your layer will always have `elempack=1`. The engine will automatically insert conversions if the preceding layer produced a packed output.\n* **Output**: Regardless of the property's value, your layer can output a `Mat` with any `elempack`. However, it is highly recommended to output a `Mat` with an adaptive `elempack` to avoid unnecessary conversions in subsequent layers.\n\n### `support_any_packing` (for CPU)\n\n* **Purpose**: An extension of `support_packing`. It declares that the layer's **CPU implementation** is flexible enough to handle a `Mat` with **any** `elempack` value (`1`, `4`, `8`, etc.).\n* **Behavior if `true`**:\n * The `ncnn` engine can pass an input `Mat` with any packing layout to your `forward` method, without forcing a conversion to the hardware's \"optimal\" `elempack`. For example, on an AVX512 system where `elempack=16` is optimal, your layer can still accept `elempack=1`, `4`, or `8`.\n * This gives the engine more flexibility to avoid unnecessary packing/unpacking conversions between layers.\n* **Behavior if `false`**: If `false` (but `support_packing` is `true`), the engine will try to provide an input `Mat` with an optimal `elempack` for the target architecture.\n* **Output**: This property does not enforce any constraint on the output `Mat`, which can have any `elempack`.\n\n### `support_vulkan_packing` (for Vulkan)\n\n* **Purpose**: This is the Vulkan equivalent of `support_packing`. It declares that the layer's **Vulkan implementation** can handle `VkMat` with `elempack=4`.\n* **Behavior if `true`**: When the input `VkMat` has a channel count that is a multiple of 4, the `ncnn` engine will provide a packed `VkMat` (with `elempack=4`) to your Vulkan `forward` methods.\n* **Behavior if `false`**: The engine will ensure the input `VkMat` has `elempack=1`.\n* **Note**: `support_packing` and `support_vulkan_packing` are independent. A layer can support packing on CPU but not on Vulkan, or vice-versa.\n\n### `support_vulkan_any_packing` (for Vulkan)\n\n* **Purpose**: An extension of `support_vulkan_packing`. It declares that the layer's **Vulkan implementation** can handle a `VkMat` with **any** supported `elempack` value (e.g., `1`, `4`).\n* **Behavior if `true`**:\n * The `ncnn` engine can pass an input `VkMat` with any supported packing layout to your Vulkan `forward` method. This allows the engine to avoid unnecessary repacking operations on the GPU.\n * This is particularly useful for optimizing shader dispatch and memory access patterns.\n* **Behavior if `false`**: If `false` (but `support_vulkan_packing` is `true`), the engine will try to provide a `VkMat` with `elempack=4` if the channel count is a multiple of 4.\n* **Note**: This property is independent of its CPU counterpart, `support_any_packing`.\n\n### `support_bf16_storage`\n\n* **Purpose**: Declares that the layer can process `bfloat16` data.\n* **Behavior if `true`**:\n * The `forward` method may receive an input `Mat` of type `bfloat16` (`elembits() == 16`) or `fp32`.\n * Inside your `forward` implementation, you must check `opt.use_bf16_storage` and `bottom_blob.elembits()` to determine whether to use a `bfloat16`-optimized code path.\n* **Behavior if `false`**: The `ncnn` engine ensures your layer will **not** receive a `bfloat16` `Mat`.\n* **Output**: Your layer can output either a `bfloat16` or `fp32` `Mat`. When `opt.use_bf16_storage` is active, outputting `bfloat16` is recommended to maintain precision and performance across the network.\n\n### `support_fp16_storage`\n\n* **Purpose**: Declares that the layer can process `float16` data for half-precision inference.\n* **Behavior if `true`**:\n * Similar to `support_bf16_storage`, the `forward` method may receive an `fp16` or `fp32` `Mat`.\n * Your implementation should check `opt.use_fp16_storage` and `bottom_blob.elembits()` to select the correct code path.\n* **Behavior if `false`**: The `ncnn` engine ensures your layer will **not** receive an `fp16` `Mat`.\n* **Output**: Your layer can output either a `fp16` or `fp32` `Mat`. When `opt.use_fp16_storage` is active, outputting an `fp16` `Mat` is recommended.\n\n### `support_int8_storage`\n\n* **Purpose**: Declares that the layer supports `int8` quantized inference.\n* **Behavior if `true`**:\n * When `opt.use_int8_inference` is `true`, the `forward` method may receive an `int8` or `fp32` `Mat`.\n * **Important**: If the input is `fp32`, your `forward` implementation is responsible for dynamically quantizing it to `int8` before performing computations.\n* **Behavior if `false`**: The `ncnn` engine ensures your layer will **not** receive an `int8` `Mat`.\n* **Output**: The output can be `int8` or `fp32`, depending on your layer's design.\n\n---\n\n## Practical Implementation and Priorities\n\n### Handling Multiple Precision Types\n\nA layer can set `support_fp16_storage` and `support_bf16_storage` to `true` simultaneously. The `ncnn` engine prioritizes these formats based on the `Option` flags. As seen in the `convert_layout` function in `src/net.cpp`, if `opt.use_bf16_storage` is true, the engine will prefer converting inputs to `bfloat16`. Otherwise, it falls back to `fp16` if `opt.use_fp16_storage` is true.\n\nThe chosen `elempack` also depends on the precision. For instance, with SIMD, the priority might be:\n* FP16: `elempack=8` (if supported), then `elempack=4`, then `1`.\n* BF16: `elempack=4`, then `1`.\n\nYour `forward` implementation should reflect this by checking `elembits()` and `elempack` to dispatch to the correct kernel.\n\n### Code Example: `Clip_arm`\n\nThe `Clip_arm` layer provides a great example of these concepts in practice.\n\n1. **Declaring Support in the Constructor**:\n It declares support for packing and, conditionally, for fp16 and bf16 storage.\n ```cpp\n // From: src/layer/arm/clip_arm.cpp\n Clip_arm::Clip_arm()\n {\n #if __ARM_NEON\n support_packing = true;\n #if NCNN_ARM82\n support_fp16_storage = cpu_support_arm_asimdhp();\n #endif\n #endif // __ARM_NEON\n\n #if NCNN_BF16\n support_bf16_storage = true;\n #endif\n }\n ```\n\n2. **Dispatching in `forward_inplace`**:\n The `forward_inplace` method acts as a dispatcher. It first checks the element size (`elembits`) and the corresponding `opt` flag to decide whether to call a specialized low-precision implementation (`fp16s` or `bf16s`). If neither is applicable, it defaults to the standard `fp32` implementation.\n\n ```cpp\n // From: src/layer/arm/clip_arm.cpp\n int Clip_arm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const\n {\n int elembits = bottom_top_blob.elembits();\n\n #if NCNN_ARM82\n if (support_fp16_storage && opt.use_fp16_storage && elembits == 16)\n return forward_inplace_fp16s(bottom_top_blob, opt);\n #endif\n\n #if NCNN_BF16\n if (opt.use_bf16_storage && elembits == 16)\n return forward_inplace_bf16s(bottom_top_blob, opt);\n #endif\n\n // Default fp32 implementation follows...\n int w = bottom_top_blob.w;\n // ...\n }\n ```\n\n### An Incremental Development Workflow\n\nAdopting a gradual approach can simplify the development of a new layer:\n\n1. **Implement the Core Algorithm**: Start with all `support_XYZ` properties set to `false`. Focus on getting the mathematical logic correct using standard `fp32` data and `elempack=1`.\n2. **Add Packing Support**: Once the core logic is validated, set `support_packing = true`. Modify your code to handle `elempack > 1` and implement SIMD optimizations (e.g., using NEON intrinsics).\n3. **Add Low-Precision Support**: Next, add support for `fp16`, `bf16`, or `int8`. Set the corresponding `support_*_storage` flags to `true` and add branches in your `forward` method to handle these data types based on the `opt` flags.\n4. **Add Vulkan Support**: Finally, if GPU acceleration is desired, set `support_vulkan = true` and implement the Vulkan-specific methods.\n\nThis incremental process allows you to tackle one challenge at a time, making it easier to develop a highly optimized and feature-rich layer.\n" }, { "path": "docs/developer-guide/low-level-operation-api.md", "content": "# implement elementwise addition with/without broadcast using BinaryOp operation\n\n* input must be fp32 storage without packing\n* output is expected to be fp32 storage without packing\n\n```cpp\nvoid binary_add(const ncnn::Mat& a, const ncnn::Mat& b, ncnn::Mat& c)\n{\n ncnn::Option opt;\n opt.num_threads = 2;\n opt.use_fp16_storage = false;\n opt.use_packing_layout = false;\n\n ncnn::Layer* op = ncnn::create_layer(\"BinaryOp\");\n\n // set param\n ncnn::ParamDict pd;\n pd.set(0, 0);// op_type\n\n op->load_param(pd);\n\n op->create_pipeline(opt);\n\n // forward\n std::vector bottoms(2);\n bottoms[0] = a;\n bottoms[1] = b;\n\n std::vector tops(1);\n op->forward(bottoms, tops, opt);\n\n c = tops[0];\n\n op->destroy_pipeline(opt);\n\n delete op;\n}\n```\n\n# implement 3x3 box blur on three channel image using ConvolutionDepthWise operation\n\n* input must be fp32 storage without packing\n* output is expected to be fp32 storage without packing\n\n```cpp\nvoid convolution_3x3_boxblur_RGB(const ncnn::Mat& rgb, ncnn::Mat& out)\n{\n ncnn::Option opt;\n opt.num_threads = 2;\n opt.use_fp16_storage = false;\n opt.use_packing_layout = false;\n\n ncnn::Layer* op = ncnn::create_layer(\"ConvolutionDepthWise\");\n\n // set param\n ncnn::ParamDict pd;\n pd.set(0, 3);// num_output\n pd.set(1, 3);// kernel_w\n pd.set(5, 0);// bias_term\n pd.set(6, 3*3*3);// weight_data_size\n pd.set(7, 3);// group\n\n op->load_param(pd);\n\n // set weights\n ncnn::Mat weights[1];\n weights[0].create(3*3*3);// weight_data\n\n for (int i=0; i<3*3*3; i++)\n {\n weights[0][i] = 1.f / 9;\n }\n\n op->load_model(ncnn::ModelBinFromMatArray(weights));\n\n op->create_pipeline(opt);\n\n // forward\n op->forward(rgb, out, opt);\n\n op->destroy_pipeline(opt);\n\n delete op;\n}\n```\n# transpose Mat, chw to cwh\n\n* input must be fp32 storage with/without packing\n* output is expected to be fp32 storage packed\n\n```cpp\nvoid transpose(const ncnn::Mat& in, ncnn::Mat& out)\n{\n ncnn::Option opt;\n opt.num_threads = 2;\n opt.use_fp16_storage = false;\n opt.use_packing_layout = true;\n\n ncnn::Layer* op = ncnn::create_layer(\"Permute\");\n\n // set param\n ncnn::ParamDict pd;\n pd.set(0, 1);// order_type\n\n op->load_param(pd);\n\n op->create_pipeline(opt);\n\n ncnn::Mat in_packed = in;\n {\n // resolve dst_elempack\n int dims = in.dims;\n int elemcount = 0;\n if (dims == 1) elemcount = in.elempack * in.w;\n if (dims == 2) elemcount = in.elempack * in.h;\n if (dims == 3) elemcount = in.elempack * in.c;\n\n int dst_elempack = 1;\n if (op->support_packing)\n {\n if (elemcount % 8 == 0 && (ncnn::cpu_support_x86_avx2() || ncnn::cpu_support_x86_avx()))\n dst_elempack = 8;\n else if (elemcount % 4 == 0)\n dst_elempack = 4;\n }\n\n if (in.elempack != dst_elempack)\n {\n convert_packing(in, in_packed, dst_elempack, opt);\n }\n }\n\n // forward\n op->forward(in_packed, out, opt);\n\n op->destroy_pipeline(opt);\n\n delete op;\n}\n```\n# apply instance normalization\n// x = (x - mean) / sqrt(var)\n\n* input can be fp32/fp16 storage with/without packing\n* output is expected to be fp16 storage packed when supported, or fp32 storage packed otherwise\n\n```cpp\nvoid normalize(const ncnn::Mat& in, ncnn::Mat& out)\n{\n ncnn::Option opt;\n opt.num_threads = 2;\n opt.use_fp16_storage = true;\n opt.use_packing_layout = true;\n\n ncnn::Layer* op = ncnn::create_layer(\"InstanceNorm\");\n\n // set param\n ncnn::ParamDict pd;\n pd.set(0, in.c);// channels\n pd.set(1, 0.f);// eps\n\n op->load_param(pd);\n\n // set weights\n ncnn::Mat weights[2];\n weights[0].create(in.c);// gamma_data\n weights[1].create(in.c);// beta_data\n\n weights[0].fill(1.f);\n weights[1].fill(0.f);\n\n op->load_model(ncnn::ModelBinFromMatArray(weights));\n\n op->create_pipeline(opt);\n\n ncnn::Mat in_fp16 = in;\n if (in.elembits() == 32 && op->support_fp16_storage)\n {\n cast_float32_to_float16(in, in_fp16, opt);\n }\n if (in.elembits() == 16 && !op->support_fp16_storage)\n {\n cast_float16_to_float32(in, in_fp16, opt);\n }\n\n ncnn::Mat in_fp16_packed = in_fp16;\n {\n // resolve dst_elempack\n int dims = in_fp16.dims;\n int elemcount = 0;\n if (dims == 1) elemcount = in_fp16.elempack * in_fp16.w;\n if (dims == 2) elemcount = in_fp16.elempack * in_fp16.h;\n if (dims == 3) elemcount = in_fp16.elempack * in_fp16.c;\n\n int dst_elempack = 1;\n if (op->support_packing)\n {\n if (elemcount % 8 == 0 && (ncnn::cpu_support_x86_avx2() || ncnn::cpu_support_x86_avx()))\n dst_elempack = 8;\n else if (elemcount % 4 == 0)\n dst_elempack = 4;\n }\n\n if (in_fp16.elempack != dst_elempack)\n {\n convert_packing(in_fp16, in_fp16_packed, dst_elempack, opt);\n }\n }\n\n // forward\n op->forward(in_fp16_packed, out, opt);\n\n op->destroy_pipeline(opt);\n\n delete op;\n}\n```\n\n# cpu -> gpu -> forward -> gpu -> cpu\n\n```cpp\nncnn::VulkanDevice* vkdev = ncnn::get_gpu_device();\n\nncnn::VkAllocator* blob_vkallocator = vkdev->acquire_blob_allocator();\nncnn::VkAllocator* staging_vkallocator = vkdev->acquire_staging_allocator();\n\nncnn::VkWeightAllocator* weight_vkallocator = new ncnn::VkWeightAllocator(vkdev);\nncnn::VkWeightStagingAllocator* weight_staging_vkallocator = new ncnn::VkWeightStagingAllocator(vkdev);\n\n// create layer\nncnn::Layer* convolution = ncnn::create_layer(\"Convolution\");\nconvolution->vkdev = vkdev;\n\n// set option\nncnn::Option opt;\nopt.num_threads = 4;\nopt.use_vulkan_compute = true;\nopt.blob_vkallocator = blob_vkallocator;\nopt.workspace_vkallocator = blob_vkallocator;\nopt.staging_vkallocator = staging_vkallocator;\n\n// load param\n{\n ncnn::ParamDict pd;\n pd.set(0, outch);\n pd.set(1, ksize);\n pd.set(6, outch*inch*ksize*ksize);\n pd.use_vulkan_compute = 1;\n\n convolution->load_param(pd);\n}\n\n// load model\n{\n ncnn::Mat weights[2];\n weights[0] = random_mat(outch*inch*ksize*ksize);\n weights[1] = random_mat(outch);\n\n ncnn::ModelBinFromMatArray mb(weights);\n convolution->load_model(mb);\n}\n\n// create pipeline\nconvolution->create_pipeline(opt);\n\n// upload model\n{\n ncnn::VkTransfer cmd(vkdev);\n\n ncnn::Option opt_upload = opt;\n opt_upload.blob_vkallocator = weight_vkallocator;\n opt_upload.workspace_vkallocator = weight_vkallocator;\n opt_upload.staging_vkallocator = weight_staging_vkallocator;\n\n convolution->upload_model(cmd, opt_upload);\n\n cmd.submit_and_wait();\n}\n\nncnn::Mat bottom = random_mat(w, h, inch);\n\nncnn::Mat top;\n\n// forward\n{\n ncnn::VkCompute cmd(vkdev);\n\n ncnn::VkMat bottom_gpu;\n cmd.record_upload(bottom, bottom_gpu, opt);\n\n ncnn::VkMat top_gpu;\n convolution->forward(bottom_gpu, top_gpu, cmd, opt);\n\n cmd.record_download(top_gpu, top, opt);\n\n cmd.submit_and_wait();\n}\n\nconvolution->destroy_pipeline(opt);\n\ndelete convolution;\n\nvkdev->reclaim_blob_allocator(blob_vkallocator);\nvkdev->reclaim_staging_allocator(staging_vkallocator);\n\nweight_vkallocator->clear();\nweight_staging_vkallocator->clear();\ndelete weight_vkallocator;\ndelete weight_staging_vkallocator;\n```\n\n" }, { "path": "docs/developer-guide/ncnn-tips-and-tricks.zh.md", "content": "### blob内存是隐含共享的\n\nncnn的blob最初直接使用opencv的cv::Mat，后发现blob最多只支持三维，因此实现了类似的Mat\nMat的data每个通道内存16字节对齐，并且有原子的引用计数，a=b不复制数据，超级快\nMat支持直接引用外部的内存块，不复制数据，加快模型加载和输入输出\n\n举个例子：split layer 将一个blob复制成n个，ncnn中实现为单纯的增加引用计数，没有任何数据复制\n\n### 只运算一部分并保留中间结果\n\nncnn的net在解决分支依赖时是自上而下深度优先的，因此当网络有多个分支时，运算只会在需要结果的那个分支中进行，节约时间\n当多个分支有重合部分时，运算其中一个分支后会自动保留其余分支所需的中间结果，隐含共享，以便运算其余分支时利用\n\n举个例子：某网络结构为 A -> B -> C1 + C2，向ncnn索要C1结果时，运算过程是 A -> B -> C1，同时B结果引用计数加1自动保留，后面还需要C2结果时，只运算C2就足够了\n\n### 开启轻模式省内存\n\n每个layer都会产生blob，除了最后的结果和多分支中间结果，大部分blob都不值得保留，开启轻模式可以在运算后自动回收，省下内存\n\n举个例子：某网络结构为 A -> B -> C，在轻模式下，向ncnn索要C结果时，A结果会在运算B时自动回收，而B结果会在运算C时自动回收，最后只保留C结果，后面再需要C结果会直接获得，满足绝大部分深度网络的使用方式\n\n### 网络和运算是分开的\n\nncnn的net是网络模型，实际使用的是extractor，也就是同个net可以有很多个运算实例，而且运算实例互不影响，中间结果保留在extractor内部，在多线程使用时共用网络的结构和参数数据，初始化网络模型和参数只需要一遍\n\n举个例子：全局静态的net实例，初始化一次后，就能不停地生成extractor使用\n\n### openmp虽快但未必合适\n\nncnn中几乎所有运算都能用上openmp多线程加速，而且性能很赞\n不过系统有时候会突然慢一下，比如手机太热自动降频，界面操作等等，ncnn耗时也会偶尔抖动变长，在计算耗时稳定性比较重要的时候建议关闭openmp，或者设置下extractor线程数\n\n举个例子：手机自拍时，用ncnn进行人脸实时定位，如果耗时突然涨一下就会感觉到掉帧，而稳定的帧率体验更好\n\n### NCNN_STDIO/NCNN_STRING禁用模型文件\n\nncnn支持加载自有的模型文件和模型内存，NCNN_STDIO控制是否需要支持加载模型文件，设成0能禁用这部分代码，从而减小库的体积，NCNN_STRING设成0能清除大部分可见的字符串和解析过程\n模型内存加载时的参数数据是直接引用的，速度更快，通常在手机上使用这种方式\n\n### 削减 ncnn 内置的层实现\n\ncmake的时候，加参数 -DWITH_LAYER_xxx=OFF 就可以完全不编译对应的内置层，这样可以进一步减小库的体积\n\n### 关于 ARM big.LITTLE 调度\n\n调用set_cpu_powersave可以把ncnn运算线程控制在特定的cpu核心上，大核心速度快耗电多，小核心速度慢点但省电，大小一起用手机热得快\n" }, { "path": "docs/developer-guide/new-model-load-api.md", "content": "## current model load api\n### Cons\n#### long and awful code\n#### two functions\n#### deal float32 float16 quantized-u8\n#### deal alignment size\n```cpp\n#if NCNN_STDIO\nint Convolution::load_model(FILE* binfp)\n{\n int nread;\n\n union\n {\n struct\n {\n unsigned char f0;\n unsigned char f1;\n unsigned char f2;\n unsigned char f3;\n };\n unsigned int tag;\n } flag_struct;\n\n nread = fread(&flag_struct, sizeof(flag_struct), 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read flag_struct failed %d\\n\", nread);\n return -1;\n }\n\n unsigned int flag = flag_struct.f0 + flag_struct.f1 + flag_struct.f2 + flag_struct.f3;\n\n weight_data.create(weight_data_size);\n if (weight_data.empty())\n return -100;\n\n if (flag_struct.tag == 0x01306B47)\n {\n // half-precision weight data\n int align_weight_data_size = alignSize(weight_data_size * sizeof(unsigned short), 4);\n std::vector float16_weights;\n float16_weights.resize(align_weight_data_size);\n nread = fread(float16_weights.data(), align_weight_data_size, 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read float16_weights failed %d\\n\", nread);\n return -1;\n }\n\n weight_data = Mat::from_float16(float16_weights.data(), weight_data_size);\n if (weight_data.empty())\n return -100;\n }\n else if (flag != 0)\n {\n // quantized weight data\n float quantization_value[256];\n nread = fread(quantization_value, 256 * sizeof(float), 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read quantization_value failed %d\\n\", nread);\n return -1;\n }\n\n int align_weight_data_size = alignSize(weight_data_size * sizeof(unsigned char), 4);\n std::vector index_array;\n index_array.resize(align_weight_data_size);\n nread = fread(index_array.data(), align_weight_data_size, 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read index_array failed %d\\n\", nread);\n return -1;\n }\n\n float* weight_data_ptr = weight_data;\n for (int i = 0; i < weight_data_size; i++)\n {\n weight_data_ptr[i] = quantization_value[ index_array[i] ];\n }\n }\n else if (flag_struct.f0 == 0)\n {\n // raw weight data\n nread = fread(weight_data, weight_data_size * sizeof(float), 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read weight_data failed %d\\n\", nread);\n return -1;\n }\n }\n\n if (bias_term)\n {\n bias_data.create(num_output);\n if (bias_data.empty())\n return -100;\n nread = fread(bias_data, num_output * sizeof(float), 1, binfp);\n if (nread != 1)\n {\n fprintf(stderr, \"Convolution read bias_data failed %d\\n\", nread);\n return -1;\n }\n }\n\n return 0;\n}\n#endif // NCNN_STDIO\n\nint Convolution::load_model(const unsigned char*& mem)\n{\n union\n {\n struct\n {\n unsigned char f0;\n unsigned char f1;\n unsigned char f2;\n unsigned char f3;\n };\n unsigned int tag;\n } flag_struct;\n\n memcpy(&flag_struct, mem, sizeof(flag_struct));\n mem += sizeof(flag_struct);\n\n unsigned int flag = flag_struct.f0 + flag_struct.f1 + flag_struct.f2 + flag_struct.f3;\n\n if (flag_struct.tag == 0x01306B47)\n {\n // half-precision weight data\n weight_data = Mat::from_float16((unsigned short*)mem, weight_data_size);\n mem += alignSize(weight_data_size * sizeof(unsigned short), 4);\n if (weight_data.empty())\n return -100;\n }\n else if (flag != 0)\n {\n // quantized weight data\n const float* quantization_value = (const float*)mem;\n mem += 256 * sizeof(float);\n\n const unsigned char* index_array = (const unsigned char*)mem;\n mem += alignSize(weight_data_size * sizeof(unsigned char), 4);\n\n weight_data.create(weight_data_size);\n if (weight_data.empty())\n return -100;\n float* weight_data_ptr = weight_data;\n for (int i = 0; i < weight_data_size; i++)\n {\n weight_data_ptr[i] = quantization_value[ index_array[i] ];\n }\n }\n else if (flag_struct.f0 == 0)\n {\n // raw weight data\n weight_data = Mat(weight_data_size, (float*)mem);\n mem += weight_data_size * sizeof(float);\n }\n\n if (bias_term)\n {\n bias_data = Mat(num_output, (float*)mem);\n mem += num_output * sizeof(float);\n }\n\n return 0;\n}\n```\n\n## new model load api proposed\n### Pros\n#### clean and simple api\n#### element type detection\n```cpp\nint Convolution::load_model(const ModelBin& mb)\n{\n // auto detect element type\n weight_data = mb.load(weight_data_size, 0);\n if (weight_data.empty())\n return -100;\n\n if (bias_term)\n {\n // certain type specified\n bias_data = mb.load(num_output, 1);\n if (bias_data.empty())\n return -100;\n }\n\n return 0;\n}\n```\n" }, { "path": "docs/developer-guide/new-param-load-api.md", "content": "## current param load api\n### Cons\n#### long and awful code\n#### three functions\n#### not extensible\n#### no default value\n#### no variable length array\n```\nMyLayer mylayer 1 1 in out 100 1.250000\n```\n```\nbinary 100\nbinary 1.250000\n```\n```cpp\n#if NCNN_STDIO\n#if NCNN_STRING\nint MyLayer::load_param(FILE* paramfp)\n{\n int nscan = fscanf(paramfp, \"%d %f\", &a, &b);\n if (nscan != 2)\n {\n fprintf(stderr, \"MyLayer load_param failed %d\\n\", nscan);\n return -1;\n }\n\n return 0;\n}\n#endif // NCNN_STRING\nint MyLayer::load_param_bin(FILE* paramfp)\n{\n fread(&a, sizeof(int), 1, paramfp);\n\n fread(&b, sizeof(float), 1, paramfp);\n\n return 0;\n}\n#endif // NCNN_STDIO\n\nint MyLayer::load_param(const unsigned char*& mem)\n{\n a = *(int*)(mem);\n mem += 4;\n\n b = *(float*)(mem);\n mem += 4;\n\n return 0;\n}\n```\n\n## new param load api proposed\n### Pros\n#### clean and simple api\n#### default value\n#### extensible\n#### variable length array\n```\n7767517\nMyLayer mylayer 1 1 in out 0=100 1=1.250000 -23303=5,0.1,0.2,0.4,0.8,1.0\n```\n```\nbinary 0xDD857600(magic)\n\nbinary 0\nbinary 100\nbinary 1\nbinary 1.250000\nbinary -23303\nbinary 5\nbinary 0.1\nbinary 0.2\nbinary 0.4\nbinary 0.8\nbinary 1.0\nbinary -233(EOP)\n```\n```cpp\nint MyLayer::load_param(const ParamDict& pd)\n{\n // pd.get( param id (seq), default value );\n a = pd.get(0, 100);\n b = pd.get(1, 1.25f);\n\n // get default value for c if not specified in param file\n c = pd.get(2, 0.001);\n\n // get array\n d = pd.get(3, Mat(len, array));\n return 0;\n}\n```\n" }, { "path": "docs/developer-guide/operation-param-weight-table.md", "content": "\n|operation|param id|param phase|default value|weight order|\n|:---:|:---:|:---:|:---:|:---:|\n|AbsVal|||\n|ArgMax|0|out_max_val|0|\n||1|topk|1|\n|BatchNorm|0|channels|0|slope mean variance bias|\n||1|eps|0.f|\n|Bias|0|bias_data_size|0|\n|BinaryOp|0|op_type|0|\n||1|with_scalar|0|\n||2|b|0.f|\n|BNLL|||\n|Cast|0|type_from|0|\n||1|type_to|0|\n|Clip|0|min|-FLT_MAX|\n||1|max|FLT_MAX|\n|Concat|0|axis|0|\n|Convolution|0|num_output|0|weight bias|\n||1|kernel_w|0|\n||2|dilation_w|1|\n||3|stride_w|1|\n||4|pad_left|0|\n||5|bias_term|0|\n||6|weight_data_size|0|\n||8|int8_scale_term|0|\n||9|activation_type|0|\n||10|activation_params|[ ]|\n||11|kernel_h|kernel_w|\n||12|dilation_h|dilation_w|\n||13|stride_h|stride_w|\n||15|pad_right|pad_left|\n||14|pad_top|pad_left|\n||16|pad_bottom|pad_top|\n||17|impl_type|0|\n||18|pad_value|0.f|\n|ConvolutionDepthWise|0|num_output|0|weight bias|\n||1|kernel_w|0|\n||2|dilation_w|1|\n||3|stride_w|1|\n||4|pad_left|0|\n||5|bias_term|0|\n||6|weight_data_size|0|\n||7|group|1|\n||8|int8_scale_term|0|\n||9|activation_type|0|\n||10|activation_params|[ ]|\n||11|kernel_h|kernel_w|\n||12|dilation_h|dilation_w|\n||13|stride_h|stride_w|\n||15|pad_right|pad_left|\n||14|pad_top|pad_left|\n||16|pad_bottom|pad_top|\n||18|pad_value|0.f|\n|Crop|0|woffset|0|\n||1|hoffset|0|\n||2|coffset|0|\n||3|outw|0|\n||4|outh|0|\n||5|outc|0|\n||6|woffset2|0|\n||7|hoffset2|0|\n||8|coffset2|0|\n||9|starts|[ ]|\n||10|ends|[ ]|\n||11|axes|[ ]|\n|Deconvolution|0|num_output|0|weight bias|\n||1|kernel_w|0|\n||2|dilation_w|1|\n||3|stride_w|1|\n||4|pad_left|0|\n||5|bias_term|0|\n||6|weight_data_size|0|\n||9|activation_type|0|\n||10|activation_params|[ ]|\n||11|kernel_h|kernel_w|\n||12|dilation_h|dilation_w|\n||13|stride_h|stride_w|\n||15|pad_right|pad_left|\n||14|pad_top|pad_left|\n||16|pad_bottom|pad_top|\n||18|output_pad_right|0|\n||19|output_pad_bottom|output_pad_right|\n||20|output_w|0|\n||21|output_h|output_w|\n|DeconvolutionDepthWise|0|num_output|0|weight bias|\n||1|kernel_w|0|\n||2|dilation_w|1|\n||3|stride_w|1|\n||4|pad_left|0|\n||5|bias_term|0|\n||6|weight_data_size|0|\n||7|group|1|\n||9|activation_type|0|\n||10|activation_params|[ ]|\n||11|kernel_h|kernel_w|\n||12|dilation_h|dilation_w|\n||13|stride_h|stride_w|\n||15|pad_right|pad_left|\n||14|pad_top|pad_left|\n||16|pad_bottom|pad_top|\n||18|output_pad_right|0|\n||19|output_pad_bottom|output_pad_right|\n||20|output_w|0|\n||21|output_h|output_w|\n|Dequantize|0|scale|1.f|bias|\n||1|bias_term|0|\n||2|bias_data_size|0|\n|DetectionOutput|0|num_class|0|\n||1|nms_threshold|0.05f|\n||2|nms_top_k|300|\n||3|keep_top_k|100|\n||4|confidence_threshold|0.5f|\n||5|variances[0]|0.1f|\n||6|variances[1]|0.1f|\n||7|variances[2]|0.2f|\n||8|variances[3]|0.2f|\n|Dropout|0|scale|1.f|\n|Eltwise|0|op_type|0|\n||1|coeffs|[ ]|\n|ELU|0|alpha|0.1f|\n|Embed|0|num_output|0|weight bias|\n||1|input_dim|0|\n||2|bias_term|0|\n||3|weight_data_size|0|\n|Exp|0|base|-1.f|\n||1|scale|1.f|\n||2|shift|0.f|\n|ExpandDims|0|expand_w|0|\n||1|expand_h|0|\n||2|expand_c|0|\n||3|axes|[ ]|\n|Flatten|||\n|HardSigmoid|0|alpha|0.2f||\n||1|beta|0.5f|\n|HardSwish|0|alpha|0.2f||\n||1|beta|0.5f|\n|InnerProduct|0|num_output|0|weight bias|\n||1|bias_term|0|\n||2|weight_data_size|0|\n||8|int8_scale_term|0|\n||9|activation_type|0|\n||10|activation_params|[ ]|\n|Input|0|w|0|\n||1|h|0|\n||2|c|0|\n|InstanceNorm|0|channels|0|gamma bias|\n||1|eps|0.001f|\n|Interp|0|resize_type|0|\n||1|height_scale|1.f|\n||2|width_scale|1.f|\n||3|output_height|0|\n||4|output_width|0|\n|Log|0|base|-1.f|\n||1|scale|1.f|\n||2|shift|0.f|\n|LRN|0|region_type|0|\n||1|local_size|5|\n||2|alpha|1.f|\n||3|beta|0.75f|\n||4|bias|1.f|\n|LSTM|0|num_output|0|\n||1|weight_data_size|1|\n||2|direction|0|\n|MemoryData|0|w|0|\n||1|h|0|\n||2|c|0|\n|Mish|||\n|MVN|0|normalize_variance|0|\n||1|across_channels|0|\n||2|eps|0.0001f|\n|Noop|||\n|Normalize|0|across_spatial|0|scale|\n||4|across_channel|0|\n||1|channel_shared|0|\n||2|eps|0.0001f|\n||9|eps_mode|0|\n||3|scale_data_size|0|\n|Packing|0|out_packing|1|\n||1|use_padding|0|\n||2|cast_type_from|0|\n||3|cast_type_to|0|\n||4|storage_type_from|0|\n||5|storage_type_to|0|\n|Padding|0|top|0|per_channel_pad_data|\n||1|bottom|0|\n||2|left|0|\n||3|right|0|\n||4|type|0|\n||5|value|0.f|\n||6|per_channel_pad_data_size|0|\n||7|front|0|\n||8|behind|0|\n|Permute|0|order_type|0|\n|PixelShuffle|0|upscale_factor|1|\n|Pooling|0|pooling_type(0: max 1: avg)|0|\n||1|kernel_w|0|\n||11|kernel_h|kernel_w|\n||2|stride_w|1|\n||12|stride_h|stride_w|\n||3|pad_left|0|\n||14|pad_right|pad_left|\n||13|pad_top|pad_left|\n||15|pad_bottom|pad_top|\n||4|global_pooling|0|\n||5|pad_mode|0|\n|Power|0|power|1.f|\n||1|scale|1.f|\n||2|shift|0.f|\n|PReLU|0|num_slope|0|slope|\n|PriorBox|0|min_sizes|[ ]|\n||1|max_sizes|[ ]|\n||2|aspect_ratios|[ ]|\n||3|varainces[0]|0.f|\n||4|varainces[1]|0.f|\n||5|varainces[2]|0.f|\n||6|varainces[3]|0.f|\n||7|flip|1|\n||8|clip|0|\n||9|image_width|0|\n||10|image_height|0|\n||11|step_width|-233.f|\n||12|step_height|-233.f|\n||13|offset|0.f|\n||14|step_mmdetection|0|\n||15|center_mmdetection|0|\n|Proposal|0|feat_stride|16|\n||1|base_size|16|\n||2|pre_nms_topN|6000|\n||3|after_nms_topN|300|\n||4|num_thresh|0.7f|\n||5|min_size|16|\n|PSROIPooling|0|pooled_width|7|\n||1|pooled_height|7|\n||2|spatial_scale|0.0625f|\n||3|output_dim|0|\n|Quantize|0|scale|1.f|\n|Reduction|0|operation|0|\n||1|dim|0|\n||2|coeff|1.f|\n||3|axes|[ ]|\n||4|keepdims|0|\n|ReLU|0|slope|0.f|\n|Reorg|0|stride|0|\n|Requantize|0|scale_in|1.f|bias|\n||1|scale_out|1.f|\n||2|bias_term|0|\n||3|bias_data_size|0|\n||4|fusion_relu|0|\n|Reshape|0|w|-233|\n||1|h|-233|\n||2|c|-233|\n||3|permute|0|\n|ROIAlign|0|pooled_width|0|\n||1|pooled_height|0|\n||2|spatial_scale|1.f|\n||3|sampling_ratio|0|\n||4|aligned|0|\n||5|version|0|\n|ROIPooling|0|pooled_width|0|\n||1|pooled_height|0|\n||2|spatial_scale|1.f|\n|Scale|0|scale_data_size|0|scale bias|\n||1|bias_term|0|\n|SELU|0|alpha|1.67326324f||\n||1|lambda|1.050700987f|\n|ShuffleChannel|0|group|1|\n|Sigmoid|||\n|Slice|0|slices|[ ]|\n||1|axis|0|\n|Softmax|0|axis|0|\n|Split|||\n|SPP|0|pooling_type|0|\n||1|pyramid_height|1|\n|Squeeze|0|squeeze_w|0|\n||1|squeeze_h|0|\n||2|squeeze_c|0|\n||3|axes|[ ]|\n|StatisticsPooling|0|include_stddev|0|\n|Swish|||\n|TanH|||\n|Threshold|0|threshold|0.f|\n|Tile|0|dim|0|\n||1|tiles|1|\n|UnaryOp|0|op_type|0|\n|YoloDetectionOutput|0|num_class|20|\n||1|num_box|5|\n||2|confidence_threshold|0.01f|\n||3|num_threshold|0.45f|\n||4|biases|[]|\n|Yolov3DetectionOutput|0|num_class|20|\n||1|num_box|5|\n||2|confidence_threshold|0.01f|\n||3|num_threshold|0.45f|\n||4|biases|[]|\n||5|mask|[]|\n||6|anchors_scale|[]|\n|RNN|0|num_output|0|\n||1|weight_data_size|0|\n||2|direction|0|\n|MultiHeadAttention|0|embed_dim|0|\n||1|num_head|1|\n||2|weight_data_size|0|\n" }, { "path": "docs/developer-guide/operators.md", "content": "\n* [AbsVal](#absval)\n* [ArgMax](#argmax)\n* [BatchNorm](#batchnorm)\n* [Bias](#bias)\n* [BinaryOp](#binaryop)\n* [BNLL](#bnll)\n* [Cast](#cast)\n* [CELU](#celu)\n* [Clip](#clip)\n* [Concat](#concat)\n* [Convolution](#convolution)\n* [Convolution1D](#convolution1d)\n* [Convolution3D](#convolution3d)\n* [ConvolutionDepthWise](#convolutiondepthwise)\n* [ConvolutionDepthWise1D](#convolutiondepthwise1d)\n* [ConvolutionDepthWise3D](#convolutiondepthwise3d)\n* [CopyTo](#copyto)\n* [Crop](#crop)\n* [CumulativeSum](#cumulativesum)\n* [Deconvolution](#deconvolution)\n* [Deconvolution1D](#deconvolution1d)\n* [Deconvolution3D](#deconvolution3d)\n* [DeconvolutionDepthWise](#deconvolutiondepthwise)\n* [DeconvolutionDepthWise1D](#deconvolutiondepthwise1d)\n* [DeconvolutionDepthWise3D](#deconvolutiondepthwise3d)\n* [DeformableConv2D](#deformableconv2d)\n* [Dequantize](#dequantize)\n* [Diag](#diag)\n* [Dropout](#dropout)\n* [Eltwise](#eltwise)\n* [ELU](#elu)\n* [Embed](#embed)\n* [Exp](#exp)\n* [ExpandDims](#expanddims)\n* [Flatten](#flatten)\n* [Flip](#flip)\n* [Fold](#fold)\n* [GELU](#gelu)\n* [GLU](#glu)\n* [Gemm](#gemm)\n* [GridSample](#gridsample)\n* [GroupNorm](#groupnorm)\n* [GRU](#gru)\n* [HardSigmoid](#hardsigmoid)\n* [HardSwish](#hardswish)\n* [InnerProduct](#innerproduct)\n* [Input](#input)\n* [InstanceNorm](#instancenorm)\n* [Interp](#interp)\n* [InverseSpectrogram](#inversespectrogram)\n* [LayerNorm](#layernorm)\n* [Log](#log)\n* [LRN](#lrn)\n* [LSTM](#lstm)\n* [MemoryData](#memorydata)\n* [Mish](#mish)\n* [MultiHeadAttention](#multiheadattention)\n* [MVN](#mvn)\n* [Noop](#noop)\n* [Normalize](#normalize)\n* [Packing](#packing)\n* [Padding](#padding)\n* [Permute](#permute)\n* [PixelShuffle](#pixelshuffle)\n* [Pooling](#pooling)\n* [Pooling1D](#pooling1d)\n* [Pooling3D](#pooling3d)\n* [Power](#power)\n* [PReLU](#prelu)\n* [Quantize](#quantize)\n* [Reduction](#reduction)\n* [ReLU](#relu)\n* [Reorg](#reorg)\n* [Requantize](#requantize)\n* [Reshape](#reshape)\n* [RMSNorm](#rmsnorm)\n* [RNN](#rnn)\n* [RotaryEmbed](#rotaryembed)\n* [Scale](#scale)\n* [SDPA](#sdpa)\n* [SELU](#selu)\n* [Shrink](#shrink)\n* [ShuffleChannel](#shufflechannel)\n* [Sigmoid](#sigmoid)\n* [Slice](#slice)\n* [Softmax](#softmax)\n* [Softplus](#softplus)\n* [Spectrogram](#spectrogram)\n* [Split](#split)\n* [Squeeze](#squeeze)\n* [Swish](#swish)\n* [TanH](#tanh)\n* [Threshold](#threshold)\n* [Tile](#tile)\n* [UnaryOp](#unaryop)\n* [Unfold](#unfold)\n\n# AbsVal\n```\ny = abs(x)\n```\n\n* one_blob_only\n* support_inplace\n\n# ArgMax\n```\ny = argmax(x, out_max_val, topk)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | out_max_val | int | 0 | |\n| 1 | topk | int | 1 | |\n\n# BatchNorm\n```\ny = (x - mean) / sqrt(var + eps) * slope + bias\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | channels | int | 0 | |\n| 1 | eps | float | 0.f | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| slope_data | float | [channels] |\n| mean_data | float | [channels] |\n| var_data | float | [channels] |\n| bias_data | float | [channels] |\n\n# Bias\n```\ny = x + bias\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | bias_data_size| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| bias_data | float | [channels] |\n\n# BinaryOp\n This operation is used for binary computation, and the calculation rule depends on the [broadcasting rule](https://github.com/Tencent/ncnn/wiki/binaryop-broadcasting).\n```\nC = binaryop(A, B)\n```\nif with_scalar = 1:\n- one_blob_only\n- support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | op_type | int | 0 | Operation type as follows |\n| 1 | with_scalar | int | 0 | with_scalar=0 B is a matrix, with_scalar=1 B is a scalar |\n| 2 | b | float | 0.f | When B is a scalar, B = b |\n\nOperation type:\n- 0 = ADD\n- 1 = SUB\n- 2 = MUL\n- 3 = DIV\n- 4 = MAX\n- 5 = MIN\n- 6 = POW\n- 7 = RSUB\n- 8 = RDIV\n- 9 = RPOW\n- 10 = ATAN2\n- 11 = RATAN2\n\n# BNLL\n```\ny = log(1 + e^(-x)) , x > 0\ny = log(1 + e^x), x < 0\n```\n\n* one_blob_only\n* support_inplace\n\n# Cast\n```\ny = cast(x)\n```\n\n* one_blob_only\n* support_packing\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | type_from | int | 0 | |\n| 1 | type_to | int | 0 | |\n\nElement type:\n- 0 = auto\n- 1 = float32\n- 2 = float16\n- 3 = int8\n- 4 = bfloat16\n\n# CELU\n```\nif x < 0 y = (exp(x / alpha) - 1.f) * alpha\nelse y = x\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 1.f | |\n\n# Clip\n```\ny = clamp(x, min, max)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | min | float | -FLT_MAX | |\n| 1 | max | float | FLT_MAX | |\n\n# Concat\n```\ny = concat(x0, x1, x2, ...) by axis\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axis | int | 0 | |\n\n# Convolution\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv(x2, weight, kernel, stride, dilation) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 8 | int8_scale_term| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 18 | pad_value | float | 0.f | |\n| 19 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, kernel_h, num_input, num_output] |\n| bias_data | float | [num_output] |\n| weight_data_int8_scales| float | [num_output] |\n| bottom_blob_int8_scales| float | [1] |\n| top_blob_int8_scales| float | [1] |\n\n# Convolution1D\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv1d(x2, weight, kernel, stride, dilation) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 15 | pad_right | int | pad_left | |\n| 18 | pad_value | float | 0.f | |\n| 19 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# Convolution3D\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv3d(x2, weight, kernel, stride, dilation) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 17 | pad_behind | int | pad_front | |\n| 18 | pad_value | float | 0.f | |\n| 21 | kernel_d | int | kernel_w | |\n| 22 | dilation_d | int | dilation_w | |\n| 23 | stride_d | int | stride_w | |\n| 24 | pad_front | int | pad_left | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, kernel_h, kernel_d, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# ConvolutionDepthWise\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv(x2, weight, kernel, stride, dilation, group) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 8 | int8_scale_term| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 18 | pad_value | float | 0.f | |\n| 19 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, kernel_h, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n| weight_data_int8_scales| float | [group] |\n| bottom_blob_int8_scales| float | [1] |\n| top_blob_int8_scales| float | [1] |\n\n# ConvolutionDepthWise1D\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv1d(x2, weight, kernel, stride, dilation, group) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 15 | pad_right | int | pad_left | |\n| 18 | pad_value | float | 0.f | |\n| 19 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n\n# ConvolutionDepthWise3D\n```\nx2 = pad(x, pads, pad_value)\nx3 = conv3d(x2, weight, kernel, stride, dilation, group) + bias\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 17 | pad_behind | int | pad_front | |\n| 18 | pad_value | float | 0.f | |\n| 21 | kernel_d | int | kernel_w | |\n| 22 | dilation_d | int | dilation_w | |\n| 23 | stride_d | int | stride_w | |\n| 24 | pad_front | int | pad_left | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, kernel_h, kernel_d, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n\n# CopyTo\n```\nself[offset] = src\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | woffset | int | 0 | |\n| 1 | hoffset | int | 0 | |\n| 13 | doffset | int | 0 | |\n| 2 | coffset | int | 0 | |\n| 9 | starts | array | [ ] | |\n| 11 | axes | array | [ ] | |\n\n# Crop\n```\ny = crop(x)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | woffset | int | 0 | |\n| 1 | hoffset | int | 0 | |\n| 13 | doffset | int | 0 | |\n| 2 | coffset | int | 0 | |\n| 3 | outw | int | 0 | |\n| 4 | outh | int | 0 | |\n| 14 | outd | int | 0 | |\n| 5 | outc | int | 0 | |\n| 6 | woffset2 | int | 0 | |\n| 7 | hoffset2 | int | 0 | |\n| 15 | doffset2 | int | 0 | |\n| 8 | coffset2 | int | 0 | |\n| 9 | starts | array | [ ] | |\n| 10 | ends | array | [ ] | |\n| 11 | axes | array | [ ] | |\n| 19 | starts_expr | str | \"\" | |\n| 20 | ends_expr | str | \"\" | |\n| 21 | axes_expr | str | \"\" | |\n\n# CumulativeSum\n\nIf axis < 0, we use axis = x.dims + axis\n\nIt implements https://pytorch.org/docs/stable/generated/torch.cumsum.html\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axis | int | 0 | |\n\n\n# Deconvolution\n```\nx2 = deconv(x, weight, kernel, stride, dilation) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 18 | output_pad_right| int | 0 | |\n| 19 | output_pad_bottom| int | output_pad_right | |\n| 20 | output_w | int | 0 | |\n| 21 | output_h | int | output_w | |\n| 28 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, kernel_h, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# Deconvolution1D\n```\nx2 = deconv1d(x, weight, kernel, stride, dilation) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 15 | pad_right | int | pad_left | |\n| 18 | output_pad_right| int | 0 | |\n| 20 | output_w | int | 0 | |\n| 28 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# Deconvolution3D\n```\nx2 = deconv3d(x, weight, kernel, stride, dilation) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 17 | pad_behind | int | pad_front | |\n| 18 | output_pad_right| int | 0 | |\n| 19 | output_pad_bottom| int | output_pad_right | |\n| 20 | output_pad_behind| int | output_pad_right | |\n| 21 | kernel_d | int | kernel_w | |\n| 22 | dilation_d | int | dilation_w | |\n| 23 | stride_d | int | stride_w | |\n| 24 | pad_front | int | pad_left | |\n| 25 | output_w | int | 0 | |\n| 26 | output_h | int | output_w | |\n| 27 | output_d | int | output_w | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, kernel_h, kernel_d, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# DeconvolutionDepthWise\n```\nx2 = deconv(x, weight, kernel, stride, dilation, group) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 18 | output_pad_right| int | 0 | |\n| 19 | output_pad_bottom| int | output_pad_right | |\n| 20 | output_w | int | 0 | |\n| 21 | output_h | int | output_w | |\n| 28 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, kernel_h, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n\n# DeconvolutionDepthWise1D\n```\nx2 = deconv1d(x, weight, kernel, stride, dilation, group) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 15 | pad_right | int | pad_left | |\n| 18 | output_pad_right| int | 0 | |\n| 20 | output_w | int | 0 | |\n| 28 | dynamic_weight| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n\n# DeconvolutionDepthWise3D\n```\nx2 = deconv3d(x, weight, kernel, stride, dilation, group) + bias\nx3 = depad(x2, pads, pad_value)\ny = activation(x3, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 7 | group | int | 1 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 17 | pad_behind | int | pad_front | |\n| 18 | output_pad_right| int | 0 | |\n| 19 | output_pad_bottom| int | output_pad_right | |\n| 20 | output_pad_behind| int | output_pad_right | |\n| 21 | kernel_d | int | kernel_w | |\n| 22 | dilation_d | int | dilation_w | |\n| 23 | stride_d | int | stride_w | |\n| 24 | pad_front | int | pad_left | |\n| 25 | output_w | int | 0 | |\n| 26 | output_h | int | output_w | |\n| 27 | output_d | int | output_w | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16 | [kernel_w, kernel_h, kernel_d, num_input / group, num_output / group, group] |\n| bias_data | float | [num_output] |\n\n# DeformableConv2D\n```\nx2 = deformableconv2d(x, offset, mask, weight, kernel, stride, dilation) + bias\ny = activation(x2, act_type, act_params)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 5 | bias_term | int | 0 | |\n| 6 | weight_data_size| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [kernel_w, kernel_h, num_input, num_output] |\n| bias_data | float | [num_output] |\n\n# Dequantize\n```\ny = x * scale + bias\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | scale_data_size| int | 1 | |\n| 1 | bias_data_size| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| scale_data | float | [scale_data_size] |\n| bias_data | float | [bias_data_size] |\n\n# Diag\n```\ny = diag(x, diagonal)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | diagonal | int | 0 | |\n\n# Dropout\n```\ny = x * scale\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | scale | float | 1.f | |\n\n# Eltwise\n```\ny = elementwise_op(x0, x1, ...)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | op_type | int | 0 | |\n| 1 | coeffs | array | [ ] | |\n\nOperation type:\n- 0 = PROD\n- 1 = SUM\n- 2 = MAX\n\n# ELU\n```\nif x < 0 y = (exp(x) - 1) * alpha\nelse y = x\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 0.1f | |\n\n# Embed\n```\ny = embedding(x)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | input_dim | int | 0 | |\n| 2 | bias_term | int | 0 | |\n| 3 | weight_data_size | int | 0 | |\n| 18 | int8_scale_term| int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float | [weight_data_size] |\n| bias_term | float | [num_output] |\n| weight_data_int8_scales| float | [1] |\n\n# Exp\n```\nif base == -1 y = exp(shift + x * scale)\nelse y = pow(base, (shift + x * scale))\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | base | float | -1.f | |\n| 1 | scale | float | 1.f | |\n| 2 | shift | float | 0.f | |\n\n# ExpandDims\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 3 | axes | array | [ ] | |\n\n# Flatten\nReshape blob to 1 dimension\n\n* one_blob_only\n\n# Flip\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axes | array | [ ] | |\n\n# Fold\n```\ny = fold(x)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n| 20 | output_w | int | 0 | |\n| 21 | output_h | int | output_w | |\n\n# GELU\n```\nif fast_gelu == 1 y = 0.5 * x * (1 + tanh(0.79788452 * (x + 0.044715 * x * x * x)));\nelse y = 0.5 * x * erfc(-0.70710678 * x)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | fast_gelu | int | 0 | use approximation |\n\n# GLU\n\nIf axis < 0, we use axis = x.dims + axis\n\nGLU(a,b)=a⊗σ(b)\n\nwhere a is the first half of the input matrix and b is the second half.\n\naxis specifies the dimension to split the input\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axis | int | 0 | |\n\n# Gemm\n```\na = transA ? transpose(x0) : x0\nb = transb ? transpose(x1) : x1\nc = x2\ny = (gemm(a, b) + c * beta) * alpha\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 1.f | |\n| 1 | beta | float | 1.f | |\n| 2 | transA | int | 0 | |\n| 3 | transb | int | 0 | |\n| 4 | constantA | int | 0 | |\n| 5 | constantB | int | 0 | |\n| 6 | constantC | int | 0 | |\n| 7 | constantM | int | 0 | |\n| 8 | constantN | int | 0 | |\n| 9 | constantK | int | 0 | |\n| 10 | constant_broadcast_type_C | int | 0 | |\n| 11 | output_N1M | int | 0 | |\n| 12 | output_elempack | int | 0 | |\n| 13 | output_elemtype | int | 0 | |\n| 14 | output_transpose | int| 0 | |\n| 18 | int8_scale_term | int | 0 | |\n| 20 | constant_TILE_M | int | 0 | |\n| 21 | constant_TILE_N | int | 0 | |\n| 22 | constant_TILE_K | int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| A_data | float/fp16/int8 | [M, K] or [K, M] |\n| B_data | float/fp16/int8 | [N, K] or [K, N] |\n| C_data | float | [1], [M] or [N] or [1, M] or [N,1] or [N, M] |\n| A_data_int8_scales| float | [M] |\n| B_data_int8_scales| float | [1] |\n\n# GridSample\n```\nGiven an input and a flow-field grid, computes the output using input values and pixel locations from grid.\n\nFor each output location output[:, h2, w2], the size-2 vector grid[h2, w2, 2] specifies input pixel[:, h1, w1] locations x and y, \nwhich are used to interpolate the output value output[:, h2, w2]\n\nThis function is often used in conjunction with affine_grid() to build Spatial Transformer Networks .\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | sample_type | int | 1 | |\n| 1 | padding_mode | int | 1 | |\n| 2 | align_corner | int | 0 | |\n| 3 | permute_fusion| int | 0 | fuse with permute |\n\n\nSample type:\n- 1 = Nearest\n- 2 = Bilinear\n- 3 = Bicubic\n\nPadding mode:\n- 1 = zeros\n- 2 = border\n- 3 = reflection\n\n\n# GroupNorm\n```\nsplit x along channel axis into group x0, x1 ...\nl2 normalize for each group x0, x1 ...\ny = x * gamma + beta\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | group | int | 1 | |\n| 1 | channels | int | 0 | |\n| 2 | eps | float | 0.001f | x = x / sqrt(var + eps) |\n| 3 | affine | int | 1 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| gamma_data | float | [channels] |\n| beta_data | float | [channels] |\n\n# GRU\nApply a single-layer GRU to a feature sequence of `T` timesteps. The input blob shape is `[w=input_size, h=T]` and the output blob shape is `[w=num_output, h=T]`.\n\n```\ny = gru(x)\ny0, hidden y1 = gru(x0, hidden x1)\n```\n\n* one_blob_only if bidirectional\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | hidden size of output |\n| 1 | weight_data_size| int | 0 | total size of weight matrix |\n| 2 | direction | int | 0 | 0=forward, 1=reverse, 2=bidirectional |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_xc_data| float/fp16/int8 | [input_size, num_output * 3, num_directions] |\n| bias_c_data | float/fp16/int8 | [num_output, 4, num_directions] |\n| weight_hc_data| float/fp16/int8 | [num_output, num_output * 3, num_directions] |\n\nDirection flag:\n- 0 = forward only\n- 1 = reverse only\n- 2 = bidirectional\n\n# HardSigmoid\n```\ny = clamp(x * alpha + beta, 0, 1)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 0.2f | |\n| 1 | beta | float | 0.5f | |\n\n# HardSwish\n```\ny = x * clamp(x * alpha + beta, 0, 1)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 0.2f | |\n| 1 | beta | float | 0.5f | |\n\n# InnerProduct\n```\nx2 = innerproduct(x, weight) + bias\ny = activation(x2, act_type, act_params)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | bias_term | int | 0 | |\n| 2 | weight_data_size| int | 0 | |\n| 8 | int8_scale_term| int | 0 | |\n| 9 | activation_type| int | 0 | |\n| 10 | activation_params| array | [ ] | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_data | float/fp16/int8 | [num_input, num_output] |\n| bias_data | float | [num_output] |\n| weight_data_int8_scales| float | [num_output] |\n| bottom_blob_int8_scales| float | [1] |\n\n# Input\n```\ny = input\n```\n\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | w | int | 0 | |\n| 1 | h | int | 0 | |\n| 11 | d | int | 0 | |\n| 2 | c | int | 0 | |\n\n# InstanceNorm\n```\nsplit x along channel axis into instance x0, x1 ...\nl2 normalize for each channel instance x0, x1 ...\ny = x * gamma + beta\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | channels | int | 0 | |\n| 1 | eps | float | 0.001f | x = x / sqrt(var + eps) |\n| 2 | affine | int | 1 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| gamma_data | float | [channels] |\n| beta_data | float | [channels] |\n\n# Interp\n```\nif dynamic_target_size == 0 y = resize(x) by fixed size or scale\nelse y = resize(x0, size(x1))\n```\n\n* one_blob_only if dynamic_target_size == 0\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | resize_type | int | 0 | |\n| 1 | height_scale | float | 1.f | |\n| 2 | width_scale | float | 1.f | |\n| 3 | output_height | int | 0 | |\n| 4 | output_width | int | 0 | |\n| 5 | dynamic_target_size| int | 0 | |\n| 6 | align_corner | int | 0 | |\n| 9 | size_expr | str | \"\" | |\n\nResize type:\n- 1 = Nearest\n- 2 = Bilinear\n- 3 = Bicubic\n\n# InverseSpectrogram\n```\nx1 = x as complex\nx1 = x1 * sqrt(norm) if normalized\ny = istft(x1)\ny1 = unpad(y) if center\n\nif returns == 0 return y1 as complex\nif returns == 1 return y1 real\nif returns == 2 return y1 imag\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | n_fft | int | 0 | |\n| 1 | returns | int | 1 | |\n| 2 | hoplen | int | n_fft / 4 | |\n| 3 | winlen | int | n_fft | |\n| 4 | window_type | int | 0 | 0=ones 1=hann 2=hamming |\n| 5 | center | int | 1 | |\n| 7 | normalized | int | 0 | 0=no 1=n_fft 2=window-l2-energy |\n\n# LayerNorm\n```\nsplit x along outmost axis into part x0, x1 ...\nl2 normalize for each part x0, x1 ...\ny = x * gamma + beta by elementwise\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | affine_size | int | 0 | |\n| 1 | eps | float | 0.001f | x = x / sqrt(var + eps) |\n| 2 | affine | int | 1 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| gamma_data | float | [affine_size] |\n| beta_data | float | [affine_size] |\n\n# Log\n```\nif base == -1 y = log(shift + x * scale)\nelse y = log(shift + x * scale) / log(base)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | base | float | -1.f | |\n| 1 | scale | float | 1.f | |\n| 2 | shift | float | 0.f | |\n\n# LRN\n```\nif region_type == ACROSS_CHANNELS square_sum = sum of channel window of local_size\nif region_type == WITHIN_CHANNEL square_sum = sum of spatial window of local_size\ny = x * pow(bias + alpha * square_sum / (local_size * local_size), -beta)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | region_type | int | 0 | |\n| 1 | local_size | int | 5 | |\n| 2 | alpha | float | 1.f | |\n| 3 | beta | float | 0.75f | |\n| 4 | bias | float | 1.f | |\n\nRegion type:\n- 0 = ACROSS_CHANNELS\n- 1 = WITHIN_CHANNEL\n\n# LSTM\nApply a single-layer LSTM to a feature sequence of `T` timesteps. The input blob shape is `[w=input_size, h=T]` and the output blob shape is `[w=num_output, h=T]`.\n\n```\ny = lstm(x)\ny0, hidden y1, cell y2 = lstm(x0, hidden x1, cell x2)\n```\n\n* one_blob_only if bidirectional\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | output size of output |\n| 1 | weight_data_size| int | 0 | total size of IFOG weight matrix |\n| 2 | direction | int | 0 | 0=forward, 1=reverse, 2=bidirectional |\n| 3 | hidden_size | int | num_output| hidden size |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_xc_data| float/fp16/int8 | [input_size, hidden_size * 4, num_directions] |\n| bias_c_data | float/fp16/int8 | [hidden_size, 4, num_directions] |\n| weight_hc_data| float/fp16/int8 | [num_output, hidden_size * 4, num_directions] |\n| weight_hr_data| float/fp16/int8 | [hidden_size, num_output, num_directions] |\n\nDirection flag:\n- 0 = forward only\n- 1 = reverse only\n- 2 = bidirectional\n\n# MemoryData\n```\ny = data\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | w | int | 0 | |\n| 1 | h | int | 0 | |\n| 11 | d | int | 0 | |\n| 2 | c | int | 0 | |\n| 21 | load_type | int | 1 | 1=fp32 |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| data | float | [w, h, d, c] |\n\n# Mish\n```\ny = x * tanh(log(exp(x) + 1))\n```\n\n* one_blob_only\n* support_inplace\n\n# MultiHeadAttention\n```\nq_affine = affine(q) / (embed_dim / num_head)\nk_affine = affine(k) or reuse kv_cache part\nv_affine = affine(v) or reuse kv_cache part\nsplit q k v into num_head part q0, k0, v0, q1, k1, v1 ...\nfor each num_head part\n qk = q * k\n qk = qk + attn_mask if attn_mask exists\n softmax(qk)\n qkv = qk * v\n merge qkv to out\ny = affine(out)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | embed_dim | int | 0 | |\n| 1 | num_heads | int | 1 | |\n| 2 | weight_data_size| int | 0 | qdim = weight_data_size / embed_dim |\n| 3 | kdim | int | embed_dim | |\n| 4 | vdim | int | embed_dim | |\n| 5 | attn_mask | int | 0 | |\n| 6 | scale | float | 1.f / sqrt(embed_dim / num_heads) | |\n| 7 | kv_cache | int | 0 | |\n| 18 | int8_scale_term | int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| q_weight_data | float/fp16/int8 | [embed_dim * qdim] |\n| q_bias_data | float | [embed_dim] |\n| k_weight_data | float/fp16/int8 | [embed_dim * kdim] |\n| k_bias_data | float | [embed_dim] |\n| v_weight_data | float/fp16/int8 | [embed_dim * vdim] |\n| v_bias_data | float | [embed_dim] |\n| out_weight_data| float/fp16/int8 | [qdim * embed_dim] |\n| out_bias_data | float | [qdim] |\n| q_weight_data_int8_scales| float | [embed_dim] |\n| k_weight_data_int8_scales| float | [embed_dim] |\n| v_weight_data_int8_scales| float | [embed_dim] |\n| out_weight_data_int8_scales| float | [1] |\n\n# MVN\n```\nif normalize_variance == 1 && across_channels == 1 y = (x - mean) / (sqrt(var) + eps) of whole blob\nif normalize_variance == 1 && across_channels == 0 y = (x - mean) / (sqrt(var) + eps) of each channel\nif normalize_variance == 0 && across_channels == 1 y = x - mean of whole blob\nif normalize_variance == 0 && across_channels == 0 y = x - mean of each channel\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | normalize_variance| int | 0 | |\n| 1 | across_channels| int | 0 | |\n| 2 | eps | float | 0.0001f | x = x / (sqrt(var) + eps) |\n\n# Noop\n```\ny = x\n```\n\n# Normalize\n```\nif across_spatial == 1 && across_channel == 1 x2 = normalize(x) of whole blob\nif across_spatial == 1 && across_channel == 0 x2 = normalize(x) of each channel\nif across_spatial == 0 && across_channel == 1 x2 = normalize(x) of each position\ny = x2 * scale\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | across_spatial| int | 0 | |\n| 1 | channel_shared| int | 0 | |\n| 2 | eps | float | 0.0001f | see eps mode |\n| 3 | scale_data_size| int | 0 | |\n| 4 | across_channel| int | 0 | |\n| 9 | eps_mode | int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| scale_data | float | [scale_data_size] |\n\nEps Mode:\n- 0 = caffe/mxnet x = x / sqrt(var + eps)\n- 1 = pytorch x = x / max(sqrt(var), eps)\n- 2 = tensorflow x = x / sqrt(max(var, eps))\n\n# Packing\n```\ny = wrap_packing(x)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | out_elempack | int | 1 | |\n| 1 | use_padding | int | 0 | |\n| 2 | cast_type_from| int | 0 | |\n| 3 | cast_type_to | int | 0 | |\n| 4 | storage_type_from| int | 0 | |\n| 5 | storage_type_to| int | 0 | |\n\n# Padding\n```\ny = pad(x, pads)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ---- | --------- | ----------------- |\n| 0 | top | int | 0 | |\n| 1 | bottom | int | 0 | |\n| 2 | left | int | 0 | |\n| 3 | right | int | 0 | |\n| 4 | type | int | 0 | |\n| 5 | value | float | 0 | |\n| 6 | per_channel_pad_data_size| int | 0 | |\n| 7 | front | int | stride_w | |\n| 8 | behind | int | pad_left | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| per_channel_pad_data| float | [per_channel_pad_data_size] |\n\nPadding type:\n- 0 = CONSTANT\n- 1 = REPLICATE\n- 2 = REFLECT\n\n# Permute\n```\ny = reorder(x)\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ---- | --------- | ----------------- |\n| 0 | order_type | int | 0 | |\n\nOrder Type:\n- 0 = WH WHC WHDC\n- 1 = HW HWC HWDC\n- 2 = WCH WDHC\n- 3 = CWH DWHC\n- 4 = HCW HDWC\n- 5 = CHW DHWC\n- 6 = WHCD\n- 7 = HWCD\n- 8 = WCHD\n- 9 = CWHD\n- 10 = HCWD\n- 11 = CHWD\n- 12 = WDCH\n- 13 = DWCH\n- 14 = WCDH\n- 15 = CWDH\n- 16 = DCWH\n- 17 = CDWH\n- 18 = HDCW\n- 19 = DHCW\n- 20 = HCDW\n- 21 = CHDW\n- 22 = DCHW\n- 23 = CDHW\n\n# PixelShuffle\n```\nif mode == 0 y = depth_to_space(x) where x channel order is sw-sh-outc\nif mode == 1 y = depth_to_space(x) where x channel order is outc-sw-sh\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ---- | --------- | ----------------- |\n| 0 | upscale_factor| int | 1 | |\n| 1 | mode | int | 0 | |\n\n# Pooling\n```\nx2 = pad(x, pads)\nx3 = pooling(x2, kernel, stride)\n```\n\n| param id | name | type | default | description |\n| --------- | --------------| ---- | --------- | ----------------- |\n| 0 | pooling_type | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | stride_w | int | 1 | |\n| 3 | pad_left | int | 0 | |\n| 4 | global_pooling| int | 0 | |\n| 5 | pad_mode | int | 0 | |\n| 6 | avgpool_count_include_pad| int | 0 | |\n| 7 | adaptive_pooling| int | 0 | |\n| 8 | out_w | int | 0 | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | stride_h | int | stride_w | |\n| 13 | pad_top | int | pad_left | |\n| 14 | pad_right | int | pad_left | |\n| 15 | pad_bottom | int | pad_top | |\n| 18 | out_h | int | out_w | |\n\nPooling type:\n- 0 = MAX\n- 1 = AVG\n\nPad mode:\n- 0 = full padding\n- 1 = valid padding\n- 2 = tensorflow padding=SAME or onnx padding=SAME_UPPER\n- 3 = onnx padding=SAME_LOWER\n\n# Pooling1D\n```\nx2 = pad(x, pads)\nx3 = pooling1d(x2, kernel, stride)\n```\n\n| param id | name | type | default | description |\n| --------- | --------------| ---- | --------- | ----------------- |\n| 0 | pooling_type | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | stride_w | int | 1 | |\n| 3 | pad_left | int | 0 | |\n| 4 | global_pooling| int | 0 | |\n| 5 | pad_mode | int | 0 | |\n| 6 | avgpool_count_include_pad| int | 0 | |\n| 7 | adaptive_pooling| int | 0 | |\n| 8 | out_w | int | 0 | |\n| 14 | pad_right | int | pad_left | |\n\nPooling type:\n- 0 = MAX\n- 1 = AVG\n\nPad mode:\n- 0 = full padding\n- 1 = valid padding\n- 2 = tensorflow padding=SAME or onnx padding=SAME_UPPER\n- 3 = onnx padding=SAME_LOWER\n\n# Pooling3D\n```\nx2 = pad(x, pads)\nx3 = pooling3d(x2, kernel, stride)\n```\n\n| param id | name | type | default | description |\n| --------- | --------------| ---- | --------- | ----------------- |\n| 0 | pooling_type | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | stride_w | int | 1 | |\n| 3 | pad_left | int | 0 | |\n| 4 | global_pooling| int | 0 | |\n| 5 | pad_mode | int | 0 | |\n| 6 | avgpool_count_include_pad| int | 0 | |\n| 7 | adaptive_pooling| int | 0 | |\n| 8 | out_w | int | 0 | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | stride_h | int | stride_w | |\n| 13 | pad_top | int | pad_left | |\n| 14 | pad_right | int | pad_left | |\n| 15 | pad_bottom | int | pad_top | |\n| 16 | pad_behind | int | pad_front | |\n| 18 | out_h | int | out_w | |\n| 21 | kernel_d | int | kernel_w | |\n| 22 | stride_d | int | stride_w | |\n| 23 | pad_front | int | pad_left | |\n| 28 | out_d | int | out_w | |\n\nPooling type:\n- 0 = MAX\n- 1 = AVG\n\nPad mode:\n- 0 = full padding\n- 1 = valid padding\n- 2 = tensorflow padding=SAME or onnx padding=SAME_UPPER\n- 3 = onnx padding=SAME_LOWER\n\n# Power\n```\ny = pow((shift + x * scale), power)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | power | float | 1.f | |\n| 1 | scale | float | 1.f | |\n| 2 | shift | float | 0.f | |\n\n# PReLU\n```\nif x < 0 y = x * slope\nelse y = x\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_slope | int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| slope_data | float | [num_slope] |\n\n# Quantize\n```\ny = float2int8(x * scale)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | scale_data_size| int | 1 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| scale_data | float | [scale_data_size] |\n\n# Reduction\n```\ny = reduce_op(x * coeff)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | operation | int | 0 | |\n| 1 | reduce_all | int | 1 | |\n| 2 | coeff | float | 1.f | |\n| 3 | axes | array | [ ] | |\n| 4 | keepdims | int | 0 | |\n| 5 | fixbug0 | int | 0 | hack for bug fix, should be 1 |\n\nOperation type:\n- 0 = SUM\n- 1 = ASUM\n- 2 = SUMSQ\n- 3 = MEAN\n- 4 = MAX\n- 5 = MIN\n- 6 = PROD\n- 7 = L1\n- 8 = L2\n- 9 = LogSum\n- 10 = LogSumExp\n\n# ReLU\n```\nif x < 0 y = x * slope\nelse y = x\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | slope | float | 0.f | |\n\n# Reorg\n```\nif mode == 0 y = space_to_depth(x) where x channel order is sw-sh-outc\nif mode == 1 y = space_to_depth(x) where x channel order is outc-sw-sh\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ---- | --------- | ----------------- |\n| 0 | stride | int | 1 | |\n| 1 | mode | int | 0 | |\n\n# Requantize\n```\nx2 = x * scale_in + bias\nx3 = activation(x2)\ny = float2int8(x3 * scale_out)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | scale_in_data_size| int | 1 | |\n| 1 | scale_out_data_size| int | 1 | |\n| 2 | bias_data_size| int | 0 | |\n| 3 | activation_type| int | 0 | |\n| 4 | activation_params| int | [ ] | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| scale_in_data | float | [scale_in_data_size] |\n| scale_out_data| float | [scale_out_data_size] |\n| bias_data | float | [bias_data_size] |\n\n# Reshape\n```\ny = reshape(x)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | w | int | -233 | |\n| 1 | h | int | -233 | |\n| 11 | d | int | -233 | |\n| 2 | c | int | -233 | |\n| 6 | shape_expr | str | \"\" | |\n\nReshape flag:\n- 0 = copy from bottom\n- -1 = remaining\n- -233 = drop this dim(default)\n\n# RMSNorm\n```\nsplit x along outmost axis into part x0, x1 ...\nroot mean square normalize for each part x0, x1 ...\ny = x * gamma by elementwise\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | affine_size | int | 0 | |\n| 1 | eps | float | 0.001f | x = x / sqrt(var + eps) |\n| 2 | affine | int | 1 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| gamma_data | float | [affine_size] |\n\n# RNN\nApply a single-layer RNN to a feature sequence of `T` timesteps. The input blob shape is `[w=input_size, h=T]` and the output blob shape is `[w=num_output, h=T]`.\n\n```\ny = rnn(x)\ny0, hidden y1 = rnn(x0, hidden x1)\n```\n\n* one_blob_only if bidirectional\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | hidden size of output |\n| 1 | weight_data_size| int | 0 | total size of weight matrix |\n| 2 | direction | int | 0 | 0=forward, 1=reverse, 2=bidirectional |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| weight_xc_data| float/fp16/int8 | [input_size, num_output, num_directions] |\n| bias_c_data | float/fp16/int8 | [num_output, 1, num_directions] |\n| weight_hc_data| float/fp16/int8 | [num_output, num_output, num_directions] |\n\nDirection flag:\n- 0 = forward only\n- 1 = reverse only\n- 2 = bidirectional\n\n# RotaryEmbed\nApply rotary positional embeddings with cos and sin cache\n\n```\ny1 = x1 * cos - x2 * sin\ny2 = x1 * sin + x2 * cos\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | interleaved | int | 0 | |\n\n# Scale\n```\nif scale_data_size == -233 y = x0 * x1\nelse y = x * scale + bias\n```\n\n* one_blob_only if scale_data_size != -233\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | scale_data_size| int | 0 | |\n| 1 | bias_term | int | 0 | |\n\n| weight | type | shape |\n| ------------- | ----- | --------------------- |\n| scale_data | float | [scale_data_size] |\n| bias_data | float | [scale_data_size] |\n\n# SDPA\n```\nscaled dot product attention\nfor each num_head part\n qk = q * k\n qk = qk + attn_mask if attn_mask exists\n softmax(qk)\n qkv = qk * v\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 5 | attn_mask | int | 0 | |\n| 6 | scale | float | 0.f | auto = 1.f / sqrt(embed_dim) |\n| 7 | kv_cache | int | 0 | |\n| 18 | int8_scale_term | int | 0 | |\n\n# SELU\n```\nif x < 0 y = (exp(x) - 1.f) * alpha * lambda\nelse y = x * lambda\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | alpha | float | 1.67326324f| |\n| 1 | lambda | float | 1.050700987f| |\n\n# Shrink\n```\nif x < -lambd y = x + bias\nif x > lambd y = x - bias\nelse y = x\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | bias | float | 0.0f | |\n| 1 | lambd | float | 0.5f | |\n\n# ShuffleChannel\n```\nif reverse == 0 y = shufflechannel(x) by group\nif reverse == 1 y = shufflechannel(x) by channel / group\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ---- | --------- | ----------------- |\n| 0 | group | int | 1 | |\n| 1 | reverse | int | 0 | |\n\n# Sigmoid\n```\ny = 1 / (1 + exp(-x))\n```\n\n* one_blob_only\n* support_inplace\n\n# Slice\n```\nsplit x along axis into slices, each part slice size is based on slices array\n```\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | slices | array | [ ] | |\n| 1 | axis | int | 0 | |\n| 2 | indices | array | [ ] | |\n\n# Softmax\n```\nsoftmax(x, axis)\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axis | int | 0 | |\n| 1 | fixbug0 | int | 0 | hack for bug fix, should be 1 |\n\n# Softplus\n```\ny = log(exp(x) + 1)\n```\n\n* one_blob_only\n* support_inplace\n\n# Spectrogram\n```\nx1 = pad(x) if center\ny = stft(x1)\ny = y / sqrt(norm) if normalized\n\nif power == 0 return y as real\nif power == 1 return magnitude\nif power == 2 return square of magnitude\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | n_fft | int | 0 | |\n| 1 | power | int | 0 | |\n| 2 | hoplen | int | n_fft / 4 | |\n| 3 | winlen | int | n_fft | |\n| 4 | window_type | int | 0 | 0=ones 1=hann 2=hamming |\n| 5 | center | int | 1 | |\n| 6 | pad_type | int | 2 | 0=CONSTANT 1=REPLICATE 2=REFLECT |\n| 7 | normalized | int | 0 | 0=no 1=n_fft 2=window-l2-energy |\n| 8 | onesided | int | 1 | |\n\n# Split\n```\ny0, y1 ... = x\n```\n\n# Squeeze\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | squeeze_w | int | 0 | |\n| 1 | squeeze_h | int | 0 | |\n| 11 | squeeze_d | int | 0 | |\n| 2 | squeeze_c | int | 0 | |\n| 3 | axes | array | [ ] | |\n\n# Swish\n```\ny = x / (1 + exp(-x))\n```\n\n* one_blob_only\n* support_inplace\n\n# TanH\n```\ny = tanh(x)\n```\n\n* one_blob_only\n* support_inplace\n\n# Threshold\n```\nif x > threshold y = 1\nelse y = 0\n```\n\n* one_blob_only\n* support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | threshold | float | 0.f | |\n\n# Tile\n```\ny = repeat tiles along axis for x\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | axis | int | 0 | |\n| 1 | tiles | int | 1 | |\n| 2 | repeats | array | [ ] | |\n\n# UnaryOp\n```\ny = unaryop(x)\n```\n\n- one_blob_only\n- support_inplace\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | op_type | int | 0 | Operation type as follows |\n\nOperation type:\n- 0 = ABS\n- 1 = NEG\n- 2 = FLOOR\n- 3 = CEIL\n- 4 = SQUARE\n- 5 = SQRT\n- 6 = RSQ\n- 7 = EXP\n- 8 = LOG\n- 9 = SIN\n- 10 = COS\n- 11 = TAN\n- 12 = ASIN\n- 13 = ACOS\n- 14 = ATAN\n- 15 = RECIPROCAL\n- 16 = TANH\n- 17 = LOG10\n- 18 = ROUND\n- 19 = TRUNC\n\n# Unfold\n```\ny = unfold(x)\n```\n\n* one_blob_only\n\n| param id | name | type | default | description |\n| --------- | ------------- | ----- | --------- | ----------------- |\n| 0 | num_output | int | 0 | |\n| 1 | kernel_w | int | 0 | |\n| 2 | dilation_w | int | 1 | |\n| 3 | stride_w | int | 1 | |\n| 4 | pad_left | int | 0 | |\n| 11 | kernel_h | int | kernel_w | |\n| 12 | dilation_h | int | dilation_w | |\n| 13 | stride_h | int | stride_w | |\n| 14 | pad_top | int | pad_left | |\n| 15 | pad_right | int | pad_left | |\n| 16 | pad_bottom | int | pad_top | |\n" }, { "path": "docs/developer-guide/param-and-model-file-structure.md", "content": "## net.param\n### example\n```\n7767517\n3 3\nInput input 0 1 data 0=4 1=4 2=1\nInnerProduct ip 1 1 data fc 0=10 1=1 2=80\nSoftmax softmax 1 1 fc prob 0=0\n```\n### overview\n```\n[magic]\n```\n* magic number : 7767517\n```\n[layer count] [blob count]\n```\n* layer count : count of the layer line follows, should be exactly the count of all layer names\n* blob count : count of all blobs, usually greater than or equals to the layer count\n### layer line\n```\n[layer type] [layer name] [input count] [output count] [input blobs] [output blobs] [layer specific params]\n```\n* layer type : type name, such as Convolution Softmax etc\n* layer name : name of this layer, must be unique among all layer names\n* input count : count of the blobs this layer needs as input\n* output count : count of the blobs this layer produces as output\n* input blobs : name list of all the input blob names, separated by space, must be unique among input blob names of all layers\n* output blobs : name list of all the output blob names, separated by space, must be unique among output blob names of all layers\n* layer specific params : key=value pair list, separated by space\n### layer param\n```\n0=1 1=2.5 -23303=2,2.0,3.0\n```\nkey index should be unique in each layer line, pair can be omitted if the default value used\n\nthe meaning of existing param key index can be looked up at [operation-param-weight-table](operation-param-weight-table)\n\n* integer or float key : index 0 ~ 19\n* integer value : int\n* float value : float\n* integer array or float array key : -23300 minus index 0 ~ 19\n* integer array value : [array size],int,int,...,int\n* float array value : [array size],float,float,...,float\n\nIn modern ncnn param file\n\n* array could be represented as `3=2.0,3.0` that is much more human friendly\n* string typed value: `4=hello` and the string is no longer than 255\n\n## net.bin\n```\n +---------+---------+---------+---------+---------+---------+\n | weight1 | weight2 | weight3 | weight4 | ....... | weightN |\n +---------+---------+---------+---------+---------+---------+\n ^ ^ ^ ^\n 0x0 0x80 0x140 0x1C0\n```\nthe model binary is the concatenation of all weight data, each weight buffer is aligned by 32bit\n\n### weight buffer\n```\n[flag] (optional)\n[raw data]\n[padding] (optional)\n```\n* flag : unsigned int, little-endian, indicating the weight storage type, 0 => float32, 0x01306B47 => float16, otherwise => quantized int8, may be omitted if the layer implementation forced the storage type explicitly\n* raw data : raw weight data, little-endian, float32 data or float16 data or quantized table and indexes depending on the storage type flag\n* padding : padding space for 32bit alignment, may be omitted if already aligned\n" }, { "path": "docs/developer-guide/preload-practice.zh.md", "content": "## 只是实践经验，没有理论，不一定正确\n\n```\nprfm pldl1keep, [x0, #256]\n```\n* 放在 ld1 [x0] 前面 0~8 条指令\n* #256 表示把 x0+256 的内容放进 L1 cache\n* ldp 也适用\n* (经验)不写 offset 不如写个 #128\n* (经验)pldl1strm 似乎没啥意思，也没 pldl1keep 快\n* (经验)x0 ~ x0+256 的内容也会进来\n* (经验)load 128bit 用 #128，256bit或更多用 #256\n* (经验)避免 pld a，pld b，load a，load b 顺序，可能相互干扰\n* (经验)提前太多会失效\n* (经验)适合连续读\n\n```\nprfm pldl2strm, [x0, #256]\n```\n* 放在 ld1 [x0] 前面 N 条指令，N 尽量大些\n* #256 表示把 x0+256 的内容放进 L2 cache\n* ldp 也适用\n* (经验)不写 offset 不如写个 #128\n* (经验)pldl2strm 效果稍好于 pldl2keep\n* (经验)x0 ~ x0+256 的内容也会进来\n* (经验)load 128bit 用 #128，256bit 用 #256\n* (经验)读很多数据，用不同 offset 连续两次 pldl2strm\n* (经验)后面不要对同位置再 pldl1keep，会变慢\n* (经验)适合提前准备要跳到很远的地方读，比如换 channel\n" }, { "path": "docs/developer-guide/tensorflow-op-combination.md", "content": "## batchnorm\n```\nInput A 0 1 A 0 0 0\nMemoryData sub/y 0 1 sub/y 16 0 0\nBinaryOp sub 2 1 A sub/y sub 1\nMemoryData div/y 0 1 div/y 16 0 0\nBinaryOp div 2 1 sub div/y div 3\nMemoryData mul/y 0 1 mul/y 16 0 0\nBinaryOp mul 2 1 div mul/y mul 2\nMemoryData BiasAdd/bias 0 1 BiasAdd/bias 16 0 0\nBinaryOp BiasAdd 2 1 mul BiasAdd/bias BiasAdd 0\n```\n## convolution\n```\nInput A 0 1 A 0 0 0\nConvolution Conv2D 1 1 A Conv2D 10 3 1 1 0 0 270\nMemoryData biases/read 0 1 biases/read 10 0 0\nBinaryOp BiasAdd 2 1 Conv2D biases/read BiasAdd 0\n```\n## innerproduct\n```\nInput A 0 1 A 0 0 0\nMemoryData biases/read 0 1 biases/read 10 0 0\nInnerProduct MatMul 1 1 A MatMul 10 0 2560\nBinaryOp conv6 2 1 MatMul biases/read conv6 0\n```\n## leakyrelu\n```\nInput A 0 1 A 0 0 0\nSplit splitncnn_0 1 2 A A_splitncnn_0 A_splitncnn_1\nMemoryData mul_1/x 0 1 mul_1/x 0 0 0\nBinaryOp mul_1 2 1 mul_1/x A_splitncnn_1 mul_1 2\nBinaryOp leaky 2 1 mul_1 A_splitncnn_0 leaky 4\n```\n## prelu\n```\nInput A 0 1 A 0 0 0\nSplit splitncnn_0 1 2 A A_splitncnn_0 A_splitncnn_1\nMemoryData prelu/alpha 0 1 prelu/alpha 10 0 0\nReLU prelu/Relu 1 1 A_splitncnn_1 prelu/Relu 0.000000\nUnaryOp prelu/Neg 1 1 A_splitncnn_0 prelu/Neg 1\nReLU prelu/Relu_1 1 1 prelu/Neg prelu/Relu_1 0.000000\nUnaryOp prelu/Neg_1 1 1 prelu/Relu_1 prelu/Neg_1 1\nBinaryOp prelu/Mul 2 1 prelu/alpha prelu/Neg_1 prelu/Mul 2\nBinaryOp prelu/add 2 1 prelu/Relu prelu/Mul prelu/add 0\n```\n## softmax\n```\nInput A 0 1 A 0 0 0\nSplit splitncnn_4 1 2 A A_splitncnn_0 A_splitncnn_1\nReduction Max 1 1 A_splitncnn_1 Max 4 -2 1.000000\nBinaryOp sub 2 1 A_splitncnn_0 Max sub 1\nUnaryOp Exp 1 1 sub Exp 7\nSplit splitncnn_5 1 2 Exp Exp_splitncnn_0 Exp_splitncnn_1\nReduction Sum 1 1 Exp_splitncnn_1 Sum 0 -2 1.000000\nBinaryOp prob 2 1 Exp_splitncnn_0 Sum prob 3\n```" }, { "path": "docs/developer-guide/vulkan-driver-loader.md", "content": "# ncnn vulkan driver loader\n\nncnn turns on the ```NCNN_SIMPLEVK``` cmake option by default, when ```NCNN_VULKAN``` is enabled\n\nsimplevk is ncnn's built-in vulkan loader. It provides vulkan function declarations and function entries that meet ncnn's needs. It allows the use and compilation of vulkan-related codes without relying on vulkan-sdk. It can dynamically load the vulkan runtime library at runtime or directly load the graphics card driver. vulkan driver. When distributing ncnn applications, it is not required that the target system has a vulkan driver.\n\nUsually you don't need to care about how simplevk loads the vulkan driver, because ncnn will automatically load and initialize when using vulkan related functions. It is sufficient to set the `Option` switch before loading the model.\n\nTypical code\n\n```cpp\nncnn::Net net;\nnet.opt.use_vulkan_compute = true;\nnet.load_param(\"model.param\");\nnet.load_param(\"model.bin\");\n```\n\nUsing the in-house vulkan loader instead of the standard libvulkan has the following benefits\n\n- Can compile ncnn vulkan code without installing vulkan-sdk\n- Can deploy and distribute applications without libvulkan linkage\n- Can load external vulkan driver instead of system driver\n- Can directly load android hal module\n- Can directly load graphics card driver files via NCNN_VULKAN_DRIVER env\n- Able to actively search for graphics card driver files in the system and load them\n- Can compile android libraries supporting vulkan under the platform of android-api<24\n\n## Create and manage gpu context\n\n```cpp\nint create_gpu_instance(const char* driver_path = 0);\n\nvoid destroy_gpu_instance();\n\nVkInstance get_gpu_instance();\n```\n\n## Loading order\n\n```\nIf driver_path == 0\n 1a from env ```VK_ICD_FILENAMES```\n 1b from env ```NCNN_VULKAN_DRIVER```\n\nIf driver_path != 0\n 1 from specified driver_path\n\n2 from vulkan-1.dll / libvulkan.so / libvulkan.dylib in system\n\n3 search driver by name nvoglv64.dll / amdvlk64.dll / libGLX_nvidia.so.0 .... and load it\n```\n\n## Load from system vulkan library or graphics driver\n\nThis is the default behavior and it should work on most systems\n\nsample usage\n```cpp\nint ret = create_gpu_instance();\n```\n\nLoad from system-installed libvulkan\n\n#### Windows\nvulkan-1.dll\n\n#### Linux Android\nlibvulkan.so\n\n#### macOS iOS and other APPLE platforms\nlibvulkan.dylib\n\nIf static moltenvk driver linked, should always succeed\n\nIf failed, it will try to find graphics driver object and load it\n\n#### Windows\nfor 64bit applications. search in ```%SystemRoot%\\System32\\DriverStore\\FileRepository```\n- nvoglv64.dll\n- amdvlk64.dll\n- igvk64.dll\n- qcvkarm64xum.dll\n\nfor 32bit applications. search in ```%SystemRoot%\\System32\\DriverStore\\FileRepository```\n- nvoglv32.dll\n- amdvlk32.dll\n- igvk32.dll\n\n#### Linux\n`dlopen()` search for\n- libGLX_nvidia.so.0\n- libvulkan_radeon.so\n- libvulkan_intel.so\n- libMaliVulkan.so.1\n- libVK_IMG.so\n\n#### Android\nfor 64bit applications\n- /vendor/lib64/hw/vulkan.adreno.so\n- /vendor/lib64/egl/libGLES_mali.so\n\nfor 32bit applications\n- /vendor/lib/hw/vulkan.adreno.so\n- /vendor/lib/egl/libGLES_mali.so\n\n#### macOS iOS and other APPLE platforms\n`dlopen()` search for\n- libMoltenVK.dylib\n- libvulkan_kosmickrisp.dylib\n\n## Load from driver_path\n\nfor advanced developer\n\nsample usage\n```cpp\nint ret = create_gpu_instance(\"libvulkan.so\");\nint ret = create_gpu_instance(\"/usr/lib64/libvulkan_radeon.so\");\nint ret = create_gpu_instance(\"/vendor/lib64/hw/vulkan.adreno.so\");\nint ret = create_gpu_instance(\"/data/local/tmp/vulkan.ad07XX.so\");\n```\n\n## Load from env VK_ICD_FILENAMES\n\nfor debug purpose\n\nsample usage\n```sh\nexport VK_ICD_FILENAMES=./vk_swiftshader_icd.json\nexport VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json\nexport VK_ICD_FILENAMES=/etc/vulkan/icd.d/nvidia_icd.json\n```\n\n## Load from env NCNN_VULKAN_DRIVER\n\nfor debug purpose\n\nsample usage\n```sh\nexport NCNN_VULKAN_DRIVER=/data/local/tmp/vulkan.ad07XX.so\n```\n" }, { "path": "docs/faq.en.md", "content": "\n\n# How to join the technical Community Groups with QQ ？\n\n- Open QQ -> click the group chat search-> search group number 637093648, enter the answer to the question: conv conv conv conv conv → join the group chat → ready to accept the Turing test(a joke)\n- Open QQ -> search Pocky group: 677104663 (lots experts), the answer to the question\n\n# How to watch the author's on live in Bilibili？\n\n- nihui：[水竹院落](https://live.bilibili.com/1264617)\n\n# Compilation\n\n- ## How to download the full source code？\n\n git clone --recursive https://github.com/Tencent/ncnn/\n\n or\n\n download [ncnn-xxxxx-full-source.zip](https://github.com/Tencent/ncnn/releases)\n\n- ## How to cross-compile？How to set the cmake toolchain？\n\n See https://github.com/Tencent/ncnn/wiki/how-to-build\n\n- ## The submodules were not downloaded! Please update submodules with \"git submodule update --init\" and try again\n\n As above, download the full source code. Or follow the prompts to execute: git submodule update --init\n\n- ## Could NOT find Protobuf (missing: Protobuf_INCLUDE_DIR)\n\n sudo apt-get install libprotobuf-dev protobuf-compiler\n\n- ## Could NOT find CUDA (missing: CUDA_TOOLKIT_ROOT_DIR CUDA_INCLUDE_DIRS CUDA_CUDART_LIBRARY)\n\n https://github.com/Tencent/ncnn/issues/1873\n\n- ## Could not find a package configuration file provided by \"OpenCV\" with any of the following names: OpenCVConfig.cmake opencv-config.cmake\n\n sudo apt-get install libopencv-dev\n\n or customized compile and install ，with set(OpenCV_DIR {the dir OpenCVConfig.cmake exist})\n\n- ## Could not find a package configuration file provided by \"ncnn\" with any of the following names: ncnnConfig.cmake ncnn-config.cmake\n\n set(ncnn_DIR { the dir ncnnConfig.cmake exist})\n\n- ## xxx.lib not found（be specified by system/compiler）\n\n undefined reference to __kmpc_for_static_init_4 __kmpc_for_static_fini __kmpc_fork_call ...\n\n Need to link openmp\n\n undefined reference to glslang::InitializeProcess() glslang::TShader::TShader(EShLanguage) ...\n\n need glslang.lib glslang-default-resource-limits.lib\n\n undefined reference to AAssetManager_fromJava AAssetManager_open AAsset_seek ...\n\n Add android to find_library and target_like_libraries\n\n find_package(ncnn)\n\n- ## undefined reference to typeinfo for ncnn::Layer\n\n opencv rtti -> opencv-mobile\n\n- ## undefined reference to __cpu_model\n\n upgrade compiler / libgcc_s libgcc\n\n- ## unrecognized command line option \"-mavx2\"\n\n upgrade gcc\n\n- ## Why is the compiled ncnn-android library so large？\n\n See https://github.com/Tencent/ncnn/wiki/build-for-android.zh and see How to trim smaller ncnn\n\n- ## ncnnoptimize and custom layer\n\n ncnnoptimize first before adding a custom layer to avoid ncnnoptimize not being able to handle custom layer saves.\n\n\n- ## rtti/exceptions Conflict\n\n The reason for the conflict is that the libraries used in the project are configured differently, so analyze whether you need to turn them on or off according to your actual situation. ncnn is ON by default, add the following two parameters when recompiling ncnn.\n - ON: -DNCNN_DISABLE_RTTI=OFF -DNCNN_DISABLE_EXCEPTION=OFF\n - OFF: -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON\n\n\n- ## error: undefined symbol: ncnn::Extractor::extract(char const*, ncnn::Mat&)\n\n Possible scenarios.\n - Try upgrading the NDK version of Android Studio\n\n\n# How do I add the ncnn library to my project and how does the cmake method work?\n\nCompile ncnn,and make install. linux/windows should set/export ncnn_DIR points to the directory containing ncnnConfig.cmake under the install directory\n\n- ## android\n\n- ## ios\n\n- ## linux\n\n- ## windows\n\n- ## macos\n\n- ## arm linux\n\n\n# Convert model issues\n\n- ## caffe\n\n `./caffe2ncnn caffe.prototxt caffe.caffemodel ncnn.param ncnn.bin`\n\n- ## mxnet\n\n ` ./mxnet2ncnn mxnet-symbol.json mxnet.params ncnn.param ncnn.bin`\n\n- ## darknet\n\n [https://github.com/xiangweizeng/darknet2ncnn](https://github.com/xiangweizeng/darknet2ncnn)\n\n- ## pytorch - onnx\n\n [use ncnn with pytorch or onnx](https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx)\n\n- ## tensorflow 1.x/2.x - keras\n\n [https://github.com/MarsTechHAN/keras2ncnn](https://github.com/MarsTechHAN/keras2ncnn) **[@MarsTechHAN](https://github.com/MarsTechHAN)**\n\n- ## tensorflow 2.x - mlir\n\n [Converting tensorflow2 models to ncnn via MLIR](https://zhuanlan.zhihu.com/p/152535430) **@[nihui](https://www.zhihu.com/people/nihui-2)**\n\n- ## netron\n\n [https://github.com/lutzroeder/netron](https://github.com/lutzroeder/netron)\n\n- ## How to generate a model with fixed shape？\n\n Input 0=w 1=h 2=c\n\n- ## why gpu can speedup\n\n- ## How to convert ncnnoptimize to fp16 model\n\n `ncnnoptimize model.param model.bin yolov5s-opt.param yolov5s-opt.bin 65536`\n\n- ## How to use ncnnoptimize checking the FLOPS / memory usage of your model\n\n- ## How to modify the model to support dynamics shape？\n\n Interp Reshape\n\n- ## How to convert a model into code embedded in a program？\n\n use ncnn2mem\n\n- ## How to encrypt the model？\n\n See https://zhuanlan.zhihu.com/p/268327784\n\n- ## The ncnn model transferred under Linux, Windows/MacOS/Android/... Can I use it directly?\n\n Yes, for all platforms\n\n- ## How to remove post-processing and export onnx？\n\n Ref：\n\n Referring to an article by UP , step 3 is to remove the post-processing and then export the onnx, where removing the post-processing can be the result of removing the subsequent steps when testing within the project.\n\n- ## pytorch layers can't export to onnx？\n\n Mode 1:\n\n ONNX_ATEN_FALLBACK\nFully customizable op, first change to one that can export (e.g. concat slice), go to ncnn and then modify param\n\n Way 2.\n\n You can try this with PNNX, see the following article for a general description:\n\n 1. [Windows/Linux/macOS steps for compiling PNNX](https://zhuanlan.zhihu.com/p/431833958)\n\n 2. [Learn in 5 minutes! Converting TorchScript models to ncnn models with PNNX](https://zhuanlan.zhihu.com/p/427512763)\n\n# Using\n\n- ## vkEnumeratePhysicalDevices failed -3\n\n- ## vkCreateInstance failed -9\n\n Please upgrade your GPU driver if you meet this crash or error.\n Here are the download sites for some brands of GPU drivers. We have provided some driver download pages here.\n [Intel](https://downloadcenter.intel.com/product/80939/Graphics-Drivers), [AMD](https://www.amd.com/en/support), [Nvidia](https://) www.nvidia.com/Download/index.aspx)\n\n- ## ModuleNotFoundError: No module named 'ncnn.ncnn'\n\n python setup.py develop\n\n- ## fopen nanodet-m.param failed\n\n path should be working dir\n\n File not found or not readable. Make sure that XYZ.param/XYZ.bin is accessible.\n\n- ## find_blob_index_by_name data / output / ... failed\n\n layer name vs blob name\n\n param.bin use xxx.id.h enum\n\n- ## parse magic failed\n\n- ## param is too old, please regenerate\n\n The model maybe has problems\n\n Your model file is being the old format converted by an old caffe2ncnn tool.\n\n Checkout the latest ncnn code, build it and regenerate param and model binary files, and that should work.\n\n Make sure that your param file starts with the magic number 7767517.\n\n you may find more info on use-ncnn-with-alexnet\n\n When adding the softmax layer yourself, you need to add 1=1\n\n- ## set_vulkan_compute failed, network use_vulkan_compute disabled\n\n Set net.opt.use_vulkan_compute = true before load_param / load_model;\n\n- ## How to execute multiple blob inputs, multiple blob outputs？\n Multiple execute `ex.input()` and `ex.extract()` like following\n ```\n ex.input(\"data1\", in_1);\n ex.input(\"data2\", in_2);\n ex.extract(\"output1\", out_1);\n ex.extract(\"output2\", out_2);\n ```\n- ## Multiple executions of Extractor extract double the calculation？\n\n No\n\n- ## How to see the elapsed time for every layer？\n\n cmake -DNCNN_BENCHMARK=ON ..\n\n- ## How to convert a cv::Mat CV_8UC3 BGR image\n\n from_pixels to_pixels\n\n- ## How to convert float data to ncnn::Mat\n\n First of all, you need to manage the memory you request yourself, at this point ncnn::Mat will not automatically free up the float data you pass over to it\n ``` c++\n std::vector testData(60, 1.0); // use std::vector to manage memory requests and releases yourself\n ncnn::Mat in1 = ncnn::Mat(60, (void*)testData.data()).reshape(4, 5, 3); // just pass the pointer to the float data as a void*, and even specify the dimension (up says it's best to use reshape to solve the channel gap)\n float* a = new float[60]; // New a piece of memory yourself, you need to release it later\n ncnn::Mat in2 = ncnn::Mat(60, (void*)a).reshape(4, 5, 3).clone(); // use the same method as above, clone() to transfer data owner\n ```\n" }, { "path": "docs/faq.md", "content": "\n\n# 如何加入技术交流QQ群？\n\n- 打开QQ→点击群聊搜索→搜索群号637093648→输入问题答案：卷卷卷卷卷→进入群聊→准备接受图灵测试（bushi）\n- 前往QQ搜索Pocky群：677104663(超多大佬)，问题答案：multi level intermediate representation\n\n# 如何看作者b站直播？\n\n- nihui的bilibili直播间：[水竹院落](https://live.bilibili.com/1264617)\n\n# 编译\n\n- ## 怎样下载完整源码？\n\n git clone --recursive https://github.com/Tencent/ncnn/\n \n 或者\n \n 下载 [ncnn-xxxxx-full-source.zip](https://github.com/Tencent/ncnn/releases)\n\n- ## 怎么交叉编译？cmake 工具链怎么设置啊？\n \n 参见 https://github.com/Tencent/ncnn/wiki/how-to-build\n\n- ## The submodules were not downloaded! Please update submodules with \"git submodule update --init\" and try again\n\n 如上，下载完整源码。或者按提示执行: git submodule update --init\n\n- ## Could NOT find Protobuf (missing: Protobuf_INCLUDE_DIR)\n \n sudo apt-get install libprotobuf-dev protobuf-compiler\n\n- ## Could NOT find CUDA (missing: CUDA_TOOLKIT_ROOT_DIR CUDA_INCLUDE_DIRS CUDA_CUDART_LIBRARY)\n\n https://github.com/Tencent/ncnn/issues/1873\n\n- ## Could not find a package configuration file provided by \"OpenCV\" with any of the following names: OpenCVConfig.cmake opencv-config.cmake\n\n sudo apt-get install libopencv-dev\n\n 或者自行编译安装，set(OpenCV_DIR {OpenCVConfig.cmake所在目录})\n\n- ## Could not find a package configuration file provided by \"ncnn\" with any of the following names: ncnnConfig.cmake ncnn-config.cmake\n\n set(ncnn_DIR {ncnnConfig.cmake所在目录})\n\n- ## 找不到库（需要根据系统/编译器指定）\n\n undefined reference to __kmpc_for_static_init_4 __kmpc_for_static_fini __kmpc_fork_call ...\n\n 需要链接openmp库 \n\n undefined reference to glslang::InitializeProcess() glslang::TShader::TShader(EShLanguage) ...\n\n 需要 glslang.lib glslang-default-resource-limits.lib\n\n undefined reference to AAssetManager_fromJava AAssetManager_open AAsset_seek ...\n\n find_library和target_like_libraries中增加 android \n\n find_package(ncnn)\n\n- ## undefined reference to typeinfo for ncnn::Layer\n\n opencv rtti -> opencv-mobile\n\n- ## undefined reference to __cpu_model\n\n 升级编译器 / libgcc_s libgcc\n\n- ## unrecognized command line option \"-mavx2\"\n\n 升级 gcc\n\n- ## 为啥自己编译的ncnn android库特别大？\n\n https://github.com/Tencent/ncnn/wiki/build-for-android.zh 以及见 如何裁剪更小的 ncnn 库\n\n- ## ncnnoptimize和自定义层\n\n 先ncnnoptimize再增加自定义层，避免ncnnoptimize不能处理自定义层保存。\n\n\n- ## rtti/exceptions冲突\n\n 产生原因是项目工程中使用的库配置不一样导致冲突，根据自己的实际情况分析是需要开启还是关闭。ncnn默认是ON，在重新编译ncnn时增加以下2个参数即可：\n - 开启：-DNCNN_DISABLE_RTTI=OFF -DNCNN_DISABLE_EXCEPTION=OFF\n - 关闭：-DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON\n\n\n- ## error: undefined symbol: ncnn::Extractor::extract(char const*, ncnn::Mat&)\n\n 可能的情况：\n - 尝试升级 Android Studio 的 NDK 版本\n\n- ## CMake 3.14.0 or higher is required. You are running version 2.8.12.2\n```shell\nwget https://github.com/Kitware/CMake/releases/download/v3.18.2/cmake-3.18.2-Linux-x86_64.tar.gz\ntar zxvf cmake-3.18.2-Linux-x86_64.tar.gz\nmv cmake-3.18.2-Linux-x86_64 /opt/cmake-3.18.2\nln -sf /opt/cmake-3.18.2/bin/* /usr/bin/\n```\n\n# 怎样添加ncnn库到项目中？cmake方式怎么用？\n\n编译ncnn，make install。linux/windows set/export ncnn_DIR 指向 install目录下包含ncnnConfig.cmake 的目录\n\n- ## android\n\n- ## ios\n\n- ## linux\n\n- ## windows\n\n- ## macos\n\n- ## arm linux\n\n\n# 转模型问题\n\n- ## caffe\n\n `./caffe2ncnn caffe.prototxt caffe.caffemodel ncnn.param ncnn.bin`\n\n- ## mxnet\n\n ` ./mxnet2ncnn mxnet-symbol.json mxnet.params ncnn.param ncnn.bin`\n\n- ## darknet\n\n [https://github.com/xiangweizeng/darknet2ncnn](https://github.com/xiangweizeng/darknet2ncnn)\n\n- ## pytorch - onnx\n\n [use ncnn with pytorch or onnx](https://github.com/Tencent/ncnn/wiki/use-ncnn-with-pytorch-or-onnx)\n\n- ## tensorflow 1.x/2.x - keras\n\n [https://github.com/MarsTechHAN/keras2ncnn](https://github.com/MarsTechHAN/keras2ncnn) **[@MarsTechHAN](https://github.com/MarsTechHAN)**\n\n- ## tensorflow 2.x - mlir\n\n [通过MLIR将tensorflow2模型转换到ncnn](https://zhuanlan.zhihu.com/p/152535430) **@[nihui](https://www.zhihu.com/people/nihui-2)**\n\n- ## netron\n\n [https://github.com/lutzroeder/netron](https://github.com/lutzroeder/netron)\n\n- ## 怎么生成有固定 shape 信息的模型？\n\n Input 0=w 1=h 2=c\n\n- ## why gpu能更快\n\n- ## ncnnoptimize 怎么转成 fp16 模型\n\n `ncnnoptimize model.param model.bin yolov5s-opt.param yolov5s-opt.bin 65536`\n\n- ## ncnnoptimize 怎样查看模型的 FLOPS / 内存占用情况\n\n- ## 怎么修改模型支持动态 shape？\n\n Interp Reshape\n\n- ## 如何将模型转换为代码内嵌到程序里？\n\n ncnn2mem\n\n- ## 如何加密模型？\n\n https://zhuanlan.zhihu.com/p/268327784\n\n- ## Linux下转的ncnn模型，Windows/MacOS/Android/.. 也能直接用吗？\n\n Yes，全平台通用\n\n- ## 如何去掉后处理，再导出 onnx？\n\n 检测：\n\n 参考up的一篇文章，步骤三就是去掉后处理,再导出onnx,其中去掉后处理可以是项目内测试时去掉后续步骤的结果。\n\n- ## pytorch 有的层导不出 onnx 怎么办？\n\n 方式一:\n\n ONNX_ATEN_FALLBACK\n完全自定义的op，先改成能导出的（如 concat slice），转到 ncnn 后再修改 param\n\n 方式二：\n\n 可以使用PNNX来试试，参考以下文章大概说明:\n\n 1. [Windows/Linux/macOS 编译 PNNX 步骤](https://zhuanlan.zhihu.com/p/431833958)\n\n 2. [5分钟学会！用 PNNX 转换 TorchScript 模型到 ncnn 模型](https://zhuanlan.zhihu.com/p/427512763)\n\n# 使用\n\n- ## vkEnumeratePhysicalDevices failed -3\n\n- ## vkCreateInstance failed -9\n\n 出现此类问题请先更新GPU驱动。Please upgrade your GPU driver if you encounter this crash or error.\n 这里提供了一些品牌的GPU驱动下载网址.We have provided some drivers' download pages here.\n [Intel](https://downloadcenter.intel.com/product/80939/Graphics-Drivers)，[AMD](https://www.amd.com/en/support)，[Nvidia](https://www.nvidia.com/Download/index.aspx)\n\n- ## docker 环境里面 nvidia-smi 能看到显卡也能跑 cuda 却不能跑 vulkan\n\n 因为这个docker环境的nvidia驱动没有安装opengl/vulkan支持\n\n 首先运行 nvidia-smi 查看当前驱动版本\n\n```\nNVIDIA-SMI 535.161.07\nDriver Version: 535.161.07\nCUDA Version: 12.2\n```\n\n然后去下载对应版本的NVIDIA驱动，安装用户态驱动文件，跳过内核部分\n\n```\nwget https://us.download.nvidia.com/tesla/535.161.07/NVIDIA-Linux-x86_64-535.161.07.run\nchmod +x NVIDIA-Linux-x86_64-535.161.07.run\n./NVIDIA-Linux-x86_64-535.161.07.run --silent --no-kernel-module\n```\n\n安装时会报一些文件权限错误，不用管，安装完成后 vulkan 支持就可用了。最后安装 vulkaninfo 查看gpu信息\n\n```\ndnf install vulkan-tools\nvulkaninfo\n```\n\n- ## ModuleNotFoundError: No module named 'ncnn.ncnn'\n\n python setup.py develop\n\n- ## fopen nanodet-m.param failed\n\n 文件路径 working dir\n\n File not found or not readable. Make sure that XYZ.param/XYZ.bin is accessible.\n\n- ## find_blob_index_by_name data / output / ... failed\n\n layer name vs blob name\n \n param.bin 应该用 xxx.id.h 的枚举\n\n- ## parse magic failed\n\n- ## param is too old, please regenerate\n\n 模型本身有问题\n\n Your model file is being the old format converted by an old caffe2ncnn tool.\n\n Checkout the latest ncnn code, build it and regenerate param and model binary files, and that should work.\n\n Make sure that your param file starts with the magic number 7767517.\n\n you may find more info on use-ncnn-with-alexnet\n \n When adding the softmax layer yourself, you need to add 1=1\n\n- ## set_vulkan_compute failed, network use_vulkan_compute disabled\n\n 你应该在 load_param / load_model 之前设置 net.opt.use_vulkan_compute = true;\n\n- ## 多个blob输入，多个blob输出，怎么做？\n 多次执行`ex.input()` 和 `ex.extract()`\n```\nex.input(\"data1\", in_1);\nex.input(\"data2\", in_2);\nex.extract(\"output1\", out_1);\nex.extract(\"output2\", out_2);\n```\n- ## Extractor extract 多次会重复计算吗？\n\n 不会\n\n- ## 如何看每一层的耗时？\n\n cmake -DNCNN_BENCHMARK=ON ..\n\n- ## 如何转换 cv::Mat CV_8UC3 BGR 图片\n\n from_pixels to_pixels\n\n- ## 如何转换 float 数据为 ncnn::Mat\n\n 首先，自己申请的内存需要自己管理，此时ncnn::Mat不会自动给你释放你传过来的float数据\n ``` c++\n std::vector testData(60, 1.0); // 利用std::vector自己管理内存的申请和释放\n ncnn::Mat in1 = ncnn::Mat(60, (void*)testData.data()).reshape(4, 5, 3); // 把float数据的指针转成void*传过去即可，甚至还可以指定维度(up说最好使用reshape用来解决channel gap)\n float* a = new float[60]; // 自己new一块内存，后续需要自己释放\n ncnn::Mat in2 = ncnn::Mat(60, (void*)a).reshape(4, 5, 3).clone(); // 使用方法和上面相同，clone() to transfer data owner\n ```\n\n- ## 如何初始化 ncnn::Mat 为全 0\n\n `mat.fill(0.f);`\n\n- ## 如何查看／获取版本号\n\n cmake时会打印\n\n c_api.h ncnn_version()\n\n 自己拼 1.0+yyyymmdd\n\n- ## 如何转换 yuv 数据\n\n yuv420sp2rgb yuv420sp2rgb_nv12\n\n **[@metarutaiga](https://github.com/metarutaiga/xxYUV)**\n\n- ## 如何 resize crop rotate 图片\n\n [efficient roi resize rotate](https://github.com/Tencent/ncnn/wiki/efficient-roi-resize-rotate)\n\n- ## 如何人脸5点对齐\n\n get_affine_transform\n\n warpaffine_bilinear_c3\n\n```c\n// 计算变换矩阵 并且求逆变换\nint type = 0; // 0->区域外填充为v[0],v[1],v[2], -233->区域外不处理\nunsigned int v = 0;\nfloat tm[6];\nfloat tm_inv[6];\n// 人脸区域在原图上的坐标和宽高\nfloat src_x = target->det.rect.x / target->det.w * pIveImageU8C3->u32Width;\nfloat src_y = target->det.rect.y / target->det.h * pIveImageU8C3->u32Height;\nfloat src_w = target->det.rect.w / target->det.w * pIveImageU8C3->u32Width;\nfloat src_h = target->det.rect.h / target->det.h * pIveImageU8C3->u32Height;\nfloat point_src[10] = {\nsrc_x + src_w * target->attr.land[0][0], src_x + src_w * target->attr.land[0][1],\nsrc_x + src_w * target->attr.land[1][0], src_x + src_w * target->attr.land[1][1],\nsrc_x + src_w * target->attr.land[2][0], src_x + src_w * target->attr.land[2][1],\nsrc_x + src_w * target->attr.land[3][0], src_x + src_w * target->attr.land[3][1],\nsrc_x + src_w * target->attr.land[4][0], src_x + src_w * target->attr.land[4][1],\n};\nfloat point_dst[10] = { // +8 是因为我们处理112*112的图\n30.2946f + 8.0f, 51.6963f,\n65.5318f + 8.0f, 51.5014f,\n48.0252f + 8.0f, 71.7366f,\n33.5493f + 8.0f, 92.3655f,\n62.7299f + 8.0f, 92.2041f,\n};\n// 第一种方式：先计算变换在求逆\nAffineTrans::get_affine_transform(point_src, point_dst, 5, tm);\nAffineTrans::invert_affine_transform(tm, tm_inv);\n// 第二种方式：直接拿到求逆的结果\n// AffineTrans::get_affine_transform(point_dst, point_src, 5, tm_inv);\n// rgb 分离的，所以要单独处理\nfor(int c = 0; c < 3; c++)\n{\n unsigned char* pSrc = malloc(xxx);\n unsigned char* pDst = malloc(xxx);\n ncnn::warpaffine_bilinear_c1(pSrc, SrcWidth, SrcHeight, SrcStride[c], pDst, DstWidth, DstHeight, DstStride[c], tm_inv, type, v);\n}\n// rgb packed则可以一次处理\nncnn::warpaffine_bilinear_c3(pSrc, SrcWidth, SrcHeight, SrcStride, pDst, DstWidth, DstHeight, DstStride, tm_inv, type, v);\n```\n\n- ## 如何获得中间层的blob输出\n \n ncnn::Mat output;\n \n ex.extract(\"your_blob_name\", output);\n\n- ## 为什么我使用GPU，但是GPU占用为0\n\n windows 10 任务管理器 - 性能选项卡 - GPU - 选择其中一个视图左上角的下拉箭头切换到 Compute_0 / Compute_1 / Cuda\n\n 你还可以安装软件：GPU-Z \n\n- ## layer XYZ not exists or registered\n\n Your network contains some operations that are not implemented in ncnn.\n\n You may implement them as custom layer followed in how-to-implement-custom-layer-step-by-step.\n\n Or you could simply register them as no-op if you are sure those operations make no sense.\n\n```\nclass Noop : public ncnn::Layer {};\nDEFINE_LAYER_CREATOR(Noop)\n\nnet.register_custom_layer(\"LinearRegressionOutput\", Noop_layer_creator);\nnet.register_custom_layer(\"MAERegressionOutput\", Noop_layer_creator);\n```\n\n- ## network graph not ready\n\n You shall call Net::load_param() first, then Net::load_model().\n\n This error may also happens when Net::load_param() failed, but not properly handled.\n\n For more information about the ncnn model load api, see ncnn-load-model\n\n- ## memory not 32-bit aligned at XYZ\n\n The pointer passed to Net::load_param() or Net::load_model() is not 32bit aligned.\n\n In practice, the head pointer of std::vector is not guaranteed to be 32bit aligned.\n\n you can store your binary buffer in ncnn::Mat structure, its internal memory is aligned.\n\n- ## crash on android with '__kmp_abort_process'\n\n This usually happens if you bundle multiple shared library with openmp linked\n\n It is actually an issue of the android ndk https://github.com/android/ndk/issues/1028\n\n On old android ndk, modify the link flags as\n\n -Wl,-Bstatic -lomp -Wl,-Bdynamic\n\n For recent ndk >= 21\n\n -fstatic-openmp\n\n- ## dlopen failed: library \"libomp.so\" not found\n Newer android ndk defaults to dynamic openmp runtime\n\n modify the link flags as\n\n -fstatic-openmp -fopenmp\n\n- ## crash when freeing a ncnn dynamic library(.dll/.so) built with openMP\n\n for optimal performance, the openmp threadpool spin waits for about a second prior to shutting down in case more work becomes available.\n\n If you unload a dynamic library that's in the process of spin-waiting, it will crash in the manner you see (most of the time).\n\n Just set OMP_WAIT_POLICY=passive in your environment, before calling loadlibrary. or Just wait a few seconds before calling freelibrary.\n\n You can also use the following method to set environment variables in your code:\n\n for msvc++:\n\n SetEnvironmentVariable(_T(\"OMP_WAIT_POLICY\"), _T(\"passive\"));\n\n for g++:\n\n setenv(\"OMP_WAIT_POLICY\", \"passive\", 1)\n \n reference: https://stackoverflow.com/questions/34439956/vc-crash-when-freeing-a-dll-built-with-openmp\n\n# 跑出来的结果对不上\n\n[ncnn-produce-wrong-result](https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result)\n\n- ## 如何打印 ncnn::Mat 的值？\n\n```C++\nvoid pretty_print(const ncnn::Mat& m)\n{\n for (int q=0; q // Don't forget this\n\nvoid pretty_print(const ncnn::Mat& m)\n{\n for (int q=0; q normed_feats(m.c);\n\n for (int i=0; i(y);\n uchar* sp = normed_feats[i].ptr(y);\n for (int x=0; x注意这里的flag指的是fp32和fp16，其中0指的是fp32，1指的是fp16\n\n 使用方式二：\n - ./ncnnoptimize ncnn.param ncnn.bin new.param new.bin flag cutstartname cutendname\n
cutstartname：模型截取的起点\n
cutendname：模型截取的终点\n\n\n- ## 如何使用量化工具？\n\n [Post Training Quantization Tools](https://github.com/Tencent/ncnn/tree/master/tools/quantize)\n\n- ## 如何设置线程数？\n\n opt.num_threads\n\n- ## 如何降低CPU占用率？\n\n net.opt.openmp_blocktime = 0;\n \n OMP_WAIT_POLICY=passive\n\n- ## 如何 batch inference？\n\n```\n int max_batch_size = vkdev->info.compute_queue_count;\n \n ncnn::Mat inputs[1000];\n ncnn::Mat outputs[1000];\n \n #pragma omp parallel for num_threads(max_batch_size)\n for (int i=0; i<1000; i++)\n {\n ncnn::Extractor ex = net1.create_extractor();\n ex.input(\"data\", inputs[i]);\n ex.extract(\"prob\", outputs[i]);\n }\n```\n\n \n\n- ## partial graph inference\n\n 先 extract 分类，判断后，再 extract bbox\n\n- ## 如何启用 bf16s 加速？\n\n```\nnet.opt.use_packing_layout = true;\nnet.opt.use_bf16_storage = true;\n```\n\n [用bf16加速ncnn](https://zhuanlan.zhihu.com/p/112564372) **@[nihui](https://www.zhihu.com/people/nihui-2)**\n\n A53\n\n- ## 如何裁剪更小的 ncnn 库？\n\n [build-minimal-library](https://github.com/Tencent/ncnn/wiki/build-minimal-library)\n\n- ## net.opt sgemm winograd fp16_storage 各是有什么作用？\n\n 对内存消耗的影响\n\n- ## 如何解决显卡进入节能模式造成的一系列问题？\n\n nVidia显卡（Intel和AMD估计也有）会在它认为的所谓空闲模式下，自动进入 `节能模式`，显存和核心频率就都会降低。\n \n 简单来说就是如果你的计算任务是 `非连续的`，那么可能会让耗时看起来非常 `不均匀`，当期间有运算空闲间隔发生，显卡进入节能模式，则会在下一次冷启动时发生计算耗时远超正常耗时几倍的情况，如下日志所示：\n\n ```cpp\n //开始播放\n Total: 162ms, Diff: 0ms, GLTex2Mat: 7ms, calc: 152ms, Mat2GLTex: 3ms\n Total: 43ms, Diff: 0ms, GLTex2Mat: 3ms, calc: 35ms, Mat2GLTex: 2ms\n Total: 45ms, Diff: 0ms, GLTex2Mat: 3ms, calc: 37ms, Mat2GLTex: 3ms\n Total: 40ms, Diff: 0ms, GLTex2Mat: 3ms, calc: 32ms, Mat2GLTex: 4ms\n //暂停3秒\n //继续播放\n Total: 190ms, Diff: 0ms, GLTex2Mat: 9ms, calc: 177ms, Mat2GLTex: 3ms\n Total: 134ms, Diff: 0ms, GLTex2Mat: 5ms, calc: 110ms, Mat2GLTex: 18ms\n Total: 40ms, Diff: 0ms, GLTex2Mat: 3ms, calc: 34ms, Mat2GLTex: 2ms\n Total: 42ms, Diff: 0ms, GLTex2Mat: 3ms, calc: 36ms, Mat2GLTex: 2ms\n Total: 47ms, Diff: 0ms, GLTex2Mat: 5ms, calc: 38ms, Mat2GLTex: 3ms\n ...\n ```\n\n 在对时间不敏感的项目上，这个问题没什么大不了的，完全可以忽略，但是有些业务场景上必须精准推估下一帧及其未来几帧的从上传、计算到渲染的耗时情况，则这种现象将会给开发者打开些许困扰。\n\n ### 3种解决方法\n * 联系显卡厂商，让其更新驱动将你的应用加入到免节能模式的白名单。\n * 优点：你什么都不用改。缺点：沟通困难，很可能显卡厂商根本不理你。\n * [显卡控制面板] - [管理3D设置] - [电源管理模式]，改成：[最高性能优先]。\n * 优点：不用改代码。缺点：如果是部署端是小白用户，需要编写手册手把手教他。\n * 可以空闲（暂停）时定期灌一些心跳计算包的任务进去（放1x1小图）让GPU维持在高性能状态。\n * 优点：需要改代码。缺点：不低碳不环保。\n\n# 白嫖项目\n\n- ## nanodet\n\n# 其他\n\n- ## up主用的什么系统/编辑器/开发环境？\n\n | 软件类型 | 软件名称 |\n | ------------| ----------- |\n | 系统 | Fedora |\n | 桌面环境 | KDE |\n | 编辑器 | Kate |\n | 画草图 | kolourpaint |\n | 画函数图像 | kmplot |\n | bilibili直播 | OBS |\n" }, { "path": "docs/how-to-build/build-mlir2ncnn.md", "content": "# mlir2ncnn\n\n## Compile\n\n**Clone LLVM**\n```bash\nhttps://github.com/llvm/llvm-project.git\ngit checkout -b mlir \n```\nCurrent working commit id is 74e6030bcbcc8e628f9a99a424342a0c656456f9:\n```bash\n$ git log\n\ncommit 74e6030bcbcc8e628f9a99a424342a0c656456f9 (HEAD -> main, origin/main, origin/HEAD)\nAuthor: Craig Topper \nDate: Thu Mar 4 22:30:38 2021 -0800\n\n [TargetLowering] Use HandleSDNodes to prevent nodes from being deleted by recursive calls in getNegatedExpression.\n```\n\nIt is determined by query lastest git commit date of `tools/mlir` directory.\n\n\n**Compile mlir**\n```bash\ncd llvm-project\nmkdir build\ncd build\ncmake -G Ninja -DCMAKE_INSTALL_PREFIX=install -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIBS=ON -DLLVM_ENABLE_PROJECTS=\"mlir\" -DLLVM_TARGETS_TO_BUILD=\"\" -DLLVM_INCLUDE_EXAMPLES=OFF -DLLVM_INCLUDE_TESTS=OFF ../llvm/\nninja -j8\nninja install\n```\n\n**Compile mlir2ncnn**\n```bash\ncd tools/mlir\nmkdir build\ncd build\ncmake .. -D LLVM_DIR=\nmake\n```\n\n## Usage\n\n**Export `.mlir`**\n\nSee https://zhuanlan.zhihu.com/p/152535430\n\n\n**Usage mlir2ncnn**\n\n```bash\n./mlir2ncnn pix2pix.mlir pix2pix.param pix2pix.bin\n```\n" }, { "path": "docs/how-to-build/how-to-build.md", "content": "### Git clone ncnn repo with submodule\n\n```\ngit clone https://github.com/Tencent/ncnn.git\ncd ncnn\ngit submodule update --init\n```\n\n- [Git clone ncnn repo with submodule](#git-clone-ncnn-repo-with-submodule)\n- [Build for Linux](#build-for-linux)\n - [Nvidia Jetson](#nvidia-jetson)\n - [Raspberry Pi](#raspberry-pi)\n - [POWER](#power)\n - [Intel oneAPI](#intel-oneapi)\n - [Cross compile: Riscv-gnu-toolchain](#cross-compile-riscv-gnu-toolchain)\n - [Verification](#verification)\n- [Build for Windows x64 using Visual Studio Community 2017](#build-for-windows-x64-using-visual-studio-community-2017)\n- [Build for Windows x64 using MinGW-w64](#build-for-windows-x64-using-mingw-w64)\n- [Build for Windows XP (x86)](#build-for-windows-xp-x86)\n - [Using MinGW-w64](#using-mingw-w64)\n - [Using Clang](#using-clang)\n - [Using Visual Studio (MSVC)](#using-visual-studio-msvc)\n- [Build for macOS](#build-for-macos)\n- [Build for ARM Cortex-A family with cross-compiling](#build-for-arm-cortex-a-family-with-cross-compiling)\n- [Build for Hisilicon platform with cross-compiling](#build-for-hisilicon-platform-with-cross-compiling)\n- [Build for AnyCloud platform with cross-compiling](#build-for-AnyCloud-platform-with-cross-compiling)\n- [Build for Android](#build-for-android)\n- [Build for iOS on macOS with xcode](#build-for-ios-on-macos-with-xcode)\n- [Build for WebAssembly](#build-for-webassembly)\n- [Build for AllWinner D1](#build-for-allwinner-d1)\n- [Build for Loongson 2K1000](#build-for-loongson-2k1000)\n- [Build for Termux on Android](#build-for-termux-on-android)\n- [Build for QNX](#build-for-qnx)\n- [Build for Nintendo 3DS Homebrew Launcher](#build-for-nintendo-3ds-homebrew-launcher)\n- [Build for HarmonyOS with cross-compiling](#build-for-harmonyos-with-cross-compiling)\n- [Build for ESP32 with cross-compiling](#build-for-esp32-with-cross-compiling)\n\n***\n\n### Build for Linux\n\nInstall required build dependencies:\n\n* git\n* g++\n* cmake\n* protocol buffer (protobuf) headers files and protobuf compiler\n* (optional) LLVM OpenMP header files # If building with Clang, and multithreaded CPU inference is desired\n* (optional) opencv # For building examples\n\nGenerally if you have Intel, AMD or Nvidia GPU from last 10 years, Vulkan can be easily used.\n\nOn some systems there are no Vulkan drivers easily available at the moment (October 2020), so you might need to disable use of Vulkan on them. This applies to Raspberry Pi 3 (but there is experimental open source Vulkan driver in the works, which is not ready yet). Nvidia Tegra series devices (like Nvidia Jetson) should support Vulkan. Ensure you have most recent software installed for best experience.\n\nOn Debian, Ubuntu, or Raspberry Pi OS, you can install all required dependencies using:\n```shell\nsudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libomp-dev libopencv-dev\n```\nOn Redhat or Centos, you can install all required dependencies using:\n```shell\nsudo yum install build-essential git cmake libprotobuf-dev protobuf-compiler libopencv-dev\n```\n\nTo use Vulkan after building ncnn later, you will also need to have Vulkan driver for your GPU. For AMD and Intel GPUs these can be found in Mesa graphics driver, which usually is installed by default on all distros (i.e. `sudo apt install mesa-vulkan-drivers` on Debian/Ubuntu). For Nvidia GPUs the proprietary Nvidia driver must be downloaded and installed (some distros will allow easier installation in some way). After installing Vulkan driver, confirm Vulkan libraries and driver are working, by using `vulkaninfo` or `vulkaninfo | grep deviceType`, it should list GPU device type. If there are more than one GPU installed (including the case of integrated GPU and discrete GPU, commonly found in laptops), you might need to note the order of devices to use later on.\n\n#### Nvidia Jetson\n\nThe Vulkan driver is a default component of the Linux For Tegra BSP release, check [the device list](https://developer.nvidia.com/embedded/vulkan).\n\n```shell\ncd ncnn\nmkdir -p build\ncd build\ncmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../toolchains/jetson.toolchain.cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..\nmake -j$(nproc)\n```\n\n#### Raspberry Pi\n\nVulkan drivers do exists, but are not mature. You are free to experiment at your own discretion, and report results and performance.\n\n```shell\ncd ncnn\nmkdir -p build\ncd build\ncmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..\nmake -j$(nproc)\n```\n\nYou can add `-GNinja` to `cmake` above to use Ninja build system (invoke build using `ninja` or `cmake --build .`).\n\nFor Raspberry Pi 3 on 32bit OS, add `-DCMAKE_TOOLCHAIN_FILE=../toolchains/pi3.toolchain.cmake` to cmake. You can also consider disabling Vulkan support as the Vulkan drivers for Raspberry Pi are still not mature, but it doesn't hurt to build the support in, but not use it.\n\n#### POWER\n\nFor POWER9 with Clang:\n\n```shell\ncd ncnn\nmkdir -p build\ncd build\ncmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../toolchains/power9le-linux-gnu-vsx.clang.toolchain.cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..\nmake -j$(nproc)\n```\n\nTo use GCC instead, use the `power9le-linux-gnu-vsx.toolchain.cmake` toolchain file instead. Note that according to benchmarks, Clang appears to produce noticeably faster CPU inference than GCC for POWER9 targets. For fastest inference, use Clang 18 or higher; earlier versions of Clang may have impaired inference speed due to [Bug 49864](https://github.com/llvm/llvm-project/issues/49864) and [Bug 64664](https://github.com/llvm/llvm-project/issues/64664).\n\nFor POWER8 instead of POWER9, use the `power8le-linux-gnu-vsx.clang.toolchain.cmake` or `power8le-linux-gnu-vsx.toolchain.cmake` toolchain file instead. POWER8 will be slower than POWER9.\n\nNote that the POWER toolchain files only support little-endian mode.\n\n#### Intel oneAPI\n\nBesides the prerequests in this section, Intel oneAPI BaseKit and HPCKit should be installed. They are available from https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html and https://www.intel.com/content/www/us/en/developer/tools/oneapi/hpc-toolkit.html freely.\n\nIntel oneAPI offers two kinds of compilers, the classic `icc/icpc` and the LLVM based `icx/icpx`. To build with these compilers, add `CC=icc CXX=icpc` or `CC=icx CXX=icpx` before the `cmake` command. When compiling with `icc/icpc`, cmake will warn that `xop`, `avx512`, and `bf16` extensions are not supported by the compiler, while `icx/icpx` works well.\n\nBoth of these compilers have been tested and passed the ncnn benchmark successfully. The results have been included in ncnn benchmark readme. Generally, `icx/icpx` are likely to show better performance than `icc/icpc` and the quantized models can benefit from the extensions `icx/icpx` supports.\n\n#### Cross compile: Riscv-gnu-toolchain\nBefore compiling the whole project, toolchain must be installed.\n[Reference: Riscv-gnu-toolchain build guide](https://github.com/riscv-collab/riscv-gnu-toolchain/blob/master/README.md)\n```shell\n\n# configure with vector extension.\n./configure --prefix=/opt/riscv --enable-multilib --with-arch=rv64gcv\n\n# configure without vector extension.\n./configure --prefix=/opt/riscv --enable-multilib --with-arch=rv64gc\n\n# it takes quite a long time:(\nsudo make linux\n\n```\nNow you can build the project:\n```shell\nmkdir build-riscv\ncd build-riscv\ncmake -DDCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../toolchains/riscv64-unknown-linux-gnu.toolchain.cmake -DNCNN_BUILD_EXAMPLES=ON ..\nmake -j$(nproc) # or `make -j2` if your cpu isn't powerful enough.\n```\n\n#### Verification\n\nVerify build by running some examples:\n\n```shell\ncd ../examples\n../build/examples/squeezenet ../images/256-ncnn.png\n[0 AMD RADV FIJI (LLVM 10.0.1)] queueC=1[4] queueG=0[1] queueT=0[1]\n[0 AMD RADV FIJI (LLVM 10.0.1)] bugsbn1=0 buglbia=0 bugcopc=0 bugihfa=0\n[0 AMD RADV FIJI (LLVM 10.0.1)] fp16p=1 fp16s=1 fp16a=0 int8s=1 int8a=1\n532 = 0.163452\n920 = 0.093140\n716 = 0.061584\n```\n\nYou can also run benchmarks (the 4th argument is a GPU device index to use, refer to `vulkaninfo`, if you have more than one GPU):\n\n```shell\ncd ../benchmark\n../build/benchmark/benchncnn 10 $(nproc) 0 0\n[0 AMD RADV FIJI (LLVM 10.0.1)] queueC=1[4] queueG=0[1] queueT=0[1]\n[0 AMD RADV FIJI (LLVM 10.0.1)] bugsbn1=0 buglbia=0 bugcopc=0 bugihfa=0\n[0 AMD RADV FIJI (LLVM 10.0.1)] fp16p=1 fp16s=1 fp16a=0 int8s=1 int8a=1\nnum_threads = 4\npowersave = 0\ngpu_device = 0\ncooling_down = 1\n squeezenet min = 4.68 max = 4.99 avg = 4.85\n squeezenet_int8 min = 38.52 max = 66.90 avg = 48.52\n...\n```\n\nTo run benchmarks on a CPU, set the 5th argument to `-1`.\n\n\n***\n\n### Build for Windows x64 using Visual Studio Community 2017\n\nDownload and Install Visual Studio Community 2017 from https://visualstudio.microsoft.com/vs/community/\n\nStart the command prompt: `Start → Programs → Visual Studio 2017 → Visual Studio Tools → x64 Native Tools Command Prompt for VS 2017`\n\n> You can also search `x64 Native Tools Command Prompt for VS 2017` directly.\n\nDownload protobuf-3.11.2 from https://github.com/google/protobuf/archive/v3.11.2.zip\n\nBuild protobuf library:\n\n```shell\ncd \nmkdir protobuf_build\ncd protobuf_build\ncmake -A x64 -DCMAKE_INSTALL_PREFIX=%cd%/install -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\nBuild ncnn library (replace `` with a proper path):\n\n```shell\ncd \nmkdir -p protobuf_build\ncd protobuf_build\ncmake -A x64 -DCMAKE_INSTALL_PREFIX=%cd%/install -Dprotobuf_DIR=/protobuf_build/install/cmake -DNCNN_VULKAN=ON ..\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\nNote: To speed up compilation process on multi core machines, configuring `cmake` to use `jom` or `ninja` using `-G` flag is recommended.\n\nNote: For protobuf >=22.0 (Take v25.3 for example):\n\nBuild zlib:\n```shell\ngit clone -b -v1.3.1 https://github.com/madler/zlib.git\ncd zlib\nmkdir build\ncd build\ncmake -A x64 -DCMAKE_INSTALL_PREFIX=%cd%/install ..\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\nBuild protobuf library (replace `` with a proper path):\n```shell\ngit clone -b v25.3 https://github.com/protocolbuffers/protobuf.git\ncd protobuf\ngit submodule update --init --recursive\n\nmkdir protobuf_build\ncd protobuf_build\ncmake -A x64 -DCMAKE_INSTALL_PREFIX=%cd%/install -DCMAKE_CXX_STANDARD=14 -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF -DZLIB_INCLUDE_DIR=\\build\\install\\include -DZLIB_LIBRARY=\\build\\install\\lib\\zlib.lib -DABSL_PROPAGATE_CXX_STD=ON ../cmake\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\nBuild ncnn library (replace `` and `` with a proper path):\n\n```shell\ncd \nmkdir -p build\ncd build\ncmake -A x64 -DCMAKE_INSTALL_PREFIX=%cd%/install -DCMAKE_PREFIX_PATH=/protobuf_build\\install\\cmake -DZLIB_INCLUDE_DIR=\\build\\install\\include -DZLIB_LIBRARY=\\build\\install\\lib\\zlib.lib -Dabsl_DIR=/protobuf_build\\install\\lib\\cmake\\absl -Dutf8_range_DIR=/protobuf_build\\install\\lib\\cmake\\utf8_range -DNCNN_VULKAN=ON ..\ncmake --build . --config Release -j 2\ncmake --build . --config Release --target install\n```\n\n***\n\n### Build for Windows x64 using MinGW-w64\n\nDownload MinGW-w64 toolchain from [winlibs](https://winlibs.com/) or [w64devkit](https://github.com/skeeto/w64devkit), add `bin` folder to environment variables.\n\nBuild ncnn library:\n\n```shell\ncd \nmkdir build\ncd build\ncmake -DNCNN_VULKAN=ON -G \"MinGW Makefiles\" ..\ncmake --build . --config Release -j 4\ncmake --build . --config Release --target install\n```\n\n***\n\n### Build for Windows XP (x86)\n\n> **Note:** Windows XP support is provided through collaborative contributions from [@Sugar-Baby](https://github.com/Sugar-Baby) and [@AtomAlpaca](https://github.com/AtomAlpaca).\n\n#### Using MinGW-w64\n\nDownload mingw toolchain targeting 32 bit from [sourceforge](https://jaist.dl.sourceforge.net/project/mingw-w64/Toolchains%20targetting%20Win32/Personal%20Builds/mingw-builds/8.1.0/threads-posix/dwarf/i686-8.1.0-release-posix-dwarf-rt_v6-rev0.7z), extract and add environment variable named `MINGW32_ROOT_PATH` valued by ``, and add `/bin` to `PATH`.\n\n```shell\nmkdir build\ncd build\ncmake -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-mingw.toolchain.cmake\" -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_AVX=OFF .. -G \"MinGW Makefiles\"\ncmake --build . --config Release -j 4\ncmake --build . --config Release --target install\n```\n\n#### Using Clang\n\nClang requires libraries from mingw. Configure mingw toolchain targeting 32-bit as described in the [MinGW-w64 section](#using-mingw-w64).\n\nInstall Clang 6.0 or later.\n\n```shell\nmkdir build\ncd build\ncmake -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-clang.toolchain.cmake\" -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_AVX=OFF .. -G \"MinGW Makefiles\"\ncmake --build . --config Release -j 4\ncmake --build . --config Release --target install\n```\n\n#### Using Visual Studio (MSVC)\n\nInstall v141_xp toolset for Windows XP:\n\n1. Bring up the Visual Studio installer (Tools → Get Tools and Features)\n2. Select Desktop development with C++\n3. Select Windows XP support for C++ from the Summary section\n4. Click Modify\n\n```shell\nmkdir build\ncd build\ncmake -A WIN32 -G \"Visual Studio 17 2022\" -T v141_xp -DNCNN_WINXP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_OPENMP=OFF -DNCNN_AVX=OFF -DNCNN_BUILD_WITH_STATIC_CRT=ON -DCMAKE_TOOLCHAIN_FILE=\"../toolchains/windows-xp-msvc.toolchain.cmake\" ..\ncmake --build . --config Release -j 4\ncmake --build . --config Release --target install\n```\n\n**Note:** The MSVC toolchain uses the `v141_xp` platform toolset for Windows XP compatibility. Vulkan is disabled for XP compatibility, and advanced CPU features (AVX, AVX2, AVX512) are disabled to ensure compatibility with older processors.\n\n***\n\n### Build for macOS\n\nWe've published ncnn to [brew](https://formulae.brew.sh/formula/ncnn#default) now, you can just use following method to install ncnn if you have the Xcode Command Line Tools installed.\n\n```shell\nbrew update\nbrew install ncnn\n```\n\nOr if you want to compile and build ncnn locally, first install Xcode or Xcode Command Line Tools according to your needs.\n\nThen install `protobuf` and `libomp` via homebrew\n\n```shell\nbrew install protobuf libomp\n```\n\nDownload and install Vulkan SDK from \n\n\n```shell\nwget https://sdk.lunarg.com/sdk/download/1.3.280.1/mac/vulkansdk-macos-1.3.280.1.dmg -O vulkansdk-macos-1.3.280.1.dmg\nhdiutil attach vulkansdk-macos-1.3.280.1.dmg\nsudo /Volumes/vulkansdk-macos-1.3.280.1/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.3.280.1 --accept-licenses --default-answer --confirm-command install\nhdiutil detach /Volumes/vulkansdk-macos-1.3.280.1\n\n# setup env\nexport VULKAN_SDK=`pwd`/vulkansdk-macos-1.3.280.1/macOS\n```\n\n```shell\ncd \nmkdir -p build\ncd build\n\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DPLATFORM=MAC -DARCHS=\"x86_64;arm64\" \\\n -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.3.280.1/macOS/lib/libMoltenVK.dylib \\\n -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..\n\ncmake --build . -j 4\ncmake --build . --target install\n```\n\n*Note: If you encounter `libomp` related errors during installation, you can also check our GitHub Actions at [here](https://github.com/Tencent/ncnn/blob/d91cccf/.github/workflows/macos-x64-gpu.yml#L50-L68) to install and use `openmp`.*\n***\n\n### Build for ARM Cortex-A family with cross-compiling\nDownload ARM toolchain from https://developer.arm.com/open-source/gnu-toolchain/gnu-a/downloads\n\n```shell\nexport PATH=\":${PATH}\"\n```\n\nAlternatively install a cross-compiler provided by the distribution (i.e. on Debian / Ubuntu, you can do `sudo apt install g++-arm-linux-gnueabi g++-arm-linux-gnueabihf g++-aarch64-linux-gnu`).\n\nDepending on your needs build one or more of the below targets.\n\nAArch32 target with soft float (arm-linux-gnueabi)\n```shell\ncd \nmkdir -p build-arm-linux-gnueabi\ncd build-arm-linux-gnueabi\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake ..\nmake -j$(nproc)\nmake install\n```\n\nAArch32 target with hard float (arm-linux-gnueabihf)\n```shell\ncd \nmkdir -p build-arm-linux-gnueabihf\ncd build-arm-linux-gnueabihf\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..\nmake -j$(nproc)\nmake install\n```\n\nAArch64 GNU/Linux target (aarch64-linux-gnu)\n```shell\ncd \nmkdir -p build-aarch64-linux-gnu\ncd build-aarch64-linux-gnu\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..\nmake -j$(nproc)\nmake install\n```\n\n***\n\n### Build for Hisilicon platform with cross-compiling\nDownload and install Hisilicon SDK. The toolchain should be in `/opt/hisi-linux/x86-arm` \nnew version of Hisilicon toolchain should be in `/opt/linux/x86-arm/` \n\n```shell\ncd \nmkdir -p build\ncd build\n\n# Choose one cmake toolchain file depends on your target platform\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv300.toolchain.cmake ..\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv500.toolchain.cmake ..\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix100.toolchain.cmake ..\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix200.toolchain.cmake ..\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix210.toolchain.cmake ..\n\nmake -j$(nproc)\nmake install\n```\n\n***\n\n### Build for AnyCloud platform with cross-compiling\nDownload and install AnyCloud SDK. And load env to set toolchain can access in shell\n\n```shell\ncd \nmkdir -p build\ncd build\n\n# Choose one cmake toolchain file depends on your target platform\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/anykav500.toolchain.cmake ..\n\nmake -j$(nproc)\nmake install\n```\n\n***\n\n### Build for Android\nYou can use the pre-build ncnn-android-lib.zip from https://github.com/Tencent/ncnn/releases\n\nDownload Android NDK from http://developer.android.com/ndk/downloads/index.html and install it, for example:\n\n```shell\nunzip android-ndk-r21d-linux-x86_64.zip\nexport ANDROID_NDK=\n```\n\n(optional) remove the hardcoded debug flag in Android NDK [android-ndk issue](https://github.com/android-ndk/ndk/issues/243)\n```\n# open $ANDROID_NDK/build/cmake/android.toolchain.cmake for ndk < r23\n# or $ANDROID_NDK/build/cmake/android-legacy.toolchain.cmake for ndk >= r23\n# delete \"-g\" line\nlist(APPEND ANDROID_COMPILER_FLAGS\n -g\n -DANDROID\n```\n\nBuild armv7 library\n\n```shell\ncd \nmkdir -p build-android-armv7\ncd build-android-armv7\n\ncmake -DCMAKE_TOOLCHAIN_FILE=\"$ANDROID_NDK/build/cmake/android.toolchain.cmake\" \\\n -DANDROID_ABI=\"armeabi-v7a\" -DANDROID_ARM_NEON=ON \\\n -DANDROID_PLATFORM=android-14 -DNCNN_VULKAN=ON ..\n\n# If you use cmake >= 3.21 and ndk-r23\n# you need to add -DANDROID_USE_LEGACY_TOOLCHAIN_FILE=False option for working optimization flags\n\nmake -j$(nproc)\nmake install\n```\n\nPick `build-android-armv7/install` folder for further JNI usage.\n\n\nBuild aarch64 library:\n\n```shell\ncd \nmkdir -p build-android-aarch64\ncd build-android-aarch64\n\ncmake -DCMAKE_TOOLCHAIN_FILE=\"$ANDROID_NDK/build/cmake/android.toolchain.cmake\"\\\n -DANDROID_ABI=\"arm64-v8a\" \\\n -DANDROID_PLATFORM=android-21 -DNCNN_VULKAN=ON ..\n\n# If you use cmake >= 3.21 and ndk-r23\n# you need to add -DANDROID_USE_LEGACY_TOOLCHAIN_FILE=False option for working optimization flags\n\nmake -j$(nproc)\nmake install\n```\n\nPick `build-android-aarch64/install` folder for further JNI usage.\n\n***\n\n### Build for iOS on macOS with xcode\nYou can use the pre-build ncnn.framework glslang.framework and openmp.framework from https://github.com/Tencent/ncnn/releases\n\nInstall xcode\n\nYou can replace ```-DENABLE_BITCODE=0``` to ```-DENABLE_BITCODE=1``` in the following cmake arguments if you want to build bitcode enabled libraries.\n\nDownload and install openmp for multithreading inference feature on iPhoneOS\n```shell\nwget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz\ntar -xf openmp-11.0.0.src.tar.xz\ncd openmp-11.0.0.src\n\n# apply some compilation fix\nsed -i'' -e '/.size __kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S\nsed -i'' -e 's/__kmp_unnamed_critical_addr/___kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S\n\nmkdir -p build-ios\ncd build-ios\n\ncmake -DCMAKE_TOOLCHAIN_FILE=/toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \\\n -DPLATFORM=OS64 -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DARCHS=\"arm64;arm64e\" \\\n -DPERL_EXECUTABLE=/usr/local/bin/perl \\\n -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..\n\ncmake --build . -j 4\ncmake --build . --target install\n\n# copy openmp library and header files to xcode toolchain sysroot\n# is usually /Applications/Xcode.app or /Applications/Xcode-beta.app depends on your Xcode version\nsudo cp install/include/* /Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/include\nsudo cp install/lib/libomp.a /Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib\n```\n\nDownload and install openmp for multithreading inference feature on iPhoneSimulator\n```shell\nwget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz\ntar -xf openmp-11.0.0.src.tar.xz\ncd openmp-11.0.0.src\n\n# apply some compilation fix\nsed -i'' -e '/.size __kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S\nsed -i'' -e 's/__kmp_unnamed_critical_addr/___kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S\n\nmkdir -p build-ios-sim\ncd build-ios-sim\n\ncmake -DCMAKE_TOOLCHAIN_FILE=/toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \\\n -DPLATFORM=SIMULATORARM64 -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DARCHS=\"x86_64;arm64\" \\\n -DPERL_EXECUTABLE=/usr/local/bin/perl \\\n -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..\n\ncmake --build . -j 4\ncmake --build . --target install\n\n# copy openmp library and header files to xcode toolchain sysroot\n# is usually /Applications/Xcode.app or /Applications/Xcode-beta.app depends on your Xcode version\nsudo cp install/include/* /Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/include\nsudo cp install/lib/libomp.a /Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib\n```\n\nPackage openmp framework:\n```shell\ncd \n\nmkdir -p openmp.framework/Versions/A/Headers\nmkdir -p openmp.framework/Versions/A/Resources\nln -s A openmp.framework/Versions/Current\nln -s Versions/Current/Headers openmp.framework/Headers\nln -s Versions/Current/Resources openmp.framework/Resources\nln -s Versions/Current/openmp openmp.framework/openmp\nlipo -create build-ios/install/lib/libomp.a build-ios-sim/install/lib/libomp.a -o openmp.framework/Versions/A/openmp\ncp -r build-ios/install/include/* openmp.framework/Versions/A/Headers/\nsed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/11.0/g' /Info.plist > openmp.framework/Versions/A/Resources/Info.plist\n```\n\nDownload and install Vulkan SDK from https://vulkan.lunarg.com/sdk/home\n```shell\nwget https://sdk.lunarg.com/sdk/download/1.2.189.0/mac/vulkansdk-macos-1.2.189.0.dmg?Human=true -O vulkansdk-macos-1.2.189.0.dmg\nhdiutil attach vulkansdk-macos-1.2.189.0.dmg\nsudo /Volumes/vulkansdk-macos-1.2.189.0/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.2.189.0 --accept-licenses --default-answer --confirm-command install\nhdiutil detach /Volumes/vulkansdk-macos-1.2.189.0\n\n# setup env\nexport VULKAN_SDK=`pwd`/vulkansdk-macos-1.2.189.0/macOS\n```\n\nBuild library for iPhoneOS:\n\n```shell\ncd \ngit submodule update --init\nmkdir -p build-ios\ncd build-ios\n\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DPLATFORM=OS64 -DARCHS=\"arm64;arm64e\" \\\n -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib/libomp.a\" \\\n -DNCNN_VULKAN=ON -DNCNN_BUILD_BENCHMARK=OFF ..\n\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nBuild library for iPhoneSimulator:\n\n```shell\ncd \nmkdir -p build-ios-sim\ncd build-ios-sim\n\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DPLATFORM=SIMULATORARM64 -DARCHS=\"x86_64;arm64\" \\\n -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \\\n -DOpenMP_C_FLAGS=\"-Xclang -fopenmp\" -DOpenMP_CXX_FLAGS=\"-Xclang -fopenmp\" \\\n -DOpenMP_C_LIB_NAMES=\"libomp\" -DOpenMP_CXX_LIB_NAMES=\"libomp\" \\\n -DOpenMP_libomp_LIBRARY=\"/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib/libomp.a\" \\\n -DNCNN_BUILD_BENCHMARK=OFF ..\n\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nPackage glslang framework for iPhoneOS:\n```shell\ncd \n\nmkdir -p glslang.framework/Versions/A/Headers\nmkdir -p glslang.framework/Versions/A/Resources\nln -s A glslang.framework/Versions/Current\nln -s Versions/Current/Headers glslang.framework/Headers\nln -s Versions/Current/Resources glslang.framework/Resources\nln -s Versions/Current/glslang glslang.framework/glslang\nlibtool -static build-ios/install/lib/libglslang.a build-ios/install/lib/libSPIRV.a -o build-ios/install/lib/libglslang_combined.a\nlipo -create build-ios/install/lib/libglslang_combined.a -o glslang.framework/Versions/A/glslang\ncp -r build/install/include/glslang glslang.framework/Versions/A/Headers/\nsed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist\n```\n\nPackage ncnn framework for iPhoneOS:\n```shell\ncd \n\nmkdir -p ncnn.framework/Versions/A/Headers\nmkdir -p ncnn.framework/Versions/A/Resources\nln -s A ncnn.framework/Versions/Current\nln -s Versions/Current/Headers ncnn.framework/Headers\nln -s Versions/Current/Resources ncnn.framework/Resources\nln -s Versions/Current/ncnn ncnn.framework/ncnn\nlipo -create build-ios/install/lib/libncnn.a -o ncnn.framework/Versions/A/ncnn\ncp -r build-ios/install/include/* ncnn.framework/Versions/A/Headers/\nsed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist\n```\n\nPick `ncnn.framework` `glslang.framework` and `openmp.framework` folder for app development.\n\n***\n\n### Build for WebAssembly\n\nInstall Emscripten\n\n```shell\ngit clone https://github.com/emscripten-core/emsdk.git\ncd emsdk\n./emsdk install 3.1.28\n./emsdk activate 3.1.28\n\nsource emsdk_env.sh\n```\n\nBuild without any extension for general compatibility:\n```shell\nmkdir -p build\ncd build\ncmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \\\n -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nBuild with WASM SIMD extension:\n```shell\nmkdir -p build-simd\ncd build-simd\ncmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \\\n -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nBuild with WASM Thread extension:\n```shell\nmkdir -p build-threads\ncd build-threads\ncmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \\\n -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nBuild with WASM SIMD and Thread extension:\n```shell\nmkdir -p build-simd-threads\ncd build-simd-threads\ncmake -DCMAKE_TOOLCHAIN_FILE=$EMSDK/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \\\n -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_SIMPLEOCV=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \\\n -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nPick `build-XYZ/install` folder for further usage.\n\n***\n\n### Build for AllWinner D1\n\nDownload c906 toolchain package from https://www.xrvm.cn/community/download?id=4453617141140230144\n\n```shell\ntar -xf Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.1.0-20250522.tar.gz\nexport RISCV_ROOT_PATH=/home/nihui/osd/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.1.0\n```\n\nBuild ncnn with riscv-v vector and simpleocv enabled:\n```shell\nmkdir -p build-c906\ncd build-c906\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v310.toolchain.cmake \\\n -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_ZFH=ON -DNCNN_ZVFH=OFF \\\n -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..\ncmake --build . -j 4\ncmake --build . --target install\n```\n\nPick `build-c906/install` folder for further usage.\n\nYou can upload binary inside `build-c906/examples` folder and run on D1 board for testing.\n\n***\n\n### Build for Loongson 2K1000\n\nFor gcc version < 8.5, you need to fix msa.h header for workaround msa fmadd/fmsub/maddv/msubv bug.\n\nOpen ```/usr/lib/gcc/mips64el-linux-gnuabi64/8/include/msa.h```, find ```__msa_fmadd``` and ```__msa_fmsub``` and apply changes as the following\n```c\n// #define __msa_fmadd_w __builtin_msa_fmadd_w\n// #define __msa_fmadd_d __builtin_msa_fmadd_d\n// #define __msa_fmsub_w __builtin_msa_fmsub_w\n// #define __msa_fmsub_d __builtin_msa_fmsub_d\n#define __msa_fmadd_w(a, b, c) __builtin_msa_fmadd_w(c, b, a)\n#define __msa_fmadd_d(a, b, c) __builtin_msa_fmadd_d(c, b, a)\n#define __msa_fmsub_w(a, b, c) __builtin_msa_fmsub_w(c, b, a)\n#define __msa_fmsub_d(a, b, c) __builtin_msa_fmsub_d(c, b, a)\n```\n\nfind ```__msa_maddv``` and ```__msa_msubv``` and apply changes as the following\n```c\n// #define __msa_maddv_b __builtin_msa_maddv_b\n// #define __msa_maddv_h __builtin_msa_maddv_h\n// #define __msa_maddv_w __builtin_msa_maddv_w\n// #define __msa_maddv_d __builtin_msa_maddv_d\n// #define __msa_msubv_b __builtin_msa_msubv_b\n// #define __msa_msubv_h __builtin_msa_msubv_h\n// #define __msa_msubv_w __builtin_msa_msubv_w\n// #define __msa_msubv_d __builtin_msa_msubv_d\n#define __msa_maddv_b(a, b, c) __builtin_msa_maddv_b(c, b, a)\n#define __msa_maddv_h(a, b, c) __builtin_msa_maddv_h(c, b, a)\n#define __msa_maddv_w(a, b, c) __builtin_msa_maddv_w(c, b, a)\n#define __msa_maddv_d(a, b, c) __builtin_msa_maddv_d(c, b, a)\n#define __msa_msubv_b(a, b, c) __builtin_msa_msubv_b(c, b, a)\n#define __msa_msubv_h(a, b, c) __builtin_msa_msubv_h(c, b, a)\n#define __msa_msubv_w(a, b, c) __builtin_msa_msubv_w(c, b, a)\n#define __msa_msubv_d(a, b, c) __builtin_msa_msubv_d(c, b, a)\n```\n\nBuild ncnn with mips msa and simpleocv enabled:\n```shell\nmkdir -p build\ncd build\ncmake -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_MSA=ON -DNCNN_MMI=ON -DNCNN_SIMPLEOCV=ON ..\ncmake --build . -j 2\ncmake --build . --target install\n```\n\nPick `build/install` folder for further usage.\n\nYou can run binary inside `build/examples` folder for testing.\n\n***\n\n### Build for Termux on Android\n\nInstall app Termux on your phone,and install Ubuntu in Termux.\n\n If you want use ssh, just install openssh in Termux\n\n```shell\npkg install proot-distro\nproot-distro install ubuntu\n```\n\nor you can see what system can be installed using `proot-distro list`\n\nwhile you install ubuntu successfully, using `proot-distro login ubuntu` to login Ubuntu.\n\nThen make ncnn,no need to install any other dependencies.\n\n```shell\ngit clone https://github.com/Tencent/ncnn.git\ncd ncnn\ngit submodule update --init\nmkdir -p build\ncd build\ncmake -DCMAKE_BUILD_TYPE=Release -DNCNN_BUILD_EXAMPLES=ON -DNCNN_PLATFORM_API=OFF -DNCNN_SIMPLEOCV=ON ..\nmake -j$(nproc)\n```\n\nThen you can run a test\n\n> on my Pixel 3 XL using Qualcomm 845,cant load `256-ncnn.png`\n\n```shell\ncd ../examples\n../build/examples/squeezenet ../images/128-ncnn.png\n```\n\n### Build for QNX\n\nRequest license and download SDP from QNX Software Center: https://www.qnx.com/products/everywhere/ .\n\nSetup QNX environment by invoking SDP's bundled script:\n\non Windows, open cmd and run\n```batch\ncall C:\\Users\\zz\\qnx800\\qnxsdp-env.bat\n```\n\non Linux, use /bin/bash and run\n```shell\nsource /home/zz/qnx800/qnxsdp-env.sh\n```\n\nIf it gives error `cannot find ld` on Linux, solve it by creaing link file:\n```shell\ncd ${QNX_HOST}/usr/bin/\nln -s aarch64-unknown-nto-qnx7.1.0-ld ld\n```\n\nBuild ncnn with cmake in same shell:\n\n```shell\ngit clone https://github.com/Tencent/ncnn.git\ncd ncnn\ngit submodule update --init\nmkdir -p build-qnx\ncd build-qnx\ncmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-qnx.toolchain.cmake ..\nmake -j$(nproc)\nmake install\n```\n\nPick `build-qnx/install` folder for further usage.\n\n### Build for Nintendo 3DS Homebrew Launcher\nInstall DevkitPRO toolchains\n- If you are working on windows, download DevkitPro installer from [DevkitPro](https://devkitpro.org/wiki/Getting_Started).\n- If you are using Ubuntu, the official guidelines from DevkitPro might not work for you. Try using the lines below to install\n```shell\nsudo apt-get update\nsudo apt-get upgrade\nwget https://apt.devkitpro.org/install-devkitpro-pacman\nchmod +x ./install-devkitpro-pacman\nsudo ./install-devkitpro-pacman\n```\n\n```shell\nexport DEVKITPRO=/opt/devkitpro\nexport DEVKITARM=/opt/devkitpro/devkitARM\nexport DEVKITPPC=/opt/devkitpro/devkitPPC\nexport export PATH=$/opt/devkitpro/tools/bin:$PATH\nsource ~/.profile\n```\n```shell\nsudo dkp-pacman -Sy\nsudo dkp-pacman -Syu\nsudo dkp-pacman -S 3ds-dev\n```\nCopy the toolchain files from [3DS-cmake](https://github.com/Xtansia/3ds-cmake)(DevitARM3DS.cmake and the cmake folder) to NCNN's toolchains folder.\n```\n├── toolchains\n│   ├── cmake\n│   │   ├── bin2s_header.h.in\n│   │   ├── FindCITRO3D.cmake\n│   │   ├── FindCTRULIB.cmake\n│   │   ├── FindFreetype.cmake\n│   │   ├── FindJPEG.cmake\n│   │   ├── FindPNG.cmake\n│   │   ├── FindSF2D.cmake\n│   │   ├── FindSFIL.cmake\n│   │   ├── FindSFTD.cmake\n│   │   ├── FindZLIB.cmake\n│   │   ├── LibFindMacros.cmake\n│   │   ├── Tools3DS.cmake\n│   │   ├── ToolsGBA.cmake\n│   │   └── try_add_imported_target.cmake\n│   ├── DevkitArm3DS.cmake\n...\n\n```\nBuild with:\n```shell\ncd ncnn\nmkdir build && cd build\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/DevkitArm3DS.cmake .. -DNCNN_SIMPLEOCV=ON -DNCNN_OPENMP=OFF -DNCNN_VFPV4=OFF ..\nmake -j4\nmake install\n```\nModify the Makefile in Homebrew example to link and use NCNN in your 3DS Homebrew app.\n\n***\n\n### Build for HarmonyOS with cross-compiling\nDownload and install HarmonyOS SDK. The sdk installation directory is `/opt/ohos-sdk/linux`\n\n```shell\ncd \nmkdir -p build\ncd build\n\nexport HM_SDK=/opt/ohos-sdk/linux\n\n# Choose HarmonyOS sdk cmake toolchain file.\n# If you want to enable vulkan, set -DNCNN_VULKAN=ON\n# The HarmonyOS sdk does not support openmp, use ncnn simpleomp instead.\n# Cross-compiling with CMake must use the one provided by the HarmonyOS SDK; otherwise, it won't recognize parameters like OHOS_PLATFORM, leading to compilation errors.\n${HM_SDK}/native/build-tools/cmake/bin/cmake -DOHOS_STL=c++_static -DOHOS_ARCH=arm64-v8a -DOHOS_PLATFORM=OHOS -DCMAKE_TOOLCHAIN_FILE=${HM_SDK}/native/build/cmake/ohos.toolchain.cmake -DNCNN_VULKAN=ON -DNCNN_SIMPLEOMP=ON ..\n\nmake -j$(nproc)\nmake install\n```\n\n***\n\n### Build for ESP32 with cross-compiling\nDownload esp-idf sdk\n```shell\ngit clone https://github.com/espressif/esp-idf\ncd esp-idf\ngit submodule update --init --recursive\n```\nInstall esp-idf sdk and configure the environment\n```shell\n./install.sh\nsource export.sh\n```\nAnd for Windows, you should use:\n```bash\ninstall.bat # or `install.ps1`\nexport.bat\n```\nNote: python>=3.8, cmake>=3.24.0\n\nBuild ncnn library:\n```shell\nmkdir build-esp32\ncd build-esp32\ncmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/esp32.toolchain.cmake -DCMAKE_BUILD_TYPE=Release ..\nmake -j 4\nmake install\n```\nNote: Make sure to compile in esp-idf environment.\n\nThe compiled ncnn library and headers can be put to the esp32 project to test.\n" }, { "path": "docs/how-to-use-and-FAQ/FAQ-ncnn-produce-wrong-result.md", "content": "### caffemodel should be row-major\n\n`caffe2ncnn` tool assumes the caffemodel is row-major (produced by c++ caffe train command).\n\nThe kernel 3x3 weights should be stored as\n```\na b c\nd e f\ng h i\n```\n\nHowever, matlab caffe produced col-major caffemodel.\n\nYou have to transpose all the kernel weights by yourself or re-training using c++ caffe train command.\n\nBesides, you may interest in https://github.com/conanhujinming/matcaffe2caffe\n\n### check input is RGB or BGR\n\nIf your caffemodel is trained using c++ caffe and opencv, then the input image should be BGR order.\n\nIf your model is trained using matlab caffe or pytorch or mxnet or tensorflow, the input image would probably be RGB order.\n\nThe channel order can be changed on-the-fly through proper pixel type enum\n```\n// construct RGB blob from rgb image\nncnn::Mat in_rgb = ncnn::Mat::from_pixels(rgb_data, ncnn::Mat::PIXEL_RGB, w, h);\n\n// construct BGR blob from bgr image\nncnn::Mat in_bgr = ncnn::Mat::from_pixels(bgr_data, ncnn::Mat::PIXEL_BGR, w, h);\n\n// construct BGR blob from rgb image\nncnn::Mat in_bgr = ncnn::Mat::from_pixels(rgb_data, ncnn::Mat::PIXEL_RGB2BGR, w, h);\n\n// construct RGB blob from bgr image\nncnn::Mat in_rgb = ncnn::Mat::from_pixels(bgr_data, ncnn::Mat::PIXEL_BGR2RGB, w, h);\n```\n\n\n### image decoding\n\nJPEG(`.jpg`,`.jpeg`) is loss compression, people may get different pixel value for same image on same position. \n\n`.bmp` images are recommended instead.\n\n### interpolation / resizing\n\nThere are several image resizing methods, which may generate different result for same input image.\n\nEven we specify same interpolation method, different frameworks/libraries and their various versions may also introduce difference.\n\nA good practice is feed same size image as the input layer expected, e.g. read a 224x244 bmp image when input layer need 224x224 size.\n\n\n### Mat::from_pixels/from_pixels_resize assume that the pixel data is continuous\n\nYou shall pass continuous pixel buffer to from_pixels family.\n\nIf your image is an opencv submat from an image roi, call clone() to get a continuous one.\n```\ncv::Mat image;// the image\ncv::Rect facerect;// the face rectangle\n\ncv::Mat faceimage = image(facerect).clone();// get a continuous sub image\n\nncnn::Mat in = ncnn::Mat::from_pixels(faceimage.data, ncnn::Mat::PIXEL_BGR, faceimage.cols, faceimage.rows);\n```\n\n### pre process\nApply pre process according to your training configuration\n\nDifferent model has different pre process config, you may find the following transform config in Data layer section\n```\ntransform_param {\n mean_value: 103.94\n mean_value: 116.78\n mean_value: 123.68\n scale: 0.017\n}\n```\nThen the corresponding code for ncnn pre process is\n```cpp\nconst float mean_vals[3] = { 103.94f, 116.78f, 123.68f };\nconst float norm_vals[3] = { 0.017f, 0.017f, 0.017f };\nin.substract_mean_normalize(mean_vals, norm_vals);\n```\n\nMean file is not supported currently\n\nSo you have to pre process the input data by yourself (use opencv or something)\n```\ntransform_param {\n mean_file: \"imagenet_mean.binaryproto\"\n}\n```\n\nFor pytorch or mxnet-gluon\n```python\ntransforms.ToTensor(),\ntransforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),\n```\nThen the corresponding code for ncnn pre process is\n```cpp\n// R' = (R / 255 - 0.485) / 0.229 = (R - 0.485 * 255) / 0.229 / 255\n// G' = (G / 255 - 0.456) / 0.224 = (G - 0.456 * 255) / 0.224 / 255\n// B' = (B / 255 - 0.406) / 0.225 = (B - 0.406 * 255) / 0.225 / 255\nconst float mean_vals[3] = {0.485f*255.f, 0.456f*255.f, 0.406f*255.f};\nconst float norm_vals[3] = {1/0.229f/255.f, 1/0.224f/255.f, 1/0.225f/255.f};\nin.substract_mean_normalize(mean_vals, norm_vals);\n```\n\n### use the desired blob\nThe blob names for input and extract are differ among models.\n\nFor example, squeezenet v1.1 use \"data\" as input blob and \"prob\" as output blob while mobilenet-ssd use \"data\" as input blob and \"detection_out\" as output blob.\n\nSome models may need multiple input or produce multiple output.\n\n```cpp\nncnn::Extractor ex = net.create_extractor();\n\nex.input(\"data\", in);// change \"data\" to yours\nex.input(\"mask\", mask);// change \"mask\" to yours\n\nex.extract(\"output1\", out1);// change \"output1\" to yours\nex.extract(\"output2\", out2);// change \"output2\" to yours\n```\n\n### blob may have channel gap\nEach channel pointer is aligned by 128bit in ncnn Mat structure.\n\nblob may have gaps between channels if (width x height) can not divided exactly by 4\n\nPrefer using ncnn::Mat::from_pixels or ncnn::Mat::from_pixels_resize for constructing input blob from image data\n\nIf you do need a continuous blob buffer, reshape the output.\n```cpp\n// out is the output blob extracted\nncnn::Mat flattened_out = out.reshape(out.w * out.h * out.c);\n\n// plain array, C-H-W\nconst float* outptr = flattened_out;\n```\n\n### create new Extractor for each image\nThe `ncnn::Extractor` object is stateful, if you reuse for different input, you will always get exact the same result cached inside.\n\nAlways create new Extractor to process images in loop unless you do know how the stateful Extractor works.\n```cpp\nfor (int i=0; i is not guaranteed to be 32bit aligned.\n\nyou can store your binary buffer in ncnn::Mat structure, its internal memory is aligned.\n\n### undefined reference to '__kmpc_XYZ_XYZ'\n\nuse clang for building android shared library\n\ncomment the following line in your Application.mk\n```\nNDK_TOOLCHAIN_VERSION := 4.9\n```\n\n### crash on android with '__kmp_abort_process'\n\nThis usually happens if you bundle multiple shared library with openmp linked\n\nIt is actually an issue of the android ndk https://github.com/android/ndk/issues/1028\n\nOn old android ndk, modify the link flags as\n\n```\n-Wl,-Bstatic -lomp -Wl,-Bdynamic\n```\n\nFor recent ndk >= 21\n\n```\n-fstatic-openmp\n```\n\n### dlopen failed: library \"libomp.so\" not found\n\nNewer android ndk defaults to dynamic openmp runtime\n\nmodify the link flags as\n\n```\n-fstatic-openmp -fopenmp\n```\n\n### crash when freeing a ncnn dynamic library(*.dll/*.so) built with openMP\n\nfor optimal performance, the openmp threadpool spin waits for about a second prior to shutting down in case more work becomes available. \n\nIf you unload a dynamic library that's in the process of spin-waiting, it will crash in the manner you see (most of the time).\n\nJust set OMP_WAIT_POLICY=passive in your environment, before calling loadlibrary. or Just wait a few seconds before calling freelibrary.\n\nYou can also use the following method to set environment variables in your code:\n\nfor msvc++:\n\n```\nSetEnvironmentVariable(_T(\"OMP_WAIT_POLICY\"), _T(\"passive\"));\n```\n\nfor g++:\n\n```\nsetenv(\"OMP_WAIT_POLICY\", \"passive\", 1)\n```\n\nreference: https://stackoverflow.com/questions/34439956/vc-crash-when-freeing-a-dll-built-with-openmp\n" }, { "path": "docs/how-to-use-and-FAQ/FAQ-ncnn-vulkan.md", "content": "### how to enable ncnn vulkan capability\n\nfollow [the build and install instruction](https://github.com/Tencent/ncnn/blob/master/docs/how-to-build/how-to-build.md)\n\nmake sure you have installed vulkan sdk from [lunarg vulkan sdk website](https://vulkan.lunarg.com/sdk/home)\n\nUsually, you can enable the vulkan compute inference feature by adding only one line of code to your application.\n\n```cpp\n// enable vulkan compute feature before loading\nncnn::Net net;\nnet.opt.use_vulkan_compute = 1;\n```\n\n### does my graphics device support vulkan\n\nSome platforms have been tested and known working. In theory, if your platform support vulkan api, either 1.0 or 1.1, it shall work.\n\n* Y = known work\n* ? = shall work, not confirmed\n* / = not applied\n\n| |windows|linux|android|mac|ios|\n|---|---|---|---|---|---|\n|intel|Y|Y|?|?|/|\n|amd|Y|Y|/|?|/|\n|nvidia|Y|Y|?|/|/|\n|qcom|/|/|Y|/|/|\n|apple|/|/|/|Y|Y|\n|arm|/|?|Y|/|/|\n\nYou can search [the vulkan database](https://vulkan.gpuinfo.org) to see if your device supports vulkan.\n\nSome old buggy drivers may produce wrong result, that are blacklisted in ncnn and treated as non-vulkan capable device.\nYou could check if your device and driver have this issue with [my conformance test here](vulkan-conformance-test).\nMost of these systems are android with version lower than 8.1.\n\n### why using vulkan over cuda/opencl/metal\n\nIn the beginning, I had no GPGPU programming experience, and I had to learn one.\n\nvulkan is considered more portable and well supported by vendors and the cross-platform low-overhead graphics api. As a contrast, cuda is only available on nvidia device, metal is only available on macos and ios, while loading opencl library is banned in android 7.0+ and does not work on ios.\n\n### I got errors like \"vkCreateComputePipelines failed -1000012000\" or random stalls or crashes\n\nUpgrade your vulkan driver.\n\n[intel https://downloadcenter.intel.com/product/80939/Graphics-Drivers](https://downloadcenter.intel.com/product/80939/Graphics-Drivers)\n\n[amd https://www.amd.com/en/support](https://www.amd.com/en/support)\n\n[nvidia https://www.nvidia.com/Download/index.aspx](https://www.nvidia.com/Download/index.aspx)\n\n### how to use ncnn vulkan on android\n\nminimum android ndk version: android-ndk-r18b\n\nminimum sdk platform api version: android-24\n\nlink your jni project with libvulkan.so\n\n[The squeezencnn example](https://github.com/Tencent/ncnn/tree/master/examples/squeezencnn) have equipped gpu inference, you could take it as reference.\n\n### how to use ncnn vulkan on ios\n\nsetup vulkan sdk (https://vulkan.lunarg.com/sdk/home#mac)\n\nmetal only works on real device with arm64 cpu (iPhone 5s and later)\n\nlink your project with MoltenVK framework and Metal\n\n### what about the layers without vulkan support\n\nThese layers have vulkan support currently\n\nAbsVal, BatchNorm, BinaryOp, Cast, Clip, Concat, Convolution, ConvolutionDepthWise, Crop, Deconvolution, DeconvolutionDepthWise, Dropout, Eltwise, Flatten, HardSigmoid, InnerProduct, Interp, LRN, Packing, Padding, Permute, Pooling(pad SAME not supported), PReLU, PriorBox, ReLU, Reorg, Reshape, Scale, ShuffleChannel, Sigmoid, Softmax, TanH, UnaryOp\n\nFor these layers without vulkan support, ncnn inference engine will automatically fallback to cpu path.\n\nThus, it is usually not a serious issue if your network only has some special head layers like SSD or YOLO. All examples in ncnn are known working properly with vulkan enabled.\n\n### my model runs slower on gpu than cpu\n\nThe current vulkan inference implementation is far from the preferred state. Many handful optimization techniques are planned, such as winograd convolution, operator fusion, fp16 storage and arithmetic etc.\n\nIt is common that your model runs slower on gpu than cpu on arm devices like mobile phones, since we have quite good arm optimization in ncnn ;)\n\n### vulkan device not found / extra high cpu utility while vulkan is enabled on nvidia gpu\n\nThere are several reasons could lead to this outcome. First please check your driver status with `nvidia-smi`. If you have correctly installed your driver, you should see something like this:\n\n```bash\n$ nvidia-smi\nSat Mar 06 19:53:16 2021\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 451.48 Driver Version: 451.48 CUDA Version: 11.0 |\n|-------------------------------+----------------------+----------------------+\n| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\n|===============================+======================+======================|\n| 0 GeForce GTX 1060 WDDM | 00000000:02:00.0 Off | N/A |\n| N/A 31C P8 5W / N/A | 90MiB / 6144MiB | 0% Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes: |\n| GPU GI CI PID Type Process name GPU Memory |\n| ID ID Usage |\n|=============================================================================|\n| No running processes found |\n+-----------------------------------------------------------------------------+\n```\n\nIf `nvidia-smi` crashes or cannot be found, please reinstall your graphics driver.\n\nIf ncnn *is* utilizing the Tesla GPU, you can see your program in the `Processes` block at the bottom. In that case, it's likely some operators are not yet supported in Vulkan, and have fallbacked to the CPU, thus leading to a low utilization of the GPU.\n\nIf you *couldn't* find your process running, plase check the active driver model, which can be found to the right of your device name. For Geforce and Titan GPUs, the default driver model is WDDM (Windows Desktop Driver Model), which supports both rendering graphics as well as computing. But for Tesla GPUs, without configuration, the driver model is defualted to TCC ([Tesla Computing Cluster](https://docs.nvidia.com/gameworks/content/developertools/desktop/tesla_compute_cluster.htm)). NVIDIA's TCC driver does not support Vulkan, so you need to use the following command to set the driver model back to WDDM, to use Vulkan:\n\n```bash\n$ nvidia-smi -g 0 -dm 0\n```\n\nThe number following `-g` is the GPU ID (which can be found to the left of your device name in `nvidia-smi` output); and `-dm` stands for driver model, 0 refers to WDDM and 1 means TCC.\n" }, { "path": "docs/how-to-use-and-FAQ/build-minimal-library.md", "content": "For some reason, if you're not happy with the binary size of the ncnn library, then here is the cheatsheet that helps you to build a minimal ncnn :P\n\n### disable c++ rtti and exceptions\n\n```\ncmake -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON ..\n```\n* Cannot use RTTI and Exceptions when ncnn functions are called.\n\n### disable vulkan support\n\n```\ncmake -DNCNN_VULKAN=OFF ..\n```\n\n* Cannot use GPU acceleration.\n\n### disable NCNN_STDIO\n\n```\ncmake -DNCNN_STDIO=OFF ..\n```\n\n* Cannot load model from files, but can load model from memory or by Android Assets.\n\n Read more [here](https://github.com/Tencent/ncnn/blob/master/docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.md#load-model).\n\n### disable NCNN_STRING\n\n```\ncmake -DNCNN_STRING=OFF ..\n```\n\n* Cannot load human-readable param files with visible strings, but can load binary param.bin files.\n\n Read more [here](https://github.com/Tencent/ncnn/blob/master/docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.md#strip-visible-string)\n\n* Cannot identify blobs by string name when calling `Extractor::input / extract`, but can identify them by enum value in `id.h`.\n\n Read more [here](https://github.com/Tencent/ncnn/blob/master/docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.md#input-and-output).\n\n### disable NCNN_BF16\n\n```\ncmake -DNCNN_BF16=OFF ..\n```\n\n* Cannot use bf16 storage type in inference.\n\n\n### disable NCNN_INT8\n\n```\ncmake -DNCNN_INT8=OFF ..\n```\n\n* Cannot use quantized int8 inference.\n\n\n### drop pixel drawing functions\n\n```\ncmake -DNCNN_PIXEL_DRAWING=OFF ..\n```\n\n* Cannot use functions doing drawing basic shape and text like `ncnn::draw_rectangle_xx / ncnn::draw_circle_xx / ncnn::draw_text_xx`, but functions like `Mat::from_pixels / from_pixels_resize` are still available.\n\n\n### drop pixel rotate and affine functions\n\n```\ncmake -DNCNN_PIXEL_ROTATE=OFF -DNCNN_PIXEL_AFFINE=OFF ..\n```\n\n* Cannot use functions doing rotatation and affine transformation like `ncnn::kanna_rotate_xx / ncnn::warpaffine_bilinear_xx`, but functions like `Mat::from_pixels / from_pixels_resize` are still available. \n\n### drop pixel functions\n\n```\ncmake -DNCNN_PIXEL=OFF ..\n```\n\n* Cannot use functions transferring from image to pixels like `Mat::from_pixels / from_pixels_resize / to_pixels / to_pixels_resize`, and need create a Mat and fill in data by hand.\n\n### disable openmp\n\n```\ncmake -DNCNN_OPENMP=OFF ..\n```\n\n* Cannot use openmp multi-threading acceleration. If you want to run a model in single thread on your target machine, it is recommended to close the option.\n\n### disable avx2 and arm82 optimized kernel\n\n```\ncmake -DNCNN_AVX2=OFF -DNCNN_ARM82=OFF ..\n```\n\n* Do not compile optimized kernels using avx2 / arm82 instruction set extensions. If your target machine does not support some of them, it is recommended to close the related options.\n\n### disable runtime cpu instruction dispatch\n\n```\ncmake -DNCNN_RUNTIME_CPU=OFF ..\n```\n\n* Cannot check supported cpu instruction set extensions and use related optimized kernels in runtime.\n* If you know which instruction set extensions are supported on your target machine like avx2 / arm82, you can open related options like `-DNCNN_AVX2=ON / -DNCNN_ARM82=ON` by hand and then sse2 / arm8 version kernels will not be compiled.\n\n### drop layers not used\n\n```\ncmake -DWITH_LAYER_absval=OFF -DWITH_LAYER_bnll=OFF ..\n```\n\n* If your model does not include some layers, taking absval / bnll as a example above, you can drop them.\n* Some key or dependency layers should not be dropped, like convolution / innerproduct, their dependency like padding / flatten, and activation like relu / clip.\n\n### disable c++ stl\n\n```\ncmake -DNCNN_SIMPLESTL=ON ..\n```\n\n* STL provided by compiler is no longer depended on, and use `simplestl` provided by ncnn as a replacement. Users also can only use `simplestl` when ncnn functions are called.\n* Usually with compiler parameters `-nodefaultlibs -fno-builtin -nostdinc++ -lc`\n* Need cmake parameters `cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_STL=system` to avoid STL conflict when compiling to Android.\n\n### drop optimized kernel not used\n\n* Modify the source code under `ncnn/src/layer/arm/` to delete unnecessary optimized kernels or replace them with empty functions.\n* You can also drop layers and related optimized kernels by `-DWITH_LAYER_absval=OFF` as mentioned above.\n\n### drop operators from BinaryOp UnaryOp\n\n* Modify `ncnn/src/layer/binaryop.cpp unaryop.cpp` and `ncnn/src/layer/arm/binaryop.cpp unaryop_arm.cpp` by hand to delete unnecessary operators.\n" }, { "path": "docs/how-to-use-and-FAQ/efficient-roi-resize-rotate.md", "content": "\n### image roi crop + convert to ncnn::Mat\n\n```\n+--------------+\n| y | /-------/\n| x +-------+ | +-------+|\n| | roih |im_h => | roih\n| +-roiw--+ | +-roiw--+/\n| |\n+-----im_w-----+\n```\n```cpp\nncnn::Mat in = ncnn::Mat::from_pixels_roi(im.data, ncnn::Mat::PIXEL_RGB, im_w, im_h, x, y, roiw, roih);\n```\nFor Android Application, it is :\n```cpp\nncnn::Mat in = ncnn::Mat::from_android_bitmap_roi(env, image, ncnn::Mat::PIXEL_RGBA2RGB, x, y, roiw, roih);\n```\n\n### image roi crop + resize + convert to ncnn::Mat\n\n```\n+--------------+\n| y | /----/\n| x +-------+ | +----+|\n| | roih |im_h => | target_h\n| +-roiw--+ | | ||\n| | +----+/\n+-----im_w-----+ target_w\n```\n```cpp\nncnn::Mat in = ncnn::Mat::from_pixels_roi_resize(im.data, ncnn::Mat::PIXEL_RGB, im_w, im_h, x, y, roiw, roih, target_w, target_h);\n```\nFor Android Application, it is :\n```cpp\nncnn::Mat in = ncnn::Mat::from_android_bitmap_roi_resize(env, image, ncnn::Mat::PIXEL_RGBA2RGB, x, y, roiw, roih, target_w, target_h);\n```\n\n### ncnn::Mat export image + offset paste\n\n```\n +--------------+\n /-------/ | y |\n+-------+| | x +-------+ |\n| h| => | | h |im_h\n+---w---+/ | +---w---+ |\n | |\n +-----im_w-----+\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nout.to_pixels(data, ncnn::Mat::PIXEL_RGB, im_w * 3);\n```\n\n### ncnn::Mat export image + resize + roi paste\n\n```\n +--------------+\n /----/ | y |\n+----+| | x +-------+ |\n| h| => | | roih|im_h\n| || | +-roiw--+ |\n+-w--+/ | |\n +-----im_w-----+\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nout.to_pixels_resize(data, ncnn::Mat::PIXEL_RGB, roiw, roih, im_w * 3);\n```\n\n### image roi crop + resize\n```\n+--------------+\n| y |\n| x +-------+ | +----+\n| | roih|im_h => | target_h\n| +-roiw--+ | | |\n| | +----+\n+-----im_w-----+ target_w\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nncnn::resize_bilinear_c3(data, roiw, roih, im_w * 3, outdata, target_w, target_h, target_w * 3);\n```\n\n### image resize + offset paste\n```\n +--------------+\n | y |\n+----+ | x +-------+ |\n| h => | | roih |im_h\n| | | +-roiw--+ |\n+-w--+ | |\n +-----im_w-----+\n```\n```cpp\nunsigned char* outdata = im.data + (y * im_w + x) * 3;\nncnn::resize_bilinear_c3(data, w, h, w * 3, outdata, roiw, roih, im_w * 3);\n```\n\n### image roi crop + resize + roi paste\n```\n+--------------+ +-----------------+\n| y | | roiy |\n| x +-------+ | |roix----------+ |\n| | h |im_h => | | target_h|outim_h\n| +---w---+ | | | | |\n| | | +-target_w-+ |\n+-----im_w-----+ +-----outim_w-----+\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nunsigned char* outdata = outim.data + (roiy * outim_w + roix) * 3;\nncnn::resize_bilinear_c3(data, w, h, im_w * 3, outdata, target_w, target_h, outim_w * 3);\n```\n\n### image roi crop + rotate\n```\n+--------------+\n| y |\n| x +-------+ | +---+\n| | < < h |im_h => | ^ |w\n| +---w---+ | | ^ |\n| | +---+\n+-----im_w-----+ h\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nncnn::kanna_rotate_c3(data, w, h, im_w * 3, outdata, h, w, h * 3, 6);\n```\n\n### image rotate + offset paste\n```\n +--------------+\n | y |\n +---+ | x +-------+ |\n | ^ |h => | | < < w |im_h\n | ^ | | +---h---+ |\n +---+ | |\n w +-----im_w-----+\n```\n```cpp\nunsigned char* outdata = im.data + (y * im_w + x) * 3;\nncnn::kanna_rotate_c3(data, w, h, w * 3, outdata, h, w, im_w * 3, 7);\n```\n\n### image roi crop + rotate + roi paste\n```\n+--------------+ +-----------------+\n| y | | roiy |\n| x +-------+ | | roix +---+ |\n| | < < h |im_h => | | ^ w |outim_h\n| +---w---+ | | | ^ | |\n| | | +-h-+ |\n+-----im_w-----+ +-----outim_w-----+\n```\n```cpp\nconst unsigned char* data = im.data + (y * im_w + x) * 3;\nunsigned char* outdata = outim.data + (roiy * outim_w + roix) * 3;\nncnn::kanna_rotate_c3(data, w, h, im_w * 3, outdata, h, w, outim_w * 3, 6);\n```\n" }, { "path": "docs/how-to-use-and-FAQ/ncnn-load-model.md", "content": "### the comprehensive model loading api table\n\n|load from|alexnet.param|alexnet.param.bin|alexnet.bin|\n|---|---|---|---|\n|file path|load_param(const char*)|load_param_bin(const char*)|load_model(const char*)|\n|file path
(wchar_t for windows)|load_param(const wchar_t*)|load_param_bin(const wchar_t*)|load_model(const wchar_t*)|\n|file descriptor|load_param(FILE*)|load_param_bin(FILE*)|load_model(FILE*)|\n|file memory|load_param_mem(const char*)|load_param(const unsigned char*)|load_model(const unsigned char*)|\n|android asset|load_param(AAsset*)|load_param_bin(AAsset*)|load_model(AAsset*)|\n|android asset path|load_param(AAssetManager*, const char*)|load_param_bin(AAssetManager*, const char*)|load_model(AAssetManager*, const char*)|\n|custom IO reader|load_param(const DataReader&)|load_param_bin(const DataReader&)|load_model(const DataReader&)|\n\n### points to note\n\n1. Either of the following combination shall be enough for loading model\n * alexnet.param + alexnet.bin\n * alexnet.param.bin + alexnet.bin\n\n2. Never modify Net opt member after loading\n\n3. Most loading functions return 0 if success, except loading alexnet.param.bin and alexnet.bin from file memory, which returns the bytes consumed after loading\n * size_t Net::load_param(const unsigned char*)\n * size_t Net::load_model(const unsigned char*)\n\n4. It is recommended to load model from Android asset directly to avoid copying them to sdcard on Android platform\n\n5. The custom IO reader interface can be used to implement on-the-fly model decryption and loading\n" }, { "path": "docs/how-to-use-and-FAQ/openmp-best-practice.md", "content": "ncnn openmp best practice\r\n\r\n### CPU loadaverage is too high with ncnn.\r\n\r\n When inference the neural network with ncnn, the cpu occupancy is very high even all CPU cores occupancy close to 100%.\r\n\r\n If there are other threads or processes that require more cpu resources, the running speed of the program will drop severely.\r\n\r\n### The root cause of high CPU usage\r\n\r\n1. ncnn uses openmp API to speed up the inference compute. the thread count equals to the cpu core count. If the computing work need to run frequently, it must consume many cpu resources.\r\n\r\n2. There is a thread pool managed by openmp, the pool size is equal to the cpu core size. (the max vulue is 15 if there are much more cpu cores?)\r\n Openmp need to sync the thread when acquiring and returning threads to the pool. In order to improve efficiency, almost all omp implementations use spinlock synchronization (except for simpleomp). \r\n The default spin time of the spinlock is 200ms. So after a thread is scheduled, the thread need to busy-wait up to 200ms.\r\n\r\n### Why the CPU usage is still high even using vulkan GPU acceleration.\r\n\r\n1. Openmp is also used when loading the param bin file, and this part runs on cpu.\r\n\r\n2. The fp32 to fp16 conversion before and after the GPU memory upload is executed on the cpu, and this part of the logic also uses openmp.\r\n\r\n### Solution\r\n```\r\n1. Bind to the specific cpu core.\r\n```\r\n If you use a device with large and small core CPUs, it is recommended to bind large or small cores through ncnn::set_cpu_powersave(int). Note that Windows does not support binding cores. By the way, it's possible to have multiple threadpool using openmp. A new threadpool will be created for a new thread scope.\r\nSuppose your platform is 2 big cores + 4 little cores, and you want to execute model A on 2 big cores and model B on 4 little cores concurrently.\r\n\r\ncreate two threads via std::thread or pthread\r\n ```\r\n void thread_1()\r\n {\r\n ncnn::set_cpu_powersave(2); // bind to big cores\r\n netA.opt.num_threads = 2;\r\n }\r\n\r\n void thread_2()\r\n {\r\n ncnn::set_cpu_powersave(1); // bind to little cores\r\n netB.opt.num_threads = 4;\r\n }\r\n ```\r\n \r\n```\r\n2. Use fewer threads.\r\n```\r\n Set the number of threads to half of the cpu cores count or less through ncnn::set_omp_num_threads(int) or change net.opt.num_threads field. If you are coding with clang libomp, it's recommended that the number of threads does not exceed 8. If you use other omp libraries, it is recommended that the number of threads does not exceed 4.\r\n```\r\n3. Reduce openmp spinlock blocktime.\r\n```\r\n You can modify openmp blocktime by call ncnn::set_kmp_blocktime(int) method or modify net.opt.openmp_blocktime field.\r\n This argument is the spin time set by the ncnn API, and the default is 20ms.You can set a smaller value according to\r\n the situation, or directly change it to 0.\r\n\r\n Limitations: At present, only the libomp library of clang is implemented. Neither vcomp nor libgomp have corresponding interfaces.\r\n If it is not compiled with clang, this value is still 200ms by default.\r\n If you use vcomp or libgomp, you can use the environment variable OMP_WAIT_POLICY=PASSIVE to disable spin time. If you use simpleomp,\r\n It's no need to set this parameter.\r\n```\r\n4. Limit the number of threads available in the openmp thread pool.\r\n```\r\n Even if the number of openmp threads is reduced, the CPU occupancy rate may still be high. This is more common on servers with\r\n particularly many CPU cores. \r\n This is because the waiting threads in the thread pool use a spinlock to busy-wait, which can be reducedby limiting the number of\r\n threads available in the thread pool.\r\n\r\n Generally, you can set the OMP_THREAD_LIMIT environment variable. simpleomp currently does not support this feature so it's no need to be set.\r\n Note that this environment variable is only valid if it is set before the program starts.\r\n```\r\n5. Disable openmp completely\r\n```\r\n If there is only one cpu core, or use the vulkan gpu acceleration, it is recommended to disable openmp, just specify -DNCNN_OPENMP=OFF\r\n when compiling with cmake." }, { "path": "docs/how-to-use-and-FAQ/openmp-best-practice.zh.md", "content": "ncnn openmp 最佳实践\r\n\r\n### ncnn占用过多cpu资源\r\n\r\n 使用ncnn推理运算，cpu占用非常高甚至所有核心占用都接近100%。\r\n\r\n 如果还有其它线程或进程需要较多的cpu资源，运行速度下降严重。\r\n\r\n### cpu占用高的根本原因\r\n\r\n1. ncnn使用openmp API控制多线程加速推理计算。默认情况下，线程数等于cpu内核数。如果推理需要高频率运行，必然占用大部分\r\n cpu资源。\r\n\r\n2. openmp内部维护一个线程池，线程池最大可用线程数等于cpu内核数。(核心过多时最大限制是15？）获取和归还线程时需要同步。\r\n\r\n 为了提高效率，几乎所有omp实现都使用了自旋锁同步(simpleomp除外)。自旋锁默认的spin time是200ms。因此一个线程被调度后，\r\n 需要忙等待最多200ms。\r\n\r\n### 为什么使用vulkan加速后cpu占用依然很高。\r\n\r\n1. 加载参数文件时也使用了openmp，这部分是在cpu上运行的。\r\n\r\n2. 显存上传前和下载后的 fp32 fp16转换是在cpu上执行的，这部分逻辑也使用了openmp。\r\n\r\n### 解决方法\r\n\r\n```\r\n1. 绑核\r\n```\r\n 如果使用有大小核cpu的设备，建议通过ncnn::set_cpu_powersave(int)绑定大核或小核，注意windows系统不支持绑核。顺便说一下，ncnn支持不同的模型运行在不同的核心。假设硬件平台有2个大核，4个小核，你想把netA运行在大核，netB运行在小核。\r\n 可以通过std::thread or pthread创建两个线程，运行如下代码：\r\n \r\n ```\r\n void thread_1()\r\n {\r\n ncnn::set_cpu_powersave(2); // bind to big cores\r\n netA.opt.num_threads = 2;\r\n }\r\n\r\n void thread_2()\r\n {\r\n ncnn::set_cpu_powersave(1); // bind to little cores\r\n netB.opt.num_threads = 4;\r\n }\r\n ```\r\n\r\n```\r\n2. 使用更少的线程数。\r\n```\r\n 通过ncnn::set_omp_num_threads(int)或者net.opt.num_threads字段设置线程数为cpu内核数的一半或更小。如果使用clang的libomp，\r\n 建议线程数不超过8，如果使用其它omp库，建议线程数不超过4。\r\n```\r\n3. 减小openmp blocktime。\r\n```\r\n 可以修改ncnn::set_kmp_blocktime(int)或者修改net.opt.openmp_blocktime，这个参数是ncnn API设置的spin time，默认是20ms。\r\n 可以根据情况设置更小的值，或者直接改为0。\r\n\r\n 局限：目前只有clang的libomp库有实现，vcomp和libgomp都没有相应接口，如果不是使用clang编译的，这个值默认还是200ms。\r\n 如果使用vcomp或libgomp, 可以使用环境变量OMP_WAIT_POLICY=PASSIVE禁用spin time，如果使用simpleomp,不需要设置这个参数。\r\n```\r\n4. 限制openmp线程池可用线程数量。\r\n```\r\n 即使减小了openmp线程数量，cpu占用率仍然可能会很高。这在cpu核心特别多的服务器上比较常见。这是因为线程池中的等待线程使用\r\n 自旋锁忙等待，可以通过限制线程池可用线程数量减轻这种影响。\r\n\r\n 一般可以通过设置OMP_THREAD_LIMIT环境变量。simpleomp目前不支持这一特性，不需要设置。注意这个环境变量仅在程序启动前设置才有效。\r\n```\r\n5. 完全禁用openmp\r\n```\r\n 如果只有一个cpu核心，或者使用vulkan加速，建议关闭openmp, cmake编译时指定-DNCNN_OPENMP=OFF即可。" }, { "path": "docs/how-to-use-and-FAQ/quantized-int8-inference.md", "content": "# Post Training Quantization Tools\n\nTo support int8 model deployment on mobile devices,we provide the universal post training quantization tools which can convert the float32 model to int8 model.\n\n## User Guide\n\nExample with mobilenet, just need three steps.\n\n### 1. Optimize model\n\nNOTE: **If your model is converted via pnnx, skip this step.**\n\n```shell\n./ncnnoptimize mobilenet.param mobilenet.bin mobilenet-opt.param mobilenet-opt.bin 0\n```\n\n### 2. Create the calibration table file\n\n#### 2.1 From image\n\nWe suggest that using the verification dataset for calibration, which is more than 5000 images.\n\nSome imagenet sample images here https://github.com/nihui/imagenet-sample-images\n\n```shell\nfind images/ -type f > imagelist.txt\n./ncnn2table mobilenet-opt.param mobilenet-opt.bin imagelist.txt mobilenet.table mean=[104,117,123] norm=[0.017,0.017,0.017] shape=[224,224,3] pixel=BGR thread=8 method=kl\n```\n\n* mean and norm are the values you passed to ```Mat::substract_mean_normalize()```\n* shape is the blob shape of your model, [w,h] or [w,h,c]\n\n>\n * if w and h both are given, image will be resized to exactly size.\n * if w and h both are zero or negative, image will not be resized.\n * if only h is zero or negative, image's width will scaled resize to w, keeping aspect ratio.\n * if only w is zero or negative, image's height will scaled resize to h\n\n* pixel is the pixel format of your model, image pixels will be converted to this type before ```Extractor::input()```\n* thread is the CPU thread count that could be used for parallel inference\n* method is the post training quantization algorithm, kl and aciq are currently supported\n\nIf your model has multiple input nodes, you can use multiple list files and other parameters\n\n```shell\n./ncnn2table mobilenet-opt.param mobilenet-opt.bin imagelist-bgr.txt,imagelist-depth.txt mobilenet.table mean=[104,117,123],[128] norm=[0.017,0.017,0.017],[0.0078125] shape=[224,224,3],[224,224,1] pixel=BGR,GRAY thread=8 method=kl\n```\n\n#### 2.2 From npy\n\nWe suggest that using the validation(development) set for calibration.\n\nUse the same preprocessing as the training set to get the input vectors, in the case of batchsize=1, store each input vector as an npy file, n inputs correspond to n npy files, the actual stored vectors to remove the batch dimension.\n\n\ntest net, shape is in NCHW format, but there's no `N`.\n```txt\nin0, shape=[512]\nin1, shape=[2, 1, 64]\nin2, shape=[2, 1, 64]\n```\n\nfilelist_in0.txt\n```txt\n0_in0.npy\n1_in0.npy\n2_in0.npy\n...\n```\n\nfilelist_in1.txt\n```txt\n0_in1.npy\n1_in1.npy\n2_in1.npy\n...\n```\n\nfilelist_in2.txt\n```txt\n0_in2.npy\n1_in2.npy\n2_in2.npy\n...\n```\n\n```shell\n./ncnn2table test.param test.bin filelist_in0.txt,filelist_in1.txt,filelist_in2.txt test.table shape=[512],[64,1,2],[64,1,2] thread=8 method=kl type=1\n```\n**Here shape is WHC, because the order of the arguments to `ncnn::Mat`.**\n\n### 3. Quantize model\n\n```shell\n./ncnn2int8 mobilenet-opt.param mobilenet-opt.bin mobilenet-int8.param mobilenet-int8.bin mobilenet.table\n```\n\nIf you don’t need static quantization, ncnn supports RNN/LSTM/GRU dynamic quantization. In this case, you can omit the table file.\n\n```shell\n./ncnn2int8 rnn-model.param rnn-model.bin rnn-model-int8.param rnn-model-int8.bin\n```\n\n## use ncnn int8 inference\n\nthe ncnn library would use int8 inference automatically, nothing changed in your code\n\n```cpp\nncnn::Net mobilenet;\nmobilenet.load_param(\"mobilenet-int8.param\");\nmobilenet.load_model(\"mobilenet-int8.bin\");\n```\n\n## mixed precision inference\n\nBefore quantize your model, comment the layer weight scale line in table file, then the layer will do the float32 inference\n\n```\nconv1_param_0 156.639840536\n```\n\n```\n#conv1_param_0 156.639840536\n```\n" }, { "path": "docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.md", "content": "We use alexnet as an example\n\n### prepare caffe prototxt and model\n\nThese files will usually generated when trained with caffe\n```\ntrain.prototxt\ndeploy.prototxt\nsnapshot_10000.caffemodel\n```\ndeploy.prototxt and caffemodel file are enough for TEST phase\n\nalexnet deploy.prototxt can be downloaded here\n\nhttps://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet\n\nalexnet caffemodel can be downloaded here\n\nhttp://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel\n\n### convert to ncnn model\n\nConvert old caffe prototxt and caffemodel to new ones using tools in caffe\n\nbecause the ncnn convert tool needs the new format\n```\nupgrade_net_proto_text [old prototxt] [new prototxt]\nupgrade_net_proto_binary [old caffemodel] [new caffemodel]\n```\n\nUse Input layer as input, set N dim as 1 since only one image can be processed each time\n```\nlayer {\n name: \"data\"\n type: \"Input\"\n top: \"data\"\n input_param { shape: { dim: 1 dim: 3 dim: 227 dim: 227 } }\n}\n```\nUse caffe2ncnn tool to convert caffe model to ncnn model\n```\ncaffe2ncnn deploy.prototxt bvlc_alexnet.caffemodel alexnet.param alexnet.bin\n```\n\n### strip visible string\n\nIt is already enough for deploying with param and bin file only, but there are visible strings in param file, it may not be suitable to distribute plain neural network information in your APP.\n\nYou can use ncnn2mem tool to convert plain model file to binary representation. It will generate alexnet.param.bin and two static array code files.\n```\nncnn2mem alexnet.param alexnet.bin alexnet.id.h alexnet.mem.h\n```\n\n### load model\n\nLoad param and bin file, the easy way\n```cpp\nncnn::Net net;\nnet.load_param(\"alexnet.param\");\nnet.load_model(\"alexnet.bin\");\n```\nLoad binary param.bin and bin file, no visible strings included, suitable for bundled as APP resource\n```cpp\nncnn::Net net;\nnet.load_param_bin(\"alexnet.param.bin\");\nnet.load_model(\"alexnet.bin\");\n```\nLoad network and model from external memory, no visible strings included, no external resource files bundled, the whole model is hardcoded in your program\n\nYou may use this way to load from android asset resource\n```cpp\n#include \"alexnet.mem.h\"\nncnn::Net net;\nnet.load_param(alexnet_param_bin);\nnet.load_model(alexnet_bin);\n```\nYou can choose either way to load model. Loading from external memory is zero-copy, which means you must keep your memory buffer during processing\n\n### unload model\n```cpp\nnet.clear();\n```\n\n### input and output\n\nncnn Mat is the data structure for input and output data\n\nInput image should be converted to Mat, and subtracted mean values and normalized when needed\n\n```cpp\n#include \"mat.h\"\nunsigned char* rgbdata;// data pointer to RGB image pixels\nint w;// image width\nint h;// image height\nncnn::Mat in = ncnn::Mat::from_pixels(rgbdata, ncnn::Mat::PIXEL_RGB, w, h);\n\nconst float mean_vals[3] = {104.f, 117.f, 123.f};\nin.substract_mean_normalize(mean_vals, 0);\n```\nExecute the network inference and retrieve the result\n```cpp\n#include \"net.h\"\nncnn::Mat in;// input blob as above\nncnn::Mat out;\nncnn::Extractor ex = net.create_extractor();\nex.input(\"data\", in);\nex.extract(\"prob\", out);\n```\nIf you load model with binary param.bin file, you should use the enum value in alexnet.id.h file instead of the blob name\n```cpp\n#include \"net.h\"\n#include \"alexnet.id.h\"\nncnn::Mat in;// input blob as above\nncnn::Mat out;\nncnn::Extractor ex = net.create_extractor();\nex.input(alexnet_param_id::BLOB_data, in);\nex.extract(alexnet_param_id::BLOB_prob, out);\n```\nRead the data in the output Mat. Iterate data to get all classification scores.\n```cpp\nncnn::Mat out_flatterned = out.reshape(out.w * out.h * out.c);\nstd::vector scores;\nscores.resize(out_flatterned.w);\nfor (int j=0; j $ cat alexnet.param.bin alexnet.bin > alexnet-all.bin\n\n```cpp\n#include \"net.h\"\nFILE* fp = fopen(\"alexnet-all.bin\", \"rb\");\nnet.load_param_bin(fp);\nnet.load_model(fp);\nfclose(fp);\n```\n" }, { "path": "docs/how-to-use-and-FAQ/use-ncnn-with-alexnet.zh.md", "content": "首先，非常感谢大家对 ncnn 组件的关注\n为了方便大家使用 ncnn 组件，up主特意写了这篇使用指北，以烂大街的 alexnet 作为例子\n\n\n### 准备caffe网络和模型\n\ncaffe 的网络和模型通常是搞深度学习的研究者训练出来的，一般来说训练完会有\n```\ntrain.prototxt\ndeploy.prototxt\nsnapshot_10000.caffemodel\n```\n部署的时候只需要 TEST 过程，所以有 deploy.prototxt 和 caffemodel 就足够了\n\nalexnet 的 deploy.prototxt 可以在这里下载\nhttps://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet\n\nalexnet 的 caffemodel 可以在这里下载\nhttp://dl.caffe.berkeleyvision.org/bvlc_alexnet.caffemodel\n\n### 转换ncnn网络和模型\n\ncaffe 自带了工具可以把老版本的 caffe 网络和模型转换为新版（ncnn的工具只认识新版\n```\nupgrade_net_proto_text [老prototxt] [新prototxt]\nupgrade_net_proto_binary [老caffemodel] [新caffemodel]\n```\n输入层改用 Input，因为每次只需要做一个图片，所以第一个 dim 设为 1\n```\nlayer {\n name: \"data\"\n type: \"Input\"\n top: \"data\"\n input_param { shape: { dim: 1 dim: 3 dim: 227 dim: 227 } }\n}\n```\n使用 caffe2ncnn 工具转换为 ncnn 的网络描述和模型\n```\ncaffe2ncnn deploy.prototxt bvlc_alexnet.caffemodel alexnet.param alexnet.bin\n```\n### 去除可见字符串\n\n有 param 和 bin 文件其实已经可以用了，但是 param 描述文件是明文的，如果放在 APP 分发出去容易被窥探到网络结构（说得好像不明文就看不到一样\n使用 ncnn2mem 工具转换为二进制描述文件和内存模型，生成 alexnet.param.bin 和两个静态数组的代码文件\n```\nncnn2mem alexnet.param alexnet.bin alexnet.id.h alexnet.mem.h\n```\n### 加载模型\n\n直接加载 param 和 bin，适合快速验证效果使用\n```cpp\nncnn::Net net;\nnet.load_param(\"alexnet.param\");\nnet.load_model(\"alexnet.bin\");\n```\n加载二进制的 param.bin 和 bin，没有可见字符串，适合 APP 分发模型资源\n```cpp\nncnn::Net net;\nnet.load_param_bin(\"alexnet.param.bin\");\nnet.load_model(\"alexnet.bin\");\n```\n从内存引用加载网络和模型，没有可见字符串，模型数据全在代码里头，没有任何外部文件\n另外，android apk 打包的资源文件读出来也是内存块\n```cpp\n#include \"alexnet.mem.h\"\nncnn::Net net;\nnet.load_param(alexnet_param_bin);\nnet.load_model(alexnet_bin);\n```\n以上三种都可以加载模型，其中内存引用方式加载是 zero-copy 的，所以使用 net 模型的来源内存块必须存在\n\n### 卸载模型\n```cpp\nnet.clear();\n```\n\n### 输入和输出\n\nncnn 用自己的数据结构 Mat 来存放输入和输出数据\n输入图像的数据要转换为 Mat，依需要减去均值和乘系数\n```cpp\n#include \"mat.h\"\nunsigned char* rgbdata;// data pointer to RGB image pixels\nint w;// image width\nint h;// image height\nncnn::Mat in = ncnn::Mat::from_pixels(rgbdata, ncnn::Mat::PIXEL_RGB, w, h);\n\nconst float mean_vals[3] = {104.f, 117.f, 123.f};\nin.substract_mean_normalize(mean_vals, 0);\n```\n执行前向网络，获得计算结果\n```cpp\n#include \"net.h\"\nncnn::Mat in;// input blob as above\nncnn::Mat out;\nncnn::Extractor ex = net.create_extractor();\nex.input(\"data\", in);\nex.extract(\"prob\", out);\n```\n如果是二进制的 param.bin 方式，没有可见字符串，利用 alexnet.id.h 的枚举来代替 blob 的名字\n```cpp\n#include \"net.h\"\n#include \"alexnet.id.h\"\nncnn::Mat in;// input blob as above\nncnn::Mat out;\nncnn::Extractor ex = net.create_extractor();\nex.input(alexnet_param_id::BLOB_data, in);\nex.extract(alexnet_param_id::BLOB_prob, out);\n```\n获取 Mat 中的输出数据，Mat 内部的数据通常是三维的，c / h / w，遍历所有获得全部分类的分数\n```cpp\nncnn::Mat out_flatterned = out.reshape(out.w * out.h * out.c);\nstd::vector scores;\nscores.resize(out_flatterned.w);\nfor (int j=0; j $ cat alexnet.param.bin alexnet.bin > alexnet-all.bin\n\n```cpp\n#include \"net.h\"\nFILE* fp = fopen(\"alexnet-all.bin\", \"rb\");\nnet.load_param_bin(fp);\nnet.load_model(fp);\nfclose(fp);\n```\n" }, { "path": "docs/how-to-use-and-FAQ/use-ncnn-with-opencv.md", "content": "### opencv to ncnn\n\n* cv::Mat CV_8UC3 -> ncnn::Mat 3 channel + swap RGB/BGR\n\n```cpp\n// cv::Mat a(h, w, CV_8UC3);\nncnn::Mat in = ncnn::Mat::from_pixels(a.data, ncnn::Mat::PIXEL_BGR2RGB, a.cols, a.rows);\n```\n\n* cv::Mat CV_8UC3 -> ncnn::Mat 3 channel + keep RGB/BGR order\n\n```cpp\n// cv::Mat a(h, w, CV_8UC3);\nncnn::Mat in = ncnn::Mat::from_pixels(a.data, ncnn::Mat::PIXEL_RGB, a.cols, a.rows);\n```\n\n* cv::Mat CV_8UC3 -> ncnn::Mat 1 channel + do RGB2GRAY/BGR2GRAY\n\n```cpp\n// cv::Mat rgb(h, w, CV_8UC3);\nncnn::Mat inrgb = ncnn::Mat::from_pixels(rgb.data, ncnn::Mat::PIXEL_RGB2GRAY, rgb.cols, rgb.rows);\n\n// cv::Mat bgr(h, w, CV_8UC3);\nncnn::Mat inbgr = ncnn::Mat::from_pixels(bgr.data, ncnn::Mat::PIXEL_BGR2GRAY, bgr.cols, bgr.rows);\n```\n\n* cv::Mat CV_8UC1 -> ncnn::Mat 1 channel\n\n```cpp\n// cv::Mat a(h, w, CV_8UC1);\nncnn::Mat in = ncnn::Mat::from_pixels(a.data, ncnn::Mat::PIXEL_GRAY, a.cols, a.rows);\n```\n\n* cv::Mat CV_32FC1 -> ncnn::Mat 1 channel\n\n * **You could construct ncnn::Mat and fill data into it directly to avoid data copy**\n\n```cpp\n// cv::Mat a(h, w, CV_32FC1);\nncnn::Mat in(a.cols, a.rows, 1, (void*)a.data);\nin = in.clone();\n```\n\n* cv::Mat CV_32FC3 -> ncnn::Mat 3 channel\n\n * **You could construct ncnn::Mat and fill data into it directly to avoid data copy**\n\n```cpp\n// cv::Mat a(h, w, CV_32FC3);\nncnn::Mat in_pack3(a.cols, a.rows, 1, (void*)a.data, (size_t)4u * 3, 3);\nncnn::Mat in;\nncnn::convert_packing(in_pack3, in, 1);\n```\n\n* std::vector < cv::Mat > + CV_32FC1 -> ncnn::Mat multiple channels\n\n * **You could construct ncnn::Mat and fill data into it directly to avoid data copy**\n\n```cpp\n// std::vector a(channels, cv::Mat(h, w, CV_32FC1));\nint channels = a.size();\nncnn::Mat in(a[0].cols, a[0].rows, channels);\nfor (int p=0; p cv::Mat CV_8UC3 + swap RGB/BGR\n\n * **You may need to call in.substract_mean_normalize() first to scale values from 0..1 to 0..255**\n\n```cpp\n// ncnn::Mat in(w, h, 3);\ncv::Mat a(in.h, in.w, CV_8UC3);\nin.to_pixels(a.data, ncnn::Mat::PIXEL_BGR2RGB);\n```\n\n* ncnn::Mat 3 channel -> cv::Mat CV_8UC3 + keep RGB/BGR order\n\n * **You may need to call in.substract_mean_normalize() first to scale values from 0..1 to 0..255**\n\n```cpp\n// ncnn::Mat in(w, h, 3);\ncv::Mat a(in.h, in.w, CV_8UC3);\nin.to_pixels(a.data, ncnn::Mat::PIXEL_RGB);\n```\n\n* ncnn::Mat 1 channel -> cv::Mat CV_8UC1\n\n * **You may need to call in.substract_mean_normalize() first to scale values from 0..1 to 0..255**\n\n```cpp\n// ncnn::Mat in(w, h, 1);\ncv::Mat a(in.h, in.w, CV_8UC1);\nin.to_pixels(a.data, ncnn::Mat::PIXEL_GRAY);\n```\n\n* ncnn::Mat 1 channel -> cv::Mat CV_32FC1\n\n * **You could consume or manipulate ncnn::Mat data directly to avoid data copy**\n\n```cpp\n// ncnn::Mat in;\ncv::Mat a(in.h, in.w, CV_32FC1);\nmemcpy((uchar*)a.data, in.data, in.w * in.h * sizeof(float));\n```\n\n* ncnn::Mat 3 channel -> cv::Mat CV_32FC3\n\n * **You could consume or manipulate ncnn::Mat data directly to avoid data copy**\n\n```cpp\n// ncnn::Mat in(w, h, 3);\nncnn::Mat in_pack3;\nncnn::convert_packing(in, in_pack3, 3);\ncv::Mat a(in.h, in.w, CV_32FC3);\nmemcpy((uchar*)a.data, in_pack3.data, in.w * in.h * 3 * sizeof(float));\n```\n\n* ncnn::Mat multiple channels -> std::vector < cv::Mat > + CV_32FC1\n\n * **You could consume or manipulate ncnn::Mat data directly to avoid data copy**\n\n```cpp\n// ncnn::Mat in(w, h, channels);\nstd::vector a(in.c);\nfor (int p=0; p/lib/cmake/ncnn\" CACHE PATH \"Directory that contains ncnnConfig.cmake\")\nfind_package(ncnn REQUIRED)\ntarget_link_libraries(my_target ncnn)\n```\nAfter this, both the header file search path (\"including directories\") and library paths are configured automatically, including vulkan related dependencies.\n\nNote: you have to change `` to your machine's directory, it is the directory that contains `ncnnConfig.cmake`.\n\nFor the prebuilt ncnn release packages, ncnnConfig is located in:\n- for `ncnn-YYYYMMDD-windows-vs2019`, it is `lib/cmake/ncnn`\n- for `ncnn-YYYYMMDD-android-vulkan`, it is `${ANDROID_ABI}/lib/cmake/ncnn` (`${ANDROID_ABI}` is defined in NDK's cmake toolchain file)\n- other prebuilt release packages are with similar condition\n\n**manually specify**\n\nYou may also manually specify ncnn library path and including directory. Note that if you use ncnn with vulkan, it is also required to specify vulkan related dependencies.\n\nFor example, on Visual Studio debug mode with vulkan required, the lib paths are:\n```\nE:\\github\\ncnn\\build\\vs2019-x64\\install\\lib\\ncnnd.lib\nE:\\github\\ncnn\\build\\vs2019-x64\\install\\lib\\glslangd.lib\n```\nAnd for its release mode, lib paths are:\n```\nE:\\github\\ncnn\\build\\vs2019-x64\\install\\lib\\ncnn.lib\nE:\\github\\ncnn\\build\\vs2019-x64\\install\\lib\\glslang.lib\n```\n" }, { "path": "docs/how-to-use-and-FAQ/use-ncnn-with-pytorch-or-onnx.md", "content": "# A Guide to Converting pytorch / onnx Models to ncnn\n\nThis guide is designed to help pytorch and onnx users use the new-generation model conversion tool, **pnnx**, to efficiently and reliably convert models to the ncnn format for high-performance inference on the edge.\n\nThis document is written and revised based on the **official pnnx documentation**.\n\n* pnnx project: https://github.com/pnnx/pnnx\n* ncnn project: https://github.com/Tencent/ncnn\n* supported pytorch operators: https://github.com/Tencent/ncnn/tree/master/tools/pnnx#supported-pytorch-operator-status\n* supported onnx operators: https://github.com/Tencent/ncnn/tree/master/tools/pnnx#supported-onnx-operator-status\n\n---\n\n## Why is pnnx Highly Recommended?\n\nRegardless of which framework you come from, pnnx offers significant advantages over traditional tools (like `onnx2ncnn`):\n\n* **Forget the Hassles of onnx**: The traditional `pytorch -> onnx -> ncnn` pipeline often fails due to onnx operator compatibility issues and dynamic shape problems. pnnx can convert directly from pytorch, completely bypassing the unstable intermediate step of onnx.\n* **Core Framework Support**: pnnx focuses on supporting **pytorch** and **onnx**, providing you with a unified and consistent conversion experience.\n* **More Stable and Powerful**: pnnx can handle a wider range of modern operators and complex model architectures, generating cleaner and more accurate ncnn graphs.\n* **Active and Continuous Development**: pnnx is under active development, constantly adding support for the latest operators and features from both source frameworks and the ncnn engine.\n* **Richer Graph Information**: pnnx preserves the original model's structural information during the conversion process, which is highly beneficial for model analysis and subsequent optimization.\n\n---\n\n## Workflow 1: Guide for pytorch Users (Recommended)\n\nFor pytorch users, converting directly from a pytorch model is the most stable and efficient path.\n\n### Method A: Direct Conversion in Python with `pnnx.export` (Most Recommended)\n\nThis is the simplest and most recommended workflow, allowing you to complete the model conversion with a single command without leaving your Python environment.\n\n#### 1. Install pnnx\n\nFirst, install the pnnx Python package. This command installs both the `pnnx` Python library and the `pnnx` command-line tool.\n\n```bash\npip3 install pnnx\n```\n\n#### 2. Call `pnnx.export` in Your Python Script\n\nCalling the `pnnx.export` function will generate both a TorchScript (`.pt`) file and the `.param` and `.bin` files required by ncnn.\n\n**Complete Code Example:**\n\n```python\nimport torch\nimport torch.nn as nn\nimport pnnx\n\n# 1. Define or load your pytorch model\nclass MyModel(nn.Module):\n def __init__(self):\n super(MyModel, self).__init__()\n self.conv1 = nn.Conv2d(3, 16, 3, 1, 1)\n self.relu = nn.ReLU()\n self.fc = nn.Linear(16 * 224 * 224, 10)\n\n def forward(self, x):\n x = self.conv1(x)\n x = self.relu(x)\n x = x.view(x.size(0), -1)\n x = self.fc(x)\n return x\n\n# 2. Instantiate the model and set it to evaluation mode\nmodel = MyModel()\nmodel.eval()\n\n# 3. Create a dummy input tensor with the correct input shape\ninput_tensor = torch.rand(1, 3, 224, 224)\n\n# 4. Call pnnx.export to export the model\npnnx.export(model, \"my_model.pt\", (input_tensor,))\n\nprint(\"Conversion complete!\")\nprint(\"Please check for the generated my_model.pt, my_model.ncnn.param, and my_model.ncnn.bin files.\")\n```\n\n### Method B: Using the Command-Line Tool (Alternative)\n\n#### 1. Get the pnnx Command-Line Tool\n\nIf you have already run `pip install pnnx`, the `pnnx` command is available, and you can proceed to the next step.\n\nFor non-Python environments or users who prefer a standalone executable, you can manually download the latest binary from the [pnnx Releases page](https://github.com/pnnx/pnnx/releases).\n\n#### 2. Export to TorchScript (Skip if you already have a .pt file)\n\n```python\nimport torch\n# ... (model definition from above)\nmodel = MyModel()\nmodel.eval()\ninput_tensor = torch.rand(1, 3, 224, 224)\ntraced_script_module = torch.jit.trace(model, input_tensor)\ntraced_script_module.save(\"my_model.pt\")\n```\n\n#### 3. Run the pnnx Command for Conversion\n\nRun the following command in your terminal.\n\n```bash\n# Syntax: pnnx \npnnx my_model.pt\n```\n\n---\n\n## Workflow 2: Guide for onnx Users\n\nFor users who already have an `.onnx` file, please use pnnx for conversion.\n\n### 1. Get the pnnx Command-Line Tool\n\n* **Method 1 (Recommended):** If you have Python in your environment, install it directly via pip.\n ```bash\n pip3 install pnnx\n ```\n The `pnnx` command will be automatically added to your system's path.\n\n* **Method 2 (Alternative):** For non-Python environments or to use a standalone program, you can download the latest executable from the [pnnx Releases page](https://github.com/pnnx/pnnx/releases).\n\n### 2. Run the Command-Line Conversion\n\nOpen a terminal, navigate to the directory containing your model file, and run the following command.\n\n**Basic Command Example:**\n\n```bash\n# Syntax: pnnx \npnnx my_model.onnx\n```\nAfter the command executes successfully, you will get the `my_model.ncnn.param` and `my_model.ncnn.bin` files, which can be directly loaded and used in your ncnn project.\n" }, { "path": "docs/how-to-use-and-FAQ/use-ncnnoptimize-to-optimize-model.md", "content": "\nthe typical usage\n```\nncnnoptimize mobilenet.param mobilenet.bin mobilenet-opt.param mobilenet-opt.bin 65536 \n```\n\noperator fusion\n* batchnorm - scale\n* convolution - batchnorm\n* convolutiondepthwise - batchnorm\n* deconvolution - batchnorm\n* deconvolutiondepthwise - batchnorm\n* innerproduct - batchnorm\n* convolution - relu\n* convolutiondepthwise - relu\n* deconvolution - relu\n* deconvolutiondepthwise - relu\n* innerproduct - relu\n\neliminate noop operator\n* innerproduct - dropout\n* flatten after global pooling\n\nprefer better operator\n* replace convolution with innerproduct after global pooling\n" }, { "path": "docs/how-to-use-and-FAQ/vulkan-notes.md", "content": "## supported platform\n\n* Y = known work\n* ? = shall work, not confirmed\n* / = not applied\n\n| |windows|linux|android|mac|ios|\n|---|---|---|---|---|---|\n|intel|Y|Y|Y|Y|/|\n|amd|Y|Y|/|Y|/|\n|nvidia|Y|Y|?|/|/|\n|qcom|/|/|Y|/|/|\n|apple|/|/|/|Y|Y|\n|arm|/|Y|Y|/|/|\n\n## enable vulkan compute support\n```\n$ cmake -DNCNN_VULKAN=ON ..\n```\n\n## enable vulkan compute inference\n```cpp\nncnn::Net net;\nnet.opt.use_vulkan_compute = 1;\n```\n\n## proper allocator usage\n```cpp\nncnn::VkAllocator* blob_vkallocator = vkdev.acquire_blob_allocator();\nncnn::VkAllocator* staging_vkallocator = vkdev.acquire_blob_allocator();\n\nnet.opt.blob_vkallocator = blob_vkallocator;\nnet.opt.workspace_vkallocator = blob_vkallocator;\nnet.opt.staging_vkallocator = staging_vkallocator;\n\n// ....\n\n// after inference\nvkdev.reclaim_blob_allocator(blob_vkallocator);\nvkdev.reclaim_staging_allocator(staging_vkallocator);\n```\n\n## select gpu device\n```cpp\n// get gpu count\nint gpu_count = ncnn::get_gpu_count();\n\n// set specified vulkan device before loading param and model\nnet.set_vulkan_device(0); // use device-0\nnet.set_vulkan_device(1); // use device-1\n\n// or set opt.vulkan_device_index field before loading param and model\nnet.opt.vulkan_device_index = 0; // use device-0\nnet.opt.vulkan_device_index = 1; // use device-1\n```\n\n## zero-copy on unified memory device\n```cpp\nncnn::VkMat blob_gpu;\nncnn::Mat mapped = blob_gpu.mapped();\n\n// use mapped.data directly\n```\n\n## hybrid cpu/gpu inference\n```cpp\nncnn::Net net_cpu;\nncnn::Net net_gpu;\nnet_cpu.opt.use_vulkan_compute = false;\nnet_gpu.opt.use_vulkan_compute = true;\nnet_cpu.load_param();\nnet_cpu.load_model();\nnet_gpu.load_param();\nnet_gpu.load_model();\n\nncnn::Extractor ex_cpu = net_cpu.create_extractor();\nncnn::Extractor ex_gpu = net_gpu.create_extractor();\n\n#pragma omp parallel sections\n{\n #pragma omp section\n {\n ex_cpu.input();\n ex_cpu.extract();\n }\n #pragma omp section\n {\n ex_gpu.input();\n ex_gpu.extract();\n }\n}\n```\n\n## zero-copy gpu inference chaining\n```cpp\nncnn::Extractor ex1 = net1.create_extractor();\nncnn::Extractor ex2 = net2.create_extractor();\n\nncnn::VkCompute cmd(&vkdev);\n\nncnn::VkMat conv1;\nncnn::VkMat conv2;\nncnn::VkMat conv3;\n\nex1.input(\"conv1\", conv1);\nex1.extract(\"conv2\", conv2, cmd);\n\nex2.input(\"conv2\", conv2);\nex2.extract(\"conv3\", conv3, cmd);\n\ncmd.submit_and_wait();\n```\n\n## batch inference\n```cpp\nint max_batch_size = vkdev->info.compute_queue_count();\n\nncnn::Mat inputs[1000];\nncnn::Mat outputs[1000];\n\n#pragma omp parallel for num_threads(max_batch_size)\nfor (int i=0; i<1000; i++)\n{\n ncnn::Extractor ex = net1.create_extractor();\n ex.input(\"data\", inputs[i]);\n ex.extract(\"prob\", outputs[i]);\n}\n```\n\n## control storage and arithmetic precision\n\ndisable all lower-precision optimizations, get full fp32 precision\n\n```cpp\nncnn::Net net;\nnet.opt.use_fp16_packed = false;\nnet.opt.use_fp16_storage = false;\nnet.opt.use_fp16_arithmetic = false;\nnet.opt.use_int8_storage = false;\nnet.opt.use_int8_arithmetic = false;\n```\n\n## debugging tips\n```cpp\n#define ENABLE_VALIDATION_LAYER 1 // modify to 1 in gpu.cpp\n```\n\n## add vulkan compute support to layer\n1. add vulkan shader in src/layer/shader/\n\n2. upload model weight data in Layer::upload_model()\n\n3. setup pipeline in Layer::create_pipeline()\n\n4. destroy pipeline in Layer::destroy_pipeline()\n\n5. record command in Layer::forward()\n\n## add optimized shader path\n1. add vulkan shader in src/layer/shader/ named XXX_abc.comp\n\n2. create pipeline with \"XXX_abc\"\n\n3. record command using XXX_abc pipeline\n\n## low-level op api\n1. create layer\n\n2. load param and load model\n\n3. upload model\n\n4. create pipeline\n\n5. new command\n\n6. record\n\n7. submit and wait\n\n" }, { "path": "examples/CMakeLists.txt", "content": "macro(ncnn_add_example name)\n add_executable(${name} ${name}.cpp)\n if(OpenCV_FOUND)\n target_include_directories(${name} PRIVATE ${OpenCV_INCLUDE_DIRS})\n target_link_libraries(${name} PRIVATE ncnn ${OpenCV_LIBS})\n elseif(NCNN_SIMPLEOCV)\n target_compile_definitions(${name} PUBLIC USE_NCNN_SIMPLEOCV)\n target_link_libraries(${name} PRIVATE ncnn)\n endif()\n\n # add test to a virtual project group\n set_property(TARGET ${name} PROPERTY FOLDER \"examples\")\nendmacro()\n\nif(NCNN_PIXEL)\n if(NOT NCNN_SIMPLEOCV)\n find_package(OpenCV QUIET COMPONENTS opencv_world)\n # for opencv 2.4 on ubuntu 16.04, there is no opencv_world but OpenCV_FOUND will be TRUE\n if(\"${OpenCV_LIBS}\" STREQUAL \"\")\n set(OpenCV_FOUND FALSE)\n endif()\n if(NOT OpenCV_FOUND)\n find_package(OpenCV QUIET COMPONENTS core highgui imgproc imgcodecs videoio)\n endif()\n if(NOT OpenCV_FOUND)\n find_package(OpenCV QUIET COMPONENTS core highgui imgproc)\n endif()\n endif()\n\n if(OpenCV_FOUND OR NCNN_SIMPLEOCV)\n if(OpenCV_FOUND)\n message(STATUS \"OpenCV library: ${OpenCV_INSTALL_PATH}\")\n message(STATUS \" version: ${OpenCV_VERSION}\")\n message(STATUS \" libraries: ${OpenCV_LIBS}\")\n message(STATUS \" include path: ${OpenCV_INCLUDE_DIRS}\")\n\n if(${OpenCV_VERSION_MAJOR} GREATER 3)\n set(CMAKE_CXX_STANDARD 11)\n endif()\n endif()\n\n include_directories(${CMAKE_CURRENT_SOURCE_DIR}/../src)\n include_directories(${CMAKE_CURRENT_BINARY_DIR}/../src)\n ncnn_add_example(arcface)\n ncnn_add_example(squeezenet)\n ncnn_add_example(squeezenet_c_api)\n ncnn_add_example(fasterrcnn)\n ncnn_add_example(rfcn)\n ncnn_add_example(yolov2)\n ncnn_add_example(yolov3)\n ncnn_add_example(yolov5)\n ncnn_add_example(yolov5_pnnx)\n ncnn_add_example(yolov7_pnnx)\n ncnn_add_example(yolov7)\n ncnn_add_example(yolov8)\n ncnn_add_example(yolov8_seg)\n ncnn_add_example(yolov8_pose)\n ncnn_add_example(yolov8_cls)\n ncnn_add_example(yolox)\n ncnn_add_example(yolo11)\n ncnn_add_example(yolo11_seg)\n ncnn_add_example(yolo11_pose)\n ncnn_add_example(yolo11_cls)\n ncnn_add_example(yoloworld)\n ncnn_add_example(mobilenetv2ssdlite)\n ncnn_add_example(mobilenetssd)\n ncnn_add_example(squeezenetssd)\n ncnn_add_example(shufflenetv2)\n ncnn_add_example(peleenetssd_seg)\n ncnn_add_example(simplepose)\n ncnn_add_example(retinaface)\n ncnn_add_example(yolact)\n ncnn_add_example(nanodet)\n ncnn_add_example(nanodetplus_pnnx)\n ncnn_add_example(scrfd)\n ncnn_add_example(scrfd_crowdhuman)\n ncnn_add_example(piper)\n ncnn_add_example(whisper)\n if(OpenCV_FOUND)\n ncnn_add_example(yolov4)\n ncnn_add_example(yolov8_obb)\n ncnn_add_example(yolo11_obb)\n ncnn_add_example(rvm)\n ncnn_add_example(p2pnet)\n ncnn_add_example(ppocrv5)\n endif()\n else()\n message(WARNING \"OpenCV not found and NCNN_SIMPLEOCV disabled, examples won't be built\")\n endif()\nelse()\n message(WARNING \"NCNN_PIXEL not enabled, examples won't be built\")\nendif()\n" }, { "path": "examples/arcface.cpp", "content": "// Copyright 2025 heabeounMKTO\n// SPDX-License-Identifier: BSD-3-Clause\n/* ncnn example using yolo-face and arcface to extract embeddings from a face\n *\n *\n * the arcface model is converted from\n * https://github.com/onnx/models/tree/main/validated/vision/body_analysis/arcface\n * 1. first simplify the arcface.onnx using onnxsim\n * 2. then convert it using ncnn's onnx exporter onnx2ncnn\n * using pnnx to convert would cause -nan output!\n *\n * the yolov8-face model is converted from\n * https://github.com/derronqi/yolov8-face\n *\n *\n * you can find the models preconverted at\n * https://drive.google.com/drive/folders/1P0RDzj9V7FHEL8w_-yqls5RHeVpO-2PS?usp=sharing\n *\n * */\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n#include \"layer.h\"\n#include \"net.h\"\n#include \"mat.h\"\n\n#ifndef ARCFACE_EXAMPLE_YOLO_INFER_SIZE\n#define ARCFACE_EXAMPLE_YOLO_INFER_SIZE 320\n#endif\n\nstruct Bbox\n{\n float x1, y1, x2, y2, confidence;\n int label;\n Bbox()\n : x1(0.0f), y1(0.0f), x2(0.0f), y2(0.0f), confidence(0.0f), label(0)\n {\n }\n Bbox(float x1,\n float y1,\n float x2,\n float y2,\n float confidence,\n int label = 0,\n std::string label_name = \"\")\n : x1(x1), y1(y1), x2(x2), y2(y2), confidence(confidence), label(label)\n {\n }\n Bbox apply_image_scale(const cv::Mat& original_image,\n const float scale_factor,\n const int pad_w,\n const int pad_h)\n {\n int img_w = original_image.cols;\n int img_h = original_image.rows;\n\n x1 = (x1 - pad_w) / scale_factor;\n y1 = (y1 - pad_h) / scale_factor;\n x2 = (x2 - pad_w) / scale_factor;\n y2 = (y2 - pad_h) / scale_factor;\n\n // clamp\n x1 = std::max(0.0f, std::min(x1, (float)img_w));\n y1 = std::max(0.0f, std::min(y1, (float)img_h));\n x2 = std::max(0.0f, std::min(x2, (float)img_w));\n y2 = std::max(0.0f, std::min(y2, (float)img_h));\n return Bbox(x1, y1, x2, y2, confidence, label);\n }\n std::string get_label_name(const std::vector& classes)\n {\n return classes[this->label];\n }\n\n /// what more do you need to know vro\n float area() const\n {\n float width = x2 - x1;\n float height = y2 - y1;\n return width * height;\n }\n cv::Mat crop_bbox(const cv::Mat& originalImage) const\n {\n // Calculate width and height\n int bbox_width = static_cast(x2 - x1);\n int bbox_height = static_cast(y2 - y1);\n\n // Ensure valid dimensions\n if (bbox_width <= 0 || bbox_height <= 0)\n {\n fprintf(stderr, \"Invalid bounding box dimensions\\n\");\n return cv::Mat();\n }\n\n // Ensure coordinates are within image bounds\n int x1_int = static_cast(x1);\n int y1_int = static_cast(y1);\n int x2_int = static_cast(x2);\n int y2_int = static_cast(y2);\n\n // Clamp to image bounds\n x1_int = std::max(0, x1_int);\n y1_int = std::max(0, y1_int);\n x2_int = std::min(originalImage.cols, x2_int);\n y2_int = std::min(originalImage.rows, y2_int);\n\n // Create ROI and return cropped image\n cv::Rect roi(x1_int, y1_int, x2_int - x1_int, y2_int - y1_int);\n return originalImage(roi).clone();\n }\n cv::Rect_ get_rect() const\n {\n int x1_int = static_cast(x1);\n int y1_int = static_cast(y1);\n int width = static_cast(x2 - x1);\n int height = static_cast(y2 - y1);\n\n // Ensure valid dimensions\n if (width <= 0 || height <= 0)\n {\n return cv::Rect(0, 0, 0, 0); // Return invalid rect\n }\n\n return cv::Rect(x1_int, y1_int, width, height);\n }\n};\n\nstatic void print_bbox(Bbox& bbox)\n{\n printf(\"Bbox(x1=%.2f, y1=%.2f, x2=%.2f, y2=%.2f, conf=%.4f, label=%d)\\n\",\n bbox.x1, bbox.y1, bbox.x2, bbox.y2, bbox.confidence, bbox.label);\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].confidence;\n\n while (i <= j)\n {\n while (faceobjects[i].confidence > p)\n i++;\n\n while (faceobjects[j].confidence < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty()) return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nfloat calculate_iou(const Bbox& box1, const Bbox& box2)\n{\n float x1 = std::max(box1.x1, box2.x1);\n float y1 = std::max(box1.y1, box2.y1);\n float x2 = std::min(box1.x2, box2.x2);\n float y2 = std::min(box1.y2, box2.y2);\n\n if (x2 <= x1 || y2 <= y1)\n {\n return 0.0f; // no intersect\n }\n\n float intersection_area = (x2 - x1) * (y2 - y1);\n float box1_area = (box1.x2 - box1.x1) * (box1.y2 - box1.y1);\n float box2_area = (box2.x2 - box2.x1) * (box2.y2 - box2.y1);\n float union_area = box1_area + box2_area - intersection_area;\n return intersection_area / union_area;\n}\n\nstatic std::vector\nnon_maximum_supression(const std::vector& bbox, float iou_thresh, bool class_agnostic = false)\n{\n std::vector picked;\n const int n = bbox.size();\n if (n == 0) return picked;\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = bbox[i].area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Bbox& a = bbox[i];\n bool keep = true;\n\n for (int j : picked)\n {\n const Bbox& b = bbox[j];\n\n // Enhanced class comparison logic using labels\n if (!class_agnostic)\n {\n if (a.label != b.label)\n {\n continue; // Different classes, don't suppress\n }\n }\n\n float iou = calculate_iou(a, b);\n if (iou > iou_thresh)\n {\n keep = false;\n break;\n }\n }\n\n if (keep)\n {\n picked.push_back(i);\n }\n }\n\n return picked;\n}\n\nstatic std::vector scale_wh(float w0, float h0, float w1, float h1)\n{\n float r = std::min(w1 / w0, h1 / h0);\n std::vector _scale_factor(3);\n _scale_factor[0] = r;\n _scale_factor[1] = (float)std::round(w0 * r);\n _scale_factor[2] = (float)std::round(h0 * r);\n return _scale_factor;\n}\n\nstruct ImagePreProcessResults\n{\n ncnn::Mat result;\n float img_scale, pad_w, pad_h;\n\n ImagePreProcessResults(ncnn::Mat result, float img_scale, float pad_w, float pad_h)\n : result(result), img_scale(img_scale), pad_w(pad_w), pad_h(pad_h)\n {\n }\n};\n\nstruct DetectionResult\n{\n std::vector bboxes;\n std::vector > keypoints;\n};\n\nstatic ImagePreProcessResults preprocess_yolo_kpts(cv::Mat& input_image, int infer_size) noexcept\n{\n float mean_vals[] = {0.f, 0.f, 0.f};\n\n float norm_vals[] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n int img_w = input_image.cols;\n int img_h = input_image.rows;\n float scale_factor, new_w, new_h;\n std::vector _scale_factor = scale_wh(img_w, img_h, (float)infer_size, (float)infer_size);\n scale_factor = _scale_factor[0];\n new_w = _scale_factor[1];\n new_h = _scale_factor[2];\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(input_image.data,\n ncnn::Mat::PIXEL_BGR2RGB, img_w,\n img_h, new_w, new_h);\n\n // padding calculation\n int pad_w = (infer_size - new_w) / 2;\n int pad_h = (infer_size - new_h) / 2;\n\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, pad_h, infer_size - new_h - pad_h, pad_w,\n infer_size - new_w - pad_w, ncnn::BORDER_CONSTANT, 114.f);\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n return ImagePreProcessResults(in_pad, scale_factor, pad_w, pad_h);\n}\n\n/// parses extra keypoints data for face mmodel\n/// the format is this:\n/// [x, y, w, h, conf, class_scores..., kp1_conf, kp1_x, kp1_y, kp2_conf, kp2_x, kp2_y, ...]\nstatic DetectionResult parse_yolo_keypoints_results(ncnn::Mat& result,\n cv::Mat& original_image,\n ImagePreProcessResults& preproc_img,\n float confidence_threshold,\n float iou_threshold,\n std::vector class_names)\n{\n cv::Mat output((int)result.w, (int)result.h, CV_32FC1);\n for (int i = 0; i < output.cols; i++)\n {\n for (int j = 0; j < output.rows; j++)\n {\n output.ptr(j)[i] = result.row(i)[j];\n }\n }\n std::vector detections;\n std::vector > all_keypoints;\n\n int num_classes = class_names.size();\n int kp_stride = 3;\n int num_keypoints = 5;\n\n for (int i = 0; i < output.rows; i++)\n {\n const float* row_ptr = output.ptr(i);\n const float* bboxes_ptr = row_ptr;\n const float* classes_ptr = row_ptr + 4;\n const float* max_s_ptr = std::max_element(classes_ptr, classes_ptr + num_classes);\n\n float score = *max_s_ptr;\n int class_id = max_s_ptr - classes_ptr;\n\n if (score >= confidence_threshold)\n {\n float x = bboxes_ptr[0];\n float y = bboxes_ptr[1];\n float w = bboxes_ptr[2];\n float h = bboxes_ptr[3];\n float x1 = x - w / 2.0f;\n float y1 = y - h / 2.0f;\n float x2 = x + w / 2.0f;\n float y2 = y + h / 2.0f;\n\n if (x2 > x1 && y2 > y1)\n {\n Bbox bbox = Bbox(x1, y1, x2, y2, score, class_id)\n .apply_image_scale(original_image, preproc_img.img_scale,\n preproc_img.pad_w, preproc_img.pad_h);\n // Parse exactly 5 keypoints for this face model\n std::vector face_keypoints;\n face_keypoints.reserve(15);\n const float* kp_ptr = row_ptr + 4 + num_classes;\n float scale = 1.0f / preproc_img.img_scale;\n\n for (int k = 0; k < num_keypoints; k++)\n {\n float kp_x = kp_ptr[k * kp_stride];\n float kp_y = kp_ptr[k * kp_stride + 1];\n float kp_conf_raw = kp_ptr[k * kp_stride + 2];\n\n // Apply sigmoid to convert logit to probability\n float kp_conf = 1.0f / (1.0f + expf(-kp_conf_raw));\n\n // Scale keypoints to original\n kp_x = (kp_x - preproc_img.pad_w) * scale;\n kp_y = (kp_y - preproc_img.pad_h) * scale;\n\n face_keypoints.push_back(kp_x);\n face_keypoints.push_back(kp_y);\n face_keypoints.push_back(kp_conf);\n }\n\n detections.push_back(bbox);\n all_keypoints.push_back(face_keypoints);\n }\n }\n }\n\n // nms\n qsort_descent_inplace(detections);\n std::vector picked = non_maximum_supression(detections, iou_threshold, false);\n DetectionResult res;\n for (size_t i = 0; i < picked.size(); i++)\n {\n int idx = picked[i];\n res.bboxes.push_back(detections[idx]);\n res.keypoints.push_back(all_keypoints[idx]);\n }\n\n return res;\n}\n\nstatic inline float get_similarity(std::vector f1, std::vector f2)\n{\n float sim = 0.0;\n for (size_t i = 0; i < f1.size(); i++)\n {\n sim += f1[i] * f2[i];\n }\n return sim;\n}\n\n// these are converted from here\n// https://github.com/deepinsight/insightface/blob/master/python-package/insightface/utils/face_align.py\nstatic int estimate_norm(float* transform_matrix, const float* lmk, int image_size = 112)\n{\n float ARCFACE_DST[] {\n 38.2946f, 51.6963f, // left eye\n 73.5318f, 51.5014f, // right eye\n 56.0252f, 71.7366f, // nose\n 41.5493f, 92.3655f, // left mouth\n 70.7299f, 92.2041f // right mouth\n };\n if (image_size % 112 != 0 && image_size % 128 != 0)\n {\n return -1;\n }\n\n float ratio, diff_x;\n if (image_size % 112 == 0)\n {\n ratio = static_cast(image_size) / 112.0f;\n diff_x = 0.0f;\n }\n else\n {\n ratio = static_cast(image_size) / 128.0f;\n diff_x = 8.0f * ratio;\n }\n\n float src_points[10];\n for (int i = 0; i < 5; i++)\n {\n src_points[i * 2] = lmk[i * 3];\n src_points[i * 2 + 1] = lmk[i * 3 + 1];\n }\n\n float dst_points[10];\n for (int i = 0; i < 5; i++)\n {\n dst_points[i * 2] = ARCFACE_DST[i * 2] * ratio + diff_x;\n dst_points[i * 2 + 1] = ARCFACE_DST[i * 2 + 1] * ratio;\n }\n\n ncnn::get_affine_transform(dst_points, src_points, 5, transform_matrix);\n\n return 0;\n}\n\nstatic int norm_crop(cv::Mat& output, const cv::Mat& input, const float* lmk, int image_size = 112)\n{\n float transform_matrix[6];\n int status = estimate_norm(transform_matrix, lmk, image_size);\n\n if (status != 0)\n {\n return status;\n }\n output = cv::Mat(image_size, image_size, CV_8UC3);\n ncnn::warpaffine_bilinear_c3(input.data, input.cols, input.rows,\n output.data, image_size, image_size,\n transform_matrix);\n return 0;\n}\n\nvoid normalize_arcface(std::vector& feature)\n{\n if (feature.empty())\n return;\n float sum = 0;\n for (auto it = feature.begin(); it != feature.end(); it++)\n sum += (float)*it * (float)*it;\n sum = sqrt(sum);\n if (sum == 0.0f)\n return;\n for (auto it = feature.begin(); it != feature.end(); it++)\n *it /= sum;\n}\n\nstatic int get_face(const cv::Mat& rgb, DetectionResult& result)\n{\n int status = 0;\n ncnn::Net yoloface;\n yoloface.opt.use_vulkan_compute = true;\n status = yoloface.load_param(\"yolov8-face.param\");\n\n if (status != 0)\n {\n fprintf(stderr, \"couldn't load params\");\n return status;\n }\n\n status = yoloface.load_model(\"yolov8-face.bin\");\n\n if (status != 0)\n {\n fprintf(stderr, \"couldn't load model\");\n return status;\n }\n\n cv::Mat input_image = rgb.clone();\n ImagePreProcessResults preproc_img = preprocess_yolo_kpts(input_image, ARCFACE_EXAMPLE_YOLO_INFER_SIZE);\n ncnn::Extractor ex = yoloface.create_extractor();\n ex.input(\"in0\", preproc_img.result);\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n std::vector class_names = {\"face\"};\n result = parse_yolo_keypoints_results(out, input_image, preproc_img, 0.5, 0.4, class_names);\n if (result.bboxes.size() < 1)\n {\n fprintf(stderr, \"no faces are found!\");\n return -1;\n }\n return 0;\n}\n\nstatic int get_embedding(const cv::Mat& rgb, std::vector& result)\n{\n ncnn::Net arcface;\n arcface.opt.use_vulkan_compute = true;\n int status = arcface.load_param(\"arcfaceresnet.param\");\n if (status != 0)\n {\n fprintf(stderr, \"couldn't load arcface params\");\n return status;\n }\n status = arcface.load_model(\"arcfaceresnet.bin\");\n if (status != 0)\n {\n fprintf(stderr, \"couldn't load arcface model\");\n return status;\n }\n\n if (rgb.empty() || rgb.type() != CV_8UC3)\n {\n fprintf(stderr, \"invalid input image!\");\n return -1;\n }\n /*\n * the arcface model provided in the link has builtin normalization layers,\n * no need to run substract_mean_normalize\n *\n * reference from .param\n BinaryOp _minusscalar0 2 1 data scalar_op2 _minusscalar0 0=1\n BinaryOp _mulscalar0 2 1 _minusscalar0 scalar_op3 _mulscalar0 0=2\n * */\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(\n rgb.data,\n ncnn::Mat::PIXEL_BGR2RGB,\n rgb.cols,\n rgb.rows,\n 112,\n 112);\n ncnn::Extractor ex = arcface.create_extractor();\n ex.input(\"data\", in);\n ncnn::Mat out;\n ex.extract(\"fc1\", out);\n const float* ptr = (const float*)out.data;\n for (int i = 0; i < 512; i++)\n {\n result[i] = ptr[i];\n }\n normalize_arcface(result);\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 3)\n {\n fprintf(stderr, \"Usage: %s \\n\", argv[0]);\n return -1;\n }\n\n const char* face1_path = argv[1];\n const char* face2_path = argv[2];\n\n int status = 0;\n cv::Mat face_img1 = cv::imread(face1_path);\n cv::Mat face_img2 = cv::imread(face2_path);\n\n if (face_img1.empty())\n {\n fprintf(stderr, \"Failed to load image: %s\\n\", face1_path);\n return -1;\n }\n if (face_img2.empty())\n {\n fprintf(stderr, \"Failed to load image: %s\\n\", face2_path);\n return -1;\n }\n\n cv::Mat input_embed1, input_embed2;\n DetectionResult res1, res2;\n std::vector embedding1(512), embedding2(512);\n\n status = get_face(face_img1, res1);\n if (status != 0)\n {\n fprintf(stderr, \"get face failed for %s!\\n\", face1_path);\n return -1;\n }\n fprintf(stdout, \"found faces in face1: %d\\n\", (int)res1.bboxes.size());\n for (size_t i = 0; i < res1.bboxes.size(); i++)\n {\n print_bbox(res1.bboxes[i]);\n }\n\n status = get_face(face_img2, res2);\n if (status != 0)\n {\n fprintf(stderr, \"get face failed for %s!\\n\", face2_path);\n return -1;\n }\n fprintf(stdout, \"found faces in face2: %d\\n\", (int)res2.bboxes.size());\n for (size_t i = 0; i < res2.bboxes.size(); i++)\n {\n print_bbox(res2.bboxes[i]);\n }\n\n status = norm_crop(input_embed1, face_img1, res1.keypoints[0].data());\n status = get_embedding(input_embed1, embedding1);\n if (status != 0)\n {\n fprintf(stderr, \"get embedding failed for %s!\\n\", face1_path);\n return -1;\n }\n\n status = norm_crop(input_embed2, face_img2, res2.keypoints[0].data());\n if (status != 0)\n {\n fprintf(stderr, \"norm_crop failed for face2!\\n\");\n return -1;\n }\n status = get_embedding(input_embed2, embedding2);\n if (status != 0)\n {\n fprintf(stderr, \"get embedding failed for face2!\\n\");\n return -1;\n }\n if (status != 0)\n {\n fprintf(stderr, \"get embedding failed for %s!\\n\", face2_path);\n return -1;\n }\n\n float similarity = get_similarity(embedding1, embedding2);\n fprintf(stdout, \"Similarity: %f\\n\", similarity);\n}\n" }, { "path": "examples/fasterrcnn.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic int detect_fasterrcnn(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net fasterrcnn;\n\n fasterrcnn.opt.use_vulkan_compute = true;\n\n // original pretrained model from https://github.com/rbgirshick/py-faster-rcnn\n // py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/faster_rcnn_test.pt\n // https://dl.dropboxusercontent.com/s/o6ii098bu51d139/faster_rcnn_models.tgz?dl=0\n // ZF_faster_rcnn_final.caffemodel\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (fasterrcnn.load_param(\"ZF_faster_rcnn_final.param\"))\n exit(-1);\n if (fasterrcnn.load_model(\"ZF_faster_rcnn_final.bin\"))\n exit(-1);\n\n // hyper parameters taken from\n // py-faster-rcnn/lib/fast_rcnn/config.py\n // py-faster-rcnn/lib/fast_rcnn/test.py\n const int target_size = 600; // __C.TEST.SCALES\n\n const int max_per_image = 100;\n const float confidence_thresh = 0.05f;\n\n const float nms_threshold = 0.3f; // __C.TEST.NMS\n\n // scale to target detect size\n int w = bgr.cols;\n int h = bgr.rows;\n float scale = 1.f;\n if (w < h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, w, h);\n\n const float mean_vals[3] = {102.9801f, 115.9465f, 122.7717f};\n in.substract_mean_normalize(mean_vals, 0);\n\n ncnn::Mat im_info(3);\n im_info[0] = h;\n im_info[1] = w;\n im_info[2] = scale;\n\n // step1, extract feature and all rois\n ncnn::Extractor ex1 = fasterrcnn.create_extractor();\n\n ex1.input(\"data\", in);\n ex1.input(\"im_info\", im_info);\n\n ncnn::Mat conv5_relu5; // feature\n ncnn::Mat rois; // all rois\n ex1.extract(\"conv5_relu5\", conv5_relu5);\n ex1.extract(\"rois\", rois);\n\n // step2, extract bbox and score for each roi\n std::vector > class_candidates;\n for (int i = 0; i < rois.c; i++)\n {\n ncnn::Extractor ex2 = fasterrcnn.create_extractor();\n\n ncnn::Mat roi = rois.channel(i); // get single roi\n ex2.input(\"conv5_relu5\", conv5_relu5);\n ex2.input(\"rois\", roi);\n\n ncnn::Mat bbox_pred;\n ncnn::Mat cls_prob;\n ex2.extract(\"bbox_pred\", bbox_pred);\n ex2.extract(\"cls_prob\", cls_prob);\n\n int num_class = cls_prob.w;\n class_candidates.resize(num_class);\n\n // find class id with highest score\n int label = 0;\n float score = 0.f;\n for (int i = 0; i < num_class; i++)\n {\n float class_score = cls_prob[i];\n if (class_score > score)\n {\n label = i;\n score = class_score;\n }\n }\n\n // ignore background or low score\n if (label == 0 || score <= confidence_thresh)\n continue;\n\n // fprintf(stderr, \"%d = %f\\n\", label, score);\n\n // unscale to image size\n float x1 = roi[0] / scale;\n float y1 = roi[1] / scale;\n float x2 = roi[2] / scale;\n float y2 = roi[3] / scale;\n\n float pb_w = x2 - x1 + 1;\n float pb_h = y2 - y1 + 1;\n\n // apply bbox regression\n float dx = bbox_pred[label * 4];\n float dy = bbox_pred[label * 4 + 1];\n float dw = bbox_pred[label * 4 + 2];\n float dh = bbox_pred[label * 4 + 3];\n\n float cx = x1 + pb_w * 0.5f;\n float cy = y1 + pb_h * 0.5f;\n\n float obj_cx = cx + pb_w * dx;\n float obj_cy = cy + pb_h * dy;\n\n float obj_w = pb_w * exp(dw);\n float obj_h = pb_h * exp(dh);\n\n float obj_x1 = obj_cx - obj_w * 0.5f;\n float obj_y1 = obj_cy - obj_h * 0.5f;\n float obj_x2 = obj_cx + obj_w * 0.5f;\n float obj_y2 = obj_cy + obj_h * 0.5f;\n\n // clip\n obj_x1 = std::max(std::min(obj_x1, (float)(bgr.cols - 1)), 0.f);\n obj_y1 = std::max(std::min(obj_y1, (float)(bgr.rows - 1)), 0.f);\n obj_x2 = std::max(std::min(obj_x2, (float)(bgr.cols - 1)), 0.f);\n obj_y2 = std::max(std::min(obj_y2, (float)(bgr.rows - 1)), 0.f);\n\n // append object\n Object obj;\n obj.rect = cv::Rect_(obj_x1, obj_y1, obj_x2 - obj_x1 + 1, obj_y2 - obj_y1 + 1);\n obj.label = label;\n obj.prob = score;\n\n class_candidates[label].push_back(obj);\n }\n\n // post process\n objects.clear();\n for (int i = 0; i < (int)class_candidates.size(); i++)\n {\n std::vector& candidates = class_candidates[i];\n\n qsort_descent_inplace(candidates);\n\n std::vector picked;\n nms_sorted_bboxes(candidates, picked, nms_threshold);\n\n for (int j = 0; j < (int)picked.size(); j++)\n {\n int z = picked[j];\n objects.push_back(candidates[z]);\n }\n }\n\n qsort_descent_inplace(objects);\n\n if (max_per_image > 0 && max_per_image < objects.size())\n {\n objects.resize(max_per_image);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_fasterrcnn(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/mobilenetssd.cpp", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_mobilenet(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net mobilenet;\n\n mobilenet.opt.use_vulkan_compute = true;\n\n // model is converted from https://github.com/chuanqi305/MobileNet-SSD\n // and can be downloaded from https://drive.google.com/open?id=0ByaKLD9QaPtucWk0Y0dha1VVY0U\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (mobilenet.load_param(\"mobilenet_ssd_voc_ncnn.param\"))\n exit(-1);\n if (mobilenet.load_model(\"mobilenet_ssd_voc_ncnn.bin\"))\n exit(-1);\n\n const int target_size = 300;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {127.5f, 127.5f, 127.5f};\n const float norm_vals[3] = {1.0 / 127.5, 1.0 / 127.5, 1.0 / 127.5};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = mobilenet.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_mobilenet(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/mobilenetv2ssdlite.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nclass Noop : public ncnn::Layer\n{\n};\nDEFINE_LAYER_CREATOR(Noop)\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_mobilenetv2(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net mobilenetv2;\n\n mobilenetv2.opt.use_vulkan_compute = true;\n\n mobilenetv2.register_custom_layer(\"Silence\", Noop_layer_creator);\n\n // original pretrained model from https://github.com/chuanqi305/MobileNetv2-SSDLite\n // https://github.com/chuanqi305/MobileNetv2-SSDLite/blob/master/ssdlite/voc/deploy.prototxt\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (mobilenetv2.load_param(\"mobilenetv2_ssdlite_voc.param\"))\n exit(-1);\n if (mobilenetv2.load_model(\"mobilenetv2_ssdlite_voc.bin\"))\n exit(-1);\n\n const int target_size = 300;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {127.5f, 127.5f, 127.5f};\n const float norm_vals[3] = {1.0 / 127.5, 1.0 / 127.5, 1.0 / 127.5};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = mobilenetv2.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_mobilenetv2(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/mobilenetv3ssdlite.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n#include \"platform.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#if NCNN_VULKAN\n#include \"gpu.h\"\n#endif // NCNN_VULKAN\n\ntemplate\nconst T& clamp(const T& v, const T& lo, const T& hi)\n{\n assert(!(hi < lo));\n return v < lo ? lo : hi < v ? hi : v;\n}\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_mobilenetv3(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net mobilenetv3;\n\n#if NCNN_VULKAN\n mobilenetv3.opt.use_vulkan_compute = true;\n#endif // NCNN_VULKAN\n\n // converted ncnn model from https://github.com/ujsyehao/mobilenetv3-ssd\n if (mobilenetv3.load_param(\"./mobilenetv3_ssdlite_voc.param\"))\n exit(-1);\n if (mobilenetv3.load_model(\"./mobilenetv3_ssdlite_voc.bin\"))\n exit(-1);\n\n const int target_size = 300;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {123.675f, 116.28f, 103.53f};\n const float norm_vals[3] = {1.0f, 1.0f, 1.0f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = mobilenetv3.create_extractor();\n\n ex.input(\"input\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n\n // filter out cross-boundary\n float x1 = clamp(values[2] * target_size, 0.f, float(target_size - 1)) / target_size * img_w;\n float y1 = clamp(values[3] * target_size, 0.f, float(target_size - 1)) / target_size * img_h;\n float x2 = clamp(values[4] * target_size, 0.f, float(target_size - 1)) / target_size * img_w;\n float y2 = clamp(values[5] * target_size, 0.f, float(target_size - 1)) / target_size * img_h;\n\n object.rect.x = x1;\n object.rect.y = y1;\n object.rect.width = x2 - x1;\n object.rect.height = y2 - y1;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n if (objects[i].prob > 0.6)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_mobilenetv3(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/nanodet.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& cls_pred, const ncnn::Mat& dis_pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int num_grid = cls_pred.h;\n\n int num_grid_x;\n int num_grid_y;\n if (in_pad.w > in_pad.h)\n {\n num_grid_x = in_pad.w / stride;\n num_grid_y = num_grid / num_grid_x;\n }\n else\n {\n num_grid_y = in_pad.h / stride;\n num_grid_x = num_grid / num_grid_y;\n }\n\n const int num_class = cls_pred.w;\n const int reg_max_1 = dis_pred.w / 4;\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n const int idx = i * num_grid_x + j;\n\n const float* scores = cls_pred.row(idx);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n if (scores[k] > score)\n {\n label = k;\n score = scores[k];\n }\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat bbox_pred(reg_max_1, 4, (void*)dis_pred.row(idx));\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(bbox_pred, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = bbox_pred.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (j + 0.5f) * stride;\n float pb_cy = (i + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic int detect_nanodet(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net nanodet;\n\n nanodet.opt.use_vulkan_compute = true;\n // nanodet.opt.use_bf16_storage = true;\n\n // original pretrained model from https://github.com/RangiLyu/nanodet\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (nanodet.load_param(\"nanodet_m.param\"))\n exit(-1);\n if (nanodet.load_model(\"nanodet_m.bin\"))\n exit(-1);\n\n int width = bgr.cols;\n int height = bgr.rows;\n\n const int target_size = 320;\n const float prob_threshold = 0.4f;\n const float nms_threshold = 0.5f;\n\n // pad to multiple of 32\n int w = width;\n int h = height;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, width, height, w, h);\n\n // pad to target_size rectangle\n int wpad = (w + 31) / 32 * 32 - w;\n int hpad = (h + 31) / 32 * 32 - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 0.f);\n\n const float mean_vals[3] = {103.53f, 116.28f, 123.675f};\n const float norm_vals[3] = {0.017429f, 0.017507f, 0.017125f};\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = nanodet.create_extractor();\n\n ex.input(\"input.1\", in_pad);\n\n std::vector proposals;\n\n // stride 8\n {\n ncnn::Mat cls_pred;\n ncnn::Mat dis_pred;\n ex.extract(\"792\", cls_pred);\n ex.extract(\"795\", dis_pred);\n\n std::vector objects8;\n generate_proposals(cls_pred, dis_pred, 8, in_pad, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat cls_pred;\n ncnn::Mat dis_pred;\n ex.extract(\"814\", cls_pred);\n ex.extract(\"817\", dis_pred);\n\n std::vector objects16;\n generate_proposals(cls_pred, dis_pred, 16, in_pad, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat cls_pred;\n ncnn::Mat dis_pred;\n ex.extract(\"836\", cls_pred);\n ex.extract(\"839\", dis_pred);\n\n std::vector objects32;\n generate_proposals(cls_pred, dis_pred, 32, in_pad, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(width - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(height - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(width - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(height - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_nanodet(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/nanodetplus_pnnx.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + exp(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int num_grid = pred.h;\n\n int num_grid_x = pred.w;\n int num_grid_y = pred.h;\n\n const int num_class = 80; // number of classes. 80 for COCO\n const int reg_max_1 = (pred.c - num_class) / 4;\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n float s = pred.channel(k).row(i)[j];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n\n if (score >= prob_threshold)\n {\n ncnn::Mat bbox_pred(reg_max_1, 4);\n for (int k = 0; k < reg_max_1 * 4; k++)\n {\n bbox_pred[k] = pred.channel(num_class + k).row(i)[j];\n }\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(bbox_pred, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = bbox_pred.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = j * stride;\n float pb_cy = i * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic int detect_nanodet(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net nanodet;\n\n nanodet.opt.use_vulkan_compute = true;\n // nanodet.opt.use_bf16_storage = true;\n\n // original pretrained model from https://github.com/RangiLyu/nanodet\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n // nanodet.load_param(\"nanodet-plus-m_320.torchscript.ncnn.param\");\n // nanodet.load_model(\"nanodet-plus-m_320.torchscript.ncnn.bin\");\n if (nanodet.load_param(\"nanodet-plus-m_416.torchscript.ncnn.param\"))\n exit(-1);\n if (nanodet.load_model(\"nanodet-plus-m_416.torchscript.ncnn.bin\"))\n exit(-1);\n\n int width = bgr.cols;\n int height = bgr.rows;\n\n // const int target_size = 320;\n const int target_size = 416;\n const float prob_threshold = 0.4f;\n const float nms_threshold = 0.5f;\n\n // pad to multiple of 32\n int w = width;\n int h = height;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, width, height, w, h);\n\n // pad to target_size rectangle\n int wpad = (w + 31) / 32 * 32 - w;\n int hpad = (h + 31) / 32 * 32 - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 0.f);\n\n const float mean_vals[3] = {103.53f, 116.28f, 123.675f};\n const float norm_vals[3] = {0.017429f, 0.017507f, 0.017125f};\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = nanodet.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n std::vector proposals;\n\n // stride 8\n {\n ncnn::Mat pred;\n ex.extract(\"231\", pred);\n\n std::vector objects8;\n generate_proposals(pred, 8, in_pad, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat pred;\n ex.extract(\"228\", pred);\n\n std::vector objects16;\n generate_proposals(pred, 16, in_pad, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat pred;\n ex.extract(\"225\", pred);\n\n std::vector objects32;\n generate_proposals(pred, 32, in_pad, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // stride 64\n {\n ncnn::Mat pred;\n ex.extract(\"222\", pred);\n\n std::vector objects64;\n generate_proposals(pred, 64, in_pad, prob_threshold, objects64);\n\n proposals.insert(proposals.end(), objects64.begin(), objects64.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(width - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(height - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(width - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(height - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_nanodet(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/p2pnet.cpp", "content": "// Copyright 2021 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n#include \n\nstruct CrowdPoint\n{\n cv::Point pt;\n float prob;\n};\n\nstatic void shift(int w, int h, int stride, std::vector anchor_points, std::vector& shifted_anchor_points)\n{\n std::vector x_, y_;\n for (int i = 0; i < w; i++)\n {\n float x = (i + 0.5) * stride;\n x_.push_back(x);\n }\n for (int i = 0; i < h; i++)\n {\n float y = (i + 0.5) * stride;\n y_.push_back(y);\n }\n\n std::vector shift_x((size_t)w * h, 0), shift_y((size_t)w * h, 0);\n for (int i = 0; i < h; i++)\n {\n for (int j = 0; j < w; j++)\n {\n shift_x[i * w + j] = x_[j];\n }\n }\n for (int i = 0; i < h; i++)\n {\n for (int j = 0; j < w; j++)\n {\n shift_y[i * w + j] = y_[i];\n }\n }\n\n std::vector shifts((size_t)w * h * 2, 0);\n for (int i = 0; i < w * h; i++)\n {\n shifts[i * 2] = shift_x[i];\n shifts[i * 2 + 1] = shift_y[i];\n }\n\n shifted_anchor_points.resize((size_t)2 * w * h * anchor_points.size() / 2, 0);\n for (int i = 0; i < w * h; i++)\n {\n for (int j = 0; j < anchor_points.size() / 2; j++)\n {\n float x = anchor_points[j * 2] + shifts[i * 2];\n float y = anchor_points[j * 2 + 1] + shifts[i * 2 + 1];\n shifted_anchor_points[i * anchor_points.size() / 2 * 2 + j * 2] = x;\n shifted_anchor_points[i * anchor_points.size() / 2 * 2 + j * 2 + 1] = y;\n }\n }\n}\nstatic void generate_anchor_points(int stride, int row, int line, std::vector& anchor_points)\n{\n float row_step = (float)stride / row;\n float line_step = (float)stride / line;\n\n std::vector x_, y_;\n for (int i = 1; i < line + 1; i++)\n {\n float x = (i - 0.5) * line_step - stride / 2;\n x_.push_back(x);\n }\n for (int i = 1; i < row + 1; i++)\n {\n float y = (i - 0.5) * row_step - stride / 2;\n y_.push_back(y);\n }\n std::vector shift_x((size_t)row * line, 0), shift_y((size_t)row * line, 0);\n for (int i = 0; i < row; i++)\n {\n for (int j = 0; j < line; j++)\n {\n shift_x[i * line + j] = x_[j];\n }\n }\n for (int i = 0; i < row; i++)\n {\n for (int j = 0; j < line; j++)\n {\n shift_y[i * line + j] = y_[i];\n }\n }\n anchor_points.resize((size_t)row * line * 2, 0);\n for (int i = 0; i < row * line; i++)\n {\n float x = shift_x[i];\n float y = shift_y[i];\n anchor_points[i * 2] = x;\n anchor_points[i * 2 + 1] = y;\n }\n}\nstatic void generate_anchor_points(int img_w, int img_h, std::vector pyramid_levels, int row, int line, std::vector& all_anchor_points)\n{\n std::vector > image_shapes;\n std::vector strides;\n for (int i = 0; i < pyramid_levels.size(); i++)\n {\n int new_h = std::floor((img_h + std::pow(2, pyramid_levels[i]) - 1) / std::pow(2, pyramid_levels[i]));\n int new_w = std::floor((img_w + std::pow(2, pyramid_levels[i]) - 1) / std::pow(2, pyramid_levels[i]));\n image_shapes.push_back(std::make_pair(new_w, new_h));\n strides.push_back(std::pow(2, pyramid_levels[i]));\n }\n\n all_anchor_points.clear();\n for (int i = 0; i < pyramid_levels.size(); i++)\n {\n std::vector anchor_points;\n generate_anchor_points(std::pow(2, pyramid_levels[i]), row, line, anchor_points);\n std::vector shifted_anchor_points;\n shift(image_shapes[i].first, image_shapes[i].second, strides[i], anchor_points, shifted_anchor_points);\n all_anchor_points.insert(all_anchor_points.end(), shifted_anchor_points.begin(), shifted_anchor_points.end());\n }\n}\n\nstatic int detect_crowd(const cv::Mat& bgr, std::vector& crowd_points)\n{\n ncnn::Option opt;\n opt.num_threads = 4;\n opt.use_vulkan_compute = false;\n opt.use_bf16_storage = false;\n\n ncnn::Net net;\n net.opt = opt;\n\n // model is converted from\n // https://github.com/TencentYoutuResearch/CrowdCounting-P2PNet\n // the ncnn model https://pan.baidu.com/s/1O1CBgvY6yJkrK8Npxx3VMg pwd: ezhx\n if (net.load_param(\"p2pnet.param\"))\n exit(-1);\n if (net.load_model(\"p2pnet.bin\"))\n exit(-1);\n\n int width = bgr.cols;\n int height = bgr.rows;\n\n int new_width = width / 128 * 128;\n int new_height = height / 128 * 128;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, width, height, new_width, new_height);\n\n std::vector pyramid_levels(1, 3);\n std::vector all_anchor_points;\n generate_anchor_points(in.w, in.h, pyramid_levels, 2, 2, all_anchor_points);\n\n ncnn::Mat anchor_points = ncnn::Mat(2, all_anchor_points.size() / 2, all_anchor_points.data());\n\n ncnn::Extractor ex = net.create_extractor();\n const float mean_vals1[3] = {123.675f, 116.28f, 103.53f};\n const float norm_vals1[3] = {0.01712475f, 0.0175f, 0.01742919f};\n\n in.substract_mean_normalize(mean_vals1, norm_vals1);\n\n ex.input(\"input\", in);\n ex.input(\"anchor\", anchor_points);\n\n ncnn::Mat score, points;\n ex.extract(\"pred_scores\", score);\n ex.extract(\"pred_points\", points);\n\n for (int i = 0; i < points.h; i++)\n {\n float* score_data = score.row(i);\n float* points_data = points.row(i);\n CrowdPoint cp;\n int x = points_data[0] / new_width * width;\n int y = points_data[1] / new_height * height;\n cp.pt = cv::Point(x, y);\n cp.prob = score_data[1];\n crowd_points.push_back(cp);\n }\n\n return 0;\n}\n\nstatic void draw_result(const cv::Mat& bgr, const std::vector& crowd_points)\n{\n cv::Mat image = bgr.clone();\n const float threshold = 0.5f;\n for (int i = 0; i < crowd_points.size(); i++)\n {\n if (crowd_points[i].prob > threshold)\n {\n cv::circle(image, crowd_points[i].pt, 4, cv::Scalar(0, 0, 255), -1, 8, 0);\n }\n }\n cv::imshow(\"image\", image);\n cv::waitKey();\n}\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat bgr = cv::imread(imagepath, 1);\n if (bgr.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector crowd_points;\n detect_crowd(bgr, crowd_points);\n draw_result(bgr, crowd_points);\n\n return 0;\n}\n" }, { "path": "examples/peleenetssd_seg.cpp", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_peleenet(const cv::Mat& bgr, std::vector& objects, ncnn::Mat& resized)\n{\n ncnn::Net peleenet;\n\n peleenet.opt.use_vulkan_compute = true;\n\n // model is converted from https://github.com/eric612/MobileNet-YOLO\n // and can be downloaded from https://drive.google.com/open?id=1Wt6jKv13sBRMHgrGAJYlOlRF-o80pC0g\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (peleenet.load_param(\"pelee.param\"))\n exit(-1);\n if (peleenet.load_model(\"pelee.bin\"))\n exit(-1);\n\n const int target_size = 304;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {103.9f, 116.7f, 123.6f};\n const float norm_vals[3] = {0.017f, 0.017f, 0.017f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = peleenet.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n ncnn::Mat seg_out;\n ex.extract(\"sigmoid\", seg_out);\n resize_bilinear(seg_out, resized, img_w, img_h);\n //resize_bicubic(seg_out,resized,img_w,img_h); // sharpness\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects, ncnn::Mat map)\n{\n static const char* class_names[] = {\"background\",\n \"person\", \"rider\", \"car\", \"bus\",\n \"truck\", \"bike\", \"motor\",\n \"traffic light\", \"traffic sign\", \"train\"\n };\n\n cv::Mat image = bgr.clone();\n const int color[] = {128, 255, 128, 244, 35, 232};\n const int color_count = sizeof(color) / sizeof(int);\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n int width = map.w;\n int height = map.h;\n int size = map.c;\n int img_index2 = 0;\n float threshold = 0.45;\n const float* ptr2 = map;\n for (int i = 0; i < height; i++)\n {\n unsigned char* ptr1 = image.ptr(i);\n int img_index1 = 0;\n for (int j = 0; j < width; j++)\n {\n float maxima = threshold;\n int index = -1;\n for (int c = 0; c < size; c++)\n {\n //const float* ptr3 = map.channel(c);\n const float* ptr3 = ptr2 + c * width * height;\n if (ptr3[img_index2] > maxima)\n {\n maxima = ptr3[img_index2];\n index = c;\n }\n }\n if (index > -1)\n {\n int color_index = (index)*3;\n if (color_index < color_count)\n {\n int b = color[color_index];\n int g = color[color_index + 1];\n int r = color[color_index + 2];\n ptr1[img_index1] = b / 2 + ptr1[img_index1] / 2;\n ptr1[img_index1 + 1] = g / 2 + ptr1[img_index1 + 1] / 2;\n ptr1[img_index1 + 2] = r / 2 + ptr1[img_index1 + 2] / 2;\n }\n }\n img_index1 += 3;\n img_index2++;\n }\n }\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n ncnn::Mat seg_out;\n detect_peleenet(m, objects, seg_out);\n\n draw_objects(m, objects, seg_out);\n\n return 0;\n}\n" }, { "path": "examples/piper.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// convert piper checkpoints to ncnn models\n// 1. checkout https://github.com/OHF-Voice/piper1-gpl (113931937cf235fc881afd1ca4be209bc6919bc7)\n// 2. apply patch piper1-gpl.patch from https://github.com/nihui/ncnn-android-piper\n// 3. setup piper with\n// python3 -m venv .venv\n// source .venv/bin/activate\n// python3 -m pip install -e .[train]\n// 4. download piper checkpoint file (*.ckpt) from https://huggingface.co/datasets/rhasspy/piper-checkpoints\n// 5. install pnnx via pip install -U pnnx\n// 6. obtain export_ncnn.py script from https://github.com/nihui/ncnn-android-piper\n// python export_ncnn.py en.ckpt\n\n// convert word list to simple phonemizer dict\n// 1. prepare word list from https://github.com/Alexir/CMUdict\n// 2. for each word, get phonemes via command \"./espeak-ng -q -v en-us --ipa word\"\n// 3. obtain config.json file from https://huggingface.co/datasets/rhasspy/piper-checkpoints\n// 4. replace phonemes with ids according to phoneme_id_map in config.json\n// 5. write dict binary\n// word1 \\0x00 ids1 \\0xff word2 \\0x00 ids2 \\0xff .....\n\n#include \"layer.h\"\n#include \"mat.h\"\n#include \"net.h\"\n\n#include \n#include \n#include \n#include \n\nclass relative_embeddings_k_module : public ncnn::Layer\n{\npublic:\n relative_embeddings_k_module()\n {\n one_blob_only = true;\n }\n\n virtual int forward(const ncnn::Mat& bottom_blob, ncnn::Mat& top_blob, const ncnn::Option& opt) const\n {\n const int window_size = 4;\n\n const int wsize = bottom_blob.w;\n const int len = bottom_blob.h;\n const int num_heads = bottom_blob.c;\n\n top_blob.create(len, len, num_heads);\n\n top_blob.fill(0.f);\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int q = 0; q < num_heads; q++)\n {\n const ncnn::Mat x0 = bottom_blob.channel(q);\n ncnn::Mat out0 = top_blob.channel(q);\n\n for (int i = 0; i < len; i++)\n {\n const float* xptr = x0.row(i) + std::max(0, window_size - i);\n float* outptr = out0.row(i) + std::max(i - window_size, 0);\n const int wsize2 = std::min(len, i - window_size + wsize) - std::max(i - window_size, 0);\n for (int j = 0; j < wsize2; j++)\n {\n *outptr++ = *xptr++;\n }\n }\n }\n\n return 0;\n }\n};\n\nDEFINE_LAYER_CREATOR(relative_embeddings_k_module)\n\nclass relative_embeddings_v_module : public ncnn::Layer\n{\npublic:\n relative_embeddings_v_module()\n {\n one_blob_only = true;\n }\n\n virtual int forward(const ncnn::Mat& bottom_blob, ncnn::Mat& top_blob, const ncnn::Option& opt) const\n {\n const int window_size = 4;\n\n const int wsize = window_size * 2 + 1;\n const int len = bottom_blob.h;\n const int num_heads = bottom_blob.c;\n\n top_blob.create(wsize, len, num_heads);\n\n top_blob.fill(0.f);\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int q = 0; q < num_heads; q++)\n {\n const ncnn::Mat x0 = bottom_blob.channel(q);\n ncnn::Mat out0 = top_blob.channel(q);\n\n for (int i = 0; i < len; i++)\n {\n const float* xptr = x0.row(i) + std::max(i - window_size, 0);\n float* outptr = out0.row(i) + std::max(0, window_size - i);\n const int wsize2 = std::min(len, i - window_size + wsize) - std::max(i - window_size, 0);\n for (int j = 0; j < wsize2; j++)\n {\n *outptr++ = *xptr++;\n }\n }\n }\n\n return 0;\n }\n};\n\nDEFINE_LAYER_CREATOR(relative_embeddings_v_module)\n\nclass piecewise_rational_quadratic_transform_module : public ncnn::Layer\n{\npublic:\n piecewise_rational_quadratic_transform_module()\n {\n one_blob_only = false;\n }\n\n virtual int forward(const std::vector& bottom_blobs, std::vector& top_blobs, const ncnn::Option& opt) const\n {\n const ncnn::Mat& h = bottom_blobs[0];\n const ncnn::Mat& x1 = bottom_blobs[1];\n ncnn::Mat& outputs = top_blobs[0];\n\n const int num_bins = 10;\n const int filter_channels = 192;\n const bool reverse = true;\n const float tail_bound = 5.0f;\n const float DEFAULT_MIN_BIN_WIDTH = 1e-3f;\n const float DEFAULT_MIN_BIN_HEIGHT = 1e-3f;\n const float DEFAULT_MIN_DERIVATIVE = 1e-3f;\n\n const int batch_size = x1.w;\n const int h_params_per_item = 2 * num_bins + (num_bins - 1); // 29\n\n outputs = x1.clone();\n\n float* out_ptr = outputs;\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int i = 0; i < batch_size; ++i)\n {\n const float current_x = ((const float*)x1)[i];\n\n const float* h_data = h.row(i);\n\n if (current_x < -tail_bound || current_x > tail_bound)\n {\n continue;\n }\n\n std::vector unnormalized_widths(num_bins);\n std::vector unnormalized_heights(num_bins);\n std::vector unnormalized_derivatives(num_bins + 1);\n\n const float inv_sqrt_filter_channels = 1.0f / sqrtf(filter_channels);\n for (int j = 0; j < num_bins; ++j)\n {\n unnormalized_widths[j] = h_data[j] * inv_sqrt_filter_channels;\n }\n for (int j = 0; j < num_bins; ++j)\n {\n unnormalized_heights[j] = h_data[num_bins + j] * inv_sqrt_filter_channels;\n }\n for (int j = 0; j < num_bins - 1; ++j)\n {\n unnormalized_derivatives[j + 1] = h_data[2 * num_bins + j];\n }\n\n const float constant = logf(expf(1.f - DEFAULT_MIN_DERIVATIVE) - 1.f);\n unnormalized_derivatives[0] = constant;\n unnormalized_derivatives[num_bins] = constant;\n\n const float left = -tail_bound, right = tail_bound;\n const float bottom = -tail_bound, top = tail_bound;\n\n // Softmax + Affine\n std::vector widths(num_bins);\n float w_max = -INFINITY;\n for (float val : unnormalized_widths) w_max = std::max(w_max, val);\n float w_sum = 0.f;\n for (int j = 0; j < num_bins; ++j)\n {\n widths[j] = expf(unnormalized_widths[j] - w_max);\n w_sum += widths[j];\n }\n for (int j = 0; j < num_bins; ++j)\n {\n widths[j] = DEFAULT_MIN_BIN_WIDTH + (1.f - DEFAULT_MIN_BIN_WIDTH * num_bins) * (widths[j] / w_sum);\n }\n\n // cumwidths\n std::vector cumwidths(num_bins + 1);\n cumwidths[0] = left;\n float current_w_sum = 0.f;\n for (int j = 0; j < num_bins - 1; ++j)\n {\n current_w_sum += widths[j];\n cumwidths[j + 1] = left + (right - left) * current_w_sum;\n }\n cumwidths[num_bins] = right;\n\n // heights\n std::vector heights(num_bins);\n float h_max = -INFINITY;\n for (float val : unnormalized_heights) h_max = std::max(h_max, val);\n float h_sum = 0.f;\n for (int j = 0; j < num_bins; ++j)\n {\n heights[j] = expf(unnormalized_heights[j] - h_max);\n h_sum += heights[j];\n }\n for (int j = 0; j < num_bins; ++j)\n {\n heights[j] = DEFAULT_MIN_BIN_HEIGHT + (1.f - DEFAULT_MIN_BIN_HEIGHT * num_bins) * (heights[j] / h_sum);\n }\n\n // cumheights\n std::vector cumheights(num_bins + 1);\n cumheights[0] = bottom;\n float current_h_sum = 0.f;\n for (int j = 0; j < num_bins - 1; ++j)\n {\n current_h_sum += heights[j];\n cumheights[j + 1] = bottom + (top - bottom) * current_h_sum;\n }\n cumheights[num_bins] = top;\n\n // Softplus\n std::vector derivatives(num_bins + 1);\n for (int j = 0; j < num_bins + 1; ++j)\n {\n float x = unnormalized_derivatives[j];\n derivatives[j] = DEFAULT_MIN_DERIVATIVE + (x > 0 ? x + logf(1.f + expf(-x)) : logf(1.f + expf(x)));\n }\n\n // bin_idx\n int bin_idx = 0;\n if (reverse)\n {\n auto it = std::upper_bound(cumheights.begin(), cumheights.end(), current_x);\n bin_idx = std::distance(cumheights.begin(), it) - 1;\n }\n else\n {\n auto it = std::upper_bound(cumwidths.begin(), cumwidths.end(), current_x);\n bin_idx = std::distance(cumwidths.begin(), it) - 1;\n }\n bin_idx = std::max(0, std::min(bin_idx, num_bins - 1));\n\n // collect coeffs\n const float input_cumwidths = cumwidths[bin_idx];\n const float input_bin_widths = cumwidths[bin_idx + 1] - cumwidths[bin_idx];\n const float input_cumheights = cumheights[bin_idx];\n const float input_heights = cumheights[bin_idx + 1] - cumheights[bin_idx];\n const float input_derivatives = derivatives[bin_idx];\n const float input_derivatives_plus_one = derivatives[bin_idx + 1];\n const float delta = input_heights / input_bin_widths;\n\n // apply transform\n if (reverse)\n {\n float a = (current_x - input_cumheights) * (input_derivatives + input_derivatives_plus_one - 2 * delta) + input_heights * (delta - input_derivatives);\n float b = input_heights * input_derivatives - (current_x - input_cumheights) * (input_derivatives + input_derivatives_plus_one - 2 * delta);\n float c = -delta * (current_x - input_cumheights);\n float discriminant = b * b - 4 * a * c;\n discriminant = std::max(0.f, discriminant);\n float root = (2 * c) / (-b - sqrtf(discriminant));\n out_ptr[i] = root * input_bin_widths + input_cumwidths;\n }\n else\n {\n float theta = (current_x - input_cumwidths) / input_bin_widths;\n float theta_one_minus_theta = theta * (1 - theta);\n float numerator = input_heights * (delta * theta * theta + input_derivatives * theta_one_minus_theta);\n float denominator = delta + ((input_derivatives + input_derivatives_plus_one - 2 * delta) * theta_one_minus_theta);\n out_ptr[i] = input_cumheights + numerator / denominator;\n }\n }\n\n return 0;\n }\n};\n\nDEFINE_LAYER_CREATOR(piecewise_rational_quadratic_transform_module)\n\nstatic bool is_word_eos(const char* word)\n{\n const char c = word[0];\n return c == ',' || c == '.' || c == ';' || c == '?' || c == '!';\n}\n\nstatic void find_word_id(const std::map >& dict, const char* word, const unsigned char*& ids)\n{\n ids = 0;\n\n unsigned char first_char = toupper(word[0]);\n if (dict.find(first_char) == dict.end())\n return;\n\n const std::vector& wordlist = dict.at(first_char);\n for (size_t i = 0; i < wordlist.size(); i++)\n {\n if (strcasecmp(wordlist[i], word) == 0)\n {\n // hit\n ids = (const unsigned char*)(wordlist[i] + strlen(wordlist[i]) + 1);\n return;\n }\n }\n}\n\nstatic void simple_phonemize(const char* text, std::vector& sequence_ids)\n{\n // this is a very simple g2p function, it works for english only\n\n // load dict buffer\n std::vector dictbinbuf;\n {\n FILE* fp = fopen(\"en-word_id.bin\", \"rb\");\n if (!fp)\n return;\n\n fseek(fp, 0, SEEK_END);\n size_t len = ftell(fp);\n rewind(fp);\n\n dictbinbuf.resize(len);\n fread(dictbinbuf.data(), 1, len, fp);\n\n fclose(fp);\n }\n\n // build dict\n std::map > dict;\n {\n const unsigned char* p = dictbinbuf.data();\n const char* word = (const char*)p;\n for (size_t i = 0; i < dictbinbuf.size(); i++)\n {\n if (dictbinbuf[i] == 0xff)\n {\n unsigned int first_char = toupper(word[0]);\n dict[first_char].push_back(word);\n word = (const char*)(p + i + 1);\n }\n }\n }\n\n // phonemize mainpart\n {\n const int ID_PAD = 0; // interleaved\n const int ID_BOS = 1; // beginning of sentence\n const int ID_EOS = 2; // end of sentence\n const int ID_SPACE = 3; // space\n\n bool last_char_is_control = false;\n bool sentence_begin = true;\n bool sentence_end = true;\n\n char word[256];\n\n const char* p = text;\n while (*p)\n {\n if (sentence_end && !last_char_is_control)\n {\n sequence_ids.push_back(ID_BOS);\n sequence_ids.push_back(ID_PAD);\n sentence_end = false;\n }\n\n if (sentence_begin || last_char_is_control)\n {\n // the very first word\n }\n else\n {\n // space id\n sequence_ids.push_back(ID_SPACE);\n sequence_ids.push_back(ID_PAD);\n }\n\n if (isalnum((unsigned char)*p))\n {\n char* pword = word;\n\n // alpha or number\n *pword++ = *p++;\n\n // consume word\n int wordlen = 1;\n while (isalnum((unsigned char)*p) && wordlen < 233)\n {\n *pword++ = *p++;\n wordlen++;\n }\n\n *pword = '\\0';\n\n if (is_word_eos(word))\n {\n if (!sentence_end)\n sequence_ids.push_back(ID_EOS);\n sentence_end = true;\n last_char_is_control = false;\n sentence_begin = false;\n continue;\n }\n\n const unsigned char* ids = 0;\n find_word_id(dict, word, ids);\n if (ids)\n {\n const unsigned char* pids = ids;\n while (*pids != 0xff)\n {\n sequence_ids.push_back(*pids);\n sequence_ids.push_back(ID_PAD);\n pids++;\n }\n }\n else\n {\n // no such word, spell alphabet one by one\n char tmp[2] = {'\\0', '\\0'};\n for (size_t i = 0; i < strlen(word); i++)\n {\n tmp[0] = word[i];\n find_word_id(dict, tmp, ids);\n if (ids)\n {\n const unsigned char* pids = ids;\n while (*pids != 0xff)\n {\n sequence_ids.push_back(*pids);\n sequence_ids.push_back(ID_PAD);\n pids++;\n }\n if (i + 1 != strlen(word))\n {\n sequence_ids.push_back(ID_SPACE);\n sequence_ids.push_back(ID_PAD);\n }\n }\n else\n {\n fprintf(stderr, \"word char %c not recognized\\n\", word[i]);\n }\n }\n }\n\n last_char_is_control = false;\n sentence_begin = false;\n continue;\n }\n else\n {\n // skip control character\n p++;\n last_char_is_control = true;\n }\n }\n\n if (!sentence_end)\n sequence_ids.push_back(ID_EOS);\n }\n}\n\nstatic void path_attention(const ncnn::Mat& logw, const ncnn::Mat& m_p, const ncnn::Mat& logs_p, float noise_scale, float length_scale, ncnn::Mat& z_p)\n{\n const int x_lengths = logw.w;\n\n // assert m_p.h == logs_p.h\n const int depth = m_p.h;\n\n std::vector w_ceil(x_lengths);\n int y_lengths = 0;\n for (int i = 0; i < x_lengths; i++)\n {\n w_ceil[i] = (int)ceilf(expf(logw[i]) * length_scale);\n y_lengths += w_ceil[i];\n }\n\n z_p.create(y_lengths, depth);\n\n for (int i = 0; i < depth; i++)\n {\n const float* m_p_ptr = m_p.row(i);\n const float* logs_p_ptr = logs_p.row(i);\n float* ptr = z_p.row(i);\n\n for (int j = 0; j < x_lengths; j++)\n {\n const float m = m_p_ptr[j];\n const float nl = expf(logs_p_ptr[j]) * noise_scale;\n const int duration = w_ceil[j];\n\n for (int k = 0; k < duration; k++)\n {\n ptr[k] = m + (rand() / (float)RAND_MAX) * nl;\n }\n ptr += duration;\n }\n }\n}\n\nstatic int tts_piper(const char* text, int speaker_id, std::vector& pcm)\n{\n // zh models could be found at\n // https://github.com/nihui/ncnn-android-piper/tree/master/app/src/main/assets\n\n // hyper parameters from https://huggingface.co/datasets/rhasspy/piper-checkpoints/blob/main/en/en_US/libritts_r/medium/config.json\n const float noise_scale = 0.333f;\n const float length_scale = 1.f;\n const float noise_scale_w = 0.333f;\n\n // phonemize\n ncnn::Mat sequence;\n {\n std::vector sequence_ids;\n simple_phonemize(text, sequence_ids);\n\n const int sequence_length = (int)sequence_ids.size();\n\n sequence.create(sequence_length);\n memcpy(sequence, sequence_ids.data(), sequence_length * sizeof(int));\n }\n\n // enc_p\n ncnn::Mat x;\n ncnn::Mat m_p;\n ncnn::Mat logs_p;\n {\n ncnn::Net enc_p;\n enc_p.opt.use_vulkan_compute = true;\n enc_p.register_custom_layer(\"piper.train.vits.attentions.relative_embeddings_k_module\", relative_embeddings_k_module_layer_creator);\n enc_p.register_custom_layer(\"piper.train.vits.attentions.relative_embeddings_v_module\", relative_embeddings_v_module_layer_creator);\n enc_p.load_param(\"en_enc_p.ncnn.param\");\n enc_p.load_model(\"en_enc_p.ncnn.bin\");\n\n ncnn::Extractor ex = enc_p.create_extractor();\n\n ex.input(\"in0\", sequence);\n\n ex.extract(\"out0\", x);\n ex.extract(\"out1\", m_p);\n ex.extract(\"out2\", logs_p);\n }\n\n // emb_g\n ncnn::Mat g;\n {\n ncnn::Net emb_g;\n emb_g.opt.use_vulkan_compute = true;\n emb_g.load_param(\"en_emb_g.ncnn.param\");\n emb_g.load_model(\"en_emb_g.ncnn.bin\");\n\n ncnn::Mat speaker_id_mat(1);\n {\n int* p = speaker_id_mat;\n p[0] = speaker_id;\n }\n\n ncnn::Extractor ex = emb_g.create_extractor();\n\n ex.input(\"in0\", speaker_id_mat);\n\n ex.extract(\"out0\", g);\n\n g = g.reshape(1, g.w);\n }\n\n // dp\n ncnn::Mat logw;\n {\n ncnn::Net dp;\n dp.opt.use_vulkan_compute = true;\n dp.register_custom_layer(\"piper.train.vits.modules.piecewise_rational_quadratic_transform_module\", piecewise_rational_quadratic_transform_module_layer_creator);\n dp.load_param(\"en_dp.ncnn.param\");\n dp.load_model(\"en_dp.ncnn.bin\");\n\n ncnn::Mat noise(x.w, 2);\n for (int i = 0; i < noise.w * noise.h; i++)\n {\n noise[i] = rand() / (float)RAND_MAX * noise_scale_w;\n }\n\n ncnn::Extractor ex = dp.create_extractor();\n\n ex.input(\"in0\", x);\n ex.input(\"in1\", noise);\n ex.input(\"in2\", g);\n\n ex.extract(\"out0\", logw);\n }\n\n // path attention\n ncnn::Mat z_p;\n {\n path_attention(logw, m_p, logs_p, noise_scale, length_scale, z_p);\n }\n\n // flow\n ncnn::Mat z;\n {\n ncnn::Net flow;\n flow.opt.use_vulkan_compute = true;\n flow.load_param(\"en_flow.ncnn.param\");\n flow.load_model(\"en_flow.ncnn.bin\");\n\n ncnn::Extractor ex = flow.create_extractor();\n\n ex.input(\"in0\", z_p);\n ex.input(\"in1\", g);\n\n ex.extract(\"out0\", z);\n }\n\n // dec\n ncnn::Mat o;\n {\n ncnn::Net dec;\n dec.opt.use_vulkan_compute = true;\n dec.load_param(\"en_dec.ncnn.param\");\n dec.load_model(\"en_dec.ncnn.bin\");\n\n ncnn::Extractor ex = dec.create_extractor();\n\n ex.input(\"in0\", z);\n ex.input(\"in1\", g);\n\n ex.extract(\"out0\", o);\n }\n\n // normalize and clip\n {\n float volume = 1.f;\n float absmax = 0.f;\n for (int i = 0; i < o.w; i++)\n {\n absmax = std::max(absmax, fabs(o[i]));\n }\n if (absmax > 1e-8)\n {\n for (int i = 0; i < o.w; i++)\n {\n float v = o[i] / absmax * volume;\n v = std::min(std::max(v, -1.f), 1.f);\n o[i] = v;\n }\n }\n }\n\n // 16bit pcm\n {\n pcm.resize(o.w);\n for (int i = 0; i < o.w; i++)\n {\n pcm[i] = (short)(o[i] * 32767);\n }\n }\n\n return 0;\n}\n\nstatic void save_pcm_to_wav(const char* path, const short* pcm, int num_samples, int sample_rate)\n{\n FILE* f = fopen(path, \"wb\");\n if (!f)\n return;\n\n // write wav header\n {\n int16_t num_channels = 1;\n int16_t bits_per_sample = 16;\n int32_t byte_rate = sample_rate * num_channels * bits_per_sample / 8;\n int16_t block_align = num_channels * bits_per_sample / 8;\n int32_t data_chunk_size = num_samples * num_channels * bits_per_sample / 8;\n int32_t chunk_size = 36 + data_chunk_size;\n\n // RIFF header\n fwrite(\"RIFF\", 1, 4, f);\n fwrite(&chunk_size, 4, 1, f);\n fwrite(\"WAVE\", 1, 4, f);\n\n // fmt subchunk\n fwrite(\"fmt \", 1, 4, f);\n int32_t subchunk1_size = 16;\n int16_t audio_format = 1; // PCM\n fwrite(&subchunk1_size, 4, 1, f);\n fwrite(&audio_format, 2, 1, f);\n fwrite(&num_channels, 2, 1, f);\n fwrite(&sample_rate, 4, 1, f);\n fwrite(&byte_rate, 4, 1, f);\n fwrite(&block_align, 2, 1, f);\n fwrite(&bits_per_sample, 2, 1, f);\n\n // data subchunk\n fwrite(\"data\", 1, 4, f);\n fwrite(&data_chunk_size, 4, 1, f);\n }\n\n fwrite(pcm, sizeof(short), num_samples, f);\n fclose(f);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 4)\n {\n fprintf(stderr, \"Usage: %s [sentences] [speaker id 0~903] [out path]\\n\", argv[0]);\n fprintf(stderr, \" %s \\\"Hello World\\\" 0 out.wav\\n\", argv[0]);\n fprintf(stderr, \" %s \\\"Happy New Year\\\" 123 out.wav\\n\", argv[0]);\n return 0;\n }\n\n const char* text = argv[1];\n const int speaker_id = atoi(argv[2]);\n const char* outpath = argv[3];\n\n std::vector pcm;\n tts_piper(text, speaker_id, pcm);\n\n // \"sample_rate\": 22050\n save_pcm_to_wav(outpath, pcm.data(), pcm.size(), 22050);\n\n return 0;\n}\n" }, { "path": "examples/ppocrv5.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// pip install paddlepaddle==3.0.0\n// pip install paddleocr==3.0.0\n// paddlex --install paddle2onnx\n// paddleocr ocr -i test.png\n// paddlex --paddle2onnx --paddle_model_dir ~/.paddlex/official_models/PP-OCRv5_mobile_det --onnx_model_dir PP-OCRv5_mobile_det\n// paddlex --paddle2onnx --paddle_model_dir ~/.paddlex/official_models/PP-OCRv5_mobile_rec --onnx_model_dir PP-OCRv5_mobile_rec\n// pnnx PP-OCRv5_mobile_det.onnx inputshape=[1,3,320,320] inputshape2=[1,3,256,256]\n// pnnx PP-OCRv5_mobile_rec.onnx inputshape=[1,3,48,160] inputshape2=[1,3,48,256]\n// pnnx PP-OCRv5_server_det.onnx inputshape=[1,3,320,320] inputshape2=[1,3,256,256] fp16=0\n// pnnx PP-OCRv5_server_rec.onnx inputshape=[1,3,48,160] inputshape2=[1,3,48,256] fp16=0\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n\n#include \"ppocrv5_dict.h\"\n\nstruct Character\n{\n int id;\n float prob;\n};\n\nstruct Object\n{\n cv::RotatedRect rrect;\n int orientation;\n float prob;\n std::vector text;\n};\n\nstatic double contour_score(const cv::Mat& binary, const std::vector& contour)\n{\n cv::Rect rect = cv::boundingRect(contour);\n if (rect.x < 0)\n rect.x = 0;\n if (rect.y < 0)\n rect.y = 0;\n if (rect.x + rect.width > binary.cols)\n rect.width = binary.cols - rect.x;\n if (rect.y + rect.height > binary.rows)\n rect.height = binary.rows - rect.y;\n\n cv::Mat binROI = binary(rect);\n\n cv::Mat mask = cv::Mat::zeros(rect.height, rect.width, CV_8U);\n std::vector roiContour;\n for (size_t i = 0; i < contour.size(); i++)\n {\n cv::Point pt = cv::Point(contour[i].x - rect.x, contour[i].y - rect.y);\n roiContour.push_back(pt);\n }\n\n std::vector > roiContours = {roiContour};\n cv::fillPoly(mask, roiContours, cv::Scalar(255));\n\n double score = cv::mean(binROI, mask).val[0];\n return score / 255.f;\n}\n\nstatic cv::Mat get_rotate_crop_image(const cv::Mat& bgr, const Object& object)\n{\n const int orientation = object.orientation;\n const float rw = object.rrect.size.width;\n const float rh = object.rrect.size.height;\n\n const int target_height = 48;\n const float target_width = rh * target_height / rw;\n\n // warpperspective shall be used to rotate the image\n // but actually they are all rectangles, so warpaffine is almost enough :P\n\n cv::Mat dst;\n\n cv::Point2f corners[4];\n object.rrect.points(corners);\n\n if (orientation == 0)\n {\n // horizontal text\n // corner points order\n // 0--------1\n // | |rw -> as angle=90\n // 3--------2\n // rh\n\n std::vector src_pts(3);\n src_pts[0] = corners[0];\n src_pts[1] = corners[1];\n src_pts[2] = corners[3];\n\n std::vector dst_pts(3);\n dst_pts[0] = cv::Point2f(0, 0);\n dst_pts[1] = cv::Point2f(target_width, 0);\n dst_pts[2] = cv::Point2f(0, target_height);\n\n cv::Mat tm = cv::getAffineTransform(src_pts, dst_pts);\n\n cv::warpAffine(bgr, dst, tm, cv::Size(target_width, target_height), cv::INTER_LINEAR, cv::BORDER_REPLICATE);\n }\n else\n {\n // vertial text\n // corner points order\n // 1----2\n // | |\n // | |\n // | |rh -> as angle=0\n // | |\n // | |\n // 0----3\n // rw\n\n std::vector src_pts(3);\n src_pts[0] = corners[2];\n src_pts[1] = corners[3];\n src_pts[2] = corners[1];\n\n std::vector dst_pts(3);\n dst_pts[0] = cv::Point2f(0, 0);\n dst_pts[1] = cv::Point2f(target_width, 0);\n dst_pts[2] = cv::Point2f(0, target_height);\n\n cv::Mat tm = cv::getAffineTransform(src_pts, dst_pts);\n\n cv::warpAffine(bgr, dst, tm, cv::Size(target_width, target_height), cv::INTER_LINEAR, cv::BORDER_REPLICATE);\n }\n\n return dst;\n}\n\nclass PPOCRv5\n{\npublic:\n void init();\n\n void detect(const cv::Mat& bgr, std::vector& objects);\n\n void recognize(const cv::Mat& bgr, Object& object);\n\nprotected:\n ncnn::Net ppocrv5_det;\n ncnn::Net ppocrv5_rec;\n};\n\nvoid PPOCRv5::init()\n{\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n // https://github.com/nihui/ncnn-android-ppocrv5/tree/master/app/src/main/assets\n\n ppocrv5_det.opt.use_vulkan_compute = true;\n // ppocrv5_det.opt.use_bf16_storage = true;\n\n // fp16 must be disabled for server model\n // ppocrv5_det.opt.use_fp16_packed = false;\n // ppocrv5_det.opt.use_fp16_storage = false;\n\n ppocrv5_det.load_param(\"PP_OCRv5_mobile_det.ncnn.param\");\n ppocrv5_det.load_model(\"PP_OCRv5_mobile_det.ncnn.bin\");\n // ppocrv5_det.load_param(\"PP_OCRv5_server_det.ncnn.param\");\n // ppocrv5_det.load_model(\"PP_OCRv5_server_det.ncnn.bin\");\n\n ppocrv5_rec.opt.use_vulkan_compute = true;\n // ppocrv5_rec.opt.use_bf16_storage = true;\n\n // fp16 must be disabled for server model\n // ppocrv5_rec.opt.use_fp16_packed = false;\n // ppocrv5_rec.opt.use_fp16_storage = false;\n\n ppocrv5_rec.load_param(\"PP_OCRv5_mobile_rec.ncnn.param\");\n ppocrv5_rec.load_model(\"PP_OCRv5_mobile_rec.ncnn.bin\");\n // ppocrv5_rec.load_param(\"PP_OCRv5_server_rec.ncnn.param\");\n // ppocrv5_rec.load_model(\"PP_OCRv5_server_rec.ncnn.bin\");\n}\n\nvoid PPOCRv5::detect(const cv::Mat& bgr, std::vector& objects)\n{\n const int target_size = 960;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n const int target_stride = 32;\n\n // letterbox pad to multiple of target_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (std::max(w, h) > target_size)\n {\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, img_w, img_h, w, h);\n\n int wpad = (w + target_stride - 1) / target_stride * target_stride - w;\n int hpad = (h + target_stride - 1) / target_stride * target_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float mean_vals[3] = {0.485f * 255.f, 0.456f * 255.f, 0.406f * 255.f};\n const float norm_vals[3] = {1 / 0.229f / 255.f, 1 / 0.224f / 255.f, 1 / 0.225f / 255.f};\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = ppocrv5_det.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n const float denorm_vals[1] = {255.f};\n out.substract_mean_normalize(0, denorm_vals);\n\n cv::Mat pred(out.h, out.w, CV_8UC1);\n out.to_pixels(pred.data, ncnn::Mat::PIXEL_GRAY);\n\n // threshold binary\n cv::Mat bitmap;\n const float threshold = 0.3f;\n cv::threshold(pred, bitmap, threshold * 255, 255, cv::THRESH_BINARY);\n\n // boxes from bitmap\n {\n // should use dbnet post process, but I think unclip process is difficult to write\n // so simply implement expansion. This may lose detection accuracy\n // original implementation can be referenced\n // https://github.com/MhLiao/DB/blob/master/structure/representers/seg_detector_representer.py\n\n const float box_thresh = 0.6f;\n const float enlarge_ratio = 1.95f;\n\n const float min_size = 3 * scale;\n const int max_candidates = 1000;\n\n std::vector > contours;\n std::vector hierarchy;\n\n cv::findContours(bitmap, contours, hierarchy, cv::RETR_LIST, cv::CHAIN_APPROX_SIMPLE);\n\n contours.resize(std::min(contours.size(), (size_t)max_candidates));\n\n for (size_t i = 0; i < contours.size(); i++)\n {\n const std::vector& contour = contours[i];\n if (contour.size() <= 2)\n continue;\n\n double score = contour_score(pred, contour);\n if (score < box_thresh)\n continue;\n\n cv::RotatedRect rrect = cv::minAreaRect(contour);\n\n float rrect_maxwh = std::max(rrect.size.width, rrect.size.height);\n if (rrect_maxwh < min_size)\n continue;\n\n int orientation = 0;\n if (rrect.angle >= -30 && rrect.angle <= 30 && rrect.size.height > rrect.size.width * 2.7)\n {\n // vertical text\n orientation = 1;\n }\n if ((rrect.angle <= -60 || rrect.angle >= 60) && rrect.size.width > rrect.size.height * 2.7)\n {\n // vertical text\n orientation = 1;\n }\n\n if (rrect.angle < -30)\n {\n // make orientation from -90 ~ -30 to 90 ~ 150\n rrect.angle += 180;\n }\n if (orientation == 0 && rrect.angle < 30)\n {\n // make it horizontal\n rrect.angle += 90;\n std::swap(rrect.size.width, rrect.size.height);\n }\n if (orientation == 1 && rrect.angle >= 60)\n {\n // make it vertical\n rrect.angle -= 90;\n std::swap(rrect.size.width, rrect.size.height);\n }\n\n // enlarge\n rrect.size.height += rrect.size.width * (enlarge_ratio - 1);\n rrect.size.width *= enlarge_ratio;\n\n // adjust offset to original unpadded\n rrect.center.x = (rrect.center.x - (wpad / 2)) / scale;\n rrect.center.y = (rrect.center.y - (hpad / 2)) / scale;\n rrect.size.width = (rrect.size.width) / scale;\n rrect.size.height = (rrect.size.height) / scale;\n\n Object obj;\n obj.rrect = rrect;\n obj.orientation = orientation;\n obj.prob = score;\n objects.push_back(obj);\n }\n }\n}\n\nvoid PPOCRv5::recognize(const cv::Mat& bgr, Object& object)\n{\n cv::Mat roi = get_rotate_crop_image(bgr, object);\n\n ncnn::Mat in = ncnn::Mat::from_pixels(roi.data, ncnn::Mat::PIXEL_BGR, roi.cols, roi.rows);\n\n // ~/.paddlex/official_models/PP-OCRv5_mobile_rec/inference.yml\n const float mean_vals[3] = {127.5, 127.5, 127.5};\n const float norm_vals[3] = {1.0 / 127.5, 1.0 / 127.5, 1.0 / 127.5};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = ppocrv5_rec.create_extractor();\n\n ex.input(\"in0\", in);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n // 18385 x len\n int last_token = 0;\n\n for (int i = 0; i < out.h; i++)\n {\n const float* p = out.row(i);\n\n int index = 0;\n float max_score = -9999.f;\n for (int j = 0; j < out.w; j++)\n {\n float score = *p++;\n if (score > max_score)\n {\n max_score = score;\n index = j;\n }\n }\n\n if (last_token == index) // CTC rule, if index is same as last one, they will be merged into one token\n continue;\n\n last_token = index;\n\n if (index <= 0)\n continue;\n\n Character ch;\n ch.id = index - 1;\n ch.prob = max_score;\n\n object.text.push_back(ch);\n }\n}\n\nstatic int detect_ppocrv5(const cv::Mat& bgr, std::vector& objects)\n{\n PPOCRv5 ppocrv5;\n\n ppocrv5.init();\n\n ppocrv5.detect(bgr, objects);\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n ppocrv5.recognize(bgr, objects[i]);\n }\n\n return 0;\n}\n\nstatic int draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const cv::Scalar colors[] = {\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 17];\n\n fprintf(stderr, \"%s %.5f at %.2f %.2f %.2f x %.2f @ %.2f = \", obj.orientation == 0 ? \"H\" : \"V\", obj.prob,\n obj.rrect.center.x, obj.rrect.center.y, obj.rrect.size.width, obj.rrect.size.height, obj.rrect.angle);\n\n cv::Point2f corners[4];\n obj.rrect.points(corners);\n cv::line(image, corners[0], corners[1], color);\n cv::line(image, corners[1], corners[2], color);\n cv::line(image, corners[2], corners[3], color);\n cv::line(image, corners[3], corners[0], color);\n\n std::string text;\n for (size_t j = 0; j < objects[i].text.size(); j++)\n {\n const Character& ch = objects[i].text[j];\n if (ch.id >= character_dict_size)\n continue;\n\n text += character_dict[ch.id];\n }\n fprintf(stderr, \"%s\\n\", text.c_str());\n }\n\n fprintf(stderr, \"opencv putText can not draw non-latin characters, you may see question marks instead\\n\");\n fprintf(stderr, \"see opencv-mobile for drawing non-latin characters\\n\");\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 17];\n\n std::string text;\n for (size_t j = 0; j < objects[i].text.size(); j++)\n {\n const Character& ch = objects[i].text[j];\n if (ch.id >= character_dict_size)\n {\n if (!text.empty() && text.back() != ' ')\n {\n text += \" \";\n }\n continue;\n }\n\n if (obj.orientation == 0)\n {\n text += character_dict[ch.id];\n }\n else\n {\n text += character_dict[ch.id];\n if (j + 1 < objects[i].text.size())\n text += \"\\n\";\n }\n }\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rrect.center.x - label_size.width / 2;\n int y = obj.rrect.center.y - label_size.height / 2 - baseLine;\n if (y < 0)\n y = 0;\n if (y + label_size.height > image.rows)\n y = image.rows - label_size.height;\n if (x < 0)\n x = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n if (obj.orientation == 0)\n {\n cv::putText(image, text, cv::Point(x, y + label_size.height), cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n else\n {\n cv::putText(image, text, cv::Point(x, y + label_size.width), cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_ppocrv5(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/ppocrv5_dict.h", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\nstatic const char* character_dict[] = {\n \"　\",\n \"一\",\n \"乙\",\n \"二\",\n \"十\",\n \"丁\",\n \"厂\",\n \"七\",\n \"卜\",\n \"八\",\n \"人\",\n \"入\",\n \"儿\",\n \"匕\",\n \"几\",\n \"九\",\n \"刁\",\n \"了\",\n \"刀\",\n \"力\",\n \"乃\",\n \"又\",\n \"三\",\n \"干\",\n \"于\",\n \"亏\",\n \"工\",\n \"土\",\n \"士\",\n \"才\",\n \"下\",\n \"寸\",\n \"大\",\n \"丈\",\n \"与\",\n \"万\",\n \"上\",\n \"小\",\n \"口\",\n \"山\",\n \"巾\",\n \"千\",\n \"乞\",\n \"川\",\n \"亿\",\n \"个\",\n \"夕\",\n \"久\",\n \"么\",\n \"勺\",\n \"凡\",\n \"丸\",\n \"及\",\n \"广\",\n \"亡\",\n \"门\",\n \"丫\",\n \"义\",\n \"之\",\n \"尸\",\n \"己\",\n \"已\",\n \"巳\",\n \"弓\",\n \"子\",\n \"卫\",\n \"也\",\n \"女\",\n \"刃\",\n \"飞\",\n \"习\",\n \"叉\",\n \"马\",\n \"乡\",\n \"丰\",\n \"王\",\n \"开\",\n \"井\",\n \"天\",\n \"夫\",\n \"元\",\n \"无\",\n \"云\",\n \"专\",\n \"丐\",\n \"扎\",\n \"艺\",\n \"木\",\n \"五\",\n \"支\",\n \"厅\",\n \"不\",\n \"犬\",\n \"太\",\n \"区\",\n \"历\",\n \"歹\",\n \"友\",\n \"尤\",\n \"匹\",\n \"车\",\n \"巨\",\n \"牙\",\n \"屯\",\n \"戈\",\n \"比\",\n \"互\",\n \"切\",\n \"瓦\",\n \"止\",\n \"少\",\n \"曰\",\n \"日\",\n \"中\",\n \"贝\",\n \"冈\",\n \"内\",\n \"水\",\n \"见\",\n \"午\",\n \"牛\",\n \"手\",\n \"气\",\n \"毛\",\n \"壬\",\n \"升\",\n \"夭\",\n \"长\",\n \"仁\",\n \"什\",\n \"片\",\n \"仆\",\n \"化\",\n \"仇\",\n \"币\",\n \"仍\",\n \"仅\",\n \"斤\",\n \"爪\",\n \"反\",\n \"介\",\n \"父\",\n \"从\",\n \"仑\",\n \"今\",\n \"凶\",\n \"分\",\n \"乏\",\n \"公\",\n \"仓\",\n \"月\",\n \"氏\",\n \"勿\",\n \"欠\",\n \"风\",\n \"丹\",\n \"匀\",\n \"乌\",\n \"勾\",\n \"凤\",\n \"六\",\n \"文\",\n \"亢\",\n \"方\",\n \"火\",\n \"为\",\n \"斗\",\n \"忆\",\n \"计\",\n \"订\",\n \"户\",\n \"认\",\n \"冗\",\n \"讥\",\n \"心\",\n \"尺\",\n \"引\",\n \"丑\",\n \"巴\",\n \"孔\",\n \"队\",\n \"办\",\n \"以\",\n \"允\",\n \"予\",\n \"邓\",\n \"劝\",\n \"双\",\n \"书\",\n \"幻\",\n \"玉\",\n \"刊\",\n \"未\",\n \"末\",\n \"示\",\n \"击\",\n \"打\",\n \"巧\",\n \"正\",\n \"扑\",\n \"卉\",\n \"扒\",\n \"功\",\n \"扔\",\n \"去\",\n \"甘\",\n \"世\",\n \"艾\",\n \"古\",\n \"节\",\n \"本\",\n \"术\",\n \"可\",\n \"丙\",\n \"左\",\n \"厉\",\n \"石\",\n \"右\",\n \"布\",\n \"夯\",\n \"戊\",\n \"龙\",\n \"平\",\n \"灭\",\n \"轧\",\n \"东\",\n \"卡\",\n \"北\",\n \"占\",\n \"凸\",\n \"卢\",\n \"业\",\n \"旧\",\n \"帅\",\n \"归\",\n \"旦\",\n \"目\",\n \"且\",\n \"叶\",\n \"甲\",\n \"申\",\n \"叮\",\n \"电\",\n \"号\",\n \"田\",\n \"由\",\n \"只\",\n \"叭\",\n \"史\",\n \"央\",\n \"兄\",\n \"叽\",\n \"叼\",\n \"叫\",\n \"叩\",\n \"叨\",\n \"另\",\n \"叹\",\n \"冉\",\n \"皿\",\n \"凹\",\n \"囚\",\n \"四\",\n \"生\",\n \"矢\",\n \"失\",\n \"乍\",\n \"禾\",\n \"丘\",\n \"付\",\n \"仗\",\n \"代\",\n \"仙\",\n \"们\",\n \"仪\",\n \"白\",\n \"仔\",\n \"他\",\n \"斥\",\n \"瓜\",\n \"乎\",\n \"丛\",\n \"令\",\n \"用\",\n \"甩\",\n \"印\",\n \"尔\",\n \"乐\",\n \"句\",\n \"匆\",\n \"册\",\n \"卯\",\n \"犯\",\n \"外\",\n \"处\",\n \"冬\",\n \"鸟\",\n \"务\",\n \"包\",\n \"饥\",\n \"主\",\n \"市\",\n \"立\",\n \"冯\",\n \"玄\",\n \"闪\",\n \"兰\",\n \"半\",\n \"汁\",\n \"汇\",\n \"头\",\n \"汉\",\n \"宁\",\n \"穴\",\n \"它\",\n \"讨\",\n \"写\",\n \"让\",\n \"礼\",\n \"训\",\n \"议\",\n \"必\",\n \"讯\",\n \"记\",\n \"永\",\n \"司\",\n \"尼\",\n \"民\",\n \"弗\",\n \"弘\",\n \"出\",\n \"辽\",\n \"奶\",\n \"奴\",\n \"召\",\n \"加\",\n \"皮\",\n \"边\",\n \"孕\",\n \"发\",\n \"圣\",\n \"对\",\n \"台\",\n \"矛\",\n \"纠\",\n \"母\",\n \"幼\",\n \"丝\",\n \"邦\",\n \"式\",\n \"迂\",\n \"刑\",\n \"戎\",\n \"动\",\n \"扛\",\n \"寺\",\n \"吉\",\n \"扣\",\n \"考\",\n \"托\",\n \"老\",\n \"巩\",\n \"圾\",\n \"执\",\n \"扩\",\n \"扫\",\n \"地\",\n \"场\",\n \"扬\",\n \"耳\",\n \"芋\",\n \"共\",\n \"芒\",\n \"亚\",\n \"芝\",\n \"朽\",\n \"朴\",\n \"机\",\n \"权\",\n \"过\",\n \"臣\",\n \"吏\",\n \"再\",\n \"协\",\n \"西\",\n \"压\",\n \"厌\",\n \"戌\",\n \"在\",\n \"百\",\n \"有\",\n \"存\",\n \"而\",\n \"页\",\n \"匠\",\n \"夸\",\n \"夺\",\n \"灰\",\n \"达\",\n \"列\",\n \"死\",\n \"成\",\n \"夹\",\n \"夷\",\n \"轨\",\n \"邪\",\n \"尧\",\n \"划\",\n \"迈\",\n \"毕\",\n \"至\",\n \"此\",\n \"贞\",\n \"师\",\n \"尘\",\n \"尖\",\n \"劣\",\n \"光\",\n \"当\",\n \"早\",\n \"吁\",\n \"吐\",\n \"吓\",\n \"虫\",\n \"曲\",\n \"团\",\n \"吕\",\n \"同\",\n \"吊\",\n \"吃\",\n \"因\",\n \"吸\",\n \"吗\",\n \"吆\",\n \"屿\",\n \"屹\",\n \"岁\",\n \"帆\",\n \"回\",\n \"岂\",\n \"则\",\n \"刚\",\n \"网\",\n \"肉\",\n \"年\",\n \"朱\",\n \"先\",\n \"丢\",\n \"廷\",\n \"舌\",\n \"竹\",\n \"迁\",\n \"乔\",\n \"迄\",\n \"伟\",\n \"传\",\n \"乒\",\n \"乓\",\n \"休\",\n \"伍\",\n \"伏\",\n \"优\",\n \"臼\",\n \"伐\",\n \"延\",\n \"仲\",\n \"件\",\n \"任\",\n \"伤\",\n \"价\",\n \"伦\",\n \"份\",\n \"华\",\n \"仰\",\n \"仿\",\n \"伙\",\n \"伪\",\n \"自\",\n \"伊\",\n \"血\",\n \"向\",\n \"似\",\n \"后\",\n \"行\",\n \"舟\",\n \"全\",\n \"会\",\n \"杀\",\n \"合\",\n \"兆\",\n \"企\",\n \"众\",\n \"爷\",\n \"伞\",\n \"创\",\n \"肌\",\n \"肋\",\n \"朵\",\n \"杂\",\n \"危\",\n \"旬\",\n \"旨\",\n \"旭\",\n \"负\",\n \"匈\",\n \"名\",\n \"各\",\n \"多\",\n \"争\",\n \"色\",\n \"壮\",\n \"冲\",\n \"妆\",\n \"冰\",\n \"庄\",\n \"庆\",\n \"亦\",\n \"刘\",\n \"齐\",\n \"交\",\n \"衣\",\n \"次\",\n \"产\",\n \"决\",\n \"亥\",\n \"充\",\n \"妄\",\n \"闭\",\n \"问\",\n \"闯\",\n \"羊\",\n \"并\",\n \"关\",\n \"米\",\n \"灯\",\n \"州\",\n \"汗\",\n \"污\",\n \"江\",\n \"汛\",\n \"池\",\n \"汝\",\n \"汤\",\n \"忙\",\n \"兴\",\n \"宇\",\n \"守\",\n \"宅\",\n \"字\",\n \"安\",\n \"讲\",\n \"讳\",\n \"军\",\n \"讶\",\n \"许\",\n \"讹\",\n \"论\",\n \"讼\",\n \"农\",\n \"讽\",\n \"设\",\n \"访\",\n \"诀\",\n \"寻\",\n \"那\",\n \"迅\",\n \"尽\",\n \"导\",\n \"异\",\n \"弛\",\n \"孙\",\n \"阵\",\n \"阳\",\n \"收\",\n \"阶\",\n \"阴\",\n \"防\",\n \"奸\",\n \"如\",\n \"妇\",\n \"妃\",\n \"好\",\n \"她\",\n \"妈\",\n \"戏\",\n \"羽\",\n \"观\",\n \"欢\",\n \"买\",\n \"红\",\n \"驮\",\n \"纤\",\n \"驯\",\n \"约\",\n \"级\",\n \"纪\",\n \"驰\",\n \"纫\",\n \"巡\",\n \"寿\",\n \"弄\",\n \"麦\",\n \"玖\",\n \"玛\",\n \"形\",\n \"进\",\n \"戒\",\n \"吞\",\n \"远\",\n \"违\",\n \"韧\",\n \"运\",\n \"扶\",\n \"抚\",\n \"坛\",\n \"技\",\n \"坏\",\n \"抠\",\n \"扰\",\n \"扼\",\n \"拒\",\n \"找\",\n \"批\",\n \"址\",\n \"扯\",\n \"走\",\n \"抄\",\n \"贡\",\n \"汞\",\n \"坝\",\n \"攻\",\n \"赤\",\n \"折\",\n \"抓\",\n \"扳\",\n \"抡\",\n \"扮\",\n \"抢\",\n \"孝\",\n \"坎\",\n \"均\",\n \"抑\",\n \"抛\",\n \"投\",\n \"坟\",\n \"坑\",\n \"抗\",\n \"坊\",\n \"抖\",\n \"护\",\n \"壳\",\n \"志\",\n \"块\",\n \"扭\",\n \"声\",\n \"把\",\n \"报\",\n \"拟\",\n \"却\",\n \"抒\",\n \"劫\",\n \"芙\",\n \"芜\",\n \"苇\",\n \"芽\",\n \"花\",\n \"芹\",\n \"芥\",\n \"芬\",\n \"苍\",\n \"芳\",\n \"严\",\n \"芦\",\n \"芯\",\n \"劳\",\n \"克\",\n \"芭\",\n \"苏\",\n \"杆\",\n \"杠\",\n \"杜\",\n \"材\",\n \"村\",\n \"杖\",\n \"杏\",\n \"杉\",\n \"巫\",\n \"极\",\n \"李\",\n \"杨\",\n \"求\",\n \"甫\",\n \"匣\",\n \"更\",\n \"束\",\n \"吾\",\n \"豆\",\n \"两\",\n \"酉\",\n \"丽\",\n \"医\",\n \"辰\",\n \"励\",\n \"否\",\n \"还\",\n \"尬\",\n \"歼\",\n \"来\",\n \"连\",\n \"轩\",\n \"步\",\n \"卤\",\n \"坚\",\n \"肖\",\n \"旱\",\n \"盯\",\n \"呈\",\n \"时\",\n \"吴\",\n \"助\",\n \"县\",\n \"里\",\n \"呆\",\n \"吱\",\n \"吠\",\n \"呕\",\n \"园\",\n \"旷\",\n \"围\",\n \"呀\",\n \"吨\",\n \"足\",\n \"邮\",\n \"男\",\n \"困\",\n \"吵\",\n \"串\",\n \"员\",\n \"呐\",\n \"听\",\n \"吟\",\n \"吩\",\n \"呛\",\n \"吻\",\n \"吹\",\n \"呜\",\n \"吭\",\n \"吧\",\n \"邑\",\n \"吼\",\n \"囤\",\n \"别\",\n \"吮\",\n \"岖\",\n \"岗\",\n \"帐\",\n \"财\",\n \"针\",\n \"钉\",\n \"牡\",\n \"告\",\n \"我\",\n \"乱\",\n \"利\",\n \"秃\",\n \"秀\",\n \"私\",\n \"每\",\n \"兵\",\n \"估\",\n \"体\",\n \"何\",\n \"佐\",\n \"佑\",\n \"但\",\n \"伸\",\n \"佃\",\n \"作\",\n \"伯\",\n \"伶\",\n \"佣\",\n \"低\",\n \"你\",\n \"住\",\n \"位\",\n \"伴\",\n \"身\",\n \"皂\",\n \"伺\",\n \"佛\",\n \"囱\",\n \"近\",\n \"彻\",\n \"役\",\n \"返\",\n \"余\",\n \"希\",\n \"坐\",\n \"谷\",\n \"妥\",\n \"含\",\n \"邻\",\n \"岔\",\n \"肝\",\n \"肛\",\n \"肚\",\n \"肘\",\n \"肠\",\n \"龟\",\n \"甸\",\n \"免\",\n \"狂\",\n \"犹\",\n \"狈\",\n \"角\",\n \"删\",\n \"条\",\n \"彤\",\n \"卵\",\n \"灸\",\n \"岛\",\n \"刨\",\n \"迎\",\n \"饭\",\n \"饮\",\n \"系\",\n \"言\",\n \"冻\",\n \"状\",\n \"亩\",\n \"况\",\n \"床\",\n \"库\",\n \"庇\",\n \"疗\",\n \"吝\",\n \"应\",\n \"这\",\n \"冷\",\n \"庐\",\n \"序\",\n \"辛\",\n \"弃\",\n \"冶\",\n \"忘\",\n \"闰\",\n \"闲\",\n \"间\",\n \"闷\",\n \"判\",\n \"兑\",\n \"灶\",\n \"灿\",\n \"灼\",\n \"弟\",\n \"汪\",\n \"沐\",\n \"沛\",\n \"汰\",\n \"沥\",\n \"沙\",\n \"汽\",\n \"沃\",\n \"沦\",\n \"汹\",\n \"泛\",\n \"沧\",\n \"没\",\n \"沟\",\n \"沪\",\n \"沈\",\n \"沉\",\n \"沁\",\n \"怀\",\n \"忧\",\n \"忱\",\n \"快\",\n \"完\",\n \"宋\",\n \"宏\",\n \"牢\",\n \"究\",\n \"穷\",\n \"灾\",\n \"良\",\n \"证\",\n \"启\",\n \"评\",\n \"补\",\n \"初\",\n \"社\",\n \"祀\",\n \"识\",\n \"诈\",\n \"诉\",\n \"罕\",\n \"诊\",\n \"词\",\n \"译\",\n \"君\",\n \"灵\",\n \"即\",\n \"层\",\n \"屁\",\n \"尿\",\n \"尾\",\n \"迟\",\n \"局\",\n \"改\",\n \"张\",\n \"忌\",\n \"际\",\n \"陆\",\n \"阿\",\n \"陈\",\n \"阻\",\n \"附\",\n \"坠\",\n \"妓\",\n \"妙\",\n \"妖\",\n \"姊\",\n \"妨\",\n \"妒\",\n \"努\",\n \"忍\",\n \"劲\",\n \"矣\",\n \"鸡\",\n \"纬\",\n \"驱\",\n \"纯\",\n \"纱\",\n \"纲\",\n \"纳\",\n \"驳\",\n \"纵\",\n \"纷\",\n \"纸\",\n \"纹\",\n \"纺\",\n \"驴\",\n \"纽\",\n \"奉\",\n \"玩\",\n \"环\",\n \"武\",\n \"青\",\n \"责\",\n \"现\",\n \"玫\",\n \"表\",\n \"规\",\n \"抹\",\n \"卦\",\n \"坷\",\n \"坯\",\n \"拓\",\n \"拢\",\n \"拔\",\n \"坪\",\n \"拣\",\n \"坦\",\n \"担\",\n \"坤\",\n \"押\",\n \"抽\",\n \"拐\",\n \"拖\",\n \"者\",\n \"拍\",\n \"顶\",\n \"拆\",\n \"拎\",\n \"拥\",\n \"抵\",\n \"拘\",\n \"势\",\n \"抱\",\n \"拄\",\n \"垃\",\n \"拉\",\n \"拦\",\n \"幸\",\n \"拌\",\n \"拧\",\n \"拂\",\n \"拙\",\n \"招\",\n \"坡\",\n \"披\",\n \"拨\",\n \"择\",\n \"抬\",\n \"拇\",\n \"拗\",\n \"其\",\n \"取\",\n \"茉\",\n \"苦\",\n \"昔\",\n \"苛\",\n \"若\",\n \"茂\",\n \"苹\",\n \"苗\",\n \"英\",\n \"苟\",\n \"苑\",\n \"苞\",\n \"范\",\n \"直\",\n \"茁\",\n \"茄\",\n \"茎\",\n \"苔\",\n \"茅\",\n \"枉\",\n \"林\",\n \"枝\",\n \"杯\",\n \"枢\",\n \"柜\",\n \"枚\",\n \"析\",\n \"板\",\n \"松\",\n \"枪\",\n \"枫\",\n \"构\",\n \"杭\",\n \"杰\",\n \"述\",\n \"枕\",\n \"丧\",\n \"或\",\n \"画\",\n \"卧\",\n \"事\",\n \"刺\",\n \"枣\",\n \"雨\",\n \"卖\",\n \"郁\",\n \"矾\",\n \"矿\",\n \"码\",\n \"厕\",\n \"奈\",\n \"奔\",\n \"奇\",\n \"奋\",\n \"态\",\n \"欧\",\n \"殴\",\n \"垄\",\n \"妻\",\n \"轰\",\n \"顷\",\n \"转\",\n \"斩\",\n \"轮\",\n \"软\",\n \"到\",\n \"非\",\n \"叔\",\n \"歧\",\n \"肯\",\n \"齿\",\n \"些\",\n \"卓\",\n \"虎\",\n \"虏\",\n \"肾\",\n \"贤\",\n \"尚\",\n \"旺\",\n \"具\",\n \"味\",\n \"果\",\n \"昆\",\n \"国\",\n \"哎\",\n \"咕\",\n \"昌\",\n \"呵\",\n \"畅\",\n \"明\",\n \"易\",\n \"咙\",\n \"昂\",\n \"迪\",\n \"典\",\n \"固\",\n \"忠\",\n \"呻\",\n \"咒\",\n \"咋\",\n \"咐\",\n \"呼\",\n \"鸣\",\n \"咏\",\n \"呢\",\n \"咄\",\n \"咖\",\n \"岸\",\n \"岩\",\n \"帖\",\n \"罗\",\n \"帜\",\n \"帕\",\n \"岭\",\n \"凯\",\n \"败\",\n \"账\",\n \"贩\",\n \"贬\",\n \"购\",\n \"贮\",\n \"图\",\n \"钓\",\n \"制\",\n \"知\",\n \"迭\",\n \"氛\",\n \"垂\",\n \"牧\",\n \"物\",\n \"乖\",\n \"刮\",\n \"秆\",\n \"和\",\n \"季\",\n \"委\",\n \"秉\",\n \"佳\",\n \"侍\",\n \"岳\",\n \"供\",\n \"使\",\n \"例\",\n \"侠\",\n \"侥\",\n \"版\",\n \"侄\",\n \"侦\",\n \"侣\",\n \"侧\",\n \"凭\",\n \"侨\",\n \"佩\",\n \"货\",\n \"侈\",\n \"依\",\n \"卑\",\n \"的\",\n \"迫\",\n \"质\",\n \"欣\",\n \"征\",\n \"往\",\n \"爬\",\n \"彼\",\n \"径\",\n \"所\",\n \"舍\",\n \"金\",\n \"刹\",\n \"命\",\n \"肴\",\n \"斧\",\n \"爸\",\n \"采\",\n \"觅\",\n \"受\",\n \"乳\",\n \"贪\",\n \"念\",\n \"贫\",\n \"忿\",\n \"肤\",\n \"肺\",\n \"肢\",\n \"肿\",\n \"胀\",\n \"朋\",\n \"股\",\n \"肮\",\n \"肪\",\n \"肥\",\n \"服\",\n \"胁\",\n \"周\",\n \"昏\",\n \"鱼\",\n \"兔\",\n \"狐\",\n \"忽\",\n \"狗\",\n \"狞\",\n \"备\",\n \"饰\",\n \"饱\",\n \"饲\",\n \"变\",\n \"京\",\n \"享\",\n \"庞\",\n \"店\",\n \"夜\",\n \"庙\",\n \"府\",\n \"底\",\n \"疟\",\n \"疙\",\n \"疚\",\n \"剂\",\n \"卒\",\n \"郊\",\n \"庚\",\n \"废\",\n \"净\",\n \"盲\",\n \"放\",\n \"刻\",\n \"育\",\n \"氓\",\n \"闸\",\n \"闹\",\n \"郑\",\n \"券\",\n \"卷\",\n \"单\",\n \"炬\",\n \"炒\",\n \"炊\",\n \"炕\",\n \"炎\",\n \"炉\",\n \"沫\",\n \"浅\",\n \"法\",\n \"泄\",\n \"沽\",\n \"河\",\n \"沾\",\n \"泪\",\n \"沮\",\n \"油\",\n \"泊\",\n \"沿\",\n \"泡\",\n \"注\",\n \"泣\",\n \"泞\",\n \"泻\",\n \"泌\",\n \"泳\",\n \"泥\",\n \"沸\",\n \"沼\",\n \"波\",\n \"泼\",\n \"泽\",\n \"治\",\n \"怔\",\n \"怯\",\n \"怖\",\n \"性\",\n \"怕\",\n \"怜\",\n \"怪\",\n \"怡\",\n \"学\",\n \"宝\",\n \"宗\",\n \"定\",\n \"宠\",\n \"宜\",\n \"审\",\n \"宙\",\n \"官\",\n \"空\",\n \"帘\",\n \"宛\",\n \"实\",\n \"试\",\n \"郎\",\n \"诗\",\n \"肩\",\n \"房\",\n \"诚\",\n \"衬\",\n \"衫\",\n \"视\",\n \"祈\",\n \"话\",\n \"诞\",\n \"诡\",\n \"询\",\n \"该\",\n \"详\",\n \"建\",\n \"肃\",\n \"录\",\n \"隶\",\n \"帚\",\n \"屉\",\n \"居\",\n \"届\",\n \"刷\",\n \"屈\",\n \"弧\",\n \"弥\",\n \"弦\",\n \"承\",\n \"孟\",\n \"陋\",\n \"陌\",\n \"孤\",\n \"陕\",\n \"降\",\n \"函\",\n \"限\",\n \"妹\",\n \"姑\",\n \"姐\",\n \"姓\",\n \"妮\",\n \"始\",\n \"姆\",\n \"迢\",\n \"驾\",\n \"叁\",\n \"参\",\n \"艰\",\n \"线\",\n \"练\",\n \"组\",\n \"绅\",\n \"细\",\n \"驶\",\n \"织\",\n \"驹\",\n \"终\",\n \"驻\",\n \"绊\",\n \"驼\",\n \"绍\",\n \"绎\",\n \"经\",\n \"贯\",\n \"契\",\n \"贰\",\n \"奏\",\n \"春\",\n \"帮\",\n \"玷\",\n \"珍\",\n \"玲\",\n \"玻\",\n \"毒\",\n \"型\",\n \"拭\",\n \"挂\",\n \"封\",\n \"持\",\n \"拷\",\n \"拱\",\n \"项\",\n \"垮\",\n \"挎\",\n \"城\",\n \"挟\",\n \"挠\",\n \"政\",\n \"赴\",\n \"赵\",\n \"挡\",\n \"拽\",\n \"哉\",\n \"挺\",\n \"括\",\n \"垢\",\n \"拴\",\n \"拾\",\n \"挑\",\n \"垛\",\n \"指\",\n \"垫\",\n \"挣\",\n \"挤\",\n \"拼\",\n \"挖\",\n \"按\",\n \"挥\",\n \"挪\",\n \"拯\",\n \"某\",\n \"甚\",\n \"荆\",\n \"茸\",\n \"革\",\n \"茬\",\n \"荐\",\n \"巷\",\n \"带\",\n \"草\",\n \"茧\",\n \"茵\",\n \"茶\",\n \"荒\",\n \"茫\",\n \"荡\",\n \"荣\",\n \"荤\",\n \"荧\",\n \"故\",\n \"胡\",\n \"荫\",\n \"荔\",\n \"南\",\n \"药\",\n \"标\",\n \"栈\",\n \"柑\",\n \"枯\",\n \"柄\",\n \"栋\",\n \"相\",\n \"查\",\n \"柏\",\n \"栅\",\n \"柳\",\n \"柱\",\n \"柿\",\n \"栏\",\n \"柠\",\n \"树\",\n \"勃\",\n \"要\",\n \"柬\",\n \"咸\",\n \"威\",\n \"歪\",\n \"研\",\n \"砖\",\n \"厘\",\n \"厚\",\n \"砌\",\n \"砂\",\n \"泵\",\n \"砚\",\n \"砍\",\n \"面\",\n \"耐\",\n \"耍\",\n \"牵\",\n \"鸥\",\n \"残\",\n \"殃\",\n \"轴\",\n \"轻\",\n \"鸦\",\n \"皆\",\n \"韭\",\n \"背\",\n \"战\",\n \"点\",\n \"虐\",\n \"临\",\n \"览\",\n \"竖\",\n \"省\",\n \"削\",\n \"尝\",\n \"昧\",\n \"盹\",\n \"是\",\n \"盼\",\n \"眨\",\n \"哇\",\n \"哄\",\n \"哑\",\n \"显\",\n \"冒\",\n \"映\",\n \"星\",\n \"昨\",\n \"咧\",\n \"昭\",\n \"畏\",\n \"趴\",\n \"胃\",\n \"贵\",\n \"界\",\n \"虹\",\n \"虾\",\n \"蚁\",\n \"思\",\n \"蚂\",\n \"虽\",\n \"品\",\n \"咽\",\n \"骂\",\n \"勋\",\n \"哗\",\n \"咱\",\n \"响\",\n \"哈\",\n \"哆\",\n \"咬\",\n \"咳\",\n \"咪\",\n \"哪\",\n \"哟\",\n \"炭\",\n \"峡\",\n \"罚\",\n \"贱\",\n \"贴\",\n \"贻\",\n \"骨\",\n \"幽\",\n \"钙\",\n \"钝\",\n \"钞\",\n \"钟\",\n \"钢\",\n \"钠\",\n \"钥\",\n \"钦\",\n \"钧\",\n \"钩\",\n \"钮\",\n \"卸\",\n \"缸\",\n \"拜\",\n \"看\",\n \"矩\",\n \"毡\",\n \"氢\",\n \"怎\",\n \"牲\",\n \"选\",\n \"适\",\n \"秒\",\n \"香\",\n \"种\",\n \"秋\",\n \"科\",\n \"重\",\n \"复\",\n \"竿\",\n \"段\",\n \"便\",\n \"俩\",\n \"贷\",\n \"顺\",\n \"修\",\n \"俏\",\n \"保\",\n \"促\",\n \"俄\",\n \"俐\",\n \"侮\",\n \"俭\",\n \"俗\",\n \"俘\",\n \"信\",\n \"皇\",\n \"泉\",\n \"鬼\",\n \"侵\",\n \"禹\",\n \"侯\",\n \"追\",\n \"俊\",\n \"盾\",\n \"待\",\n \"徊\",\n \"衍\",\n \"律\",\n \"很\",\n \"须\",\n \"叙\",\n \"剑\",\n \"逃\",\n \"食\",\n \"盆\",\n \"胚\",\n \"胧\",\n \"胆\",\n \"胜\",\n \"胞\",\n \"胖\",\n \"脉\",\n \"胎\",\n \"勉\",\n \"狭\",\n \"狮\",\n \"独\",\n \"狰\",\n \"狡\",\n \"狱\",\n \"狠\",\n \"贸\",\n \"怨\",\n \"急\",\n \"饵\",\n \"饶\",\n \"蚀\",\n \"饺\",\n \"饼\",\n \"峦\",\n \"弯\",\n \"将\",\n \"奖\",\n \"哀\",\n \"亭\",\n \"亮\",\n \"度\",\n \"迹\",\n \"庭\",\n \"疮\",\n \"疯\",\n \"疫\",\n \"疤\",\n \"咨\",\n \"姿\",\n \"亲\",\n \"音\",\n \"帝\",\n \"施\",\n \"闺\",\n \"闻\",\n \"闽\",\n \"阀\",\n \"阁\",\n \"差\",\n \"养\",\n \"美\",\n \"姜\",\n \"叛\",\n \"送\",\n \"类\",\n \"迷\",\n \"籽\",\n \"娄\",\n \"前\",\n \"首\",\n \"逆\",\n \"兹\",\n \"总\",\n \"炼\",\n \"炸\",\n \"烁\",\n \"炮\",\n \"炫\",\n \"烂\",\n \"剃\",\n \"洼\",\n \"洁\",\n \"洪\",\n \"洒\",\n \"柒\",\n \"浇\",\n \"浊\",\n \"洞\",\n \"测\",\n \"洗\",\n \"活\",\n \"派\",\n \"洽\",\n \"染\",\n \"洛\",\n \"浏\",\n \"济\",\n \"洋\",\n \"洲\",\n \"浑\",\n \"浓\",\n \"津\",\n \"恃\",\n \"恒\",\n \"恢\",\n \"恍\",\n \"恬\",\n \"恤\",\n \"恰\",\n \"恼\",\n \"恨\",\n \"举\",\n \"觉\",\n \"宣\",\n \"宦\",\n \"室\",\n \"宫\",\n \"宪\",\n \"突\",\n \"穿\",\n \"窃\",\n \"客\",\n \"诫\",\n \"冠\",\n \"诬\",\n \"语\",\n \"扁\",\n \"袄\",\n \"祖\",\n \"神\",\n \"祝\",\n \"祠\",\n \"误\",\n \"诱\",\n \"诲\",\n \"说\",\n \"诵\",\n \"垦\",\n \"退\",\n \"既\",\n \"屋\",\n \"昼\",\n \"屏\",\n \"屎\",\n \"费\",\n \"陡\",\n \"逊\",\n \"眉\",\n \"孩\",\n \"陨\",\n \"除\",\n \"险\",\n \"院\",\n \"娃\",\n \"姥\",\n \"姨\",\n \"姻\",\n \"娇\",\n \"姚\",\n \"娜\",\n \"怒\",\n \"架\",\n \"贺\",\n \"盈\",\n \"勇\",\n \"怠\",\n \"癸\",\n \"蚤\",\n \"柔\",\n \"垒\",\n \"绑\",\n \"绒\",\n \"结\",\n \"绕\",\n \"骄\",\n \"绘\",\n \"给\",\n \"绚\",\n \"骆\",\n \"络\",\n \"绝\",\n \"绞\",\n \"骇\",\n \"统\",\n \"耕\",\n \"耘\",\n \"耗\",\n \"耙\",\n \"艳\",\n \"泰\",\n \"秦\",\n \"珠\",\n \"班\",\n \"素\",\n \"匿\",\n \"蚕\",\n \"顽\",\n \"盏\",\n \"匪\",\n \"捞\",\n \"栽\",\n \"捕\",\n \"埂\",\n \"捂\",\n \"振\",\n \"载\",\n \"赶\",\n \"起\",\n \"盐\",\n \"捎\",\n \"捍\",\n \"捏\",\n \"埋\",\n \"捉\",\n \"捆\",\n \"捐\",\n \"损\",\n \"袁\",\n \"捌\",\n \"都\",\n \"哲\",\n \"逝\",\n \"捡\",\n \"挫\",\n \"换\",\n \"挽\",\n \"挚\",\n \"热\",\n \"恐\",\n \"捣\",\n \"壶\",\n \"捅\",\n \"埃\",\n \"挨\",\n \"耻\",\n \"耿\",\n \"耽\",\n \"聂\",\n \"恭\",\n \"莽\",\n \"莱\",\n \"莲\",\n \"莫\",\n \"莉\",\n \"荷\",\n \"获\",\n \"晋\",\n \"恶\",\n \"莹\",\n \"莺\",\n \"真\",\n \"框\",\n \"梆\",\n \"桂\",\n \"桔\",\n \"栖\",\n \"档\",\n \"桐\",\n \"株\",\n \"桥\",\n \"桦\",\n \"栓\",\n \"桃\",\n \"格\",\n \"桩\",\n \"校\",\n \"核\",\n \"样\",\n \"根\",\n \"索\",\n \"哥\",\n \"速\",\n \"逗\",\n \"栗\",\n \"贾\",\n \"酌\",\n \"配\",\n \"翅\",\n \"辱\",\n \"唇\",\n \"夏\",\n \"砸\",\n \"砰\",\n \"砾\",\n \"础\",\n \"破\",\n \"原\",\n \"套\",\n \"逐\",\n \"烈\",\n \"殊\",\n \"殉\",\n \"顾\",\n \"轿\",\n \"较\",\n \"顿\",\n \"毙\",\n \"致\",\n \"柴\",\n \"桌\",\n \"虑\",\n \"监\",\n \"紧\",\n \"党\",\n \"逞\",\n \"晒\",\n \"眠\",\n \"晓\",\n \"哮\",\n \"唠\",\n \"鸭\",\n \"晃\",\n \"哺\",\n \"晌\",\n \"剔\",\n \"晕\",\n \"蚌\",\n \"畔\",\n \"蚣\",\n \"蚊\",\n \"蚪\",\n \"蚓\",\n \"哨\",\n \"哩\",\n \"圃\",\n \"哭\",\n \"哦\",\n \"恩\",\n \"鸯\",\n \"唤\",\n \"唁\",\n \"哼\",\n \"唧\",\n \"啊\",\n \"唉\",\n \"唆\",\n \"罢\",\n \"峭\",\n \"峨\",\n \"峰\",\n \"圆\",\n \"峻\",\n \"贼\",\n \"贿\",\n \"赂\",\n \"赃\",\n \"钱\",\n \"钳\",\n \"钻\",\n \"钾\",\n \"铁\",\n \"铃\",\n \"铅\",\n \"缺\",\n \"氧\",\n \"氨\",\n \"特\",\n \"牺\",\n \"造\",\n \"乘\",\n \"敌\",\n \"秤\",\n \"租\",\n \"积\",\n \"秧\",\n \"秩\",\n \"称\",\n \"秘\",\n \"透\",\n \"笔\",\n \"笑\",\n \"笋\",\n \"债\",\n \"借\",\n \"值\",\n \"倚\",\n \"俺\",\n \"倾\",\n \"倒\",\n \"倘\",\n \"俱\",\n \"倡\",\n \"候\",\n \"赁\",\n \"俯\",\n \"倍\",\n \"倦\",\n \"健\",\n \"臭\",\n \"射\",\n \"躬\",\n \"息\",\n \"倔\",\n \"徒\",\n \"徐\",\n \"殷\",\n \"舰\",\n \"舱\",\n \"般\",\n \"航\",\n \"途\",\n \"拿\",\n \"耸\",\n \"爹\",\n \"舀\",\n \"爱\",\n \"豺\",\n \"豹\",\n \"颁\",\n \"颂\",\n \"翁\",\n \"胰\",\n \"脆\",\n \"脂\",\n \"胸\",\n \"胳\",\n \"脏\",\n \"脐\",\n \"胶\",\n \"脑\",\n \"脓\",\n \"逛\",\n \"狸\",\n \"狼\",\n \"卿\",\n \"逢\",\n \"鸵\",\n \"留\",\n \"鸳\",\n \"皱\",\n \"饿\",\n \"馁\",\n \"凌\",\n \"凄\",\n \"恋\",\n \"桨\",\n \"浆\",\n \"衰\",\n \"衷\",\n \"高\",\n \"郭\",\n \"席\",\n \"准\",\n \"座\",\n \"症\",\n \"病\",\n \"疾\",\n \"斋\",\n \"疹\",\n \"疼\",\n \"疲\",\n \"脊\",\n \"效\",\n \"离\",\n \"紊\",\n \"唐\",\n \"瓷\",\n \"资\",\n \"凉\",\n \"站\",\n \"剖\",\n \"竞\",\n \"部\",\n \"旁\",\n \"旅\",\n \"畜\",\n \"阅\",\n \"羞\",\n \"羔\",\n \"瓶\",\n \"拳\",\n \"粉\",\n \"料\",\n \"益\",\n \"兼\",\n \"烤\",\n \"烘\",\n \"烦\",\n \"烧\",\n \"烛\",\n \"烟\",\n \"烙\",\n \"递\",\n \"涛\",\n \"浙\",\n \"涝\",\n \"浦\",\n \"酒\",\n \"涉\",\n \"消\",\n \"涡\",\n \"浩\",\n \"海\",\n \"涂\",\n \"浴\",\n \"浮\",\n \"涣\",\n \"涤\",\n \"流\",\n \"润\",\n \"涧\",\n \"涕\",\n \"浪\",\n \"浸\",\n \"涨\",\n \"烫\",\n \"涩\",\n \"涌\",\n \"悖\",\n \"悟\",\n \"悄\",\n \"悍\",\n \"悔\",\n \"悯\",\n \"悦\",\n \"害\",\n \"宽\",\n \"家\",\n \"宵\",\n \"宴\",\n \"宾\",\n \"窍\",\n \"窄\",\n \"容\",\n \"宰\",\n \"案\",\n \"请\",\n \"朗\",\n \"诸\",\n \"诺\",\n \"读\",\n \"扇\",\n \"诽\",\n \"袜\",\n \"袖\",\n \"袍\",\n \"被\",\n \"祥\",\n \"课\",\n \"冥\",\n \"谁\",\n \"调\",\n \"冤\",\n \"谅\",\n \"谆\",\n \"谈\",\n \"谊\",\n \"剥\",\n \"恳\",\n \"展\",\n \"剧\",\n \"屑\",\n \"弱\",\n \"陵\",\n \"祟\",\n \"陶\",\n \"陷\",\n \"陪\",\n \"娱\",\n \"娟\",\n \"恕\",\n \"娥\",\n \"娘\",\n \"通\",\n \"能\",\n \"难\",\n \"预\",\n \"桑\",\n \"绢\",\n \"绣\",\n \"验\",\n \"继\",\n \"骏\",\n \"球\",\n \"琐\",\n \"理\",\n \"琉\",\n \"琅\",\n \"捧\",\n \"堵\",\n \"措\",\n \"描\",\n \"域\",\n \"捺\",\n \"掩\",\n \"捷\",\n \"排\",\n \"焉\",\n \"掉\",\n \"捶\",\n \"赦\",\n \"堆\",\n \"推\",\n \"埠\",\n \"掀\",\n \"授\",\n \"捻\",\n \"教\",\n \"掏\",\n \"掐\",\n \"掠\",\n \"掂\",\n \"培\",\n \"接\",\n \"掷\",\n \"控\",\n \"探\",\n \"据\",\n \"掘\",\n \"掺\",\n \"职\",\n \"基\",\n \"聆\",\n \"勘\",\n \"聊\",\n \"娶\",\n \"著\",\n \"菱\",\n \"勒\",\n \"黄\",\n \"菲\",\n \"萌\",\n \"萝\",\n \"菌\",\n \"萎\",\n \"菜\",\n \"萄\",\n \"菊\",\n \"菩\",\n \"萍\",\n \"菠\",\n \"萤\",\n \"营\",\n \"乾\",\n \"萧\",\n \"萨\",\n \"菇\",\n \"械\",\n \"彬\",\n \"梦\",\n \"婪\",\n \"梗\",\n \"梧\",\n \"梢\",\n \"梅\",\n \"检\",\n \"梳\",\n \"梯\",\n \"桶\",\n \"梭\",\n \"救\",\n \"曹\",\n \"副\",\n \"票\",\n \"酝\",\n \"酗\",\n \"厢\",\n \"戚\",\n \"硅\",\n \"硕\",\n \"奢\",\n \"盔\",\n \"爽\",\n \"聋\",\n \"袭\",\n \"盛\",\n \"匾\",\n \"雪\",\n \"辅\",\n \"辆\",\n \"颅\",\n \"虚\",\n \"彪\",\n \"雀\",\n \"堂\",\n \"常\",\n \"眶\",\n \"匙\",\n \"晨\",\n \"睁\",\n \"眯\",\n \"眼\",\n \"悬\",\n \"野\",\n \"啪\",\n \"啦\",\n \"曼\",\n \"晦\",\n \"晚\",\n \"啄\",\n \"啡\",\n \"距\",\n \"趾\",\n \"啃\",\n \"跃\",\n \"略\",\n \"蚯\",\n \"蛀\",\n \"蛇\",\n \"唬\",\n \"累\",\n \"鄂\",\n \"唱\",\n \"患\",\n \"啰\",\n \"唾\",\n \"唯\",\n \"啤\",\n \"啥\",\n \"啸\",\n \"崖\",\n \"崎\",\n \"崭\",\n \"逻\",\n \"崔\",\n \"帷\",\n \"崩\",\n \"崇\",\n \"崛\",\n \"婴\",\n \"圈\",\n \"铐\",\n \"铛\",\n \"铝\",\n \"铜\",\n \"铭\",\n \"铲\",\n \"银\",\n \"矫\",\n \"甜\",\n \"秸\",\n \"梨\",\n \"犁\",\n \"秽\",\n \"移\",\n \"笨\",\n \"笼\",\n \"笛\",\n \"笙\",\n \"符\",\n \"第\",\n \"敏\",\n \"做\",\n \"袋\",\n \"悠\",\n \"偿\",\n \"偶\",\n \"偎\",\n \"偷\",\n \"您\",\n \"售\",\n \"停\",\n \"偏\",\n \"躯\",\n \"兜\",\n \"假\",\n \"衅\",\n \"徘\",\n \"徙\",\n \"得\",\n \"衔\",\n \"盘\",\n \"舶\",\n \"船\",\n \"舵\",\n \"斜\",\n \"盒\",\n \"鸽\",\n \"敛\",\n \"悉\",\n \"欲\",\n \"彩\",\n \"领\",\n \"脚\",\n \"脖\",\n \"脯\",\n \"豚\",\n \"脸\",\n \"脱\",\n \"象\",\n \"够\",\n \"逸\",\n \"猜\",\n \"猪\",\n \"猎\",\n \"猫\",\n \"凰\",\n \"猖\",\n \"猛\",\n \"祭\",\n \"馅\",\n \"馆\",\n \"凑\",\n \"减\",\n \"毫\",\n \"烹\",\n \"庶\",\n \"麻\",\n \"庵\",\n \"痊\",\n \"痒\",\n \"痕\",\n \"廊\",\n \"康\",\n \"庸\",\n \"鹿\",\n \"盗\",\n \"章\",\n \"竟\",\n \"商\",\n \"族\",\n \"旋\",\n \"望\",\n \"率\",\n \"阎\",\n \"阐\",\n \"着\",\n \"羚\",\n \"盖\",\n \"眷\",\n \"粘\",\n \"粗\",\n \"粒\",\n \"断\",\n \"剪\",\n \"兽\",\n \"焊\",\n \"焕\",\n \"清\",\n \"添\",\n \"鸿\",\n \"淋\",\n \"涯\",\n \"淹\",\n \"渠\",\n \"渐\",\n \"淑\",\n \"淌\",\n \"混\",\n \"淮\",\n \"淆\",\n \"渊\",\n \"淫\",\n \"渔\",\n \"淘\",\n \"淳\",\n \"液\",\n \"淤\",\n \"淡\",\n \"淀\",\n \"深\",\n \"涮\",\n \"涵\",\n \"婆\",\n \"梁\",\n \"渗\",\n \"情\",\n \"惜\",\n \"惭\",\n \"悼\",\n \"惧\",\n \"惕\",\n \"惟\",\n \"惊\",\n \"惦\",\n \"悴\",\n \"惋\",\n \"惨\",\n \"惯\",\n \"寇\",\n \"寅\",\n \"寄\",\n \"寂\",\n \"宿\",\n \"窒\",\n \"窑\",\n \"密\",\n \"谋\",\n \"谍\",\n \"谎\",\n \"谐\",\n \"袱\",\n \"祷\",\n \"祸\",\n \"谓\",\n \"谚\",\n \"谜\",\n \"逮\",\n \"敢\",\n \"尉\",\n \"屠\",\n \"弹\",\n \"隋\",\n \"堕\",\n \"随\",\n \"蛋\",\n \"隅\",\n \"隆\",\n \"隐\",\n \"婚\",\n \"婶\",\n \"婉\",\n \"颇\",\n \"颈\",\n \"绩\",\n \"绪\",\n \"续\",\n \"骑\",\n \"绰\",\n \"绳\",\n \"维\",\n \"绵\",\n \"绷\",\n \"绸\",\n \"综\",\n \"绽\",\n \"绿\",\n \"缀\",\n \"巢\",\n \"琴\",\n \"琳\",\n \"琢\",\n \"琼\",\n \"斑\",\n \"替\",\n \"揍\",\n \"款\",\n \"堪\",\n \"塔\",\n \"搭\",\n \"堰\",\n \"揩\",\n \"越\",\n \"趁\",\n \"趋\",\n \"超\",\n \"揽\",\n \"堤\",\n \"提\",\n \"博\",\n \"揭\",\n \"喜\",\n \"彭\",\n \"揣\",\n \"插\",\n \"揪\",\n \"搜\",\n \"煮\",\n \"援\",\n \"搀\",\n \"裁\",\n \"搁\",\n \"搓\",\n \"搂\",\n \"搅\",\n \"壹\",\n \"握\",\n \"搔\",\n \"揉\",\n \"斯\",\n \"期\",\n \"欺\",\n \"联\",\n \"葫\",\n \"散\",\n \"惹\",\n \"葬\",\n \"募\",\n \"葛\",\n \"董\",\n \"葡\",\n \"敬\",\n \"葱\",\n \"蒋\",\n \"蒂\",\n \"落\",\n \"韩\",\n \"朝\",\n \"辜\",\n \"葵\",\n \"棒\",\n \"棱\",\n \"棋\",\n \"椰\",\n \"植\",\n \"森\",\n \"焚\",\n \"椅\",\n \"椒\",\n \"棵\",\n \"棍\",\n \"椎\",\n \"棉\",\n \"棚\",\n \"棕\",\n \"棺\",\n \"榔\",\n \"椭\",\n \"惠\",\n \"惑\",\n \"逼\",\n \"粟\",\n \"棘\",\n \"酣\",\n \"酥\",\n \"厨\",\n \"厦\",\n \"硬\",\n \"硝\",\n \"确\",\n \"硫\",\n \"雁\",\n \"殖\",\n \"裂\",\n \"雄\",\n \"颊\",\n \"雳\",\n \"暂\",\n \"雅\",\n \"翘\",\n \"辈\",\n \"悲\",\n \"紫\",\n \"凿\",\n \"辉\",\n \"敞\",\n \"棠\",\n \"赏\",\n \"掌\",\n \"晴\",\n \"睐\",\n \"暑\",\n \"最\",\n \"晰\",\n \"量\",\n \"鼎\",\n \"喷\",\n \"喳\",\n \"晶\",\n \"喇\",\n \"遇\",\n \"喊\",\n \"遏\",\n \"晾\",\n \"景\",\n \"畴\",\n \"践\",\n \"跋\",\n \"跌\",\n \"跑\",\n \"跛\",\n \"遗\",\n \"蛙\",\n \"蛛\",\n \"蜓\",\n \"蜒\",\n \"蛤\",\n \"喝\",\n \"鹃\",\n \"喂\",\n \"喘\",\n \"喉\",\n \"喻\",\n \"啼\",\n \"喧\",\n \"嵌\",\n \"幅\",\n \"帽\",\n \"赋\",\n \"赌\",\n \"赎\",\n \"赐\",\n \"赔\",\n \"黑\",\n \"铸\",\n \"铺\",\n \"链\",\n \"销\",\n \"锁\",\n \"锄\",\n \"锅\",\n \"锈\",\n \"锋\",\n \"锌\",\n \"锐\",\n \"甥\",\n \"掰\",\n \"短\",\n \"智\",\n \"氮\",\n \"毯\",\n \"氯\",\n \"鹅\",\n \"剩\",\n \"稍\",\n \"程\",\n \"稀\",\n \"税\",\n \"筐\",\n \"等\",\n \"筑\",\n \"策\",\n \"筛\",\n \"筒\",\n \"筏\",\n \"答\",\n \"筋\",\n \"筝\",\n \"傲\",\n \"傅\",\n \"牌\",\n \"堡\",\n \"集\",\n \"焦\",\n \"傍\",\n \"储\",\n \"皓\",\n \"皖\",\n \"粤\",\n \"奥\",\n \"街\",\n \"惩\",\n \"御\",\n \"循\",\n \"艇\",\n \"舒\",\n \"逾\",\n \"番\",\n \"释\",\n \"禽\",\n \"腊\",\n \"脾\",\n \"腋\",\n \"腔\",\n \"腕\",\n \"鲁\",\n \"猩\",\n \"猬\",\n \"猾\",\n \"猴\",\n \"惫\",\n \"然\",\n \"馈\",\n \"馋\",\n \"装\",\n \"蛮\",\n \"就\",\n \"敦\",\n \"斌\",\n \"痘\",\n \"痢\",\n \"痪\",\n \"痛\",\n \"童\",\n \"竣\",\n \"阔\",\n \"善\",\n \"翔\",\n \"羡\",\n \"普\",\n \"粪\",\n \"尊\",\n \"奠\",\n \"道\",\n \"遂\",\n \"曾\",\n \"焰\",\n \"港\",\n \"滞\",\n \"湖\",\n \"湘\",\n \"渣\",\n \"渤\",\n \"渺\",\n \"湿\",\n \"温\",\n \"渴\",\n \"溃\",\n \"溅\",\n \"滑\",\n \"湃\",\n \"渝\",\n \"湾\",\n \"渡\",\n \"游\",\n \"滋\",\n \"渲\",\n \"溉\",\n \"愤\",\n \"慌\",\n \"惰\",\n \"愕\",\n \"愣\",\n \"惶\",\n \"愧\",\n \"愉\",\n \"慨\",\n \"割\",\n \"寒\",\n \"富\",\n \"寓\",\n \"窜\",\n \"窝\",\n \"窖\",\n \"窗\",\n \"窘\",\n \"遍\",\n \"雇\",\n \"裕\",\n \"裤\",\n \"裙\",\n \"禅\",\n \"禄\",\n \"谢\",\n \"谣\",\n \"谤\",\n \"谦\",\n \"犀\",\n \"属\",\n \"屡\",\n \"强\",\n \"粥\",\n \"疏\",\n \"隔\",\n \"隙\",\n \"隘\",\n \"媒\",\n \"絮\",\n \"嫂\",\n \"媚\",\n \"婿\",\n \"登\",\n \"缅\",\n \"缆\",\n \"缉\",\n \"缎\",\n \"缓\",\n \"缔\",\n \"缕\",\n \"骗\",\n \"编\",\n \"骚\",\n \"缘\",\n \"瑟\",\n \"鹉\",\n \"瑞\",\n \"瑰\",\n \"瑙\",\n \"魂\",\n \"肆\",\n \"摄\",\n \"摸\",\n \"填\",\n \"搏\",\n \"塌\",\n \"鼓\",\n \"摆\",\n \"携\",\n \"搬\",\n \"摇\",\n \"搞\",\n \"塘\",\n \"摊\",\n \"聘\",\n \"斟\",\n \"蒜\",\n \"勤\",\n \"靴\",\n \"靶\",\n \"鹊\",\n \"蓝\",\n \"墓\",\n \"幕\",\n \"蓬\",\n \"蓄\",\n \"蒲\",\n \"蓉\",\n \"蒙\",\n \"蒸\",\n \"献\",\n \"椿\",\n \"禁\",\n \"楚\",\n \"楷\",\n \"榄\",\n \"想\",\n \"槐\",\n \"榆\",\n \"楼\",\n \"概\",\n \"赖\",\n \"酪\",\n \"酬\",\n \"感\",\n \"碍\",\n \"碘\",\n \"碑\",\n \"碎\",\n \"碰\",\n \"碗\",\n \"碌\",\n \"尴\",\n \"雷\",\n \"零\",\n \"雾\",\n \"雹\",\n \"辐\",\n \"辑\",\n \"输\",\n \"督\",\n \"频\",\n \"龄\",\n \"鉴\",\n \"睛\",\n \"睹\",\n \"睦\",\n \"瞄\",\n \"睫\",\n \"睡\",\n \"睬\",\n \"嗜\",\n \"鄙\",\n \"嗦\",\n \"愚\",\n \"暖\",\n \"盟\",\n \"歇\",\n \"暗\",\n \"暇\",\n \"照\",\n \"畸\",\n \"跨\",\n \"跷\",\n \"跳\",\n \"跺\",\n \"跪\",\n \"路\",\n \"跤\",\n \"跟\",\n \"遣\",\n \"蜈\",\n \"蜗\",\n \"蛾\",\n \"蜂\",\n \"蜕\",\n \"嗅\",\n \"嗡\",\n \"嗓\",\n \"署\",\n \"置\",\n \"罪\",\n \"罩\",\n \"蜀\",\n \"幌\",\n \"错\",\n \"锚\",\n \"锡\",\n \"锣\",\n \"锤\",\n \"锥\",\n \"锦\",\n \"键\",\n \"锯\",\n \"锰\",\n \"矮\",\n \"辞\",\n \"稚\",\n \"稠\",\n \"颓\",\n \"愁\",\n \"筹\",\n \"签\",\n \"简\",\n \"筷\",\n \"毁\",\n \"舅\",\n \"鼠\",\n \"催\",\n \"傻\",\n \"像\",\n \"躲\",\n \"魁\",\n \"衙\",\n \"微\",\n \"愈\",\n \"遥\",\n \"腻\",\n \"腰\",\n \"腥\",\n \"腮\",\n \"腹\",\n \"腺\",\n \"鹏\",\n \"腾\",\n \"腿\",\n \"鲍\",\n \"猿\",\n \"颖\",\n \"触\",\n \"解\",\n \"煞\",\n \"雏\",\n \"馍\",\n \"馏\",\n \"酱\",\n \"禀\",\n \"痹\",\n \"廓\",\n \"痴\",\n \"痰\",\n \"廉\",\n \"靖\",\n \"新\",\n \"韵\",\n \"意\",\n \"誊\",\n \"粮\",\n \"数\",\n \"煎\",\n \"塑\",\n \"慈\",\n \"煤\",\n \"煌\",\n \"满\",\n \"漠\",\n \"滇\",\n \"源\",\n \"滤\",\n \"滥\",\n \"滔\",\n \"溪\",\n \"溜\",\n \"漓\",\n \"滚\",\n \"溢\",\n \"溯\",\n \"滨\",\n \"溶\",\n \"溺\",\n \"粱\",\n \"滩\",\n \"慎\",\n \"誉\",\n \"塞\",\n \"寞\",\n \"窥\",\n \"窟\",\n \"寝\",\n \"谨\",\n \"褂\",\n \"裸\",\n \"福\",\n \"谬\",\n \"群\",\n \"殿\",\n \"辟\",\n \"障\",\n \"媳\",\n \"嫉\",\n \"嫌\",\n \"嫁\",\n \"叠\",\n \"缚\",\n \"缝\",\n \"缠\",\n \"缤\",\n \"剿\",\n \"静\",\n \"碧\",\n \"璃\",\n \"赘\",\n \"熬\",\n \"墙\",\n \"墟\",\n \"嘉\",\n \"摧\",\n \"赫\",\n \"截\",\n \"誓\",\n \"境\",\n \"摘\",\n \"摔\",\n \"撇\",\n \"聚\",\n \"慕\",\n \"暮\",\n \"摹\",\n \"蔓\",\n \"蔑\",\n \"蔡\",\n \"蔗\",\n \"蔽\",\n \"蔼\",\n \"熙\",\n \"蔚\",\n \"兢\",\n \"模\",\n \"槛\",\n \"榴\",\n \"榜\",\n \"榨\",\n \"榕\",\n \"歌\",\n \"遭\",\n \"酵\",\n \"酷\",\n \"酿\",\n \"酸\",\n \"碟\",\n \"碱\",\n \"碳\",\n \"磁\",\n \"愿\",\n \"需\",\n \"辖\",\n \"辗\",\n \"雌\",\n \"裳\",\n \"颗\",\n \"瞅\",\n \"墅\",\n \"嗽\",\n \"踊\",\n \"蜻\",\n \"蜡\",\n \"蝇\",\n \"蜘\",\n \"蝉\",\n \"嘛\",\n \"嘀\",\n \"赚\",\n \"锹\",\n \"锻\",\n \"镀\",\n \"舞\",\n \"舔\",\n \"稳\",\n \"熏\",\n \"箕\",\n \"算\",\n \"箩\",\n \"管\",\n \"箫\",\n \"舆\",\n \"僚\",\n \"僧\",\n \"鼻\",\n \"魄\",\n \"魅\",\n \"貌\",\n \"膜\",\n \"膊\",\n \"膀\",\n \"鲜\",\n \"疑\",\n \"孵\",\n \"馒\",\n \"裹\",\n \"敲\",\n \"豪\",\n \"膏\",\n \"遮\",\n \"腐\",\n \"瘩\",\n \"瘟\",\n \"瘦\",\n \"辣\",\n \"彰\",\n \"竭\",\n \"端\",\n \"旗\",\n \"精\",\n \"粹\",\n \"歉\",\n \"弊\",\n \"熄\",\n \"熔\",\n \"煽\",\n \"潇\",\n \"漆\",\n \"漱\",\n \"漂\",\n \"漫\",\n \"滴\",\n \"漾\",\n \"演\",\n \"漏\",\n \"慢\",\n \"慷\",\n \"寨\",\n \"赛\",\n \"寡\",\n \"察\",\n \"蜜\",\n \"寥\",\n \"谭\",\n \"肇\",\n \"褐\",\n \"褪\",\n \"谱\",\n \"隧\",\n \"嫩\",\n \"翠\",\n \"熊\",\n \"凳\",\n \"骡\",\n \"缩\",\n \"慧\",\n \"撵\",\n \"撕\",\n \"撒\",\n \"撩\",\n \"趣\",\n \"趟\",\n \"撑\",\n \"撮\",\n \"撬\",\n \"播\",\n \"擒\",\n \"墩\",\n \"撞\",\n \"撤\",\n \"增\",\n \"撰\",\n \"聪\",\n \"鞋\",\n \"鞍\",\n \"蕉\",\n \"蕊\",\n \"蔬\",\n \"蕴\",\n \"横\",\n \"槽\",\n \"樱\",\n \"橡\",\n \"樟\",\n \"橄\",\n \"敷\",\n \"豌\",\n \"飘\",\n \"醋\",\n \"醇\",\n \"醉\",\n \"磕\",\n \"磊\",\n \"磅\",\n \"碾\",\n \"震\",\n \"霄\",\n \"霉\",\n \"瞒\",\n \"题\",\n \"暴\",\n \"瞎\",\n \"嘻\",\n \"嘶\",\n \"嘲\",\n \"嘹\",\n \"影\",\n \"踢\",\n \"踏\",\n \"踩\",\n \"踪\",\n \"蝶\",\n \"蝴\",\n \"蝠\",\n \"蝎\",\n \"蝌\",\n \"蝗\",\n \"蝙\",\n \"嘿\",\n \"嘱\",\n \"幢\",\n \"墨\",\n \"镇\",\n \"镐\",\n \"镑\",\n \"靠\",\n \"稽\",\n \"稻\",\n \"黎\",\n \"稿\",\n \"稼\",\n \"箱\",\n \"篓\",\n \"箭\",\n \"篇\",\n \"僵\",\n \"躺\",\n \"僻\",\n \"德\",\n \"艘\",\n \"膝\",\n \"膛\",\n \"鲤\",\n \"鲫\",\n \"熟\",\n \"摩\",\n \"褒\",\n \"瘪\",\n \"瘤\",\n \"瘫\",\n \"凛\",\n \"颜\",\n \"毅\",\n \"糊\",\n \"遵\",\n \"憋\",\n \"潜\",\n \"澎\",\n \"潮\",\n \"潭\",\n \"鲨\",\n \"澳\",\n \"潘\",\n \"澈\",\n \"澜\",\n \"澄\",\n \"懂\",\n \"憔\",\n \"懊\",\n \"憎\",\n \"额\",\n \"翩\",\n \"褥\",\n \"谴\",\n \"鹤\",\n \"憨\",\n \"慰\",\n \"劈\",\n \"履\",\n \"豫\",\n \"缭\",\n \"撼\",\n \"擂\",\n \"操\",\n \"擅\",\n \"燕\",\n \"蕾\",\n \"薯\",\n \"薛\",\n \"薇\",\n \"擎\",\n \"薪\",\n \"薄\",\n \"颠\",\n \"翰\",\n \"噩\",\n \"橱\",\n \"橙\",\n \"橘\",\n \"整\",\n \"融\",\n \"瓢\",\n \"醒\",\n \"霍\",\n \"霎\",\n \"辙\",\n \"冀\",\n \"餐\",\n \"嘴\",\n \"踱\",\n \"蹄\",\n \"蹂\",\n \"蟆\",\n \"螃\",\n \"器\",\n \"噪\",\n \"鹦\",\n \"赠\",\n \"默\",\n \"黔\",\n \"镜\",\n \"赞\",\n \"穆\",\n \"篮\",\n \"篡\",\n \"篷\",\n \"篱\",\n \"儒\",\n \"邀\",\n \"衡\",\n \"膨\",\n \"雕\",\n \"鲸\",\n \"磨\",\n \"瘾\",\n \"瘸\",\n \"凝\",\n \"辨\",\n \"辩\",\n \"糙\",\n \"糖\",\n \"糕\",\n \"燃\",\n \"濒\",\n \"澡\",\n \"激\",\n \"懒\",\n \"憾\",\n \"懈\",\n \"窿\",\n \"壁\",\n \"避\",\n \"缰\",\n \"缴\",\n \"戴\",\n \"擦\",\n \"藉\",\n \"鞠\",\n \"藏\",\n \"藐\",\n \"檬\",\n \"檐\",\n \"檀\",\n \"礁\",\n \"磷\",\n \"霜\",\n \"霞\",\n \"瞭\",\n \"瞧\",\n \"瞬\",\n \"瞳\",\n \"瞩\",\n \"瞪\",\n \"曙\",\n \"蹋\",\n \"蹈\",\n \"螺\",\n \"蟋\",\n \"蟀\",\n \"嚎\",\n \"赡\",\n \"穗\",\n \"魏\",\n \"簧\",\n \"簇\",\n \"繁\",\n \"徽\",\n \"爵\",\n \"朦\",\n \"臊\",\n \"鳄\",\n \"癌\",\n \"辫\",\n \"赢\",\n \"糟\",\n \"糠\",\n \"燥\",\n \"懦\",\n \"豁\",\n \"臀\",\n \"臂\",\n \"翼\",\n \"骤\",\n \"藕\",\n \"鞭\",\n \"藤\",\n \"覆\",\n \"瞻\",\n \"蹦\",\n \"嚣\",\n \"镰\",\n \"翻\",\n \"鳍\",\n \"鹰\",\n \"瀑\",\n \"襟\",\n \"璧\",\n \"戳\",\n \"孽\",\n \"警\",\n \"蘑\",\n \"藻\",\n \"攀\",\n \"曝\",\n \"蹲\",\n \"蹭\",\n \"蹬\",\n \"巅\",\n \"簸\",\n \"簿\",\n \"蟹\",\n \"颤\",\n \"靡\",\n \"癣\",\n \"瓣\",\n \"羹\",\n \"鳖\",\n \"爆\",\n \"疆\",\n \"鬓\",\n \"壤\",\n \"馨\",\n \"耀\",\n \"躁\",\n \"蠕\",\n \"嚼\",\n \"嚷\",\n \"巍\",\n \"籍\",\n \"鳞\",\n \"魔\",\n \"糯\",\n \"灌\",\n \"譬\",\n \"蠢\",\n \"霸\",\n \"露\",\n \"霹\",\n \"躏\",\n \"黯\",\n \"髓\",\n \"赣\",\n \"囊\",\n \"镶\",\n \"瓤\",\n \"罐\",\n \"矗\",\n \"乂\",\n \"乜\",\n \"兀\",\n \"弋\",\n \"孑\",\n \"孓\",\n \"幺\",\n \"亓\",\n \"韦\",\n \"廿\",\n \"丏\",\n \"卅\",\n \"仄\",\n \"厄\",\n \"仃\",\n \"仉\",\n \"仂\",\n \"兮\",\n \"刈\",\n \"爻\",\n \"卞\",\n \"闩\",\n \"讣\",\n \"尹\",\n \"夬\",\n \"爿\",\n \"毋\",\n \"邗\",\n \"邛\",\n \"艽\",\n \"艿\",\n \"札\",\n \"叵\",\n \"匝\",\n \"丕\",\n \"匜\",\n \"劢\",\n \"卟\",\n \"叱\",\n \"叻\",\n \"仨\",\n \"仕\",\n \"仟\",\n \"仡\",\n \"仫\",\n \"仞\",\n \"卮\",\n \"氐\",\n \"犰\",\n \"刍\",\n \"邝\",\n \"邙\",\n \"汀\",\n \"讦\",\n \"讧\",\n \"讪\",\n \"讫\",\n \"尻\",\n \"阡\",\n \"尕\",\n \"弁\",\n \"驭\",\n \"匡\",\n \"耒\",\n \"玎\",\n \"玑\",\n \"邢\",\n \"圩\",\n \"圬\",\n \"圭\",\n \"扦\",\n \"圪\",\n \"圳\",\n \"圹\",\n \"扪\",\n \"圮\",\n \"圯\",\n \"芊\",\n \"芍\",\n \"芄\",\n \"芨\",\n \"芑\",\n \"芎\",\n \"芗\",\n \"亘\",\n \"厍\",\n \"夼\",\n \"戍\",\n \"尥\",\n \"乩\",\n \"旯\",\n \"曳\",\n \"岌\",\n \"屺\",\n \"凼\",\n \"囡\",\n \"钇\",\n \"缶\",\n \"氘\",\n \"氖\",\n \"牝\",\n \"伎\",\n \"伛\",\n \"伢\",\n \"佤\",\n \"仵\",\n \"伥\",\n \"伧\",\n \"伉\",\n \"伫\",\n \"囟\",\n \"汆\",\n \"刖\",\n \"夙\",\n \"旮\",\n \"刎\",\n \"犷\",\n \"犸\",\n \"舛\",\n \"凫\",\n \"邬\",\n \"饧\",\n \"汕\",\n \"汔\",\n \"汐\",\n \"汲\",\n \"汜\",\n \"汊\",\n \"忖\",\n \"忏\",\n \"讴\",\n \"讵\",\n \"祁\",\n \"讷\",\n \"聿\",\n \"艮\",\n \"厾\",\n \"阱\",\n \"阮\",\n \"阪\",\n \"丞\",\n \"妁\",\n \"牟\",\n \"纡\",\n \"纣\",\n \"纥\",\n \"纨\",\n \"玕\",\n \"玙\",\n \"抟\",\n \"抔\",\n \"圻\",\n \"坂\",\n \"坍\",\n \"坞\",\n \"抃\",\n \"抉\",\n \"㧐\",\n \"芫\",\n \"邯\",\n \"芸\",\n \"芾\",\n \"苈\",\n \"苣\",\n \"芷\",\n \"芮\",\n \"苋\",\n \"芼\",\n \"苌\",\n \"苁\",\n \"芩\",\n \"芪\",\n \"芡\",\n \"芟\",\n \"苄\",\n \"苎\",\n \"苡\",\n \"杌\",\n \"杓\",\n \"杞\",\n \"杈\",\n \"忑\",\n \"孛\",\n \"邴\",\n \"邳\",\n \"矶\",\n \"奁\",\n \"豕\",\n \"忒\",\n \"欤\",\n \"轫\",\n \"迓\",\n \"邶\",\n \"忐\",\n \"卣\",\n \"邺\",\n \"旰\",\n \"呋\",\n \"呒\",\n \"呓\",\n \"呔\",\n \"呖\",\n \"呃\",\n \"旸\",\n \"吡\",\n \"町\",\n \"虬\",\n \"呗\",\n \"吽\",\n \"吣\",\n \"吲\",\n \"帏\",\n \"岐\",\n \"岈\",\n \"岘\",\n \"岑\",\n \"岚\",\n \"兕\",\n \"囵\",\n \"囫\",\n \"钊\",\n \"钋\",\n \"钌\",\n \"迕\",\n \"氙\",\n \"氚\",\n \"牤\",\n \"佞\",\n \"邱\",\n \"攸\",\n \"佚\",\n \"佝\",\n \"佟\",\n \"佗\",\n \"伽\",\n \"彷\",\n \"佘\",\n \"佥\",\n \"孚\",\n \"豸\",\n \"坌\",\n \"肟\",\n \"邸\",\n \"奂\",\n \"劬\",\n \"狄\",\n \"狁\",\n \"鸠\",\n \"邹\",\n \"饨\",\n \"饩\",\n \"饪\",\n \"饫\",\n \"饬\",\n \"亨\",\n \"庑\",\n \"庋\",\n \"疔\",\n \"疖\",\n \"肓\",\n \"闱\",\n \"闳\",\n \"闵\",\n \"羌\",\n \"炀\",\n \"沣\",\n \"沅\",\n \"沔\",\n \"沤\",\n \"沌\",\n \"沏\",\n \"沚\",\n \"汩\",\n \"汨\",\n \"沂\",\n \"汾\",\n \"沨\",\n \"汴\",\n \"汶\",\n \"沆\",\n \"沩\",\n \"泐\",\n \"怃\",\n \"怄\",\n \"忡\",\n \"忤\",\n \"忾\",\n \"怅\",\n \"忻\",\n \"忪\",\n \"怆\",\n \"忭\",\n \"忸\",\n \"诂\",\n \"诃\",\n \"诅\",\n \"诋\",\n \"诌\",\n \"诏\",\n \"诒\",\n \"孜\",\n \"陇\",\n \"陀\",\n \"陂\",\n \"陉\",\n \"妍\",\n \"妩\",\n \"妪\",\n \"妣\",\n \"妊\",\n \"妗\",\n \"妫\",\n \"妞\",\n \"姒\",\n \"妤\",\n \"邵\",\n \"劭\",\n \"刭\",\n \"甬\",\n \"邰\",\n \"纭\",\n \"纰\",\n \"纴\",\n \"纶\",\n \"纾\",\n \"玮\",\n \"玡\",\n \"玭\",\n \"玠\",\n \"玢\",\n \"玥\",\n \"玦\",\n \"盂\",\n \"忝\",\n \"匦\",\n \"坩\",\n \"抨\",\n \"拤\",\n \"坫\",\n \"拈\",\n \"垆\",\n \"抻\",\n \"劼\",\n \"拃\",\n \"拊\",\n \"坼\",\n \"坻\",\n \"㧟\",\n \"坨\",\n \"坭\",\n \"抿\",\n \"坳\",\n \"耶\",\n \"苷\",\n \"苯\",\n \"苤\",\n \"茏\",\n \"苫\",\n \"苜\",\n \"苴\",\n \"苒\",\n \"苘\",\n \"茌\",\n \"苻\",\n \"苓\",\n \"茚\",\n \"茆\",\n \"茑\",\n \"茓\",\n \"茔\",\n \"茕\",\n \"茀\",\n \"苕\",\n \"枥\",\n \"枇\",\n \"杪\",\n \"杳\",\n \"枧\",\n \"杵\",\n \"枨\",\n \"枞\",\n \"枋\",\n \"杻\",\n \"杷\",\n \"杼\",\n \"矸\",\n \"砀\",\n \"刳\",\n \"奄\",\n \"瓯\",\n \"殁\",\n \"郏\",\n \"轭\",\n \"郅\",\n \"鸢\",\n \"盱\",\n \"昊\",\n \"昙\",\n \"杲\",\n \"昃\",\n \"咂\",\n \"呸\",\n \"昕\",\n \"昀\",\n \"旻\",\n \"昉\",\n \"炅\",\n \"咔\",\n \"畀\",\n \"虮\",\n \"咀\",\n \"呷\",\n \"黾\",\n \"呱\",\n \"呤\",\n \"咚\",\n \"咆\",\n \"咛\",\n \"呶\",\n \"呣\",\n \"呦\",\n \"咝\",\n \"岢\",\n \"岿\",\n \"岬\",\n \"岫\",\n \"帙\",\n \"岣\",\n \"峁\",\n \"刿\",\n \"迥\",\n \"岷\",\n \"剀\",\n \"帔\",\n \"峄\",\n \"沓\",\n \"囹\",\n \"罔\",\n \"钍\",\n \"钎\",\n \"钏\",\n \"钒\",\n \"钕\",\n \"钗\",\n \"邾\",\n \"迮\",\n \"牦\",\n \"竺\",\n \"迤\",\n \"佶\",\n \"佬\",\n \"佰\",\n \"侑\",\n \"侉\",\n \"臾\",\n \"岱\",\n \"侗\",\n \"侃\",\n \"侏\",\n \"侩\",\n \"佻\",\n \"佾\",\n \"侪\",\n \"佼\",\n \"佯\",\n \"侬\",\n \"帛\",\n \"阜\",\n \"侔\",\n \"徂\",\n \"刽\",\n \"郄\",\n \"怂\",\n \"籴\",\n \"瓮\",\n \"戗\",\n \"肼\",\n \"䏝\",\n \"肽\",\n \"肱\",\n \"肫\",\n \"剁\",\n \"迩\",\n \"郇\",\n \"狙\",\n \"狎\",\n \"狍\",\n \"狒\",\n \"咎\",\n \"炙\",\n \"枭\",\n \"饯\",\n \"饴\",\n \"冽\",\n \"冼\",\n \"庖\",\n \"疠\",\n \"疝\",\n \"疡\",\n \"兖\",\n \"妾\",\n \"劾\",\n \"炜\",\n \"𬉼\",\n \"炖\",\n \"炘\",\n \"炝\",\n \"炔\",\n \"泔\",\n \"沭\",\n \"泷\",\n \"泸\",\n \"泱\",\n \"泅\",\n \"泗\",\n \"泠\",\n \"泺\",\n \"泖\",\n \"泫\",\n \"泮\",\n \"沱\",\n \"泯\",\n \"泓\",\n \"泾\",\n \"怙\",\n \"怵\",\n \"怦\",\n \"怛\",\n \"怏\",\n \"怍\",\n \"㤘\",\n \"怩\",\n \"怫\",\n \"怿\",\n \"宕\",\n \"穹\",\n \"宓\",\n \"诓\",\n \"诔\",\n \"诖\",\n \"诘\",\n \"戾\",\n \"诙\",\n \"戽\",\n \"郓\",\n \"衩\",\n \"祆\",\n \"祎\",\n \"祉\",\n \"祇\",\n \"诛\",\n \"诜\",\n \"诟\",\n \"诠\",\n \"诣\",\n \"诤\",\n \"诧\",\n \"诨\",\n \"诩\",\n \"戕\",\n \"孢\",\n \"亟\",\n \"陔\",\n \"妲\",\n \"妯\",\n \"姗\",\n \"帑\",\n \"弩\",\n \"孥\",\n \"驽\",\n \"虱\",\n \"迦\",\n \"迨\",\n \"绀\",\n \"绁\",\n \"绂\",\n \"驷\",\n \"驸\",\n \"绉\",\n \"绌\",\n \"驿\",\n \"骀\",\n \"甾\",\n \"珏\",\n \"珐\",\n \"珂\",\n \"珑\",\n \"玳\",\n \"珀\",\n \"顸\",\n \"珉\",\n \"珈\",\n \"拮\",\n \"垭\",\n \"挝\",\n \"垣\",\n \"挞\",\n \"垤\",\n \"赳\",\n \"贲\",\n \"垱\",\n \"垌\",\n \"郝\",\n \"垧\",\n \"垓\",\n \"挦\",\n \"垠\",\n \"茜\",\n \"荚\",\n \"荑\",\n \"贳\",\n \"荜\",\n \"莒\",\n \"茼\",\n \"茴\",\n \"茱\",\n \"莛\",\n \"荞\",\n \"茯\",\n \"荏\",\n \"荇\",\n \"荃\",\n \"荟\",\n \"荀\",\n \"茗\",\n \"荠\",\n \"茭\",\n \"茨\",\n \"垩\",\n \"荥\",\n \"荦\",\n \"荨\",\n \"荩\",\n \"剋\",\n \"荪\",\n \"茹\",\n \"荬\",\n \"荮\",\n \"柰\",\n \"栉\",\n \"柯\",\n \"柘\",\n \"栊\",\n \"柩\",\n \"枰\",\n \"栌\",\n \"柙\",\n \"枵\",\n \"柚\",\n \"枳\",\n \"柞\",\n \"柝\",\n \"栀\",\n \"柢\",\n \"栎\",\n \"枸\",\n \"柈\",\n \"柁\",\n \"枷\",\n \"柽\",\n \"剌\",\n \"酊\",\n \"郦\",\n \"甭\",\n \"砗\",\n \"砘\",\n \"砒\",\n \"斫\",\n \"砭\",\n \"砜\",\n \"奎\",\n \"耷\",\n \"虺\",\n \"殂\",\n \"殇\",\n \"殄\",\n \"殆\",\n \"轱\",\n \"轲\",\n \"轳\",\n \"轶\",\n \"轸\",\n \"虿\",\n \"毖\",\n \"觇\",\n \"尜\",\n \"哐\",\n \"眄\",\n \"眍\",\n \"𠳐\",\n \"郢\",\n \"眇\",\n \"眊\",\n \"眈\",\n \"禺\",\n \"哂\",\n \"咴\",\n \"曷\",\n \"昴\",\n \"昱\",\n \"昵\",\n \"咦\",\n \"哓\",\n \"哔\",\n \"畎\",\n \"毗\",\n \"呲\",\n \"胄\",\n \"畋\",\n \"畈\",\n \"虼\",\n \"虻\",\n \"盅\",\n \"咣\",\n \"哕\",\n \"剐\",\n \"郧\",\n \"咻\",\n \"囿\",\n \"咿\",\n \"哌\",\n \"哙\",\n \"哚\",\n \"咯\",\n \"咩\",\n \"咤\",\n \"哝\",\n \"哏\",\n \"哞\",\n \"峙\",\n \"峣\",\n \"罘\",\n \"帧\",\n \"峒\",\n \"峤\",\n \"峋\",\n \"峥\",\n \"贶\",\n \"钚\",\n \"钛\",\n \"钡\",\n \"钣\",\n \"钤\",\n \"钨\",\n \"钫\",\n \"钯\",\n \"氡\",\n \"氟\",\n \"牯\",\n \"郜\",\n \"秕\",\n \"秭\",\n \"竽\",\n \"笈\",\n \"笃\",\n \"俦\",\n \"俨\",\n \"俅\",\n \"俪\",\n \"叟\",\n \"垡\",\n \"牮\",\n \"俣\",\n \"俚\",\n \"皈\",\n \"俑\",\n \"俟\",\n \"逅\",\n \"徇\",\n \"徉\",\n \"舢\",\n \"俞\",\n \"郗\",\n \"俎\",\n \"郤\",\n \"爰\",\n \"郛\",\n \"瓴\",\n \"胨\",\n \"胪\",\n \"胛\",\n \"胂\",\n \"胙\",\n \"胍\",\n \"胗\",\n \"胝\",\n \"朐\",\n \"胫\",\n \"鸨\",\n \"匍\",\n \"狨\",\n \"狯\",\n \"飑\",\n \"狩\",\n \"狲\",\n \"訇\",\n \"逄\",\n \"昝\",\n \"饷\",\n \"饸\",\n \"饹\",\n \"胤\",\n \"孪\",\n \"娈\",\n \"弈\",\n \"奕\",\n \"庥\",\n \"疬\",\n \"疣\",\n \"疥\",\n \"疭\",\n \"庠\",\n \"竑\",\n \"彦\",\n \"飒\",\n \"闼\",\n \"闾\",\n \"闿\",\n \"阂\",\n \"羑\",\n \"迸\",\n \"籼\",\n \"酋\",\n \"炳\",\n \"炻\",\n \"炽\",\n \"炯\",\n \"烀\",\n \"炷\",\n \"烃\",\n \"洱\",\n \"洹\",\n \"洧\",\n \"洌\",\n \"浃\",\n \"洇\",\n \"洄\",\n \"洙\",\n \"涎\",\n \"洎\",\n \"洫\",\n \"浍\",\n \"洮\",\n \"洵\",\n \"浒\",\n \"浔\",\n \"浕\",\n \"洳\",\n \"恸\",\n \"恓\",\n \"恹\",\n \"恫\",\n \"恺\",\n \"恻\",\n \"恂\",\n \"恪\",\n \"恽\",\n \"宥\",\n \"扃\",\n \"衲\",\n \"衽\",\n \"衿\",\n \"袂\",\n \"祛\",\n \"祜\",\n \"祓\",\n \"祚\",\n \"诮\",\n \"祗\",\n \"祢\",\n \"诰\",\n \"诳\",\n \"鸩\",\n \"昶\",\n \"郡\",\n \"咫\",\n \"弭\",\n \"牁\",\n \"胥\",\n \"陛\",\n \"陟\",\n \"娅\",\n \"姮\",\n \"娆\",\n \"姝\",\n \"姣\",\n \"姘\",\n \"姹\",\n \"怼\",\n \"羿\",\n \"炱\",\n \"矜\",\n \"绔\",\n \"骁\",\n \"骅\",\n \"绗\",\n \"绛\",\n \"骈\",\n \"耖\",\n \"挈\",\n \"珥\",\n \"珙\",\n \"顼\",\n \"珰\",\n \"珩\",\n \"珧\",\n \"珣\",\n \"珞\",\n \"琤\",\n \"珲\",\n \"敖\",\n \"恚\",\n \"埔\",\n \"埕\",\n \"埘\",\n \"埙\",\n \"埚\",\n \"挹\",\n \"耆\",\n \"耄\",\n \"埒\",\n \"捋\",\n \"贽\",\n \"垸\",\n \"捃\",\n \"盍\",\n \"荸\",\n \"莆\",\n \"莳\",\n \"莴\",\n \"莪\",\n \"莠\",\n \"莓\",\n \"莜\",\n \"莅\",\n \"荼\",\n \"莩\",\n \"荽\",\n \"莸\",\n \"荻\",\n \"莘\",\n \"莎\",\n \"莞\",\n \"莨\",\n \"渇\",\n \"鸪\",\n \"莼\",\n \"栲\",\n \"栳\",\n \"郴\",\n \"桓\",\n \"桡\",\n \"桎\",\n \"桢\",\n \"桤\",\n \"梃\",\n \"栝\",\n \"桕\",\n \"桁\",\n \"桧\",\n \"桅\",\n \"栟\",\n \"桉\",\n \"栩\",\n \"逑\",\n \"逋\",\n \"彧\",\n \"鬲\",\n \"豇\",\n \"酐\",\n \"逦\",\n \"厝\",\n \"孬\",\n \"砝\",\n \"砹\",\n \"砺\",\n \"砧\",\n \"砷\",\n \"砟\",\n \"砼\",\n \"砥\",\n \"砣\",\n \"剞\",\n \"砻\",\n \"轼\",\n \"轾\",\n \"辂\",\n \"鸫\",\n \"趸\",\n \"龀\",\n \"鸬\",\n \"虔\",\n \"逍\",\n \"眬\",\n \"唛\",\n \"晟\",\n \"眩\",\n \"眙\",\n \"哧\",\n \"哽\",\n \"唔\",\n \"晁\",\n \"晏\",\n \"鸮\",\n \"趵\",\n \"趿\",\n \"畛\",\n \"蚨\",\n \"蚜\",\n \"蚍\",\n \"蚋\",\n \"蚬\",\n \"蚝\",\n \"蚧\",\n \"唢\",\n \"圄\",\n \"唣\",\n \"唏\",\n \"盎\",\n \"唑\",\n \"崂\",\n \"崃\",\n \"罡\",\n \"罟\",\n \"峪\",\n \"觊\",\n \"赅\",\n \"钰\",\n \"钲\",\n \"钴\",\n \"钵\",\n \"钹\",\n \"钺\",\n \"钽\",\n \"钼\",\n \"钿\",\n \"铀\",\n \"铂\",\n \"铄\",\n \"铆\",\n \"铈\",\n \"铉\",\n \"铊\",\n \"铋\",\n \"铌\",\n \"铍\",\n \"䥽\",\n \"铎\",\n \"氩\",\n \"氤\",\n \"氦\",\n \"毪\",\n \"舐\",\n \"秣\",\n \"秫\",\n \"盉\",\n \"笄\",\n \"笕\",\n \"笊\",\n \"笏\",\n \"笆\",\n \"俸\",\n \"倩\",\n \"俵\",\n \"偌\",\n \"俳\",\n \"俶\",\n \"倬\",\n \"倏\",\n \"恁\",\n \"倭\",\n \"倪\",\n \"俾\",\n \"倜\",\n \"隼\",\n \"隽\",\n \"倌\",\n \"倥\",\n \"臬\",\n \"皋\",\n \"郫\",\n \"倨\",\n \"衄\",\n \"颀\",\n \"徕\",\n \"舫\",\n \"釜\",\n \"奚\",\n \"衾\",\n \"胯\",\n \"胱\",\n \"胴\",\n \"胭\",\n \"脍\",\n \"胼\",\n \"朕\",\n \"脒\",\n \"胺\",\n \"鸱\",\n \"玺\",\n \"鸲\",\n \"狷\",\n \"猁\",\n \"狳\",\n \"猃\",\n \"狺\",\n \"逖\",\n \"桀\",\n \"袅\",\n \"饽\",\n \"凇\",\n \"栾\",\n \"挛\",\n \"亳\",\n \"疳\",\n \"疴\",\n \"疸\",\n \"疽\",\n \"痈\",\n \"疱\",\n \"痂\",\n \"痉\",\n \"衮\",\n \"凋\",\n \"颃\",\n \"恣\",\n \"旆\",\n \"旄\",\n \"旃\",\n \"阃\",\n \"阄\",\n \"訚\",\n \"阆\",\n \"恙\",\n \"粑\",\n \"朔\",\n \"郸\",\n \"烜\",\n \"烨\",\n \"烩\",\n \"烊\",\n \"剡\",\n \"郯\",\n \"烬\",\n \"涑\",\n \"浯\",\n \"涞\",\n \"涟\",\n \"娑\",\n \"涅\",\n \"涠\",\n \"浞\",\n \"涓\",\n \"浥\",\n \"涔\",\n \"浜\",\n \"浠\",\n \"浣\",\n \"浚\",\n \"悚\",\n \"悭\",\n \"悝\",\n \"悒\",\n \"悌\",\n \"悛\",\n \"宸\",\n \"窈\",\n \"剜\",\n \"诹\",\n \"冢\",\n \"诼\",\n \"袒\",\n \"袢\",\n \"祯\",\n \"诿\",\n \"谀\",\n \"谂\",\n \"谄\",\n \"谇\",\n \"屐\",\n \"屙\",\n \"陬\",\n \"勐\",\n \"奘\",\n \"牂\",\n \"蚩\",\n \"陲\",\n \"姬\",\n \"娠\",\n \"娌\",\n \"娉\",\n \"娲\",\n \"娩\",\n \"娴\",\n \"娣\",\n \"娓\",\n \"婀\",\n \"畚\",\n \"逡\",\n \"绠\",\n \"骊\",\n \"绡\",\n \"骋\",\n \"绥\",\n \"绦\",\n \"绨\",\n \"骎\",\n \"邕\",\n \"鸶\",\n \"彗\",\n \"耜\",\n \"焘\",\n \"舂\",\n \"琏\",\n \"琇\",\n \"麸\",\n \"揶\",\n \"埴\",\n \"埯\",\n \"捯\",\n \"掳\",\n \"掴\",\n \"埸\",\n \"埵\",\n \"赧\",\n \"埤\",\n \"捭\",\n \"逵\",\n \"埝\",\n \"堋\",\n \"堍\",\n \"掬\",\n \"鸷\",\n \"掖\",\n \"捽\",\n \"掊\",\n \"堉\",\n \"掸\",\n \"捩\",\n \"掮\",\n \"悫\",\n \"埭\",\n \"埽\",\n \"掇\",\n \"掼\",\n \"聃\",\n \"菁\",\n \"萁\",\n \"菘\",\n \"堇\",\n \"萘\",\n \"萋\",\n \"菽\",\n \"菖\",\n \"萜\",\n \"萸\",\n \"萑\",\n \"棻\",\n \"菔\",\n \"菟\",\n \"萏\",\n \"萃\",\n \"菏\",\n \"菹\",\n \"菪\",\n \"菅\",\n \"菀\",\n \"萦\",\n \"菰\",\n \"菡\",\n \"梵\",\n \"梿\",\n \"梏\",\n \"觋\",\n \"桴\",\n \"桷\",\n \"梓\",\n \"棁\",\n \"桫\",\n \"棂\",\n \"啬\",\n \"郾\",\n \"匮\",\n \"敕\",\n \"豉\",\n \"鄄\",\n \"酞\",\n \"酚\",\n \"戛\",\n \"硎\",\n \"硭\",\n \"硒\",\n \"硖\",\n \"硗\",\n \"硐\",\n \"硇\",\n \"硌\",\n \"鸸\",\n \"瓠\",\n \"匏\",\n \"厩\",\n \"龚\",\n \"殒\",\n \"殓\",\n \"殍\",\n \"赉\",\n \"雩\",\n \"辄\",\n \"堑\",\n \"眭\",\n \"眦\",\n \"啧\",\n \"晡\",\n \"晤\",\n \"眺\",\n \"眵\",\n \"眸\",\n \"圊\",\n \"喏\",\n \"喵\",\n \"啉\",\n \"勖\",\n \"晞\",\n \"唵\",\n \"晗\",\n \"冕\",\n \"啭\",\n \"畦\",\n \"趺\",\n \"啮\",\n \"跄\",\n \"蚶\",\n \"蛄\",\n \"蛎\",\n \"蛆\",\n \"蚰\",\n \"蛊\",\n \"圉\",\n \"蚱\",\n \"蛉\",\n \"蛏\",\n \"蚴\",\n \"啁\",\n \"啕\",\n \"唿\",\n \"啐\",\n \"唼\",\n \"唷\",\n \"啖\",\n \"啵\",\n \"啶\",\n \"啷\",\n \"唳\",\n \"唰\",\n \"啜\",\n \"帻\",\n \"崚\",\n \"崦\",\n \"帼\",\n \"崮\",\n \"崤\",\n \"崆\",\n \"赇\",\n \"赈\",\n \"赊\",\n \"铑\",\n \"铒\",\n \"铗\",\n \"铙\",\n \"铟\",\n \"铠\",\n \"铡\",\n \"铢\",\n \"铣\",\n \"铤\",\n \"铧\",\n \"铨\",\n \"铩\",\n \"铪\",\n \"铫\",\n \"铬\",\n \"铮\",\n \"铯\",\n \"铰\",\n \"铱\",\n \"铳\",\n \"铵\",\n \"铷\",\n \"氪\",\n \"牾\",\n \"鸹\",\n \"秾\",\n \"逶\",\n \"笺\",\n \"筇\",\n \"笸\",\n \"笪\",\n \"笮\",\n \"笠\",\n \"笥\",\n \"笤\",\n \"笳\",\n \"笾\",\n \"笞\",\n \"偾\",\n \"偃\",\n \"偕\",\n \"偈\",\n \"傀\",\n \"偬\",\n \"偻\",\n \"皑\",\n \"皎\",\n \"鸻\",\n \"徜\",\n \"舸\",\n \"舻\",\n \"舴\",\n \"舷\",\n \"龛\",\n \"翎\",\n \"脬\",\n \"脘\",\n \"脲\",\n \"匐\",\n \"猗\",\n \"猡\",\n \"猞\",\n \"猝\",\n \"斛\",\n \"猕\",\n \"馗\",\n \"馃\",\n \"馄\",\n \"鸾\",\n \"孰\",\n \"庹\",\n \"庾\",\n \"痔\",\n \"痍\",\n \"疵\",\n \"翊\",\n \"旌\",\n \"旎\",\n \"袤\",\n \"阇\",\n \"阈\",\n \"阉\",\n \"阊\",\n \"阋\",\n \"阍\",\n \"阏\",\n \"羟\",\n \"粝\",\n \"粕\",\n \"敝\",\n \"焐\",\n \"烯\",\n \"焓\",\n \"烽\",\n \"焖\",\n \"烷\",\n \"焗\",\n \"渍\",\n \"渚\",\n \"淇\",\n \"淅\",\n \"淞\",\n \"渎\",\n \"涿\",\n \"淖\",\n \"挲\",\n \"淠\",\n \"涸\",\n \"渑\",\n \"淦\",\n \"淝\",\n \"淬\",\n \"涪\",\n \"淙\",\n \"涫\",\n \"渌\",\n \"淄\",\n \"惬\",\n \"悻\",\n \"悱\",\n \"惝\",\n \"惘\",\n \"悸\",\n \"惆\",\n \"惚\",\n \"惇\",\n \"惮\",\n \"窕\",\n \"谌\",\n \"谏\",\n \"扈\",\n \"皲\",\n \"谑\",\n \"裆\",\n \"袷\",\n \"裉\",\n \"谒\",\n \"谔\",\n \"谕\",\n \"谖\",\n \"谗\",\n \"谙\",\n \"谛\",\n \"谝\",\n \"逯\",\n \"郿\",\n \"隈\",\n \"粜\",\n \"隍\",\n \"隗\",\n \"婧\",\n \"婊\",\n \"婕\",\n \"娼\",\n \"婢\",\n \"婵\",\n \"胬\",\n \"袈\",\n \"翌\",\n \"恿\",\n \"欸\",\n \"绫\",\n \"骐\",\n \"绮\",\n \"绯\",\n \"绱\",\n \"骒\",\n \"绲\",\n \"骓\",\n \"绶\",\n \"绺\",\n \"绻\",\n \"绾\",\n \"骖\",\n \"缁\",\n \"耠\",\n \"琫\",\n \"琵\",\n \"琶\",\n \"琪\",\n \"瑛\",\n \"琦\",\n \"琥\",\n \"琨\",\n \"靓\",\n \"琰\",\n \"琮\",\n \"琯\",\n \"琬\",\n \"琛\",\n \"琚\",\n \"辇\",\n \"鼋\",\n \"揳\",\n \"堞\",\n \"搽\",\n \"揸\",\n \"揠\",\n \"堙\",\n \"趄\",\n \"揖\",\n \"颉\",\n \"塄\",\n \"揿\",\n \"耋\",\n \"揄\",\n \"蛩\",\n \"蛰\",\n \"塆\",\n \"摒\",\n \"揆\",\n \"掾\",\n \"聒\",\n \"葑\",\n \"葚\",\n \"靰\",\n \"靸\",\n \"葳\",\n \"葺\",\n \"葸\",\n \"萼\",\n \"葆\",\n \"葩\",\n \"葶\",\n \"蒌\",\n \"萱\",\n \"戟\",\n \"葭\",\n \"楮\",\n \"棼\",\n \"椟\",\n \"棹\",\n \"椤\",\n \"棰\",\n \"赍\",\n \"椋\",\n \"椁\",\n \"椪\",\n \"棣\",\n \"椐\",\n \"鹁\",\n \"覃\",\n \"酤\",\n \"酢\",\n \"酡\",\n \"鹂\",\n \"厥\",\n \"殚\",\n \"殛\",\n \"雯\",\n \"雱\",\n \"辊\",\n \"辋\",\n \"椠\",\n \"辍\",\n \"辎\",\n \"斐\",\n \"睄\",\n \"睑\",\n \"睇\",\n \"睃\",\n \"戢\",\n \"喋\",\n \"嗒\",\n \"喃\",\n \"喱\",\n \"喹\",\n \"晷\",\n \"喈\",\n \"跖\",\n \"跗\",\n \"跞\",\n \"跚\",\n \"跎\",\n \"跏\",\n \"跆\",\n \"蛱\",\n \"蛲\",\n \"蛭\",\n \"蛳\",\n \"蛐\",\n \"蛔\",\n \"蛞\",\n \"蛴\",\n \"蛟\",\n \"蛘\",\n \"喁\",\n \"喟\",\n \"啾\",\n \"嗖\",\n \"喑\",\n \"嗟\",\n \"喽\",\n \"嗞\",\n \"喀\",\n \"喔\",\n \"喙\",\n \"嵘\",\n \"嵖\",\n \"崴\",\n \"遄\",\n \"詈\",\n \"嵎\",\n \"崽\",\n \"嵬\",\n \"嵛\",\n \"嵯\",\n \"嵝\",\n \"嵫\",\n \"幄\",\n \"嵋\",\n \"赕\",\n \"铻\",\n \"铼\",\n \"铿\",\n \"锃\",\n \"锂\",\n \"锆\",\n \"锇\",\n \"锉\",\n \"锏\",\n \"锑\",\n \"锒\",\n \"锔\",\n \"锕\",\n \"掣\",\n \"矬\",\n \"氰\",\n \"毳\",\n \"毽\",\n \"犊\",\n \"犄\",\n \"犋\",\n \"鹄\",\n \"犍\",\n \"嵇\",\n \"黍\",\n \"稃\",\n \"稂\",\n \"筚\",\n \"筵\",\n \"筌\",\n \"傣\",\n \"傈\",\n \"舄\",\n \"牍\",\n \"傥\",\n \"傧\",\n \"遑\",\n \"傩\",\n \"遁\",\n \"徨\",\n \"媭\",\n \"畲\",\n \"弑\",\n \"颌\",\n \"翕\",\n \"釉\",\n \"鹆\",\n \"舜\",\n \"貂\",\n \"腈\",\n \"腌\",\n \"腓\",\n \"腆\",\n \"腴\",\n \"腑\",\n \"腚\",\n \"腱\",\n \"鱿\",\n \"鲀\",\n \"鲂\",\n \"颍\",\n \"猢\",\n \"猹\",\n \"猥\",\n \"飓\",\n \"觞\",\n \"觚\",\n \"猱\",\n \"颎\",\n \"飧\",\n \"馇\",\n \"馊\",\n \"亵\",\n \"脔\",\n \"裒\",\n \"痣\",\n \"痨\",\n \"痦\",\n \"痞\",\n \"痤\",\n \"痫\",\n \"痧\",\n \"赓\",\n \"竦\",\n \"瓿\",\n \"啻\",\n \"颏\",\n \"鹇\",\n \"阑\",\n \"阒\",\n \"阕\",\n \"粞\",\n \"遒\",\n \"孳\",\n \"焯\",\n \"焜\",\n \"焙\",\n \"焱\",\n \"鹈\",\n \"湛\",\n \"渫\",\n \"湮\",\n \"湎\",\n \"湜\",\n \"渭\",\n \"湍\",\n \"湫\",\n \"溲\",\n \"湟\",\n \"溆\",\n \"湲\",\n \"湔\",\n \"湉\",\n \"渥\",\n \"湄\",\n \"滁\",\n \"愠\",\n \"惺\",\n \"愦\",\n \"惴\",\n \"愀\",\n \"愎\",\n \"愔\",\n \"喾\",\n \"寐\",\n \"谟\",\n \"扉\",\n \"裢\",\n \"裎\",\n \"裥\",\n \"祾\",\n \"祺\",\n \"谠\",\n \"幂\",\n \"谡\",\n \"谥\",\n \"谧\",\n \"遐\",\n \"孱\",\n \"弼\",\n \"巽\",\n \"骘\",\n \"媪\",\n \"媛\",\n \"婷\",\n \"巯\",\n \"翚\",\n \"皴\",\n \"婺\",\n \"骛\",\n \"缂\",\n \"缃\",\n \"缄\",\n \"彘\",\n \"缇\",\n \"缈\",\n \"缌\",\n \"缑\",\n \"缒\",\n \"缗\",\n \"飨\",\n \"耢\",\n \"瑚\",\n \"瑁\",\n \"瑜\",\n \"瑗\",\n \"瑄\",\n \"瑕\",\n \"遨\",\n \"骜\",\n \"韫\",\n \"髡\",\n \"塬\",\n \"鄢\",\n \"趔\",\n \"趑\",\n \"摅\",\n \"摁\",\n \"蜇\",\n \"搋\",\n \"搪\",\n \"搐\",\n \"搛\",\n \"搠\",\n \"摈\",\n \"彀\",\n \"毂\",\n \"搦\",\n \"搡\",\n \"蓁\",\n \"戡\",\n \"蓍\",\n \"鄞\",\n \"靳\",\n \"蓐\",\n \"蓦\",\n \"鹋\",\n \"蒽\",\n \"蓓\",\n \"蓖\",\n \"蓊\",\n \"蒯\",\n \"蓟\",\n \"蓑\",\n \"蒿\",\n \"蒺\",\n \"蓠\",\n \"蒟\",\n \"蒡\",\n \"蒹\",\n \"蒴\",\n \"蒗\",\n \"蓥\",\n \"颐\",\n \"楔\",\n \"楠\",\n \"楂\",\n \"楝\",\n \"楫\",\n \"楸\",\n \"椴\",\n \"槌\",\n \"楯\",\n \"皙\",\n \"榈\",\n \"槎\",\n \"榉\",\n \"楦\",\n \"楣\",\n \"楹\",\n \"椽\",\n \"裘\",\n \"剽\",\n \"甄\",\n \"酮\",\n \"酰\",\n \"酯\",\n \"酩\",\n \"蜃\",\n \"碛\",\n \"碓\",\n \"硼\",\n \"碉\",\n \"碚\",\n \"碇\",\n \"碜\",\n \"鹌\",\n \"辏\",\n \"龃\",\n \"龅\",\n \"訾\",\n \"粲\",\n \"虞\",\n \"睚\",\n \"嗪\",\n \"韪\",\n \"嗷\",\n \"嗉\",\n \"睨\",\n \"睢\",\n \"雎\",\n \"睥\",\n \"嘟\",\n \"嗑\",\n \"嗫\",\n \"嗬\",\n \"嗔\",\n \"嗝\",\n \"戥\",\n \"嗄\",\n \"煦\",\n \"暄\",\n \"遢\",\n \"暌\",\n \"跬\",\n \"跶\",\n \"跸\",\n \"跐\",\n \"跣\",\n \"跹\",\n \"跻\",\n \"蛸\",\n \"蜊\",\n \"蜍\",\n \"蜉\",\n \"蜣\",\n \"畹\",\n \"蛹\",\n \"嗣\",\n \"嗯\",\n \"嗥\",\n \"嗲\",\n \"嗳\",\n \"嗌\",\n \"嗍\",\n \"嗨\",\n \"嗐\",\n \"嗤\",\n \"嗵\",\n \"罨\",\n \"嵊\",\n \"嵩\",\n \"嵴\",\n \"骰\",\n \"锗\",\n \"锛\",\n \"锜\",\n \"锝\",\n \"锞\",\n \"锟\",\n \"锢\",\n \"锨\",\n \"锩\",\n \"锭\",\n \"锱\",\n \"雉\",\n \"氲\",\n \"犏\",\n \"歃\",\n \"稞\",\n \"稗\",\n \"稔\",\n \"筠\",\n \"筢\",\n \"筮\",\n \"筲\",\n \"筱\",\n \"牒\",\n \"煲\",\n \"敫\",\n \"徭\",\n \"愆\",\n \"艄\",\n \"觎\",\n \"毹\",\n \"貊\",\n \"貅\",\n \"貉\",\n \"颔\",\n \"腠\",\n \"腩\",\n \"腼\",\n \"腭\",\n \"腧\",\n \"塍\",\n \"媵\",\n \"詹\",\n \"鲅\",\n \"鲆\",\n \"鲇\",\n \"鲈\",\n \"稣\",\n \"鲋\",\n \"鲐\",\n \"肄\",\n \"鹐\",\n \"飕\",\n \"觥\",\n \"遛\",\n \"馐\",\n \"鹑\",\n \"亶\",\n \"瘃\",\n \"痱\",\n \"痼\",\n \"痿\",\n \"瘐\",\n \"瘁\",\n \"瘆\",\n \"麂\",\n \"裔\",\n \"歆\",\n \"旒\",\n \"雍\",\n \"阖\",\n \"阗\",\n \"阙\",\n \"羧\",\n \"豢\",\n \"粳\",\n \"猷\",\n \"煳\",\n \"煜\",\n \"煨\",\n \"煅\",\n \"煊\",\n \"煸\",\n \"煺\",\n \"滟\",\n \"溱\",\n \"溘\",\n \"漭\",\n \"滢\",\n \"溥\",\n \"溧\",\n \"溽\",\n \"裟\",\n \"溻\",\n \"溷\",\n \"滗\",\n \"滫\",\n \"溴\",\n \"滏\",\n \"滃\",\n \"滦\",\n \"溏\",\n \"滂\",\n \"滓\",\n \"溟\",\n \"滪\",\n \"愫\",\n \"慑\",\n \"慊\",\n \"鲎\",\n \"骞\",\n \"窦\",\n \"窠\",\n \"窣\",\n \"裱\",\n \"褚\",\n \"裨\",\n \"裾\",\n \"裰\",\n \"禊\",\n \"谩\",\n \"谪\",\n \"媾\",\n \"嫫\",\n \"媲\",\n \"嫒\",\n \"嫔\",\n \"媸\",\n \"缙\",\n \"缜\",\n \"缛\",\n \"辔\",\n \"骝\",\n \"缟\",\n \"缡\",\n \"缢\",\n \"缣\",\n \"骟\",\n \"耥\",\n \"璈\",\n \"瑶\",\n \"瑭\",\n \"獒\",\n \"觏\",\n \"慝\",\n \"嫠\",\n \"韬\",\n \"叆\",\n \"髦\",\n \"摽\",\n \"墁\",\n \"撂\",\n \"摞\",\n \"撄\",\n \"翥\",\n \"踅\",\n \"摭\",\n \"墉\",\n \"墒\",\n \"榖\",\n \"綦\",\n \"蔫\",\n \"蔷\",\n \"靺\",\n \"靼\",\n \"鞅\",\n \"靿\",\n \"甍\",\n \"蔸\",\n \"蔟\",\n \"蔺\",\n \"戬\",\n \"蕖\",\n \"蔻\",\n \"蓿\",\n \"斡\",\n \"鹕\",\n \"蓼\",\n \"榛\",\n \"榧\",\n \"榻\",\n \"榫\",\n \"榭\",\n \"槔\",\n \"榱\",\n \"槁\",\n \"槟\",\n \"槠\",\n \"榷\",\n \"僰\",\n \"酽\",\n \"酶\",\n \"酹\",\n \"厮\",\n \"碡\",\n \"碴\",\n \"碣\",\n \"碲\",\n \"磋\",\n \"臧\",\n \"豨\",\n \"殡\",\n \"霆\",\n \"霁\",\n \"辕\",\n \"蜚\",\n \"裴\",\n \"翡\",\n \"龇\",\n \"龈\",\n \"睿\",\n \"䁖\",\n \"睽\",\n \"嘞\",\n \"嘈\",\n \"嘌\",\n \"嘁\",\n \"嘎\",\n \"暧\",\n \"暝\",\n \"踌\",\n \"踉\",\n \"蜞\",\n \"蜥\",\n \"蜮\",\n \"蝈\",\n \"蜴\",\n \"蜱\",\n \"蜩\",\n \"蜷\",\n \"蜿\",\n \"螂\",\n \"蜢\",\n \"嘘\",\n \"嘡\",\n \"鹗\",\n \"嘣\",\n \"嘤\",\n \"嘚\",\n \"嗾\",\n \"嘧\",\n \"罴\",\n \"罱\",\n \"幔\",\n \"嶂\",\n \"幛\",\n \"赙\",\n \"罂\",\n \"骷\",\n \"骶\",\n \"鹘\",\n \"锲\",\n \"锴\",\n \"锶\",\n \"锷\",\n \"锸\",\n \"锵\",\n \"镁\",\n \"镂\",\n \"犒\",\n \"箐\",\n \"箦\",\n \"箧\",\n \"箍\",\n \"箸\",\n \"箬\",\n \"箅\",\n \"箪\",\n \"箔\",\n \"箜\",\n \"箢\",\n \"箓\",\n \"毓\",\n \"僖\",\n \"儆\",\n \"僳\",\n \"僭\",\n \"劁\",\n \"僮\",\n \"魃\",\n \"魆\",\n \"睾\",\n \"艋\",\n \"鄱\",\n \"膈\",\n \"膑\",\n \"鲑\",\n \"鲔\",\n \"鲚\",\n \"鲛\",\n \"鲟\",\n \"獐\",\n \"觫\",\n \"雒\",\n \"夤\",\n \"馑\",\n \"銮\",\n \"塾\",\n \"麽\",\n \"瘌\",\n \"瘊\",\n \"瘘\",\n \"瘙\",\n \"廖\",\n \"韶\",\n \"旖\",\n \"膂\",\n \"阚\",\n \"鄯\",\n \"鲞\",\n \"粿\",\n \"粼\",\n \"粽\",\n \"糁\",\n \"槊\",\n \"鹚\",\n \"熘\",\n \"熥\",\n \"潢\",\n \"漕\",\n \"滹\",\n \"漯\",\n \"漶\",\n \"潋\",\n \"潴\",\n \"漪\",\n \"漉\",\n \"漳\",\n \"漩\",\n \"澉\",\n \"潍\",\n \"慵\",\n \"搴\",\n \"窨\",\n \"寤\",\n \"綮\",\n \"谮\",\n \"褡\",\n \"褙\",\n \"褓\",\n \"褛\",\n \"褊\",\n \"谯\",\n \"谰\",\n \"谲\",\n \"暨\",\n \"屣\",\n \"鹛\",\n \"嫣\",\n \"嫱\",\n \"嫖\",\n \"嫦\",\n \"嫚\",\n \"嫘\",\n \"嫡\",\n \"鼐\",\n \"翟\",\n \"瞀\",\n \"鹜\",\n \"骠\",\n \"缥\",\n \"缦\",\n \"缧\",\n \"缨\",\n \"骢\",\n \"缪\",\n \"缫\",\n \"耦\",\n \"耧\",\n \"瑾\",\n \"璜\",\n \"璀\",\n \"璎\",\n \"璁\",\n \"璋\",\n \"璇\",\n \"奭\",\n \"髯\",\n \"髫\",\n \"撷\",\n \"撅\",\n \"赭\",\n \"撸\",\n \"鋆\",\n \"撙\",\n \"撺\",\n \"墀\",\n \"聩\",\n \"觐\",\n \"鞑\",\n \"蕙\",\n \"鞒\",\n \"蕈\",\n \"蕨\",\n \"蕤\",\n \"蕞\",\n \"蕺\",\n \"瞢\",\n \"蕃\",\n \"蕲\",\n \"赜\",\n \"槿\",\n \"樯\",\n \"槭\",\n \"樗\",\n \"樘\",\n \"樊\",\n \"槲\",\n \"醌\",\n \"醅\",\n \"靥\",\n \"魇\",\n \"餍\",\n \"磔\",\n \"磙\",\n \"霈\",\n \"辘\",\n \"龉\",\n \"龊\",\n \"觑\",\n \"瞌\",\n \"瞋\",\n \"瞑\",\n \"嘭\",\n \"噎\",\n \"噶\",\n \"颙\",\n \"暹\",\n \"噘\",\n \"踔\",\n \"踝\",\n \"踟\",\n \"踒\",\n \"踬\",\n \"踮\",\n \"踯\",\n \"踺\",\n \"踞\",\n \"蝽\",\n \"蝾\",\n \"蝻\",\n \"蝰\",\n \"蝮\",\n \"螋\",\n \"蝓\",\n \"蝣\",\n \"蝼\",\n \"噗\",\n \"嘬\",\n \"颚\",\n \"噍\",\n \"噢\",\n \"噙\",\n \"噜\",\n \"噌\",\n \"噔\",\n \"颛\",\n \"幞\",\n \"幡\",\n \"嶙\",\n \"嶝\",\n \"骺\",\n \"骼\",\n \"骸\",\n \"镊\",\n \"镉\",\n \"镌\",\n \"镍\",\n \"镏\",\n \"镒\",\n \"镓\",\n \"镔\",\n \"稷\",\n \"箴\",\n \"篑\",\n \"篁\",\n \"篌\",\n \"篆\",\n \"牖\",\n \"儋\",\n \"徵\",\n \"磐\",\n \"虢\",\n \"鹞\",\n \"膘\",\n \"滕\",\n \"鲠\",\n \"鲡\",\n \"鲢\",\n \"鲣\",\n \"鲥\",\n \"鲧\",\n \"鲩\",\n \"獗\",\n \"獠\",\n \"觯\",\n \"馓\",\n \"馔\",\n \"麾\",\n \"廛\",\n \"瘛\",\n \"瘼\",\n \"瘢\",\n \"瘠\",\n \"齑\",\n \"羯\",\n \"羰\",\n \"𥻗\",\n \"遴\",\n \"糌\",\n \"糍\",\n \"糅\",\n \"熜\",\n \"熵\",\n \"熠\",\n \"澍\",\n \"澌\",\n \"潸\",\n \"潦\",\n \"潲\",\n \"鋈\",\n \"潟\",\n \"潼\",\n \"潺\",\n \"憬\",\n \"憧\",\n \"寮\",\n \"窳\",\n \"谳\",\n \"褴\",\n \"褟\",\n \"褫\",\n \"谵\",\n \"熨\",\n \"屦\",\n \"嬉\",\n \"勰\",\n \"戮\",\n \"蝥\",\n \"缬\",\n \"缮\",\n \"缯\",\n \"骣\",\n \"畿\",\n \"耩\",\n \"耨\",\n \"耪\",\n \"璞\",\n \"璟\",\n \"靛\",\n \"璠\",\n \"璘\",\n \"聱\",\n \"螯\",\n \"髻\",\n \"髭\",\n \"髹\",\n \"擀\",\n \"熹\",\n \"甏\",\n \"擞\",\n \"縠\",\n \"磬\",\n \"颞\",\n \"蕻\",\n \"鞘\",\n \"颟\",\n \"薤\",\n \"薨\",\n \"檠\",\n \"薏\",\n \"薮\",\n \"薜\",\n \"薅\",\n \"樾\",\n \"橛\",\n \"橇\",\n \"樵\",\n \"檎\",\n \"橹\",\n \"樽\",\n \"樨\",\n \"橼\",\n \"墼\",\n \"橐\",\n \"翮\",\n \"醛\",\n \"醐\",\n \"醍\",\n \"醚\",\n \"磲\",\n \"赝\",\n \"飙\",\n \"殪\",\n \"霖\",\n \"霏\",\n \"霓\",\n \"錾\",\n \"辚\",\n \"臻\",\n \"遽\",\n \"氅\",\n \"瞟\",\n \"瞠\",\n \"瞰\",\n \"嚄\",\n \"嚆\",\n \"噤\",\n \"暾\",\n \"蹀\",\n \"踹\",\n \"踵\",\n \"踽\",\n \"蹉\",\n \"蹁\",\n \"螨\",\n \"蟒\",\n \"螈\",\n \"螅\",\n \"螭\",\n \"螠\",\n \"螟\",\n \"噱\",\n \"噬\",\n \"噫\",\n \"噻\",\n \"噼\",\n \"罹\",\n \"圜\",\n \"䦃\",\n \"镖\",\n \"镗\",\n \"镘\",\n \"镚\",\n \"镛\",\n \"镝\",\n \"镞\",\n \"镠\",\n \"氇\",\n \"氆\",\n \"憩\",\n \"穑\",\n \"篝\",\n \"篥\",\n \"篦\",\n \"篪\",\n \"篙\",\n \"盥\",\n \"劓\",\n \"翱\",\n \"魉\",\n \"魈\",\n \"徼\",\n \"歙\",\n \"膳\",\n \"膦\",\n \"膙\",\n \"鲮\",\n \"鲱\",\n \"鲲\",\n \"鲳\",\n \"鲴\",\n \"鲵\",\n \"鲷\",\n \"鲻\",\n \"獴\",\n \"獭\",\n \"獬\",\n \"邂\",\n \"鹧\",\n \"廨\",\n \"赟\",\n \"瘰\",\n \"廪\",\n \"瘿\",\n \"瘵\",\n \"瘴\",\n \"癃\",\n \"瘳\",\n \"斓\",\n \"麇\",\n \"麈\",\n \"嬴\",\n \"壅\",\n \"羲\",\n \"糗\",\n \"瞥\",\n \"甑\",\n \"燎\",\n \"燠\",\n \"燔\",\n \"燧\",\n \"濑\",\n \"濉\",\n \"潞\",\n \"澧\",\n \"澹\",\n \"澥\",\n \"澶\",\n \"濂\",\n \"褰\",\n \"寰\",\n \"窸\",\n \"褶\",\n \"禧\",\n \"嬖\",\n \"犟\",\n \"隰\",\n \"嬗\",\n \"颡\",\n \"缱\",\n \"缲\",\n \"缳\",\n \"璨\",\n \"璩\",\n \"璐\",\n \"璪\",\n \"螫\",\n \"擤\",\n \"壕\",\n \"觳\",\n \"罄\",\n \"擢\",\n \"薹\",\n \"鞡\",\n \"鞬\",\n \"薷\",\n \"薰\",\n \"藓\",\n \"藁\",\n \"檄\",\n \"檩\",\n \"懋\",\n \"醢\",\n \"翳\",\n \"礅\",\n \"磴\",\n \"鹩\",\n \"龋\",\n \"龌\",\n \"豳\",\n \"壑\",\n \"黻\",\n \"嚏\",\n \"嚅\",\n \"蹑\",\n \"蹒\",\n \"蹊\",\n \"蟥\",\n \"螬\",\n \"螵\",\n \"疃\",\n \"螳\",\n \"蟑\",\n \"嚓\",\n \"羁\",\n \"罽\",\n \"罾\",\n \"嶷\",\n \"黜\",\n \"黝\",\n \"髁\",\n \"髀\",\n \"镡\",\n \"镢\",\n \"镣\",\n \"镦\",\n \"镧\",\n \"镩\",\n \"镪\",\n \"镫\",\n \"罅\",\n \"黏\",\n \"簌\",\n \"篾\",\n \"篼\",\n \"簖\",\n \"簋\",\n \"鼢\",\n \"黛\",\n \"儡\",\n \"鹪\",\n \"鼾\",\n \"皤\",\n \"魍\",\n \"龠\",\n \"繇\",\n \"貘\",\n \"邈\",\n \"貔\",\n \"臌\",\n \"膻\",\n \"臆\",\n \"臃\",\n \"鲼\",\n \"鲽\",\n \"鳀\",\n \"鳃\",\n \"鳅\",\n \"鳇\",\n \"鳊\",\n \"螽\",\n \"燮\",\n \"鹫\",\n \"襄\",\n \"糜\",\n \"縻\",\n \"膺\",\n \"癍\",\n \"麋\",\n \"懑\",\n \"濡\",\n \"濮\",\n \"濞\",\n \"濠\",\n \"濯\",\n \"蹇\",\n \"謇\",\n \"邃\",\n \"襁\",\n \"檗\",\n \"擘\",\n \"孺\",\n \"隳\",\n \"嬷\",\n \"蟊\",\n \"鹬\",\n \"鍪\",\n \"鏊\",\n \"鳌\",\n \"鬈\",\n \"鬃\",\n \"瞽\",\n \"鞯\",\n \"鞨\",\n \"鞫\",\n \"鞧\",\n \"鞣\",\n \"藜\",\n \"藠\",\n \"藩\",\n \"醪\",\n \"蹙\",\n \"礓\",\n \"燹\",\n \"餮\",\n \"瞿\",\n \"曛\",\n \"颢\",\n \"曜\",\n \"躇\",\n \"蹚\",\n \"鹭\",\n \"蟛\",\n \"蟪\",\n \"蟠\",\n \"蟮\",\n \"鹮\",\n \"黠\",\n \"黟\",\n \"髅\",\n \"髂\",\n \"镬\",\n \"镭\",\n \"镯\",\n \"馥\",\n \"簟\",\n \"簪\",\n \"鼬\",\n \"雠\",\n \"艟\",\n \"鳎\",\n \"鳏\",\n \"鳐\",\n \"癞\",\n \"癔\",\n \"癜\",\n \"癖\",\n \"糨\",\n \"蹩\",\n \"鎏\",\n \"懵\",\n \"彝\",\n \"邋\",\n \"鬏\",\n \"攉\",\n \"攒\",\n \"鞲\",\n \"鞴\",\n \"藿\",\n \"蘧\",\n \"蘅\",\n \"麓\",\n \"醮\",\n \"醯\",\n \"酃\",\n \"霪\",\n \"霭\",\n \"霨\",\n \"黼\",\n \"嚯\",\n \"蹰\",\n \"蹶\",\n \"蹽\",\n \"蹼\",\n \"蹴\",\n \"蹾\",\n \"蹿\",\n \"蠖\",\n \"蠓\",\n \"蟾\",\n \"蠊\",\n \"黢\",\n \"髋\",\n \"髌\",\n \"镲\",\n \"籀\",\n \"籁\",\n \"齁\",\n \"魑\",\n \"艨\",\n \"鳓\",\n \"鳔\",\n \"鳕\",\n \"鳗\",\n \"鳙\",\n \"麒\",\n \"鏖\",\n \"羸\",\n \"㸆\",\n \"瀚\",\n \"瀣\",\n \"瀛\",\n \"襦\",\n \"谶\",\n \"襞\",\n \"骥\",\n \"缵\",\n \"瓒\",\n \"攘\",\n \"蘩\",\n \"蘖\",\n \"醴\",\n \"霰\",\n \"酆\",\n \"矍\",\n \"曦\",\n \"躅\",\n \"鼍\",\n \"巉\",\n \"黩\",\n \"黥\",\n \"黪\",\n \"镳\",\n \"镴\",\n \"黧\",\n \"纂\",\n \"璺\",\n \"鼯\",\n \"臜\",\n \"鳜\",\n \"鳝\",\n \"鳟\",\n \"獾\",\n \"孀\",\n \"骧\",\n \"瓘\",\n \"鼙\",\n \"醺\",\n \"礴\",\n \"颦\",\n \"曩\",\n \"鳢\",\n \"癫\",\n \"麝\",\n \"夔\",\n \"爝\",\n \"灏\",\n \"禳\",\n \"鐾\",\n \"羼\",\n \"蠡\",\n \"耱\",\n \"懿\",\n \"蘸\",\n \"鹳\",\n \"霾\",\n \"氍\",\n \"饕\",\n \"躐\",\n \"髑\",\n \"镵\",\n \"穰\",\n \"饔\",\n \"鬻\",\n \"鬟\",\n \"趱\",\n \"攫\",\n \"攥\",\n \"颧\",\n \"躜\",\n \"鼹\",\n \"癯\",\n \"麟\",\n \"蠲\",\n \"蠹\",\n \"躞\",\n \"衢\",\n \"鑫\",\n \"灞\",\n \"襻\",\n \"纛\",\n \"鬣\",\n \"攮\",\n \"囔\",\n \"馕\",\n \"戆\",\n \"爨\",\n \"齉\",\n \"亍\",\n \"尢\",\n \"彳\",\n \"卬\",\n \"殳\",\n \"𠙶\",\n \"毌\",\n \"邘\",\n \"戋\",\n \"圢\",\n \"氕\",\n \"伋\",\n \"仝\",\n \"冮\",\n \"氿\",\n \"汈\",\n \"氾\",\n \"忉\",\n \"宄\",\n \"讱\",\n \"扞\",\n \"圲\",\n \"圫\",\n \"芏\",\n \"芃\",\n \"朳\",\n \"朸\",\n \"𨙸\",\n \"邨\",\n \"吒\",\n \"吖\",\n \"屼\",\n \"屾\",\n \"辿\",\n \"钆\",\n \"仳\",\n \"伣\",\n \"伈\",\n \"癿\",\n \"甪\",\n \"邠\",\n \"犴\",\n \"冱\",\n \"邡\",\n \"闫\",\n \"汋\",\n \"䜣\",\n \"讻\",\n \"孖\",\n \"纩\",\n \"玒\",\n \"玓\",\n \"玘\",\n \"玚\",\n \"刬\",\n \"坜\",\n \"坉\",\n \"扽\",\n \"坋\",\n \"扺\",\n \"㧑\",\n \"毐\",\n \"芰\",\n \"芣\",\n \"苊\",\n \"苉\",\n \"芘\",\n \"芴\",\n \"芠\",\n \"芤\",\n \"杕\",\n \"杙\",\n \"杄\",\n \"杧\",\n \"杩\",\n \"尪\",\n \"尨\",\n \"轪\",\n \"坒\",\n \"芈\",\n \"旴\",\n \"旵\",\n \"呙\",\n \"㕮\",\n \"岍\",\n \"岠\",\n \"岜\",\n \"呇\",\n \"冏\",\n \"觃\",\n \"岙\",\n \"伾\",\n \"㑇\",\n \"伭\",\n \"佖\",\n \"伲\",\n \"佁\",\n \"飏\",\n \"狃\",\n \"闶\",\n \"汧\",\n \"汫\",\n \"𣲘\",\n \"𣲗\",\n \"沄\",\n \"沘\",\n \"汭\",\n \"㳇\",\n \"沇\",\n \"忮\",\n \"忳\",\n \"忺\",\n \"祃\",\n \"诇\",\n \"邲\",\n \"诎\",\n \"诐\",\n \"屃\",\n \"岊\",\n \"阽\",\n \"䢺\",\n \"阼\",\n \"妧\",\n \"妘\",\n \"𨚕\",\n \"纮\",\n \"驲\",\n \"纻\",\n \"纼\",\n \"玤\",\n \"玞\",\n \"玱\",\n \"玟\",\n \"邽\",\n \"邿\",\n \"坥\",\n \"坰\",\n \"坬\",\n \"坽\",\n \"弆\",\n \"耵\",\n \"䢼\",\n \"𦭜\",\n \"茋\",\n \"苧\",\n \"苾\",\n \"苠\",\n \"枅\",\n \"㭎\",\n \"枘\",\n \"枍\",\n \"矼\",\n \"矻\",\n \"匼\",\n \"旿\",\n \"昇\",\n \"昄\",\n \"昒\",\n \"昈\",\n \"咉\",\n \"咇\",\n \"咍\",\n \"岵\",\n \"岽\",\n \"岨\",\n \"岞\",\n \"峂\",\n \"㟃\",\n \"囷\",\n \"钐\",\n \"钔\",\n \"钖\",\n \"牥\",\n \"佴\",\n \"垈\",\n \"侁\",\n \"侹\",\n \"佸\",\n \"佺\",\n \"隹\",\n \"㑊\",\n \"侂\",\n \"佽\",\n \"侘\",\n \"郈\",\n \"舠\",\n \"郐\",\n \"郃\",\n \"攽\",\n \"肭\",\n \"肸\",\n \"肷\",\n \"狉\",\n \"狝\",\n \"饳\",\n \"忞\",\n \"於\",\n \"炌\",\n \"炆\",\n \"泙\",\n \"沺\",\n \"泂\",\n \"泜\",\n \"泃\",\n \"泇\",\n \"怊\",\n \"峃\",\n \"穸\",\n \"祋\",\n \"祊\",\n \"鸤\",\n \"弢\",\n \"弨\",\n \"陑\",\n \"陎\",\n \"卺\",\n \"乸\",\n \"妭\",\n \"姈\",\n \"迳\",\n \"叕\",\n \"驵\",\n \"䌹\",\n \"驺\",\n \"绋\",\n \"绐\",\n \"砉\",\n \"耔\",\n \"㛃\",\n \"玶\",\n \"珇\",\n \"珅\",\n \"珋\",\n \"玹\",\n \"珌\",\n \"玿\",\n \"韨\",\n \"垚\",\n \"垯\",\n \"垙\",\n \"垲\",\n \"埏\",\n \"垍\",\n \"耇\",\n \"垎\",\n \"垴\",\n \"垟\",\n \"垞\",\n \"挓\",\n \"垵\",\n \"垏\",\n \"拶\",\n \"荖\",\n \"荁\",\n \"荙\",\n \"荛\",\n \"茈\",\n \"茽\",\n \"荄\",\n \"茺\",\n \"荓\",\n \"茳\",\n \"𦰡\",\n \"茛\",\n \"荭\",\n \"㭕\",\n \"柷\",\n \"柃\",\n \"柊\",\n \"枹\",\n \"栐\",\n \"柖\",\n \"郚\",\n \"剅\",\n \"䴓\",\n \"迺\",\n \"厖\",\n \"砆\",\n \"砑\",\n \"砄\",\n \"耏\",\n \"奓\",\n \"䶮\",\n \"轵\",\n \"轷\",\n \"轹\",\n \"轺\",\n \"昺\",\n \"昽\",\n \"盷\",\n \"咡\",\n \"咺\",\n \"昳\",\n \"昣\",\n \"哒\",\n \"昤\",\n \"昫\",\n \"昡\",\n \"咥\",\n \"昪\",\n \"虷\",\n \"虸\",\n \"哃\",\n \"峘\",\n \"耑\",\n \"峛\",\n \"峗\",\n \"峧\",\n \"帡\",\n \"钘\",\n \"钜\",\n \"钪\",\n \"钬\",\n \"钭\",\n \"矧\",\n \"秬\",\n \"俫\",\n \"舁\",\n \"俜\",\n \"俙\",\n \"俍\",\n \"垕\",\n \"衎\",\n \"舣\",\n \"弇\",\n \"侴\",\n \"鸧\",\n \"䏡\",\n \"胠\",\n \"𦙶\",\n \"胈\",\n \"胩\",\n \"胣\",\n \"朏\",\n \"飐\",\n \"訄\",\n \"饻\",\n \"庤\",\n \"疢\",\n \"炣\",\n \"炟\",\n \"㶲\",\n \"洭\",\n \"洘\",\n \"洓\",\n \"洿\",\n \"㳚\",\n \"泚\",\n \"浈\",\n \"浉\",\n \"洸\",\n \"洑\",\n \"洢\",\n \"洈\",\n \"洚\",\n \"洺\",\n \"洨\",\n \"浐\",\n \"㳘\",\n \"洴\",\n \"洣\",\n \"恔\",\n \"宬\",\n \"窀\",\n \"扂\",\n \"袆\",\n \"祏\",\n \"祐\",\n \"祕\",\n \"叚\",\n \"陧\",\n \"陞\",\n \"娀\",\n \"姞\",\n \"姱\",\n \"姤\",\n \"姶\",\n \"姽\",\n \"枲\",\n \"绖\",\n \"骃\",\n \"彖\",\n \"骉\",\n \"恝\",\n \"珪\",\n \"珛\",\n \"珹\",\n \"琊\",\n \"玼\",\n \"珖\",\n \"珽\",\n \"珦\",\n \"珫\",\n \"珒\",\n \"珢\",\n \"珕\",\n \"珝\",\n \"埗\",\n \"垾\",\n \"垺\",\n \"埆\",\n \"垿\",\n \"埌\",\n \"埇\",\n \"莰\",\n \"茝\",\n \"鄀\",\n \"莶\",\n \"莝\",\n \"䓖\",\n \"莙\",\n \"栻\",\n \"桠\",\n \"桄\",\n \"梠\",\n \"栴\",\n \"梴\",\n \"栒\",\n \"酎\",\n \"酏\",\n \"砵\",\n \"砠\",\n \"砫\",\n \"砬\",\n \"硁\",\n \"恧\",\n \"翃\",\n \"郪\",\n \"𨐈\",\n \"辀\",\n \"辁\",\n \"剕\",\n \"赀\",\n \"哢\",\n \"晅\",\n \"晊\",\n \"唝\",\n \"哳\",\n \"哱\",\n \"冔\",\n \"晔\",\n \"晐\",\n \"晖\",\n \"畖\",\n \"蚄\",\n \"蚆\",\n \"帱\",\n \"崁\",\n \"峿\",\n \"崄\",\n \"帨\",\n \"崀\",\n \"赆\",\n \"钷\",\n \"眚\",\n \"甡\",\n \"笫\",\n \"倻\",\n \"倴\",\n \"脩\",\n \"倮\",\n \"倕\",\n \"倞\",\n \"倓\",\n \"倧\",\n \"衃\",\n \"虒\",\n \"舭\",\n \"舯\",\n \"舥\",\n \"瓞\",\n \"鬯\",\n \"鸰\",\n \"脎\",\n \"朓\",\n \"胲\",\n \"虓\",\n \"鱽\",\n \"狴\",\n \"峱\",\n \"狻\",\n \"眢\",\n \"勍\",\n \"痄\",\n \"疰\",\n \"痃\",\n \"竘\",\n \"羖\",\n \"羓\",\n \"桊\",\n \"敉\",\n \"烠\",\n \"烔\",\n \"烶\",\n \"烻\",\n \"涍\",\n \"浡\",\n \"浭\",\n \"浬\",\n \"涄\",\n \"涢\",\n \"涐\",\n \"浰\",\n \"浟\",\n \"浛\",\n \"浼\",\n \"浲\",\n \"涘\",\n \"悈\",\n \"悃\",\n \"悢\",\n \"宧\",\n \"窅\",\n \"窊\",\n \"窎\",\n \"扅\",\n \"扆\",\n \"袪\",\n \"袗\",\n \"袯\",\n \"祧\",\n \"隺\",\n \"堲\",\n \"疍\",\n \"𨺙\",\n \"陴\",\n \"烝\",\n \"砮\",\n \"㛚\",\n \"哿\",\n \"翀\",\n \"翂\",\n \"剟\",\n \"绤\",\n \"骍\",\n \"䂮\",\n \"琎\",\n \"珸\",\n \"珵\",\n \"琄\",\n \"琈\",\n \"琀\",\n \"珺\",\n \"掭\",\n \"堎\",\n \"堐\",\n \"埼\",\n \"掎\",\n \"埫\",\n \"堌\",\n \"晢\",\n \"掞\",\n \"埪\",\n \"壸\",\n \"㙍\",\n \"聍\",\n \"菝\",\n \"萚\",\n \"菥\",\n \"莿\",\n \"䓫\",\n \"勚\",\n \"䓬\",\n \"萆\",\n \"菂\",\n \"菍\",\n \"菼\",\n \"萣\",\n \"䓨\",\n \"菉\",\n \"䓛\",\n \"梼\",\n \"梽\",\n \"桲\",\n \"梾\",\n \"桯\",\n \"梣\",\n \"梌\",\n \"桹\",\n \"敔\",\n \"厣\",\n \"硔\",\n \"硙\",\n \"硚\",\n \"硊\",\n \"硍\",\n \"勔\",\n \"䴕\",\n \"龁\",\n \"逴\",\n \"唪\",\n \"啫\",\n \"翈\",\n \"㫰\",\n \"晙\",\n \"畤\",\n \"趼\",\n \"跂\",\n \"蛃\",\n \"蚲\",\n \"蚺\",\n \"啴\",\n \"䎃\",\n \"崧\",\n \"崟\",\n \"崞\",\n \"崒\",\n \"崌\",\n \"崡\",\n \"铏\",\n \"铕\",\n \"铖\",\n \"铘\",\n \"铚\",\n \"铞\",\n \"铥\",\n \"铴\",\n \"牻\",\n \"牿\",\n \"稆\",\n \"笱\",\n \"笯\",\n \"偰\",\n \"偡\",\n \"鸺\",\n \"偭\",\n \"偲\",\n \"偁\",\n \"㿠\",\n \"鄅\",\n \"偓\",\n \"徛\",\n \"衒\",\n \"舳\",\n \"舲\",\n \"鸼\",\n \"悆\",\n \"鄃\",\n \"瓻\",\n \"䝙\",\n \"脶\",\n \"脞\",\n \"脟\",\n \"䏲\",\n \"鱾\",\n \"猇\",\n \"猊\",\n \"猄\",\n \"觖\",\n \"𠅤\",\n \"庱\",\n \"庼\",\n \"庳\",\n \"痓\",\n \"䴔\",\n \"竫\",\n \"堃\",\n \"阌\",\n \"羝\",\n \"羕\",\n \"焆\",\n \"烺\",\n \"焌\",\n \"淏\",\n \"淟\",\n \"淜\",\n \"淴\",\n \"淯\",\n \"湴\",\n \"涴\",\n \"㥄\",\n \"惛\",\n \"惔\",\n \"悰\",\n \"惙\",\n \"寁\",\n \"逭\",\n \"袼\",\n \"裈\",\n \"祲\",\n \"谞\",\n \"艴\",\n \"弸\",\n \"弶\",\n \"隃\",\n \"婞\",\n \"娵\",\n \"婼\",\n \"媖\",\n \"婳\",\n \"婍\",\n \"婌\",\n \"婫\",\n \"婤\",\n \"婘\",\n \"婠\",\n \"绹\",\n \"骕\",\n \"絜\",\n \"珷\",\n \"琲\",\n \"琡\",\n \"琟\",\n \"琔\",\n \"琭\",\n \"堾\",\n \"堼\",\n \"揕\",\n \"㙘\",\n \"堧\",\n \"喆\",\n \"堨\",\n \"塅\",\n \"堠\",\n \"絷\",\n \"𡎚\",\n \"葜\",\n \"惎\",\n \"萳\",\n \"葙\",\n \"靬\",\n \"葴\",\n \"蒇\",\n \"蒈\",\n \"鄚\",\n \"蒉\",\n \"蓇\",\n \"萩\",\n \"蒐\",\n \"葰\",\n \"葎\",\n \"鄑\",\n \"蒎\",\n \"葖\",\n \"蒄\",\n \"萹\",\n \"棤\",\n \"棽\",\n \"棫\",\n \"椓\",\n \"椑\",\n \"鹀\",\n \"椆\",\n \"棓\",\n \"棬\",\n \"棪\",\n \"椀\",\n \"楗\",\n \"甦\",\n \"酦\",\n \"觌\",\n \"奡\",\n \"皕\",\n \"硪\",\n \"欹\",\n \"詟\",\n \"辌\",\n \"棐\",\n \"龂\",\n \"黹\",\n \"牚\",\n \"睎\",\n \"晫\",\n \"晪\",\n \"晱\",\n \"𧿹\",\n \"蛑\",\n \"畯\",\n \"斝\",\n \"喤\",\n \"崶\",\n \"嵁\",\n \"崾\",\n \"嵅\",\n \"崿\",\n \"嵚\",\n \"翙\",\n \"圌\",\n \"圐\",\n \"赑\",\n \"淼\",\n \"赒\",\n \"铹\",\n \"铽\",\n \"𨱇\",\n \"锊\",\n \"锍\",\n \"锎\",\n \"锓\",\n \"犇\",\n \"颋\",\n \"稌\",\n \"筀\",\n \"筘\",\n \"筜\",\n \"筥\",\n \"筅\",\n \"傃\",\n \"傉\",\n \"翛\",\n \"傒\",\n \"傕\",\n \"舾\",\n \"畬\",\n \"脿\",\n \"腘\",\n \"䐃\",\n \"腙\",\n \"腒\",\n \"鲃\",\n \"猰\",\n \"猯\",\n \"㺄\",\n \"馉\",\n \"鄗\",\n \"廋\",\n \"廆\",\n \"鄌\",\n \"粢\",\n \"遆\",\n \"旐\",\n \"焞\",\n \"欻\",\n \"𣸣\",\n \"溚\",\n \"溁\",\n \"湝\",\n \"渰\",\n \"湓\",\n \"㴔\",\n \"渟\",\n \"溠\",\n \"渼\",\n \"溇\",\n \"湣\",\n \"湑\",\n \"溞\",\n \"愐\",\n \"愃\",\n \"敩\",\n \"甯\",\n \"棨\",\n \"扊\",\n \"裣\",\n \"祼\",\n \"婻\",\n \"媆\",\n \"媞\",\n \"㛹\",\n \"媓\",\n \"媂\",\n \"媄\",\n \"毵\",\n \"矞\",\n \"缊\",\n \"缐\",\n \"骙\",\n \"瑃\",\n \"瑓\",\n \"瑅\",\n \"瑆\",\n \"䴖\",\n \"瑖\",\n \"瑝\",\n \"瑔\",\n \"瑀\",\n \"𤧛\",\n \"瑳\",\n \"瑂\",\n \"嶅\",\n \"瑑\",\n \"遘\",\n \"髢\",\n \"塥\",\n \"堽\",\n \"赪\",\n \"摛\",\n \"塝\",\n \"搒\",\n \"搌\",\n \"蒱\",\n \"蒨\",\n \"蓏\",\n \"蔀\",\n \"蓢\",\n \"蓂\",\n \"蒻\",\n \"蓣\",\n \"椹\",\n \"楪\",\n \"榃\",\n \"榅\",\n \"楒\",\n \"楞\",\n \"楩\",\n \"榇\",\n \"椸\",\n \"楙\",\n \"歅\",\n \"碃\",\n \"碏\",\n \"碈\",\n \"䃅\",\n \"硿\",\n \"鄠\",\n \"辒\",\n \"龆\",\n \"觜\",\n \"䣘\",\n \"暕\",\n \"鹍\",\n \"㬊\",\n \"暅\",\n \"跱\",\n \"蜐\",\n \"蜎\",\n \"嵲\",\n \"赗\",\n \"骱\",\n \"锖\",\n \"锘\",\n \"锳\",\n \"锧\",\n \"锪\",\n \"锫\",\n \"锬\",\n \"稑\",\n \"稙\",\n \"䅟\",\n \"筻\",\n \"筼\",\n \"筶\",\n \"筦\",\n \"筤\",\n \"傺\",\n \"鹎\",\n \"僇\",\n \"艅\",\n \"艉\",\n \"谼\",\n \"貆\",\n \"腽\",\n \"腨\",\n \"腯\",\n \"鲉\",\n \"鲊\",\n \"鲌\",\n \"䲟\",\n \"鲏\",\n \"雊\",\n \"猺\",\n \"飔\",\n \"觟\",\n \"𦝼\",\n \"馌\",\n \"裛\",\n \"廒\",\n \"瘀\",\n \"瘅\",\n \"鄘\",\n \"鹒\",\n \"鄜\",\n \"麀\",\n \"鄣\",\n \"阘\",\n \"煁\",\n \"煃\",\n \"煴\",\n \"煋\",\n \"煟\",\n \"煓\",\n \"滠\",\n \"溍\",\n \"溹\",\n \"滆\",\n \"滉\",\n \"溦\",\n \"溵\",\n \"漷\",\n \"滧\",\n \"滘\",\n \"滍\",\n \"愭\",\n \"慥\",\n \"慆\",\n \"塱\",\n \"裼\",\n \"禋\",\n \"禔\",\n \"禘\",\n \"禒\",\n \"谫\",\n \"鹔\",\n \"愍\",\n \"嫄\",\n \"媱\",\n \"戤\",\n \"戣\",\n \"缞\",\n \"耤\",\n \"瑧\",\n \"瑨\",\n \"瑱\",\n \"瑷\",\n \"瑢\",\n \"斠\",\n \"摏\",\n \"墕\",\n \"墈\",\n \"墐\",\n \"墘\",\n \"摴\",\n \"銎\",\n \"𡐓\",\n \"墚\",\n \"撖\",\n \"靽\",\n \"鞁\",\n \"蔌\",\n \"蔈\",\n \"蓰\",\n \"蔹\",\n \"蔊\",\n \"嘏\",\n \"榰\",\n \"榑\",\n \"槚\",\n \"𣗋\",\n \"槜\",\n \"榍\",\n \"疐\",\n \"酺\",\n \"酾\",\n \"酲\",\n \"酴\",\n \"碶\",\n \"䃎\",\n \"碨\",\n \"𥔲\",\n \"碹\",\n \"碥\",\n \"劂\",\n \"䴗\",\n \"夥\",\n \"瞍\",\n \"鹖\",\n \"㬎\",\n \"跽\",\n \"蜾\",\n \"幖\",\n \"嶍\",\n \"圙\",\n \"𨱏\",\n \"锺\",\n \"锼\",\n \"锽\",\n \"锾\",\n \"锿\",\n \"镃\",\n \"镄\",\n \"镅\",\n \"馝\",\n \"鹙\",\n \"箨\",\n \"箖\",\n \"劄\",\n \"僬\",\n \"僦\",\n \"僔\",\n \"僎\",\n \"槃\",\n \"㙦\",\n \"鲒\",\n \"鲕\",\n \"鲖\",\n \"鲗\",\n \"鲘\",\n \"鲙\",\n \"𩽾\",\n \"夐\",\n \"獍\",\n \"飗\",\n \"凘\",\n \"廑\",\n \"廙\",\n \"瘗\",\n \"瘥\",\n \"瘕\",\n \"鲝\",\n \"鄫\",\n \"熇\",\n \"漹\",\n \"漖\",\n \"潆\",\n \"漤\",\n \"潩\",\n \"漼\",\n \"漴\",\n \"㽏\",\n \"漈\",\n \"漋\",\n \"漻\",\n \"慬\",\n \"窬\",\n \"窭\",\n \"㮾\",\n \"褕\",\n \"禛\",\n \"禚\",\n \"隩\",\n \"嫕\",\n \"嫭\",\n \"嫜\",\n \"嫪\",\n \"㻬\",\n \"麹\",\n \"璆\",\n \"漦\",\n \"叇\",\n \"墣\",\n \"墦\",\n \"墡\",\n \"劐\",\n \"薁\",\n \"蕰\",\n \"蔃\",\n \"鼒\",\n \"槱\",\n \"鹝\",\n \"磏\",\n \"磉\",\n \"殣\",\n \"慭\",\n \"霅\",\n \"暵\",\n \"暲\",\n \"暶\",\n \"踦\",\n \"踣\",\n \"䗖\",\n \"蝘\",\n \"蝲\",\n \"蝤\",\n \"噇\",\n \"噂\",\n \"噀\",\n \"罶\",\n \"嶲\",\n \"嶓\",\n \"㠇\",\n \"嶟\",\n \"嶒\",\n \"镆\",\n \"镈\",\n \"镋\",\n \"镎\",\n \"镕\",\n \"稹\",\n \"儇\",\n \"皞\",\n \"皛\",\n \"䴘\",\n \"艎\",\n \"艏\",\n \"鹟\",\n \"𩾃\",\n \"鲦\",\n \"鲪\",\n \"鲬\",\n \"橥\",\n \"觭\",\n \"鹠\",\n \"鹡\",\n \"糇\",\n \"糈\",\n \"翦\",\n \"鹢\",\n \"鹣\",\n \"熛\",\n \"潖\",\n \"潵\",\n \"㵐\",\n \"澂\",\n \"澛\",\n \"瑬\",\n \"潽\",\n \"潾\",\n \"潏\",\n \"憭\",\n \"憕\",\n \"戭\",\n \"褯\",\n \"禤\",\n \"嫽\",\n \"遹\",\n \"璥\",\n \"璲\",\n \"璒\",\n \"憙\",\n \"擐\",\n \"鄹\",\n \"薳\",\n \"鞔\",\n \"黇\",\n \"蕗\",\n \"薢\",\n \"蕹\",\n \"橞\",\n \"橑\",\n \"橦\",\n \"醑\",\n \"觱\",\n \"磡\",\n \"𥕢\",\n \"磜\",\n \"豮\",\n \"鹾\",\n \"虤\",\n \"暿\",\n \"曌\",\n \"曈\",\n \"㬚\",\n \"蹅\",\n \"踶\",\n \"䗛\",\n \"螗\",\n \"疁\",\n \"㠓\",\n \"幪\",\n \"嶦\",\n \"𨱑\",\n \"馞\",\n \"穄\",\n \"篚\",\n \"篯\",\n \"簉\",\n \"鼽\",\n \"衠\",\n \"盦\",\n \"螣\",\n \"縢\",\n \"鲭\",\n \"鲯\",\n \"鲰\",\n \"鲺\",\n \"鲹\",\n \"亸\",\n \"癀\",\n \"瘭\",\n \"羱\",\n \"糒\",\n \"燋\",\n \"熻\",\n \"燊\",\n \"燚\",\n \"燏\",\n \"濩\",\n \"濋\",\n \"澪\",\n \"澽\",\n \"澴\",\n \"澭\",\n \"澼\",\n \"憷\",\n \"憺\",\n \"懔\",\n \"黉\",\n \"嬛\",\n \"鹨\",\n \"翯\",\n \"璱\",\n \"𤩽\",\n \"璬\",\n \"璮\",\n \"髽\",\n \"擿\",\n \"薿\",\n \"薸\",\n \"檑\",\n \"櫆\",\n \"檞\",\n \"醨\",\n \"繄\",\n \"磹\",\n \"磻\",\n \"瞫\",\n \"瞵\",\n \"蹐\",\n \"蟏\",\n \"㘎\",\n \"镤\",\n \"镥\",\n \"镨\",\n \"𨱔\",\n \"矰\",\n \"穙\",\n \"穜\",\n \"穟\",\n \"簕\",\n \"簃\",\n \"簏\",\n \"儦\",\n \"魋\",\n \"斶\",\n \"艚\",\n \"谿\",\n \"䲠\",\n \"鲾\",\n \"鲿\",\n \"鳁\",\n \"鳂\",\n \"鳈\",\n \"鳉\",\n \"獯\",\n \"䗪\",\n \"馘\",\n \"襕\",\n \"襚\",\n \"螱\",\n \"甓\",\n \"嬬\",\n \"嬥\",\n \"𦈡\",\n \"瓀\",\n \"釐\",\n \"鬶\",\n \"爇\",\n \"鞳\",\n \"鞮\",\n \"藟\",\n \"藦\",\n \"藨\",\n \"鹲\",\n \"檫\",\n \"黡\",\n \"礞\",\n \"礌\",\n \"𥖨\",\n \"蹢\",\n \"蹜\",\n \"蟫\",\n \"䗴\",\n \"嚚\",\n \"髃\",\n \"镮\",\n \"镱\",\n \"酂\",\n \"馧\",\n \"簠\",\n \"簝\",\n \"簰\",\n \"鼫\",\n \"鼩\",\n \"皦\",\n \"臑\",\n \"䲢\",\n \"鳑\",\n \"鳒\",\n \"鹱\",\n \"鹯\",\n \"癗\",\n \"𦒍\",\n \"旞\",\n \"翷\",\n \"冁\",\n \"䎖\",\n \"瀔\",\n \"瀍\",\n \"瀌\",\n \"襜\",\n \"䴙\",\n \"嚭\",\n \"㰀\",\n \"鬷\",\n \"醭\",\n \"蹯\",\n \"蠋\",\n \"翾\",\n \"鳘\",\n \"儳\",\n \"儴\",\n \"鼗\",\n \"𩾌\",\n \"鳚\",\n \"鳛\",\n \"麑\",\n \"麖\",\n \"蠃\",\n \"彟\",\n \"嬿\",\n \"鬒\",\n \"蘘\",\n \"欂\",\n \"醵\",\n \"颥\",\n \"甗\",\n \"𨟠\",\n \"巇\",\n \"酅\",\n \"髎\",\n \"犨\",\n \"𨭉\",\n \"㸌\",\n \"爔\",\n \"瀱\",\n \"瀹\",\n \"瀼\",\n \"瀵\",\n \"襫\",\n \"孅\",\n \"骦\",\n \"耰\",\n \"𤫉\",\n \"瓖\",\n \"鬘\",\n \"趯\",\n \"罍\",\n \"鼱\",\n \"鳠\",\n \"鳡\",\n \"鳣\",\n \"爟\",\n \"爚\",\n \"灈\",\n \"韂\",\n \"糵\",\n \"蘼\",\n \"礵\",\n \"鹴\",\n \"躔\",\n \"皭\",\n \"龢\",\n \"鳤\",\n \"亹\",\n \"籥\",\n \"鼷\",\n \"玃\",\n \"醾\",\n \"齇\",\n \"觿\",\n \"蠼\",\n \"𬣙\",\n \"𬇕\",\n \"𬣞\",\n \"𬘓\",\n \"𫭟\",\n \"𫭢\",\n \"𫇭\",\n \"𫐄\",\n \"𫵷\",\n \"𬇙\",\n \"𬣡\",\n \"𫸩\",\n \"𫘜\",\n \"𬘘\",\n \"𫘝\",\n \"𬨂\",\n \"𬀩\",\n \"𬀪\",\n \"𬬩\",\n \"𫍣\",\n \"𬣳\",\n \"𬩽\",\n \"𬮿\",\n \"𬯀\",\n \"𫰛\",\n \"𬳵\",\n \"𬳶\",\n \"𫠊\",\n \"𬍛\",\n \"鿍\",\n \"𬜬\",\n \"𪾢\",\n \"𪨰\",\n \"𫓧\",\n \"𬬮\",\n \"𬬱\",\n \"𬬭\",\n \"𬘡\",\n \"𬳽\",\n \"𬘩\",\n \"𫄧\",\n \"𪟝\",\n \"𬍤\",\n \"𫭼\",\n \"𬜯\",\n \"𬂩\",\n \"𫠆\",\n \"𬌗\",\n \"𫑡\",\n \"𪨶\",\n \"𬬸\",\n \"𬬻\",\n \"𬬹\",\n \"𬬿\",\n \"𬭁\",\n \"𫢸\",\n \"𫗧\",\n \"𬊈\",\n \"𬒈\",\n \"𬳿\",\n \"𫄨\",\n \"𬘫\",\n \"𫮃\",\n \"鿎\",\n \"𬱖\",\n \"𬟽\",\n \"𫓯\",\n \"𫟹\",\n \"𫟼\",\n \"𬇹\",\n \"𬍡\",\n \"𬤇\",\n \"𫍯\",\n \"𬤊\",\n \"𫍲\",\n \"𬯎\",\n \"𬘬\",\n \"𬘭\",\n \"𬴂\",\n \"𫘦\",\n \"𫟅\",\n \"𬘯\",\n \"𫘧\",\n \"𪣻\",\n \"𬃊\",\n \"𬷕\",\n \"𫐐\",\n \"𬹼\",\n \"𫶇\",\n \"𫖮\",\n \"鿏\",\n \"𬭊\",\n \"𫓶\",\n \"𬭎\",\n \"𫖯\",\n \"𬱟\",\n \"𫛭\",\n \"𫷷\",\n \"𬮱\",\n \"𬊤\",\n \"𬴃\",\n \"𫘨\",\n \"𬪩\",\n \"𬒔\",\n \"𬨎\",\n \"𫐓\",\n \"𫫇\",\n \"𫓹\",\n \"𬭚\",\n \"𬭛\",\n \"𬕂\",\n \"𬶋\",\n \"𬶍\",\n \"𫔶\",\n \"𫌀\",\n \"𫖳\",\n \"𫘪\",\n \"𫘬\",\n \"𫞩\",\n \"𪤗\",\n \"𬸘\",\n \"𬒗\",\n \"𫚖\",\n \"𬭤\",\n \"𫚕\",\n \"𬶐\",\n \"𬶏\",\n \"𬸚\",\n \"𬤝\",\n \"𬙂\",\n \"𬭩\",\n \"𬸣\",\n \"𫍽\",\n \"𬴊\",\n \"𬞟\",\n \"𫟦\",\n \"𬺈\",\n \"𫠜\",\n \"𪩘\",\n \"𬭬\",\n \"𬭯\",\n \"𫗴\",\n \"𬸦\",\n \"𫄷\",\n \"𬭳\",\n \"𬭶\",\n \"𫔍\",\n \"𬭸\",\n \"𬭼\",\n \"𫔎\",\n \"𬸪\",\n \"𬶟\",\n \"𬶠\",\n \"𬶨\",\n \"𫄸\",\n \"𬟁\",\n \"𬙊\",\n \"𬶭\",\n \"𬶮\",\n \"𬙋\",\n \"𬺓\",\n \"𫚭\",\n \"廠\",\n \"蔔\",\n \"兒\",\n \"幾\",\n \"幹\",\n \"虧\",\n \"纔\",\n \"與\",\n \"萬\",\n \"韆\",\n \"億\",\n \"個\",\n \"廣\",\n \"門\",\n \"義\",\n \"衛\",\n \"飛\",\n \"習\",\n \"馬\",\n \"鄉\",\n \"豐\",\n \"開\",\n \"無\",\n \"雲\",\n \"專\",\n \"藝\",\n \"廳\",\n \"區\",\n \"歷\",\n \"曆\",\n \"車\",\n \"貝\",\n \"岡\",\n \"見\",\n \"氣\",\n \"長\",\n \"僕\",\n \"幣\",\n \"僅\",\n \"從\",\n \"侖\",\n \"倉\",\n \"風\",\n \"烏\",\n \"鳳\",\n \"爲\",\n \"鬥\",\n \"憶\",\n \"計\",\n \"訂\",\n \"認\",\n \"譏\",\n \"醜\",\n \"隊\",\n \"辦\",\n \"鄧\",\n \"勸\",\n \"雙\",\n \"書\",\n \"擊\",\n \"撲\",\n \"節\",\n \"術\",\n \"厲\",\n \"龍\",\n \"滅\",\n \"軋\",\n \"東\",\n \"盧\",\n \"業\",\n \"舊\",\n \"帥\",\n \"歸\",\n \"葉\",\n \"電\",\n \"號\",\n \"衹\",\n \"隻\",\n \"嘰\",\n \"嘆\",\n \"們\",\n \"儀\",\n \"叢\",\n \"爾\",\n \"樂\",\n \"處\",\n \"鼕\",\n \"鳥\",\n \"務\",\n \"飢\",\n \"饑\",\n \"馮\",\n \"閃\",\n \"蘭\",\n \"匯\",\n \"彙\",\n \"頭\",\n \"漢\",\n \"寧\",\n \"討\",\n \"寫\",\n \"讓\",\n \"禮\",\n \"訓\",\n \"議\",\n \"訊\",\n \"記\",\n \"齣\",\n \"遼\",\n \"邊\",\n \"發\",\n \"髮\",\n \"聖\",\n \"對\",\n \"臺\",\n \"颱\",\n \"檯\",\n \"糾\",\n \"絲\",\n \"動\",\n \"鞏\",\n \"執\",\n \"擴\",\n \"掃\",\n \"場\",\n \"揚\",\n \"亞\",\n \"樸\",\n \"機\",\n \"權\",\n \"過\",\n \"協\",\n \"壓\",\n \"厭\",\n \"頁\",\n \"誇\",\n \"奪\",\n \"達\",\n \"夾\",\n \"軌\",\n \"堯\",\n \"劃\",\n \"邁\",\n \"畢\",\n \"貞\",\n \"師\",\n \"塵\",\n \"當\",\n \"噹\",\n \"籲\",\n \"嚇\",\n \"蟲\",\n \"麯\",\n \"團\",\n \"糰\",\n \"嗎\",\n \"嶼\",\n \"歲\",\n \"迴\",\n \"豈\",\n \"則\",\n \"剛\",\n \"網\",\n \"硃\",\n \"遷\",\n \"喬\",\n \"偉\",\n \"傳\",\n \"優\",\n \"傷\",\n \"價\",\n \"倫\",\n \"華\",\n \"僞\",\n \"嚮\",\n \"後\",\n \"會\",\n \"殺\",\n \"閤\",\n \"衆\",\n \"爺\",\n \"傘\",\n \"創\",\n \"雜\",\n \"負\",\n \"壯\",\n \"衝\",\n \"妝\",\n \"莊\",\n \"慶\",\n \"劉\",\n \"齊\",\n \"産\",\n \"閉\",\n \"問\",\n \"闖\",\n \"關\",\n \"燈\",\n \"湯\",\n \"興\",\n \"講\",\n \"諱\",\n \"軍\",\n \"訝\",\n \"許\",\n \"訛\",\n \"論\",\n \"訟\",\n \"農\",\n \"諷\",\n \"設\",\n \"訪\",\n \"訣\",\n \"尋\",\n \"盡\",\n \"儘\",\n \"導\",\n \"孫\",\n \"陣\",\n \"陽\",\n \"階\",\n \"陰\",\n \"婦\",\n \"媽\",\n \"戲\",\n \"觀\",\n \"歡\",\n \"買\",\n \"紅\",\n \"馱\",\n \"纖\",\n \"縴\",\n \"馴\",\n \"約\",\n \"級\",\n \"紀\",\n \"馳\",\n \"紉\",\n \"壽\",\n \"麥\",\n \"瑪\",\n \"進\",\n \"遠\",\n \"違\",\n \"韌\",\n \"運\",\n \"撫\",\n \"壇\",\n \"罎\",\n \"壞\",\n \"摳\",\n \"擾\",\n \"貢\",\n \"垻\",\n \"壩\",\n \"摺\",\n \"掄\",\n \"搶\",\n \"墳\",\n \"護\",\n \"殻\",\n \"塊\",\n \"聲\",\n \"報\",\n \"擬\",\n \"蕪\",\n \"葦\",\n \"蒼\",\n \"嚴\",\n \"蘆\",\n \"勞\",\n \"蘇\",\n \"囌\",\n \"極\",\n \"楊\",\n \"兩\",\n \"麗\",\n \"醫\",\n \"勵\",\n \"還\",\n \"殲\",\n \"來\",\n \"連\",\n \"軒\",\n \"鹵\",\n \"滷\",\n \"堅\",\n \"時\",\n \"縣\",\n \"裏\",\n \"嘔\",\n \"園\",\n \"曠\",\n \"圍\",\n \"噸\",\n \"郵\",\n \"睏\",\n \"員\",\n \"聽\",\n \"嗆\",\n \"嗚\",\n \"彆\",\n \"嶇\",\n \"崗\",\n \"帳\",\n \"財\",\n \"針\",\n \"釘\",\n \"亂\",\n \"體\",\n \"傭\",\n \"徹\",\n \"餘\",\n \"穀\",\n \"鄰\",\n \"腸\",\n \"龜\",\n \"猶\",\n \"狽\",\n \"條\",\n \"島\",\n \"飯\",\n \"飲\",\n \"係\",\n \"繫\",\n \"凍\",\n \"狀\",\n \"畝\",\n \"庫\",\n \"療\",\n \"應\",\n \"這\",\n \"廬\",\n \"閏\",\n \"閑\",\n \"間\",\n \"悶\",\n \"竈\",\n \"燦\",\n \"瀝\",\n \"淪\",\n \"滄\",\n \"溝\",\n \"滬\",\n \"瀋\",\n \"懷\",\n \"憂\",\n \"窮\",\n \"證\",\n \"啓\",\n \"評\",\n \"補\",\n \"識\",\n \"詐\",\n \"訴\",\n \"診\",\n \"詞\",\n \"譯\",\n \"靈\",\n \"層\",\n \"遲\",\n \"張\",\n \"際\",\n \"陸\",\n \"陳\",\n \"墜\",\n \"勁\",\n \"鷄\",\n \"緯\",\n \"驅\",\n \"純\",\n \"紗\",\n \"綱\",\n \"納\",\n \"駁\",\n \"縱\",\n \"紛\",\n \"紙\",\n \"紋\",\n \"紡\",\n \"驢\",\n \"紐\",\n \"環\",\n \"責\",\n \"現\",\n \"錶\",\n \"規\",\n \"攏\",\n \"揀\",\n \"擔\",\n \"頂\",\n \"擁\",\n \"勢\",\n \"攔\",\n \"擰\",\n \"撥\",\n \"擇\",\n \"蘋\",\n \"範\",\n \"莖\",\n \"樞\",\n \"櫃\",\n \"闆\",\n \"鬆\",\n \"槍\",\n \"楓\",\n \"構\",\n \"喪\",\n \"畫\",\n \"棗\",\n \"賣\",\n \"鬱\",\n \"礬\",\n \"礦\",\n \"碼\",\n \"厠\",\n \"奮\",\n \"態\",\n \"歐\",\n \"毆\",\n \"壟\",\n \"轟\",\n \"頃\",\n \"轉\",\n \"斬\",\n \"輪\",\n \"軟\",\n \"齒\",\n \"虜\",\n \"腎\",\n \"賢\",\n \"國\",\n \"暢\",\n \"嚨\",\n \"鳴\",\n \"羅\",\n \"幟\",\n \"嶺\",\n \"凱\",\n \"敗\",\n \"賬\",\n \"販\",\n \"貶\",\n \"購\",\n \"貯\",\n \"圖\",\n \"釣\",\n \"製\",\n \"颳\",\n \"俠\",\n \"僥\",\n \"偵\",\n \"側\",\n \"憑\",\n \"僑\",\n \"貨\",\n \"質\",\n \"徑\",\n \"捨\",\n \"覓\",\n \"貪\",\n \"貧\",\n \"膚\",\n \"腫\",\n \"脹\",\n \"骯\",\n \"脅\",\n \"魚\",\n \"獰\",\n \"備\",\n \"飾\",\n \"飽\",\n \"飼\",\n \"變\",\n \"龐\",\n \"廟\",\n \"瘧\",\n \"劑\",\n \"廢\",\n \"閘\",\n \"鬧\",\n \"鄭\",\n \"捲\",\n \"單\",\n \"爐\",\n \"淺\",\n \"濘\",\n \"瀉\",\n \"潑\",\n \"澤\",\n \"憐\",\n \"學\",\n \"寶\",\n \"寵\",\n \"審\",\n \"簾\",\n \"實\",\n \"試\",\n \"詩\",\n \"誠\",\n \"襯\",\n \"視\",\n \"話\",\n \"誕\",\n \"詭\",\n \"詢\",\n \"該\",\n \"詳\",\n \"肅\",\n \"録\",\n \"隸\",\n \"彌\",\n \"瀰\",\n \"陝\",\n \"駕\",\n \"參\",\n \"艱\",\n \"綫\",\n \"練\",\n \"組\",\n \"紳\",\n \"細\",\n \"駛\",\n \"織\",\n \"駒\",\n \"終\",\n \"駐\",\n \"絆\",\n \"駝\",\n \"紹\",\n \"繹\",\n \"經\",\n \"貫\",\n \"貳\",\n \"幫\",\n \"項\",\n \"挾\",\n \"撓\",\n \"趙\",\n \"擋\",\n \"墊\",\n \"擠\",\n \"揮\",\n \"薦\",\n \"帶\",\n \"繭\",\n \"蕩\",\n \"榮\",\n \"葷\",\n \"熒\",\n \"鬍\",\n \"蔭\",\n \"藥\",\n \"標\",\n \"棧\",\n \"棟\",\n \"欄\",\n \"檸\",\n \"樹\",\n \"鹹\",\n \"磚\",\n \"硯\",\n \"麵\",\n \"牽\",\n \"鷗\",\n \"殘\",\n \"軸\",\n \"輕\",\n \"鴉\",\n \"戰\",\n \"點\",\n \"臨\",\n \"覽\",\n \"竪\",\n \"嘗\",\n \"啞\",\n \"顯\",\n \"貴\",\n \"蝦\",\n \"蟻\",\n \"螞\",\n \"雖\",\n \"駡\",\n \"勛\",\n \"嘩\",\n \"響\",\n \"喲\",\n \"峽\",\n \"罰\",\n \"賤\",\n \"貼\",\n \"貽\",\n \"鈣\",\n \"鈍\",\n \"鈔\",\n \"鍾\",\n \"鐘\",\n \"鋼\",\n \"鈉\",\n \"鑰\",\n \"欽\",\n \"鈞\",\n \"鈎\",\n \"鈕\",\n \"氈\",\n \"氫\",\n \"選\",\n \"適\",\n \"種\",\n \"鞦\",\n \"復\",\n \"複\",\n \"倆\",\n \"貸\",\n \"順\",\n \"儉\",\n \"須\",\n \"鬚\",\n \"劍\",\n \"朧\",\n \"膽\",\n \"勝\",\n \"狹\",\n \"獅\",\n \"獨\",\n \"獄\",\n \"貿\",\n \"餌\",\n \"饒\",\n \"蝕\",\n \"餃\",\n \"餅\",\n \"巒\",\n \"彎\",\n \"將\",\n \"奬\",\n \"瘡\",\n \"瘋\",\n \"親\",\n \"閨\",\n \"聞\",\n \"閩\",\n \"閥\",\n \"閣\",\n \"養\",\n \"薑\",\n \"類\",\n \"婁\",\n \"總\",\n \"煉\",\n \"爍\",\n \"爛\",\n \"窪\",\n \"潔\",\n \"灑\",\n \"澆\",\n \"濁\",\n \"測\",\n \"瀏\",\n \"濟\",\n \"渾\",\n \"濃\",\n \"惱\",\n \"舉\",\n \"覺\",\n \"憲\",\n \"竊\",\n \"誡\",\n \"誣\",\n \"語\",\n \"襖\",\n \"誤\",\n \"誘\",\n \"誨\",\n \"説\",\n \"誦\",\n \"墾\",\n \"晝\",\n \"費\",\n \"遜\",\n \"隕\",\n \"險\",\n \"嬌\",\n \"賀\",\n \"壘\",\n \"綁\",\n \"絨\",\n \"結\",\n \"繞\",\n \"驕\",\n \"繪\",\n \"給\",\n \"絢\",\n \"駱\",\n \"絡\",\n \"絶\",\n \"絞\",\n \"駭\",\n \"統\",\n \"艷\",\n \"蠶\",\n \"頑\",\n \"盞\",\n \"撈\",\n \"載\",\n \"趕\",\n \"鹽\",\n \"損\",\n \"撿\",\n \"摯\",\n \"剝\",\n \"熱\",\n \"搗\",\n \"壺\",\n \"聶\",\n \"萊\",\n \"蓮\",\n \"獲\",\n \"穫\",\n \"惡\",\n \"噁\",\n \"瑩\",\n \"鶯\",\n \"檔\",\n \"橋\",\n \"樺\",\n \"樁\",\n \"樣\",\n \"賈\",\n \"礫\",\n \"礎\",\n \"顧\",\n \"轎\",\n \"較\",\n \"頓\",\n \"斃\",\n \"緻\",\n \"慮\",\n \"監\",\n \"緊\",\n \"黨\",\n \"曬\",\n \"曉\",\n \"嘮\",\n \"鴨\",\n \"暈\",\n \"鴦\",\n \"罷\",\n \"圓\",\n \"賊\",\n \"賄\",\n \"賂\",\n \"贜\",\n \"錢\",\n \"鉗\",\n \"鑽\",\n \"鉀\",\n \"鐵\",\n \"鈴\",\n \"鉛\",\n \"犧\",\n \"敵\",\n \"積\",\n \"稱\",\n \"筆\",\n \"債\",\n \"傾\",\n \"賃\",\n \"艦\",\n \"艙\",\n \"聳\",\n \"愛\",\n \"頒\",\n \"頌\",\n \"臟\",\n \"髒\",\n \"臍\",\n \"膠\",\n \"腦\",\n \"膿\",\n \"鴕\",\n \"鴛\",\n \"皺\",\n \"餓\",\n \"餒\",\n \"戀\",\n \"槳\",\n \"漿\",\n \"準\",\n \"癥\",\n \"齋\",\n \"離\",\n \"資\",\n \"競\",\n \"閲\",\n \"煩\",\n \"燒\",\n \"燭\",\n \"遞\",\n \"濤\",\n \"澇\",\n \"渦\",\n \"塗\",\n \"滌\",\n \"潤\",\n \"澗\",\n \"漲\",\n \"燙\",\n \"澀\",\n \"憫\",\n \"寬\",\n \"傢\",\n \"賓\",\n \"竅\",\n \"請\",\n \"諸\",\n \"諾\",\n \"讀\",\n \"誹\",\n \"襪\",\n \"課\",\n \"誰\",\n \"調\",\n \"諒\",\n \"諄\",\n \"談\",\n \"誼\",\n \"懇\",\n \"劇\",\n \"難\",\n \"預\",\n \"絹\",\n \"綉\",\n \"驗\",\n \"繼\",\n \"駿\",\n \"瑣\",\n \"擲\",\n \"據\",\n \"摻\",\n \"職\",\n \"蘿\",\n \"螢\",\n \"營\",\n \"蕭\",\n \"薩\",\n \"夢\",\n \"檢\",\n \"醖\",\n \"碩\",\n \"聾\",\n \"襲\",\n \"輔\",\n \"輛\",\n \"顱\",\n \"懸\",\n \"躍\",\n \"纍\",\n \"囉\",\n \"嘯\",\n \"嶄\",\n \"邏\",\n \"嬰\",\n \"銬\",\n \"鐺\",\n \"鋁\",\n \"銅\",\n \"銘\",\n \"鏟\",\n \"銀\",\n \"矯\",\n \"穢\",\n \"籠\",\n \"償\",\n \"軀\",\n \"釁\",\n \"銜\",\n \"盤\",\n \"鴿\",\n \"斂\",\n \"領\",\n \"臉\",\n \"獵\",\n \"餡\",\n \"館\",\n \"癢\",\n \"鏇\",\n \"閻\",\n \"闡\",\n \"蓋\",\n \"斷\",\n \"獸\",\n \"鴻\",\n \"漸\",\n \"淵\",\n \"漁\",\n \"澱\",\n \"滲\",\n \"慚\",\n \"懼\",\n \"驚\",\n \"慘\",\n \"慣\",\n \"謀\",\n \"諜\",\n \"謊\",\n \"諧\",\n \"禱\",\n \"禍\",\n \"謂\",\n \"諺\",\n \"謎\",\n \"彈\",\n \"墮\",\n \"隨\",\n \"隱\",\n \"嬸\",\n \"頗\",\n \"頸\",\n \"績\",\n \"緒\",\n \"續\",\n \"騎\",\n \"綽\",\n \"繩\",\n \"維\",\n \"綿\",\n \"綳\",\n \"綢\",\n \"綜\",\n \"綻\",\n \"緑\",\n \"綴\",\n \"瓊\",\n \"趨\",\n \"攬\",\n \"攙\",\n \"擱\",\n \"摟\",\n \"攪\",\n \"聯\",\n \"蔣\",\n \"韓\",\n \"橢\",\n \"確\",\n \"頰\",\n \"靂\",\n \"暫\",\n \"翹\",\n \"輩\",\n \"鑿\",\n \"輝\",\n \"賞\",\n \"睞\",\n \"噴\",\n \"疇\",\n \"踐\",\n \"遺\",\n \"鵑\",\n \"賦\",\n \"賭\",\n \"贖\",\n \"賜\",\n \"賠\",\n \"鑄\",\n \"鋪\",\n \"鏈\",\n \"銷\",\n \"鎖\",\n \"鋤\",\n \"鍋\",\n \"銹\",\n \"鋒\",\n \"鋅\",\n \"鋭\",\n \"鵝\",\n \"築\",\n \"篩\",\n \"儲\",\n \"懲\",\n \"禦\",\n \"釋\",\n \"臘\",\n \"魯\",\n \"憊\",\n \"饋\",\n \"饞\",\n \"裝\",\n \"蠻\",\n \"闊\",\n \"糞\",\n \"滯\",\n \"濕\",\n \"潰\",\n \"濺\",\n \"灣\",\n \"憤\",\n \"竄\",\n \"窩\",\n \"褲\",\n \"禪\",\n \"謝\",\n \"謡\",\n \"謗\",\n \"謙\",\n \"屬\",\n \"屢\",\n \"緬\",\n \"纜\",\n \"緝\",\n \"緞\",\n \"緩\",\n \"締\",\n \"縷\",\n \"騙\",\n \"編\",\n \"騷\",\n \"緣\",\n \"鵡\",\n \"攝\",\n \"擺\",\n \"襬\",\n \"攤\",\n \"鵲\",\n \"藍\",\n \"濛\",\n \"懞\",\n \"矇\",\n \"獻\",\n \"欖\",\n \"樓\",\n \"賴\",\n \"礙\",\n \"尷\",\n \"霧\",\n \"輻\",\n \"輯\",\n \"輸\",\n \"頻\",\n \"齡\",\n \"鑒\",\n \"蹺\",\n \"蝸\",\n \"錯\",\n \"錨\",\n \"錫\",\n \"鑼\",\n \"錘\",\n \"錐\",\n \"錦\",\n \"鍵\",\n \"鋸\",\n \"錳\",\n \"辭\",\n \"頽\",\n \"籌\",\n \"簽\",\n \"籤\",\n \"簡\",\n \"膩\",\n \"鵬\",\n \"騰\",\n \"鮑\",\n \"穎\",\n \"觸\",\n \"雛\",\n \"饃\",\n \"餾\",\n \"醬\",\n \"謄\",\n \"糧\",\n \"數\",\n \"滿\",\n \"濾\",\n \"濫\",\n \"灕\",\n \"濱\",\n \"灘\",\n \"譽\",\n \"窺\",\n \"寢\",\n \"謹\",\n \"謬\",\n \"闢\",\n \"縛\",\n \"縫\",\n \"纏\",\n \"繽\",\n \"贅\",\n \"墻\",\n \"衊\",\n \"藹\",\n \"檻\",\n \"釀\",\n \"願\",\n \"轄\",\n \"輾\",\n \"顆\",\n \"踴\",\n \"蠟\",\n \"蠅\",\n \"蟬\",\n \"賺\",\n \"鍬\",\n \"鍛\",\n \"鍍\",\n \"穩\",\n \"籮\",\n \"簫\",\n \"輿\",\n \"鮮\",\n \"饅\",\n \"瀟\",\n \"賽\",\n \"譚\",\n \"譜\",\n \"騾\",\n \"縮\",\n \"攆\",\n \"聰\",\n \"藴\",\n \"櫻\",\n \"飄\",\n \"黴\",\n \"瞞\",\n \"題\",\n \"囑\",\n \"鎮\",\n \"鎬\",\n \"鎊\",\n \"簍\",\n \"鯉\",\n \"鯽\",\n \"癟\",\n \"癱\",\n \"顔\",\n \"鯊\",\n \"瀾\",\n \"額\",\n \"譴\",\n \"鶴\",\n \"繚\",\n \"顛\",\n \"轍\",\n \"鸚\",\n \"贈\",\n \"鏡\",\n \"贊\",\n \"籃\",\n \"籬\",\n \"鯨\",\n \"癮\",\n \"辯\",\n \"瀕\",\n \"懶\",\n \"繮\",\n \"繳\",\n \"矚\",\n \"贍\",\n \"鰐\",\n \"辮\",\n \"贏\",\n \"驟\",\n \"囂\",\n \"鐮\",\n \"鰭\",\n \"鷹\",\n \"巔\",\n \"顫\",\n \"癬\",\n \"鱉\",\n \"鬢\",\n \"鱗\",\n \"躪\",\n \"贛\",\n \"鑲\",\n \"韋\",\n \"閂\",\n \"訃\",\n \"勱\",\n \"芻\",\n \"鄺\",\n \"訐\",\n \"訌\",\n \"訕\",\n \"訖\",\n \"馭\",\n \"璣\",\n \"壙\",\n \"捫\",\n \"薌\",\n \"厙\",\n \"釔\",\n \"傴\",\n \"倀\",\n \"傖\",\n \"獷\",\n \"獁\",\n \"鳬\",\n \"鄔\",\n \"餳\",\n \"懺\",\n \"謳\",\n \"詎\",\n \"訥\",\n \"紆\",\n \"紂\",\n \"紇\",\n \"紈\",\n \"璵\",\n \"摶\",\n \"塢\",\n \"㩳\",\n \"蕓\",\n \"藶\",\n \"莧\",\n \"萇\",\n \"蓯\",\n \"磯\",\n \"奩\",\n \"歟\",\n \"軔\",\n \"鄴\",\n \"嘸\",\n \"囈\",\n \"嚦\",\n \"暘\",\n \"唄\",\n \"幃\",\n \"峴\",\n \"嵐\",\n \"圇\",\n \"釗\",\n \"釙\",\n \"釕\",\n \"僉\",\n \"鳩\",\n \"鄒\",\n \"飩\",\n \"餼\",\n \"飪\",\n \"飫\",\n \"飭\",\n \"廡\",\n \"癤\",\n \"闈\",\n \"閎\",\n \"閔\",\n \"煬\",\n \"灃\",\n \"漚\",\n \"渢\",\n \"潙\",\n \"憮\",\n \"慪\",\n \"愾\",\n \"悵\",\n \"愴\",\n \"詁\",\n \"訶\",\n \"詛\",\n \"詆\",\n \"謅\",\n \"詔\",\n \"詒\",\n \"隴\",\n \"陘\",\n \"嫵\",\n \"嫗\",\n \"嬀\",\n \"剄\",\n \"紜\",\n \"紕\",\n \"紝\",\n \"綸\",\n \"紓\",\n \"瑋\",\n \"匭\",\n \"壚\",\n \"擓\",\n \"蘢\",\n \"蔦\",\n \"塋\",\n \"煢\",\n \"櫪\",\n \"梘\",\n \"棖\",\n \"樅\",\n \"碭\",\n \"甌\",\n \"郟\",\n \"軛\",\n \"鳶\",\n \"曇\",\n \"蟣\",\n \"黽\",\n \"嚀\",\n \"噝\",\n \"巋\",\n \"劌\",\n \"剴\",\n \"嶧\",\n \"釷\",\n \"釺\",\n \"釧\",\n \"釩\",\n \"釹\",\n \"釵\",\n \"儈\",\n \"儕\",\n \"儂\",\n \"劊\",\n \"慫\",\n \"糴\",\n \"戧\",\n \"膞\",\n \"邇\",\n \"梟\",\n \"餞\",\n \"飴\",\n \"癘\",\n \"瘍\",\n \"煒\",\n \"熰\",\n \"熗\",\n \"瀧\",\n \"瀘\",\n \"濼\",\n \"涇\",\n \"㥮\",\n \"懌\",\n \"誆\",\n \"誄\",\n \"詿\",\n \"詰\",\n \"詼\",\n \"鄆\",\n \"禕\",\n \"誅\",\n \"詵\",\n \"詬\",\n \"詮\",\n \"詣\",\n \"諍\",\n \"詫\",\n \"諢\",\n \"詡\",\n \"駑\",\n \"紺\",\n \"紲\",\n \"紱\",\n \"駟\",\n \"駙\",\n \"縐\",\n \"絀\",\n \"驛\",\n \"駘\",\n \"瓏\",\n \"頇\",\n \"埡\",\n \"撾\",\n \"撻\",\n \"賁\",\n \"壋\",\n \"撏\",\n \"莢\",\n \"貰\",\n \"蓽\",\n \"蕎\",\n \"薈\",\n \"薺\",\n \"堊\",\n \"滎\",\n \"犖\",\n \"蕁\",\n \"藎\",\n \"蓀\",\n \"蕒\",\n \"葤\",\n \"櫛\",\n \"櫳\",\n \"櫨\",\n \"櫟\",\n \"檉\",\n \"酈\",\n \"硨\",\n \"碸\",\n \"殤\",\n \"軲\",\n \"軻\",\n \"轤\",\n \"軼\",\n \"軫\",\n \"蠆\",\n \"覘\",\n \"瞘\",\n \"嘵\",\n \"嗶\",\n \"噦\",\n \"剮\",\n \"鄖\",\n \"噲\",\n \"噥\",\n \"嶢\",\n \"幀\",\n \"嶠\",\n \"貺\",\n \"鈈\",\n \"鈦\",\n \"鋇\",\n \"鈑\",\n \"鈐\",\n \"鎢\",\n \"鈁\",\n \"鈀\",\n \"篤\",\n \"儔\",\n \"儼\",\n \"儷\",\n \"腖\",\n \"臚\",\n \"脛\",\n \"鴇\",\n \"獪\",\n \"颮\",\n \"猻\",\n \"餉\",\n \"餄\",\n \"餎\",\n \"孿\",\n \"孌\",\n \"癧\",\n \"瘲\",\n \"颯\",\n \"闥\",\n \"閭\",\n \"闓\",\n \"閡\",\n \"熾\",\n \"烴\",\n \"浹\",\n \"澮\",\n \"滸\",\n \"潯\",\n \"濜\",\n \"慟\",\n \"懨\",\n \"愷\",\n \"惻\",\n \"惲\",\n \"誚\",\n \"禰\",\n \"誥\",\n \"誑\",\n \"鴆\",\n \"婭\",\n \"嬈\",\n \"懟\",\n \"絝\",\n \"驍\",\n \"驊\",\n \"絎\",\n \"絳\",\n \"駢\",\n \"頊\",\n \"璫\",\n \"琿\",\n \"塒\",\n \"塤\",\n \"堝\",\n \"贄\",\n \"蒔\",\n \"萵\",\n \"蕕\",\n \"鴣\",\n \"蒓\",\n \"橈\",\n \"楨\",\n \"榿\",\n \"檜\",\n \"邐\",\n \"礪\",\n \"礱\",\n \"軾\",\n \"輊\",\n \"輅\",\n \"鶇\",\n \"躉\",\n \"齔\",\n \"鸕\",\n \"矓\",\n \"嘜\",\n \"鴞\",\n \"蜆\",\n \"嗩\",\n \"嶗\",\n \"崍\",\n \"覬\",\n \"賅\",\n \"鈺\",\n \"鉦\",\n \"鈷\",\n \"鉢\",\n \"鈸\",\n \"鉞\",\n \"鉭\",\n \"鉬\",\n \"鈿\",\n \"鈾\",\n \"鉑\",\n \"鑠\",\n \"鉚\",\n \"鈰\",\n \"鉉\",\n \"鉈\",\n \"鉍\",\n \"鈮\",\n \"鈹\",\n \"鏺\",\n \"鐸\",\n \"氬\",\n \"筧\",\n \"頎\",\n \"徠\",\n \"膾\",\n \"鴟\",\n \"璽\",\n \"鴝\",\n \"獫\",\n \"裊\",\n \"餑\",\n \"欒\",\n \"攣\",\n \"癰\",\n \"痙\",\n \"頏\",\n \"閫\",\n \"鬮\",\n \"誾\",\n \"閬\",\n \"鄲\",\n \"燁\",\n \"燴\",\n \"燼\",\n \"淶\",\n \"漣\",\n \"潿\",\n \"慳\",\n \"諏\",\n \"諑\",\n \"禎\",\n \"諉\",\n \"諛\",\n \"諗\",\n \"諂\",\n \"誶\",\n \"媧\",\n \"嫻\",\n \"綆\",\n \"驪\",\n \"綃\",\n \"騁\",\n \"綏\",\n \"縧\",\n \"綈\",\n \"駸\",\n \"鷥\",\n \"燾\",\n \"璉\",\n \"麩\",\n \"擄\",\n \"摑\",\n \"鷙\",\n \"撣\",\n \"慤\",\n \"摜\",\n \"縈\",\n \"槤\",\n \"覡\",\n \"欞\",\n \"嗇\",\n \"匱\",\n \"硤\",\n \"磽\",\n \"鴯\",\n \"龔\",\n \"殞\",\n \"殮\",\n \"賚\",\n \"輒\",\n \"塹\",\n \"嘖\",\n \"囀\",\n \"嚙\",\n \"蹌\",\n \"蠣\",\n \"蠱\",\n \"蟶\",\n \"幘\",\n \"幗\",\n \"賕\",\n \"賑\",\n \"賒\",\n \"銠\",\n \"鉺\",\n \"鋏\",\n \"鐃\",\n \"銦\",\n \"鎧\",\n \"鍘\",\n \"銖\",\n \"銑\",\n \"鋌\",\n \"鏵\",\n \"銓\",\n \"鎩\",\n \"鉿\",\n \"銚\",\n \"鉻\",\n \"錚\",\n \"銫\",\n \"鉸\",\n \"銥\",\n \"銃\",\n \"銨\",\n \"銣\",\n \"鴰\",\n \"穠\",\n \"箋\",\n \"籩\",\n \"僨\",\n \"僂\",\n \"皚\",\n \"鴴\",\n \"艫\",\n \"龕\",\n \"玀\",\n \"獼\",\n \"餜\",\n \"餛\",\n \"鸞\",\n \"闍\",\n \"閾\",\n \"閹\",\n \"閶\",\n \"鬩\",\n \"閽\",\n \"閼\",\n \"羥\",\n \"糲\",\n \"燜\",\n \"漬\",\n \"瀆\",\n \"澠\",\n \"愜\",\n \"憚\",\n \"諶\",\n \"諫\",\n \"皸\",\n \"謔\",\n \"襠\",\n \"謁\",\n \"諤\",\n \"諭\",\n \"諼\",\n \"讒\",\n \"諳\",\n \"諦\",\n \"諞\",\n \"糶\",\n \"嬋\",\n \"綾\",\n \"騏\",\n \"綺\",\n \"緋\",\n \"緔\",\n \"騍\",\n \"緄\",\n \"騅\",\n \"綬\",\n \"綹\",\n \"綣\",\n \"綰\",\n \"驂\",\n \"緇\",\n \"靚\",\n \"輦\",\n \"黿\",\n \"頡\",\n \"撳\",\n \"蟄\",\n \"壪\",\n \"蔞\",\n \"櫝\",\n \"欏\",\n \"賫\",\n \"鵓\",\n \"鸝\",\n \"殫\",\n \"輥\",\n \"輞\",\n \"槧\",\n \"輟\",\n \"輜\",\n \"瞼\",\n \"躒\",\n \"蛺\",\n \"蟯\",\n \"螄\",\n \"蠐\",\n \"嘍\",\n \"嶸\",\n \"嶁\",\n \"賧\",\n \"鋙\",\n \"錸\",\n \"鏗\",\n \"鋥\",\n \"鋰\",\n \"鋯\",\n \"鋨\",\n \"銼\",\n \"鐧\",\n \"銻\",\n \"鋃\",\n \"鋦\",\n \"錒\",\n \"犢\",\n \"鵠\",\n \"篳\",\n \"牘\",\n \"儻\",\n \"儐\",\n \"儺\",\n \"嬃\",\n \"頜\",\n \"鵒\",\n \"魷\",\n \"魨\",\n \"魴\",\n \"潁\",\n \"颶\",\n \"觴\",\n \"熲\",\n \"餷\",\n \"餿\",\n \"褻\",\n \"臠\",\n \"癆\",\n \"癇\",\n \"賡\",\n \"頦\",\n \"鷳\",\n \"闌\",\n \"闃\",\n \"闋\",\n \"鵜\",\n \"憒\",\n \"嚳\",\n \"謨\",\n \"褳\",\n \"襇\",\n \"讜\",\n \"謖\",\n \"謚\",\n \"謐\",\n \"騭\",\n \"巰\",\n \"翬\",\n \"騖\",\n \"緙\",\n \"緗\",\n \"緘\",\n \"緹\",\n \"緲\",\n \"緦\",\n \"緱\",\n \"縋\",\n \"緡\",\n \"饗\",\n \"耮\",\n \"驁\",\n \"韞\",\n \"攄\",\n \"擯\",\n \"轂\",\n \"驀\",\n \"鶓\",\n \"薊\",\n \"蘺\",\n \"鎣\",\n \"頤\",\n \"櫚\",\n \"櫸\",\n \"磧\",\n \"磣\",\n \"鵪\",\n \"輳\",\n \"齟\",\n \"齙\",\n \"韙\",\n \"囁\",\n \"躂\",\n \"蹕\",\n \"躚\",\n \"躋\",\n \"噯\",\n \"鍺\",\n \"錛\",\n \"錡\",\n \"鍀\",\n \"錁\",\n \"錕\",\n \"錮\",\n \"鍁\",\n \"錈\",\n \"錠\",\n \"錙\",\n \"覦\",\n \"頷\",\n \"鮁\",\n \"鮃\",\n \"鮎\",\n \"鱸\",\n \"穌\",\n \"鮒\",\n \"鮐\",\n \"鵮\",\n \"颼\",\n \"饈\",\n \"鶉\",\n \"瘮\",\n \"闔\",\n \"闐\",\n \"闕\",\n \"灧\",\n \"瀅\",\n \"潷\",\n \"灤\",\n \"澦\",\n \"懾\",\n \"鱟\",\n \"騫\",\n \"竇\",\n \"謾\",\n \"謫\",\n \"嬡\",\n \"嬪\",\n \"縉\",\n \"縝\",\n \"縟\",\n \"轡\",\n \"騮\",\n \"縞\",\n \"縭\",\n \"縊\",\n \"縑\",\n \"騸\",\n \"覯\",\n \"韜\",\n \"靉\",\n \"攖\",\n \"薔\",\n \"藺\",\n \"鶘\",\n \"檳\",\n \"櫧\",\n \"釅\",\n \"殯\",\n \"霽\",\n \"轅\",\n \"齜\",\n \"齦\",\n \"瞜\",\n \"曖\",\n \"躊\",\n \"蟈\",\n \"鶚\",\n \"嚶\",\n \"羆\",\n \"賻\",\n \"罌\",\n \"鶻\",\n \"鍥\",\n \"鍇\",\n \"鍶\",\n \"鍔\",\n \"鍤\",\n \"鏘\",\n \"鎂\",\n \"鏤\",\n \"簀\",\n \"篋\",\n \"簞\",\n \"籙\",\n \"臏\",\n \"鮭\",\n \"鮪\",\n \"鱭\",\n \"鮫\",\n \"鱘\",\n \"饉\",\n \"鑾\",\n \"瘻\",\n \"闞\",\n \"鮝\",\n \"糝\",\n \"鷀\",\n \"瀲\",\n \"濰\",\n \"譖\",\n \"褸\",\n \"譙\",\n \"讕\",\n \"譎\",\n \"鶥\",\n \"嬙\",\n \"鶩\",\n \"驃\",\n \"縹\",\n \"縵\",\n \"縲\",\n \"纓\",\n \"驄\",\n \"繆\",\n \"繅\",\n \"耬\",\n \"瓔\",\n \"擷\",\n \"擼\",\n \"攛\",\n \"聵\",\n \"覲\",\n \"韃\",\n \"鞽\",\n \"蘄\",\n \"賾\",\n \"檣\",\n \"靨\",\n \"魘\",\n \"饜\",\n \"轆\",\n \"齬\",\n \"齪\",\n \"覷\",\n \"顒\",\n \"躓\",\n \"躑\",\n \"蠑\",\n \"螻\",\n \"顎\",\n \"嚕\",\n \"顓\",\n \"鑷\",\n \"鎘\",\n \"鎸\",\n \"鎳\",\n \"鎦\",\n \"鎰\",\n \"鎵\",\n \"鑌\",\n \"簣\",\n \"鷂\",\n \"鯁\",\n \"鱺\",\n \"鰱\",\n \"鰹\",\n \"鰣\",\n \"鯀\",\n \"鯇\",\n \"觶\",\n \"饊\",\n \"饌\",\n \"齏\",\n \"讞\",\n \"襤\",\n \"譫\",\n \"屨\",\n \"纈\",\n \"繕\",\n \"繒\",\n \"驏\",\n \"擻\",\n \"顳\",\n \"顢\",\n \"藪\",\n \"櫓\",\n \"櫞\",\n \"贋\",\n \"飆\",\n \"鏨\",\n \"轔\",\n \"蟎\",\n \"鐯\",\n \"鏢\",\n \"鏜\",\n \"鏝\",\n \"鏰\",\n \"鏞\",\n \"鏑\",\n \"鏃\",\n \"鏐\",\n \"氌\",\n \"穡\",\n \"魎\",\n \"鯪\",\n \"鯡\",\n \"鯤\",\n \"鯧\",\n \"鯝\",\n \"鯢\",\n \"鯛\",\n \"鯔\",\n \"獺\",\n \"鷓\",\n \"贇\",\n \"癭\",\n \"斕\",\n \"瀨\",\n \"顙\",\n \"繾\",\n \"繰\",\n \"繯\",\n \"蘚\",\n \"鷯\",\n \"齲\",\n \"齷\",\n \"躡\",\n \"蹣\",\n \"羈\",\n \"鐔\",\n \"鐝\",\n \"鐐\",\n \"鐓\",\n \"鑭\",\n \"鑹\",\n \"鏹\",\n \"鐙\",\n \"籪\",\n \"鷦\",\n \"鱝\",\n \"鰈\",\n \"鯷\",\n \"鰓\",\n \"鰍\",\n \"鰉\",\n \"鯿\",\n \"鷲\",\n \"懣\",\n \"鷸\",\n \"鰲\",\n \"韉\",\n \"顥\",\n \"鷺\",\n \"䴉\",\n \"髏\",\n \"鑊\",\n \"鐳\",\n \"鐲\",\n \"讎\",\n \"鰨\",\n \"鰥\",\n \"鰩\",\n \"癩\",\n \"攢\",\n \"靄\",\n \"躥\",\n \"髖\",\n \"髕\",\n \"鑔\",\n \"籟\",\n \"鰳\",\n \"鰾\",\n \"鱈\",\n \"鰻\",\n \"鱅\",\n \"讖\",\n \"驥\",\n \"纘\",\n \"瓚\",\n \"鼉\",\n \"黷\",\n \"黲\",\n \"鑣\",\n \"鑞\",\n \"臢\",\n \"鱖\",\n \"鱔\",\n \"鱒\",\n \"驤\",\n \"顰\",\n \"鱧\",\n \"癲\",\n \"灝\",\n \"鸛\",\n \"鑱\",\n \"趲\",\n \"顴\",\n \"躦\",\n \"饢\",\n \"戇\",\n \"戔\",\n \"訏\",\n \"訒\",\n \"釓\",\n \"俔\",\n \"閆\",\n \"澫\",\n \"訢\",\n \"訩\",\n \"詝\",\n \"紃\",\n \"纊\",\n \"瑒\",\n \"剗\",\n \"塸\",\n \"壢\",\n \"埨\",\n \"撝\",\n \"蔿\",\n \"榪\",\n \"軑\",\n \"軏\",\n \"咼\",\n \"㠣\",\n \"覎\",\n \"㑳\",\n \"颺\",\n \"閌\",\n \"潕\",\n \"湋\",\n \"澐\",\n \"浿\",\n \"諓\",\n \"禡\",\n \"詗\",\n \"詘\",\n \"詖\",\n \"屓\",\n \"彄\",\n \"紘\",\n \"馹\",\n \"馼\",\n \"紵\",\n \"紞\",\n \"駃\",\n \"紖\",\n \"瑲\",\n \"薴\",\n \"棡\",\n \"軝\",\n \"暐\",\n \"晛\",\n \"崬\",\n \"釴\",\n \"釤\",\n \"鍆\",\n \"鍚\",\n \"鄶\",\n \"獮\",\n \"飿\",\n \"嶨\",\n \"詷\",\n \"詪\",\n \"鄩\",\n \"鳲\",\n \"隑\",\n \"隮\",\n \"娙\",\n \"逕\",\n \"駓\",\n \"駔\",\n \"駉\",\n \"絅\",\n \"騶\",\n \"䮄\",\n \"紼\",\n \"紿\",\n \"瓅\",\n \"韍\",\n \"墶\",\n \"塏\",\n \"薘\",\n \"蕘\",\n \"蔄\",\n \"葒\",\n \"鳾\",\n \"龑\",\n \"軹\",\n \"軤\",\n \"轢\",\n \"軺\",\n \"睍\",\n \"曨\",\n \"噠\",\n \"鈃\",\n \"鈇\",\n \"鉅\",\n \"鋹\",\n \"釿\",\n \"錀\",\n \"鈧\",\n \"鈥\",\n \"鈄\",\n \"倈\",\n \"艤\",\n \"鶬\",\n \"颭\",\n \"餏\",\n \"湞\",\n \"溮\",\n \"滻\",\n \"褘\",\n \"絰\",\n \"駰\",\n \"絪\",\n \"駪\",\n \"綎\",\n \"綖\",\n \"驫\",\n \"勣\",\n \"璕\",\n \"𡑍\",\n \"䓣\",\n \"薟\",\n \"藭\",\n \"椏\",\n \"梜\",\n \"頍\",\n \"硜\",\n \"輄\",\n \"輈\",\n \"輇\",\n \"貲\",\n \"嗊\",\n \"曄\",\n \"暉\",\n \"鄳\",\n \"幬\",\n \"輋\",\n \"嶮\",\n \"贐\",\n \"鉥\",\n \"鉕\",\n \"鑪\",\n \"鉮\",\n \"鉊\",\n \"鉧\",\n \"僤\",\n \"鴒\",\n \"魛\",\n \"餗\",\n \"燖\",\n \"溳\",\n \"礐\",\n \"窵\",\n \"襏\",\n \"駼\",\n \"絺\",\n \"綌\",\n \"騂\",\n \"綄\",\n \"璡\",\n \"墠\",\n \"壼\",\n \"聹\",\n \"蘀\",\n \"勩\",\n \"罃\",\n \"檮\",\n \"棶\",\n \"厴\",\n \"䃮\",\n \"磑\",\n \"礄\",\n \"鴷\",\n \"齕\",\n \"頔\",\n \"廼\",\n \"凢\",\n \"亾\",\n \"枒\",\n \"屍\",\n \"匃\",\n \"匄\",\n \"紥\",\n \"紮\",\n \"疋\",\n \"殀\",\n \"讐\",\n \"觔\",\n \"兇\",\n \"宂\",\n \"㕥\",\n \"㠯\",\n \"栞\",\n \"佈\",\n \"佔\",\n \"呌\",\n \"敂\",\n \"冄\",\n \"坵\",\n \"僊\",\n \"怱\",\n \"悤\",\n \"冊\",\n \"夘\",\n \"戼\",\n \"牠\",\n \"妳\",\n \"嬭\",\n \"摃\",\n \"釦\",\n \"攷\",\n \"託\",\n \"衺\",\n \"衕\",\n \"弔\",\n \"喫\",\n \"囙\",\n \"㠶\",\n \"颿\",\n \"秊\",\n \"倣\",\n \"髣\",\n \"佀\",\n \"朶\",\n \"氷\",\n \"決\",\n \"併\",\n \"並\",\n \"竝\",\n \"汙\",\n \"汚\",\n \"異\",\n \"姦\",\n \"廵\",\n \"挵\",\n \"衖\",\n \"搤\",\n \"阯\",\n \"撦\",\n \"埳\",\n \"阬\",\n \"誌\",\n \"㕁\",\n \"卻\",\n \"刦\",\n \"刧\",\n \"刼\",\n \"芲\",\n \"蘤\",\n \"桿\",\n \"槓\",\n \"荳\",\n \"獃\",\n \"唫\",\n \"脗\",\n \"皁\",\n \"彿\",\n \"髴\",\n \"疘\",\n \"刪\",\n \"鉋\",\n \"鑤\",\n \"況\",\n \"牀\",\n \"恡\",\n \"棄\",\n \"洶\",\n \"汎\",\n \"災\",\n \"烖\",\n \"菑\",\n \"禩\",\n \"侷\",\n \"跼\",\n \"坿\",\n \"玅\",\n \"姉\",\n \"妬\",\n \"翫\",\n \"搨\",\n \"柺\",\n \"拕\",\n \"牴\",\n \"觝\",\n \"倖\",\n \"抝\",\n \"盃\",\n \"桮\",\n \"傑\",\n \"逩\",\n \"肎\",\n \"菓\",\n \"崐\",\n \"崑\",\n \"呪\",\n \"虖\",\n \"嘑\",\n \"謼\",\n \"詠\",\n \"㟁\",\n \"嵒\",\n \"巗\",\n \"巖\",\n \"雰\",\n \"稈\",\n \"咊\",\n \"嶽\",\n \"妷\",\n \"姪\",\n \"廹\",\n \"徃\",\n \"餚\",\n \"採\",\n \"寀\",\n \"唸\",\n \"週\",\n \"昬\",\n \"兎\",\n \"兔\",\n \"亯\",\n \"亱\",\n \"䘚\",\n \"淨\",\n \"劵\",\n \"匟\",\n \"㳒\",\n \"灋\",\n \"洩\",\n \"霑\",\n \"淚\",\n \"註\",\n \"恠\",\n \"箒\",\n \"屆\",\n \"絃\",\n \"圅\",\n \"旾\",\n \"珎\",\n \"掛\",\n \"垜\",\n \"艸\",\n \"茘\",\n \"査\",\n \"栢\",\n \"柵\",\n \"栁\",\n \"桺\",\n \"柹\",\n \"韮\",\n \"揹\",\n \"昰\",\n \"閧\",\n \"鬨\",\n \"冐\",\n \"暎\",\n \"嚥\",\n \"倃\",\n \"𠴰\",\n \"偺\",\n \"喒\",\n \"齩\",\n \"欬\",\n \"榘\",\n \"㑺\",\n \"儁\",\n \"敍\",\n \"敘\",\n \"肧\",\n \"脈\",\n \"䘑\",\n \"衇\",\n \"跡\",\n \"蹟\",\n \"砲\",\n \"礮\",\n \"薙\",\n \"鬀\",\n \"恆\",\n \"怳\",\n \"卹\",\n \"䘏\",\n \"賉\",\n \"婣\",\n \"畊\",\n \"揑\",\n \"綑\",\n \"輓\",\n \"恥\",\n \"躭\",\n \"晉\",\n \"棲\",\n \"覈\",\n \"慄\",\n \"翄\",\n \"脣\",\n \"槕\",\n \"㨪\",\n \"螡\",\n \"蟁\",\n \"㤙\",\n \"陗\",\n \"峩\",\n \"峯\",\n \"乗\",\n \"椉\",\n \"咲\",\n \"筍\",\n \"俛\",\n \"頫\",\n \"勌\",\n \"䠶\",\n \"躳\",\n \"慇\",\n \"拏\",\n \"㧱\",\n \"挐\",\n \"脃\",\n \"胷\",\n \"肐\",\n \"貍\",\n \"㽞\",\n \"畱\",\n \"淒\",\n \"悽\",\n \"蓆\",\n \"効\",\n \"傚\",\n \"涼\",\n \"缾\",\n \"菸\",\n \"煙\",\n \"淛\",\n \"湧\",\n \"誖\",\n \"猂\",\n \"醼\",\n \"讌\",\n \"㝠\",\n \"寃\",\n \"孃\",\n \"桒\",\n \"毬\",\n \"瑠\",\n \"璢\",\n \"瑯\",\n \"㨗\",\n \"搥\",\n \"搯\",\n \"蔆\",\n \"惏\",\n \"楳\",\n \"槑\",\n \"捄\",\n \"廂\",\n \"慽\",\n \"慼\",\n \"瞇\",\n \"埜\",\n \"畧\",\n \"虵\",\n \"稭\",\n \"棃\",\n \"犂\",\n \"迻\",\n \"媮\",\n \"兠\",\n \"舩\",\n \"慾\",\n \"綵\",\n \"腳\",\n \"𩓐\",\n \"夠\",\n \"豬\",\n \"貓\",\n \"湊\",\n \"減\",\n \"庻\",\n \"蔴\",\n \"菴\",\n \"朢\",\n \"睠\",\n \"觕\",\n \"麤\",\n \"釬\",\n \"銲\",\n \"痳\",\n \"殽\",\n \"婬\",\n \"滛\",\n \"湻\",\n \"㴱\",\n \"樑\",\n \"顇\",\n \"㝛\",\n \"窰\",\n \"窯\",\n \"琹\",\n \"欵\",\n \"墖\",\n \"趂\",\n \"隄\",\n \"愽\",\n \"揷\",\n \"揫\",\n \"煑\",\n \"朞\",\n \"㪚\",\n \"塟\",\n \"蔥\",\n \"蔕\",\n \"稜\",\n \"棊\",\n \"碁\",\n \"椶\",\n \"偪\",\n \"㕑\",\n \"廚\",\n \"廈\",\n \"鴈\",\n \"冣\",\n \"㝡\",\n \"晳\",\n \"鼃\",\n \"餧\",\n \"餵\",\n \"嗁\",\n \"諠\",\n \"㡌\",\n \"賸\",\n \"筴\",\n \"筞\",\n \"筩\",\n \"栰\",\n \"暠\",\n \"皜\",\n \"踰\",\n \"蝟\",\n \"㪟\",\n \"燄\",\n \"遊\",\n \"媿\",\n \"嘅\",\n \"庽\",\n \"窓\",\n \"牎\",\n \"牕\",\n \"窻\",\n \"徧\",\n \"僱\",\n \"帬\",\n \"裠\",\n \"強\",\n \"彊\",\n \"疎\",\n \"壻\",\n \"瓌\",\n \"䰟\",\n \"皷\",\n \"擕\",\n \"㩗\",\n \"㩦\",\n \"攜\",\n \"懃\",\n \"鞾\",\n \"幙\",\n \"㮣\",\n \"酧\",\n \"詶\",\n \"醻\",\n \"掽\",\n \"踫\",\n \"㼝\",\n \"盌\",\n \"磟\",\n \"覩\",\n \"倸\",\n \"㬉\",\n \"煗\",\n \"煖\",\n \"晻\",\n \"闇\",\n \"炤\",\n \"跥\",\n \"䗬\",\n \"蠭\",\n \"寘\",\n \"辠\",\n \"稺\",\n \"穉\",\n \"燬\",\n \"譭\",\n \"瘉\",\n \"癒\",\n \"顋\",\n \"骽\",\n \"猨\",\n \"蝯\",\n \"稟\",\n \"痺\",\n \"癡\",\n \"亷\",\n \"㢘\",\n \"韻\",\n \"泝\",\n \"遡\",\n \"昚\",\n \"躶\",\n \"臝\",\n \"羣\",\n \"㬪\",\n \"曡\",\n \"疊\",\n \"勦\",\n \"琍\",\n \"瓈\",\n \"𤋮\",\n \"熈\",\n \"牓\",\n \"搾\",\n \"謌\",\n \"堿\",\n \"鹻\",\n \"鹼\",\n \"矁\",\n \"燻\",\n \"髈\",\n \"𤺥\",\n \"辢\",\n \"旂\",\n \"𡚁\",\n \"潄\",\n \"砦\",\n \"詧\",\n \"嫰\",\n \"櫈\",\n \"撐\",\n \"墪\",\n \"譔\",\n \"鞵\",\n \"鞌\",\n \"蕋\",\n \"橤\",\n \"蘂\",\n \"醕\",\n \"譆\",\n \"跴\",\n \"蹤\",\n \"蜨\",\n \"蠍\",\n \"稾\",\n \"殭\",\n \"惪\",\n \"厀\",\n \"襃\",\n \"癅\",\n \"䊀\",\n \"餬\",\n \"潛\",\n \"癄\",\n \"顦\",\n \"鷰\",\n \"藷\",\n \"櫥\",\n \"螎\",\n \"蹏\",\n \"蟇\",\n \"譟\",\n \"簒\",\n \"彫\",\n \"琱\",\n \"鵰\",\n \"餹\",\n \"餻\",\n \"簷\",\n \"粦\",\n \"燐\",\n \"緐\",\n \"幑\",\n \"蹧\",\n \"粇\",\n \"穅\",\n \"臋\",\n \"籐\",\n \"繙\",\n \"飜\",\n \"孼\",\n \"蠏\",\n \"燿\",\n \"蝡\",\n \"稬\",\n \"穤\",\n \"惷\",\n \"覇\",\n \"鑵\",\n \"戹\",\n \"阨\",\n \"剳\",\n \"帀\",\n \"巵\",\n \"亙\",\n \"佇\",\n \"竚\",\n \"穽\",\n \"岅\",\n \"虯\",\n \"𦍑\",\n \"羗\",\n \"啎\",\n \"姙\",\n \"㘭\",\n \"袟\",\n \"袠\",\n \"逈\",\n \"㒺\",\n \"犛\",\n \"氂\",\n \"偘\",\n \"甕\",\n \"罋\",\n \"冺\",\n \"姍\",\n \"蝨\",\n \"琺\",\n \"瑇\",\n \"尅\",\n \"梔\",\n \"斮\",\n \"斲\",\n \"斵\",\n \"暱\",\n \"毘\",\n \"蝱\",\n \"吚\",\n \"哶\",\n \"峝\",\n \"粃\",\n \"竢\",\n \"狥\",\n \"秈\",\n \"烱\",\n \"㳄\",\n \"袵\",\n \"盇\",\n \"涖\",\n \"蒞\",\n \"碪\",\n \"蠔\",\n \"唕\",\n \"倐\",\n \"儵\",\n \"雋\",\n \"皐\",\n \"臯\",\n \"衂\",\n \"䶊\",\n \"臙\",\n \"獧\",\n \"痾\",\n \"皰\",\n \"湼\",\n \"澣\",\n \"濬\",\n \"塚\",\n \"襢\",\n \"娿\",\n \"勅\",\n \"勑\",\n \"戞\",\n \"廐\",\n \"廄\",\n \"眥\",\n \"覜\",\n \"勗\",\n \"啗\",\n \"噉\",\n \"傯\",\n \"挱\",\n \"㥫\",\n \"惥\",\n \"慂\",\n \"陻\",\n \"蕚\",\n \"萲\",\n \"蕿\",\n \"蘐\",\n \"藼\",\n \"櫂\",\n \"箠\",\n \"槨\",\n \"啑\",\n \"蹠\",\n \"蚘\",\n \"痐\",\n \"蛕\",\n \"蜖\",\n \"瘖\",\n \"遯\",\n \"醃\",\n \"飱\",\n \"冪\",\n \"簑\",\n \"枏\",\n \"柟\",\n \"檝\",\n \"楥\",\n \"矴\",\n \"椗\",\n \"嘷\",\n \"獋\",\n \"粺\",\n \"䈰\",\n \"諐\",\n \"齶\",\n \"堘\",\n \"疿\",\n \"雝\",\n \"秔\",\n \"稉\",\n \"槀\",\n \"搉\",\n \"廝\",\n \"叡\",\n \"嘠\",\n \"蜋\",\n \"筯\",\n \"篛\",\n \"麞\",\n \"糉\",\n \"緥\",\n \"璿\",\n \"髥\",\n \"臕\",\n \"餈\",\n \"剹\",\n \"橜\",\n \"罇\",\n \"蜺\",\n \"矙\",\n \"憇\",\n \"翺\",\n \"饍\",\n \"瞖\",\n \"羴\",\n \"羶\",\n \"爕\",\n \"繦\",\n \"騌\",\n \"鬉\",\n \"騣\",\n \"蔾\",\n \"䠀\",\n \"簮\",\n \"躕\",\n \"蹵\",\n \"䝔\",\n \"貛\",\n \"鼴\",\n \"麐\",\n \"塡\",\n \"あ\",\n \"い\",\n \"う\",\n \"え\",\n \"お\",\n \"か\",\n \"き\",\n \"く\",\n \"け\",\n \"こ\",\n \"さ\",\n \"し\",\n \"す\",\n \"せ\",\n \"そ\",\n \"た\",\n \"ち\",\n \"つ\",\n \"て\",\n \"と\",\n \"な\",\n \"に\",\n \"ぬ\",\n \"ね\",\n \"の\",\n \"は\",\n \"ひ\",\n \"ふ\",\n \"へ\",\n \"ほ\",\n \"ま\",\n \"み\",\n \"む\",\n \"め\",\n \"も\",\n \"や\",\n \"ゆ\",\n \"よ\",\n \"ら\",\n \"り\",\n \"る\",\n \"れ\",\n \"ろ\",\n \"わ\",\n \"を\",\n \"ん\",\n \"が\",\n \"ぎ\",\n \"ぐ\",\n \"げ\",\n \"ご\",\n \"ざ\",\n \"じ\",\n \"ず\",\n \"ぜ\",\n \"ぞ\",\n \"だ\",\n \"ぢ\",\n \"づ\",\n \"で\",\n \"ど\",\n \"ば\",\n \"び\",\n \"ぶ\",\n \"べ\",\n \"ぼ\",\n \"ぱ\",\n \"ぴ\",\n \"ぷ\",\n \"ぺ\",\n \"ぽ\",\n \"ぁ\",\n \"ぃ\",\n \"ぅ\",\n \"ぇ\",\n \"ぉ\",\n \"っ\",\n \"ゃ\",\n \"ゅ\",\n \"ょ\",\n \"ゎ\",\n \"ゕ\",\n \"ゖ\",\n \"ア\",\n \"イ\",\n \"ウ\",\n \"エ\",\n \"オ\",\n \"カ\",\n \"キ\",\n \"ク\",\n \"ケ\",\n \"コ\",\n \"サ\",\n \"シ\",\n \"ス\",\n \"セ\",\n \"ソ\",\n \"タ\",\n \"チ\",\n \"ツ\",\n \"テ\",\n \"ト\",\n \"ナ\",\n \"ニ\",\n \"ヌ\",\n \"ネ\",\n \"ノ\",\n \"ハ\",\n \"ヒ\",\n \"フ\",\n \"ヘ\",\n \"ホ\",\n \"マ\",\n \"ミ\",\n \"ム\",\n \"メ\",\n \"モ\",\n \"ヤ\",\n \"ユ\",\n \"ヨ\",\n \"ラ\",\n \"リ\",\n \"ル\",\n \"レ\",\n \"ロ\",\n \"ワ\",\n \"ヲ\",\n \"ン\",\n \"ガ\",\n \"ギ\",\n \"グ\",\n \"ゲ\",\n \"ゴ\",\n \"ザ\",\n \"ジ\",\n \"ズ\",\n \"ゼ\",\n \"ゾ\",\n \"ダ\",\n \"ヂ\",\n \"ヅ\",\n \"デ\",\n \"ド\",\n \"バ\",\n \"ビ\",\n \"ブ\",\n \"ベ\",\n \"ボ\",\n \"パ\",\n \"ピ\",\n \"プ\",\n \"ペ\",\n \"ポ\",\n \"ァ\",\n \"ィ\",\n \"ゥ\",\n \"ェ\",\n \"ォ\",\n \"ッ\",\n \"ャ\",\n \"ュ\",\n \"ョ\",\n \"ヮ\",\n \"ヵ\",\n \"ヶ\",\n \"ヷ\",\n \"ヸ\",\n \"ヹ\",\n \"ヺ\",\n \"・\",\n \"ー\",\n \"ヽ\",\n \"ヾ\",\n \"ヿ\",\n \"ｱ\",\n \"ｲ\",\n \"ｳ\",\n \"ｴ\",\n \"ｵ\",\n \"ｶ\",\n \"ｷ\",\n \"ｸ\",\n \"ｹ\",\n \"ｺ\",\n \"ｻ\",\n \"ｼ\",\n \"ｽ\",\n \"ｾ\",\n \"ｿ\",\n \"ﾀ\",\n \"ﾁ\",\n \"ﾂ\",\n \"ﾃ\",\n \"ﾄ\",\n \"ﾅ\",\n \"ﾆ\",\n \"ﾇ\",\n \"ﾈ\",\n \"ﾉ\",\n \"ﾊ\",\n \"ﾋ\",\n \"ﾌ\",\n \"ﾍ\",\n \"ﾎ\",\n \"ﾏ\",\n \"ﾐ\",\n \"ﾑ\",\n \"ﾒ\",\n \"ﾓ\",\n \"ﾔ\",\n \"ﾕ\",\n \"ﾖ\",\n \"ﾗ\",\n \"ﾘ\",\n \"ﾙ\",\n \"ﾚ\",\n \"ﾛ\",\n \"ﾜ\",\n \"ｦ\",\n \"ﾝ\",\n \"ﾞ\",\n \"ﾟ\",\n \"ｧ\",\n \"ｨ\",\n \"ｩ\",\n \"ｪ\",\n \"ｫ\",\n \"ｯ\",\n \"ｬ\",\n \"ｭ\",\n \"ｮ\",\n \"円\",\n \"気\",\n \"糸\",\n \"絵\",\n \"楽\",\n \"帰\",\n \"戸\",\n \"広\",\n \"黒\",\n \"図\",\n \"線\",\n \"読\",\n \"売\",\n \"歩\",\n \"毎\",\n \"亜\",\n \"悪\",\n \"圧\",\n \"扱\",\n \"囲\",\n \"為\",\n \"壱\",\n \"隠\",\n \"栄\",\n \"営\",\n \"駅\",\n \"塩\",\n \"縁\",\n \"艶\",\n \"応\",\n \"桜\",\n \"穏\",\n \"仮\",\n \"価\",\n \"箇\",\n \"ゑ\",\n \"ゝ\",\n \"ゞ\",\n \"ヰ\",\n \"ヴ\",\n \"㈱\",\n \"両\",\n \"丼\",\n \"丿\",\n \"亀\",\n \"仏\",\n \"伝\",\n \"侶\",\n \"俤\",\n \"値\",\n \"倶\",\n \"倹\",\n \"偐\",\n \"偽\",\n \"働\",\n \"儛\",\n \"兌\",\n \"児\",\n \"冑\",\n \"冨\",\n \"凞\",\n \"処\",\n \"凪\",\n \"別\",\n \"剣\",\n \"剤\",\n \"剰\",\n \"劔\",\n \"労\",\n \"勧\",\n \"勲\",\n \"匁\",\n \"匂\",\n \"匲\",\n \"卍\",\n \"単\",\n \"厳\",\n \"収\",\n \"呂\",\n \"呉\",\n \"呑\",\n \"呰\",\n \"唖\",\n \"喚\",\n \"喩\",\n \"喰\",\n \"噛\",\n \"噺\",\n \"嚢\",\n \"囃\",\n \"団\",\n \"圀\",\n \"圏\",\n \"堀\",\n \"堺\",\n \"塀\",\n \"塁\",\n \"塙\",\n \"増\",\n \"墺\",\n \"壊\",\n \"壌\",\n \"壷\",\n \"変\",\n \"奨\",\n \"姫\",\n \"娯\",\n \"嫐\",\n \"嬢\",\n \"嬾\",\n \"孁\",\n \"宍\",\n \"実\",\n \"宮\",\n \"寔\",\n \"寛\",\n \"対\",\n \"専\",\n \"尭\",\n \"峠\",\n \"崋\",\n \"嶋\",\n \"巀\",\n \"巌\",\n \"巣\",\n \"巻\",\n \"帯\",\n \"幇\",\n \"庁\",\n \"廃\",\n \"廻\",\n \"弉\",\n \"弌\",\n \"弐\",\n \"弖\",\n \"弾\",\n \"従\",\n \"徳\",\n \"徴\",\n \"忯\",\n \"恵\",\n \"悩\",\n \"惣\",\n \"懐\",\n \"懽\",\n \"戦\",\n \"戯\",\n \"戻\",\n \"払\",\n \"抜\",\n \"択\",\n \"拝\",\n \"拠\",\n \"拡\",\n \"拵\",\n \"挙\",\n \"挿\",\n \"捗\",\n \"捜\",\n \"掟\",\n \"掲\",\n \"掻\",\n \"揃\",\n \"換\",\n \"揺\",\n \"摂\",\n \"撃\",\n \"撹\",\n \"斉\",\n \"斎\",\n \"旛\",\n \"旡\",\n \"晧\",\n \"晩\",\n \"暁\",\n \"暦\",\n \"曽\",\n \"杁\",\n \"杢\",\n \"杣\",\n \"杮\",\n \"枓\",\n \"枠\",\n \"枡\",\n \"柾\",\n \"栂\",\n \"栃\",\n \"桝\",\n \"桟\",\n \"桾\",\n \"梛\",\n \"梱\",\n \"梲\",\n \"梶\",\n \"椙\",\n \"検\",\n \"椥\",\n \"楕\",\n \"楡\",\n \"楢\",\n \"榊\",\n \"榎\",\n \"槇\",\n \"様\",\n \"槙\",\n \"槻\",\n \"樋\",\n \"権\",\n \"樫\",\n \"橿\",\n \"檥\",\n \"欅\",\n \"歎\",\n \"歓\",\n \"歯\",\n \"歳\",\n \"歴\",\n \"毀\",\n \"沖\",\n \"沢\",\n \"浄\",\n \"涙\",\n \"済\",\n \"渉\",\n \"渋\",\n \"渓\",\n \"渕\",\n \"満\",\n \"滝\",\n \"漑\",\n \"潅\",\n \"澁\",\n \"瀞\",\n \"瀬\",\n \"焔\",\n \"焼\",\n \"煇\",\n \"煕\",\n \"煥\",\n \"燗\",\n \"爼\",\n \"犠\",\n \"狛\",\n \"猟\",\n \"獏\",\n \"獣\",\n \"珊\",\n \"瑤\",\n \"甞\",\n \"畑\",\n \"畠\",\n \"畳\",\n \"畷\",\n \"畺\",\n \"痩\",\n \"癪\",\n \"発\",\n \"県\",\n \"眞\",\n \"砕\",\n \"碕\",\n \"礒\",\n \"禖\",\n \"禿\",\n \"稲\",\n \"穂\",\n \"穣\",\n \"竃\",\n \"竜\",\n \"竴\",\n \"笹\",\n \"筈\",\n \"筬\",\n \"筰\",\n \"箆\",\n \"箏\",\n \"箙\",\n \"篠\",\n \"篭\",\n \"簺\",\n \"籾\",\n \"粂\",\n \"粋\",\n \"粛\",\n \"粧\",\n \"糺\",\n \"紬\",\n \"絁\",\n \"経\",\n \"絖\",\n \"絣\",\n \"絽\",\n \"継\",\n \"続\",\n \"綟\",\n \"総\",\n \"縄\",\n \"縅\",\n \"縒\",\n \"縦\",\n \"繊\",\n \"繋\",\n \"繍\",\n \"繝\",\n \"繧\",\n \"纐\",\n \"纒\",\n \"罠\",\n \"罧\",\n \"罵\",\n \"羂\",\n \"羇\",\n \"羨\",\n \"聟\",\n \"聡\",\n \"聨\",\n \"聴\",\n \"脇\",\n \"脳\",\n \"膣\",\n \"膵\",\n \"臈\",\n \"臓\",\n \"臥\",\n \"舎\",\n \"舖\",\n \"舗\",\n \"舘\",\n \"芿\",\n \"苅\",\n \"茲\",\n \"荊\",\n \"荘\",\n \"莬\",\n \"莵\",\n \"菫\",\n \"萠\",\n \"蔵\",\n \"薗\",\n \"薫\",\n \"薬\",\n \"薭\",\n \"蘊\",\n \"蛍\",\n \"蝋\",\n \"蝿\",\n \"蟷\",\n \"衞\",\n \"衵\",\n \"袙\",\n \"袞\",\n \"袰\",\n \"袴\",\n \"袿\",\n \"裃\",\n \"裡\",\n \"裲\",\n \"褄\",\n \"褌\",\n \"襴\",\n \"襷\",\n \"覗\",\n \"覚\",\n \"覧\",\n \"観\",\n \"訳\",\n \"証\",\n \"諌\",\n \"諚\",\n \"諟\",\n \"諡\",\n \"諮\",\n \"譛\",\n \"譲\",\n \"讃\",\n \"豅\",\n \"豊\",\n \"豎\",\n \"賎\",\n \"賛\",\n \"贔\",\n \"躙\",\n \"躰\",\n \"転\",\n \"軽\",\n \"輌\",\n \"辥\",\n \"辺\",\n \"辻\",\n \"込\",\n \"逓\",\n \"遅\",\n \"遙\",\n \"邉\",\n \"郷\",\n \"酔\",\n \"醗\",\n \"醤\",\n \"醸\",\n \"釈\",\n \"鉄\",\n \"鉇\",\n \"鉤\",\n \"鉱\",\n \"鉾\",\n \"銈\",\n \"銕\",\n \"銭\",\n \"鋲\",\n \"鋳\",\n \"鋺\",\n \"錆\",\n \"錍\",\n \"錣\",\n \"錬\",\n \"錵\",\n \"鍑\",\n \"鍮\",\n \"鍼\",\n \"鎌\",\n \"鎗\",\n \"鎚\",\n \"鎹\",\n \"鐇\",\n \"鐚\",\n \"鐡\",\n \"鑁\",\n \"鑑\",\n \"鑚\",\n \"鑢\",\n \"閇\",\n \"関\",\n \"閦\",\n \"闘\",\n \"陥\",\n \"険\",\n \"隣\",\n \"隷\",\n \"雑\",\n \"雫\",\n \"霊\",\n \"靜\",\n \"靫\",\n \"靭\",\n \"靱\",\n \"鞄\",\n \"鞆\",\n \"頚\",\n \"頬\",\n \"頴\",\n \"頼\",\n \"顕\",\n \"顗\",\n \"餝\",\n \"饂\",\n \"駄\",\n \"駆\",\n \"駈\",\n \"騒\",\n \"験\",\n \"騨\",\n \"髄\",\n \"髙\",\n \"髪\",\n \"髷\",\n \"鯖\",\n \"鯰\",\n \"鯱\",\n \"鰒\",\n \"鰯\",\n \"鰰\",\n \"鳰\",\n \"鴎\",\n \"鴫\",\n \"鵄\",\n \"鵞\",\n \"鵺\",\n \"鶏\",\n \"鹸\",\n \"麁\",\n \"麺\",\n \"麿\",\n \"黌\",\n \"黙\",\n \"鼈\",\n \"齢\",\n \"龗\",\n \"縯\",\n \"蟅\",\n \"坖\",\n \"祂\",\n \"鼂\",\n \"鱚\",\n \"蛻\",\n \"屌\",\n \"呾\",\n \"煔\",\n \"吶\",\n \"扥\",\n \"蚖\",\n \"銂\",\n \"尃\",\n \"夋\",\n \"鵼\",\n \"徬\",\n \"寳\",\n \"彡\",\n \"舨\",\n \"湳\",\n \"麼\",\n \"鍈\",\n \"崈\",\n \"鱣\",\n \"盺\",\n \"拺\",\n \"瑥\",\n \"茷\",\n \"焻\",\n \"奀\",\n \"驎\",\n \"鱰\",\n \"砢\",\n \"痟\",\n \"廱\",\n \"僜\",\n \"瘺\",\n \"鱊\",\n \"擥\",\n \"嶰\",\n \"淓\",\n \"跅\",\n \"浵\",\n \"媗\",\n \"璦\",\n \"煠\",\n \"檊\",\n \"媃\",\n \"峅\",\n \"躄\",\n \"鉟\",\n \"塽\",\n \"蟴\",\n \"鯮\",\n \"弍\",\n \"烒\",\n \"鵵\",\n \"妑\",\n \"孋\",\n \"蚡\",\n \"恊\",\n \"輭\",\n \"廞\",\n \"產\",\n \"曅\",\n \"盜\",\n \"騤\",\n \"囪\",\n \"鱀\",\n \"茇\",\n \"葊\",\n \"逹\",\n \"狓\",\n \"崢\",\n \"趖\",\n \"凃\",\n \"羙\",\n \"鮸\",\n \"昞\",\n \"楿\",\n \"渽\",\n \"圗\",\n \"麪\",\n \"屇\",\n \"鍉\",\n \"葝\",\n \"沯\",\n \"爭\",\n \"幵\",\n \"筭\",\n \"寊\",\n \"銋\",\n \"貮\",\n \"鎭\",\n \"熺\",\n \"昜\",\n \"鍱\",\n \"墬\",\n \"愒\",\n \"磺\",\n \"嚈\",\n \"稘\",\n \"珮\",\n \"釆\",\n \"殑\",\n \"鍩\",\n \"䲁\",\n \"蕷\",\n \"鐿\",\n \"僡\",\n \"佹\",\n \"輶\",\n \"冴\",\n \"襶\",\n \"賔\",\n \"猙\",\n \"辧\",\n \"絛\",\n \"磾\",\n \"韁\",\n \"螔\",\n \"譳\",\n \"礑\",\n \"鋱\",\n \"魩\",\n \"嚗\",\n \"棆\",\n \"牆\",\n \"敟\",\n \"柶\",\n \"瓛\",\n \"魣\",\n \"巎\",\n \"轘\",\n \"襌\",\n \"枼\",\n \"鸌\",\n \"逺\",\n \"錏\",\n \"縡\",\n \"帢\",\n \"騄\",\n \"媼\",\n \"埅\",\n \"鄤\",\n \"萐\",\n \"祙\",\n \"旼\",\n \"詥\",\n \"鶲\",\n \"燉\",\n \"卲\",\n \"銱\",\n \"庲\",\n \"伱\",\n \"氽\",\n \"嵿\",\n \"挻\",\n \"煵\",\n \"窋\",\n \"鐤\",\n \"鮊\",\n \"鱬\",\n \"鰧\",\n \"嬤\",\n \"譞\",\n \"諲\",\n \"脭\",\n \"悳\",\n \"崘\",\n \"阭\",\n \"內\",\n \"袾\",\n \"冚\",\n \"壐\",\n \"咗\",\n \"礠\",\n \"孮\",\n \"痲\",\n \"埈\",\n \"肹\",\n \"鰮\",\n \"鮓\",\n \"濊\",\n \"塜\",\n \"凜\",\n \"蒢\",\n \"噰\",\n \"桼\",\n \"峍\",\n \"焴\",\n \"鶒\",\n \"鋮\",\n \"綠\",\n \"鶹\",\n \"熿\",\n \"毴\",\n \"咟\",\n \"嘥\",\n \"睺\",\n \"繡\",\n \"郎\",\n \"瘞\",\n \"鉶\",\n \"蔎\",\n \"秠\",\n \"緤\",\n \"蝀\",\n \"躝\",\n \"蟜\",\n \"繃\",\n \"囮\",\n \"墫\",\n \"乭\",\n \"胊\",\n \"濙\",\n \"瘓\",\n \"榣\",\n \"鑛\",\n \"鐫\",\n \"嶴\",\n \"甹\",\n \"坮\",\n \"銾\",\n \"蒭\",\n \"睜\",\n \"俋\",\n \"餠\",\n \"榢\",\n \"蓳\",\n \"盋\",\n \"堷\",\n \"鍏\",\n \"苝\",\n \"巛\",\n \"蚵\",\n \"暏\",\n \"熤\",\n \"嬨\",\n \"墎\",\n \"鏽\",\n \"戶\",\n \"菺\",\n \"膮\",\n \"熖\",\n \"睪\",\n \"栜\",\n \"捱\",\n \"榗\",\n \"鍷\",\n \"曧\",\n \"犽\",\n \"韑\",\n \"袓\",\n \"䖝\",\n \"焄\",\n \"喦\",\n \"髲\",\n \"疌\",\n \"㴪\",\n \"侊\",\n \"貐\",\n \"蕅\",\n \"禠\",\n \"蕑\",\n \"囯\",\n \"暊\",\n \"儞\",\n \"佋\",\n \"柎\",\n \"㐱\",\n \"鰤\",\n \"苳\",\n \"鱥\",\n \"謤\",\n \"遶\",\n \"眀\",\n \"鑀\",\n \"羋\",\n \"顏\",\n \"陜\",\n \"銩\",\n \"黶\",\n \"苼\",\n \"蒤\",\n \"棛\",\n \"儫\",\n \"咁\",\n \"抦\",\n \"衚\",\n \"棩\",\n \"焿\",\n \"脫\",\n \"麅\",\n \"玏\",\n \"埧\",\n \"淸\",\n \"黁\",\n \"淽\",\n \"彠\",\n \"鮨\",\n \"沜\",\n \"糀\",\n \"厓\",\n \"楧\",\n \"嶌\",\n \"簹\",\n \"檵\",\n \"鱇\",\n \"嶬\",\n \"廸\",\n \"卽\",\n \"樀\",\n \"贌\",\n \"酼\",\n \"籛\",\n \"沒\",\n \"晸\",\n \"諪\",\n \"蕡\",\n \"妏\",\n \"鄋\",\n \"蒍\",\n \"奧\",\n \"抇\",\n \"蓨\",\n \"薆\",\n \"鱷\",\n \"巘\",\n \"䝉\",\n \"亰\",\n \"寈\",\n \"槩\",\n \"誒\",\n \"麴\",\n \"蕟\",\n \"溎\",\n \"蘗\",\n \"榦\",\n \"斿\",\n \"暟\",\n \"炲\",\n \"拚\",\n \"娖\",\n \"繖\",\n \"橚\",\n \"寜\",\n \"爀\",\n \"饟\",\n \"悅\",\n \"鯏\",\n \"彜\",\n \"眾\",\n \"葯\",\n \"嬝\",\n \"埮\",\n \"獇\",\n \"馛\",\n \"溙\",\n \"瀦\",\n \"熼\",\n \"硓\",\n \"鈢\",\n \"樆\",\n \"輬\",\n \"鰜\",\n \"蔘\",\n \"渙\",\n \"澔\",\n \"嗮\",\n \"旉\",\n \"籜\",\n \"媊\",\n \"燘\",\n \"儚\",\n \"頹\",\n \"缽\",\n \"俽\",\n \"逨\",\n \"鱓\",\n \"郞\",\n \"歊\",\n \"杴\",\n \"珡\",\n \"杋\",\n \"醁\",\n \"鰏\",\n \"鵾\",\n \"鐽\",\n \"鮋\",\n \"巶\",\n \"荅\",\n \"薾\",\n \"囓\",\n \"蹻\",\n \"獎\",\n \"禑\",\n \"鎓\",\n \"榲\",\n \"僴\",\n \"綞\",\n \"尓\",\n \"敭\",\n \"曔\",\n \"褔\",\n \"鬅\",\n \"亊\",\n \"鏦\",\n \"蓘\",\n \"裬\",\n \"鱲\",\n \"薡\",\n \"鰗\",\n \"箑\",\n \"鬪\",\n \"縂\",\n \"璸\",\n \"甙\",\n \"茮\",\n \"辵\",\n \"岻\",\n \"覿\",\n \"滈\",\n \"鯶\",\n \"鑂\",\n \"囶\",\n \"舺\",\n \"溋\",\n \"拋\",\n \"菾\",\n \"敾\",\n \"虨\",\n \"綝\",\n \"蝍\",\n \"醂\",\n \"禨\",\n \"賹\",\n \"廧\",\n \"絕\",\n \"槗\",\n \"徫\",\n \"鎔\",\n \"曮\",\n \"蠂\",\n \"捒\",\n \"堈\",\n \"莕\",\n \"蓪\",\n \"敎\",\n \"禃\",\n \"櫱\",\n \"綧\",\n \"瀶\",\n \"逌\",\n \"浤\",\n \"碻\",\n \"刄\",\n \"逤\",\n \"剏\",\n \"氹\",\n \"菈\",\n \"娫\",\n \"蜛\",\n \"嵗\",\n \"糎\",\n \"螶\",\n \"譓\",\n \"鏳\",\n \"嵙\",\n \"瑊\",\n \"隲\",\n \"檨\",\n \"緈\",\n \"畵\",\n \"砯\",\n \"簗\",\n \"彅\",\n \"鰺\",\n \"騋\",\n \"窶\",\n \"嚒\",\n \"嵻\",\n \"尙\",\n \"頵\",\n \"槰\",\n \"虉\",\n \"醞\",\n \"巂\",\n \"彔\",\n \"偊\",\n \"畇\",\n \"鱨\",\n \"妸\",\n \"塲\",\n \"畐\",\n \"鈫\",\n \"錟\",\n \"磪\",\n \"摠\",\n \"彥\",\n \"璙\",\n \"囝\",\n \"寗\",\n \"耎\",\n \"鮡\",\n \"蘓\",\n \"弅\",\n \"焃\",\n \"飥\",\n \"戙\",\n \"塰\",\n \"儱\",\n \"槺\",\n \"噏\",\n \"魟\",\n \"禵\",\n \"佧\",\n \"咘\",\n \"盪\",\n \"瑈\",\n \"鉲\",\n \"睭\",\n \"鏌\",\n \"鼇\",\n \"郋\",\n \"魮\",\n \"朖\",\n \"滽\",\n \"渃\",\n \"滙\",\n \"熯\",\n \"醿\",\n \"鎅\",\n \"褀\",\n \"鬬\",\n \"巄\",\n \"螥\",\n \"眜\",\n \"釚\",\n \"柉\",\n \"壎\",\n \"峇\",\n \"姸\",\n \"唭\",\n \"鮜\",\n \"鈖\",\n \"嫈\",\n \"壄\",\n \"洤\",\n \"黃\",\n \"伕\",\n \"堦\",\n \"嶔\",\n \"鮰\",\n \"鞞\",\n \"漎\",\n \"鉓\",\n \"鮗\",\n \"壴\",\n \"阝\",\n \"妀\",\n \"矽\",\n \"獢\",\n \"倗\",\n \"銪\",\n \"鴓\",\n \"橒\",\n \"凈\",\n \"哖\",\n \"屚\",\n \"偍\",\n \"瑺\",\n \"媯\",\n \"淍\",\n \"驌\",\n \"椇\",\n \"赬\",\n \"薐\",\n \"糹\",\n \"碽\",\n \"濲\",\n \"釭\",\n \"晭\",\n \"纕\",\n \"寖\",\n \"閞\",\n \"歿\",\n \"呎\",\n \"鶆\",\n \"屄\",\n \"櫿\",\n \"犎\",\n \"旲\",\n \"㙟\",\n \"龎\",\n \"翜\",\n \"螾\",\n \"說\",\n \"衜\",\n \"泆\",\n \"軎\",\n \"鵂\",\n \"荎\",\n \"嚧\",\n \"硂\",\n \"桖\",\n \"褭\",\n \"筊\",\n \"鰷\",\n \"秳\",\n \"戩\",\n \"轀\",\n \"鬹\",\n \"飬\",\n \"卋\",\n \"暸\",\n \"狦\",\n \"搢\",\n \"娋\",\n \"鏴\",\n \"溫\",\n \"毉\",\n \"淰\",\n \"謩\",\n \"餺\",\n \"鵙\",\n \"鳽\",\n \"鮀\",\n \"狶\",\n \"氻\",\n \"轝\",\n \"妺\",\n \"袛\",\n \"蓭\",\n \"梂\",\n \"娛\",\n \"牼\",\n \"稅\",\n \"兿\",\n \"玾\",\n \"煚\",\n \"僩\",\n \"鶿\",\n \"鬄\",\n \"崠\",\n \"鉆\",\n \"鯓\",\n \"蚢\",\n \"庀\",\n \"鵟\",\n \"坣\",\n \"殼\",\n \"悞\",\n \"熅\",\n \"敻\",\n \"鍠\",\n \"曶\",\n \"愼\",\n \"搳\",\n \"姃\",\n \"砳\",\n \"槼\",\n \"臞\",\n \"韾\",\n \"靑\",\n \"鸊\",\n \"薲\",\n \"虛\",\n \"蠄\",\n \"啟\",\n \"鶺\",\n \"苺\",\n \"滾\",\n \"褞\",\n \"仺\",\n \"胇\",\n \"憻\",\n \"郳\",\n \"烉\",\n \"驩\",\n \"冇\",\n \"枖\",\n \"夌\",\n \"搵\",\n \"匸\",\n \"盨\",\n \"櫾\",\n \"霤\",\n \"麊\",\n \"貒\",\n \"噓\",\n \"嗢\",\n \"笩\",\n \"晈\",\n \"冂\",\n \"銳\",\n \"毿\",\n \"慜\",\n \"囧\",\n \"閜\",\n \"娸\",\n \"庢\",\n \"壆\",\n \"馯\",\n \"桱\",\n \"兗\",\n \"葃\",\n \"侅\",\n \"煐\",\n \"鐦\",\n \"藸\",\n \"鷎\",\n \"嵰\",\n \"逎\",\n \"弒\",\n \"匋\",\n \"鐭\",\n \"廔\",\n \"砩\",\n \"孆\",\n \"灴\",\n \"伷\",\n \"兪\",\n \"鴗\",\n \"澯\",\n \"幚\",\n \"旙\",\n \"勻\",\n \"礽\",\n \"婑\",\n \"鱮\",\n \"娍\",\n \"銶\",\n \"吳\",\n \"鍟\",\n \"仼\",\n \"鳧\",\n \"彞\",\n \"娽\",\n \"昛\",\n \"鰼\",\n \"剎\",\n \"佉\",\n \"鉏\",\n \"偸\",\n \"鰆\",\n \"讙\",\n \"橪\",\n \"啱\",\n \"岀\",\n \"孻\",\n \"釪\",\n \"乹\",\n \"鈳\",\n \"漇\",\n \"檦\",\n \"埻\",\n \"祿\",\n \"爌\",\n \"禇\",\n \"鱵\",\n \"㸃\",\n \"梉\",\n \"燝\",\n \"霙\",\n \"炁\",\n \"飮\",\n \"蠙\",\n \"勷\",\n \"鵎\",\n \"儥\",\n \"鐠\",\n \"唻\",\n \"廰\",\n \"嚿\",\n \"嵕\",\n \"墱\",\n \"紑\",\n \"搖\",\n \"瘜\",\n \"皝\",\n \"鸑\",\n \"瀁\",\n \"粵\",\n \"撚\",\n \"巑\",\n \"梀\",\n \"啯\",\n \"眛\",\n \"諴\",\n \"夊\",\n \"僙\",\n \"鍝\",\n \"裖\",\n \"鮣\",\n \"凬\",\n \"飡\",\n \"灊\",\n \"橓\",\n \"嫳\",\n \"筳\",\n \"咑\",\n \"粍\",\n \"瓑\",\n \"璌\",\n \"伃\",\n \"閰\",\n \"傜\",\n \"黐\",\n \"謢\",\n \"驒\",\n \"橫\",\n \"蛯\",\n \"寕\",\n \"蠵\",\n \"瞓\",\n \"旳\",\n \"翏\",\n \"硏\",\n \"寯\",\n \"韡\",\n \"楤\",\n \"鰃\",\n \"朿\",\n \"侞\",\n \"鵯\",\n \"愨\",\n \"祹\",\n \"厔\",\n \"丌\",\n \"盩\",\n \"謏\",\n \"魕\",\n \"啣\",\n \"閱\",\n \"曺\",\n \"枛\",\n \"罉\",\n \"卐\",\n \"樻\",\n \"鷉\",\n \"鯒\",\n \"鋡\",\n \"磱\",\n \"枱\",\n \"攴\",\n \"蠷\",\n \"穈\",\n \"嚟\",\n \"檽\",\n \"趐\",\n \"奐\",\n \"鋐\",\n \"檇\",\n \"薀\",\n \"峼\",\n \"咭\",\n \"訔\",\n \"韠\",\n \"鑴\",\n \"鸐\",\n \"唃\",\n \"捦\",\n \"鸜\",\n \"誴\",\n \"罳\",\n \"璄\",\n \"暃\",\n \"夀\",\n \"賨\",\n \"鞥\",\n \"鈊\",\n \"灡\",\n \"鮍\",\n \"懮\",\n \"籣\",\n \"昐\",\n \"陁\",\n \"襾\",\n \"鮠\",\n \"鈏\",\n \"囍\",\n \"婯\",\n \"艔\",\n \"貭\",\n \"䰾\",\n \"姁\",\n \"禼\",\n \"堖\",\n \"鋶\",\n \"仛\",\n \"鏷\",\n \"謜\",\n \"鑅\",\n \"忬\",\n \"蘶\",\n \"謠\",\n \"觙\",\n \"奫\",\n \"狟\",\n \"泩\",\n \"桙\",\n \"飈\",\n \"垰\",\n \"啍\",\n \"嚞\",\n \"鯕\",\n \"蒧\",\n \"榞\",\n \"徸\",\n \"璹\",\n \"揔\",\n \"欉\",\n \"魞\",\n \"菶\",\n \"玧\",\n \"鳯\",\n \"廍\",\n \"侚\",\n \"岰\",\n \"岧\",\n \"鋕\",\n \"凵\",\n \"彣\",\n \"崱\",\n \"媜\",\n \"倢\",\n \"鵐\",\n \"砋\",\n \"鷚\",\n \"鱠\",\n \"鮻\",\n \"繻\",\n \"摵\",\n \"贓\",\n \"磵\",\n \"錻\",\n \"痠\",\n \"粩\",\n \"胅\",\n \"奣\",\n \"塨\",\n \"瀠\",\n \"鸘\",\n \"啚\",\n \"娳\",\n \"霶\",\n \"壔\",\n \"峚\",\n \"甂\",\n \"廁\",\n \"覌\",\n \"鰂\",\n \"猳\",\n \"鱻\",\n \"盫\",\n \"裿\",\n \"杬\",\n \"歛\",\n \"澋\",\n \"蘞\",\n \"嵜\",\n \"尐\",\n \"旽\",\n \"鉌\",\n \"鎛\",\n \"豿\",\n \"凖\",\n \"榤\",\n \"禓\",\n \"龝\",\n \"悧\",\n \"鷟\",\n \"鮟\",\n \"吋\",\n \"喢\",\n \"岪\",\n \"吥\",\n \"漵\",\n \"頠\",\n \"豔\",\n \"巿\",\n \"鑨\",\n \"醣\",\n \"熳\",\n \"懍\",\n \"湥\",\n \"檡\",\n \"韺\",\n \"戱\",\n \"緖\",\n \"鐈\",\n \"凉\",\n \"緃\",\n \"鮹\",\n \"媐\",\n \"爯\",\n \"巆\",\n \"褍\",\n \"鐬\",\n \"昍\",\n \"扙\",\n \"鍳\",\n \"芛\",\n \"蟳\",\n \"嬅\",\n \"糬\",\n \"吔\",\n \"塭\",\n \"譿\",\n \"冧\",\n \"鏓\",\n \"嶪\",\n \"嗹\",\n \"椵\",\n \"姀\",\n \"閿\",\n \"褧\",\n \"錞\",\n \"玆\",\n \"笘\",\n \"篔\",\n \"萡\",\n \"鶡\",\n \"螐\",\n \"鮄\",\n \"鰟\",\n \"脷\",\n \"啲\",\n \"杤\",\n \"蓚\",\n \"尗\",\n \"娎\",\n \"殟\",\n \"淥\",\n \"蝚\",\n \"蓧\",\n \"彐\",\n \"嚤\",\n \"銍\",\n \"囒\",\n \"坶\",\n \"淩\",\n \"鶼\",\n \"鱂\",\n \"喼\",\n \"燫\",\n \"肏\",\n \"姵\",\n \"廌\",\n \"禟\",\n \"籝\",\n \"迵\",\n \"嵨\",\n \"堮\",\n \"蟌\",\n \"憍\",\n \"廕\",\n \"蜑\",\n \"緁\",\n \"唘\",\n \"竩\",\n \"崙\",\n \"璚\",\n \"粄\",\n \"栨\",\n \"罈\",\n \"梫\",\n \"貤\",\n \"藔\",\n \"蜯\",\n \"訁\",\n \"斖\",\n \"煶\",\n \"馦\",\n \"妠\",\n \"閟\",\n \"疕\",\n \"夆\",\n \"鎪\",\n \"膥\",\n \"澻\",\n \"嘢\",\n \"嚐\",\n \"靁\",\n \"鎻\",\n \"鰛\",\n \"穵\",\n \"烋\",\n \"縕\",\n \"褎\",\n \"疒\",\n \"壠\",\n \"溼\",\n \"圂\",\n \"咅\",\n \"鯭\",\n \"鯙\",\n \"磘\",\n \"玨\",\n \"珤\",\n \"朊\",\n \"蚼\",\n \"濶\",\n \"薞\",\n \"嚩\",\n \"丟\",\n \"嫺\",\n \"鯻\",\n \"椲\",\n \"鰕\",\n \"刂\",\n \"蠘\",\n \"踎\",\n \"瀴\",\n \"琁\",\n \"鰶\",\n \"瑴\",\n \"肜\",\n \"㐂\",\n \"欥\",\n \"媺\",\n \"竻\",\n \"讚\",\n \"𣇉\",\n \"裵\",\n \"緜\",\n \"廩\",\n \"齧\",\n \"叄\",\n \"俌\",\n \"厰\",\n \"滀\",\n \"錄\",\n \"鷫\",\n \"鯗\",\n \"攞\",\n \"姌\",\n \"蔝\",\n \"幷\",\n \"縤\",\n \"屻\",\n \"鯃\",\n \"雞\",\n \"纁\",\n \"嫲\",\n \"嵮\",\n \"屭\",\n \"嶃\",\n \"跩\",\n \"鋗\",\n \"蕢\",\n \"篊\",\n \"俬\",\n \"淎\",\n \"暻\",\n \"鏻\",\n \"憓\",\n \"玗\",\n \"溈\",\n \"笭\",\n \"糢\",\n \"勳\",\n \"閒\",\n \"沍\",\n \"咾\",\n \"鉷\",\n \"蘵\",\n \"俁\",\n \"崵\",\n \"毸\",\n \"苪\",\n \"掙\",\n \"鴡\",\n \"萭\",\n \"俴\",\n \"屜\",\n \"蒾\",\n \"艹\",\n \"剷\",\n \"慍\",\n \"朮\",\n \"枴\",\n \"氳\",\n \"猓\",\n \"甽\",\n \"箝\",\n \"譁\",\n \"贗\",\n \"迆\",\n \"鈽\",\n \"鍊\",\n \"鍰\",\n \"鏍\",\n \"靦\",\n \"餽\",\n \"丮\",\n \"丱\",\n \"仜\",\n \"仩\",\n \"伬\",\n \"伔\",\n \"仱\",\n \"伀\",\n \"伻\",\n \"佢\",\n \"佒\",\n \"侀\",\n \"侇\",\n \"佷\",\n \"佌\",\n \"佪\",\n \"侐\",\n \"侜\",\n \"俓\",\n \"侲\",\n \"俉\",\n \"侻\",\n \"侳\",\n \"俇\",\n \"倅\",\n \"倇\",\n \"倰\",\n \"倛\",\n \"倳\",\n \"倷\",\n \"俷\",\n \"倠\",\n \"偯\",\n \"偞\",\n \"偠\",\n \"偋\",\n \"偝\",\n \"偛\",\n \"偢\",\n \"偅\",\n \"偟\",\n \"偩\",\n \"偫\",\n \"傛\",\n \"傔\",\n \"傞\",\n \"傋\",\n \"傌\",\n \"傎\",\n \"傝\",\n \"偨\",\n \"傂\",\n \"傽\",\n \"傿\",\n \"僆\",\n \"傮\",\n \"僄\",\n \"僈\",\n \"傰\",\n \"僁\",\n \"傱\",\n \"僋\",\n \"僗\",\n \"僛\",\n \"僪\",\n \"僝\",\n \"僓\",\n \"僿\",\n \"儃\",\n \"儰\",\n \"僸\",\n \"僶\",\n \"僾\",\n \"儌\",\n \"僽\",\n \"儜\",\n \"儓\",\n \"儗\",\n \"儑\",\n \"儢\",\n \"儤\",\n \"儠\",\n \"儸\",\n \"儹\",\n \"儽\",\n \"冓\",\n \"冘\",\n \"冞\",\n \"凊\",\n \"凅\",\n \"凔\",\n \"刌\",\n \"刉\",\n \"刓\",\n \"刜\",\n \"刞\",\n \"刵\",\n \"刲\",\n \"剆\",\n \"刱\",\n \"剉\",\n \"剚\",\n \"剒\",\n \"剫\",\n \"剭\",\n \"剬\",\n \"剺\",\n \"剸\",\n \"剻\",\n \"剼\",\n \"劀\",\n \"劋\",\n \"劖\",\n \"劘\",\n \"劗\",\n \"劙\",\n \"劦\",\n \"勴\",\n \"匊\",\n \"匢\",\n \"匰\",\n \"匴\",\n \"匷\",\n \"匽\",\n \"卌\",\n \"卼\",\n \"厎\",\n \"厒\",\n \"厗\",\n \"厞\",\n \"厜\",\n \"厤\",\n \"厬\",\n \"厹\",\n \"吰\",\n \"吷\",\n \"吪\",\n \"呿\",\n \"咈\",\n \"呫\",\n \"呺\",\n \"呥\",\n \"呬\",\n \"呴\",\n \"茍\",\n \"咷\",\n \"咮\",\n \"咶\",\n \"哅\",\n \"咠\",\n \"咢\",\n \"唦\",\n \"唗\",\n \"唒\",\n \"哤\",\n \"唚\",\n \"唈\",\n \"哫\",\n \"唅\",\n \"唴\",\n \"啢\",\n \"唶\",\n \"啒\",\n \"啅\",\n \"唌\",\n \"唲\",\n \"喨\",\n \"喥\",\n \"喭\",\n \"噅\",\n \"喓\",\n \"喣\",\n \"啽\",\n \"喌\",\n \"嗃\",\n \"嗛\",\n \"嗋\",\n \"嗀\",\n \"喿\",\n \"喍\",\n \"嗏\",\n \"嗕\",\n \"嗈\",\n \"嘕\",\n \"嘒\",\n \"嗼\",\n \"嘐\",\n \"嘓\",\n \"嘂\",\n \"嗺\",\n \"嘝\",\n \"嘄\",\n \"嗿\",\n \"噈\",\n \"噊\",\n \"噆\",\n \"噚\",\n \"嘳\",\n \"嘽\",\n \"嘾\",\n \"噮\",\n \"噳\",\n \"噣\",\n \"噭\",\n \"噞\",\n \"嚌\",\n \"嚍\",\n \"嚃\",\n \"嚘\",\n \"嚜\",\n \"嚫\",\n \"嚪\",\n \"嚬\",\n \"嚲\",\n \"嚵\",\n \"嚽\",\n \"嚾\",\n \"囆\",\n \"囅\",\n \"囋\",\n \"囗\",\n \"圁\",\n \"圞\",\n \"圠\",\n \"坁\",\n \"坅\",\n \"坲\",\n \"坱\",\n \"垀\",\n \"坴\",\n \"垗\",\n \"垝\",\n \"垔\",\n \"垘\",\n \"垽\",\n \"垼\",\n \"埢\",\n \"埶\",\n \"堩\",\n \"堣\",\n \"塈\",\n \"堥\",\n \"塓\",\n \"塉\",\n \"塯\",\n \"塕\",\n \"塼\",\n \"墆\",\n \"塿\",\n \"塴\",\n \"墋\",\n \"塺\",\n \"墝\",\n \"墯\",\n \"壈\",\n \"墽\",\n \"壖\",\n \"壝\",\n \"壛\",\n \"壾\",\n \"壿\",\n \"夃\",\n \"夎\",\n \"夒\",\n \"夗\",\n \"奅\",\n \"奊\",\n \"奰\",\n \"奲\",\n \"奼\",\n \"妦\",\n \"妎\",\n \"妢\",\n \"妐\",\n \"妵\",\n \"姏\",\n \"姎\",\n \"㚷\",\n \"姡\",\n \"姺\",\n \"姼\",\n \"娭\",\n \"婐\",\n \"婟\",\n \"婥\",\n \"婓\",\n \"婗\",\n \"媔\",\n \"媟\",\n \"媢\",\n \"婸\",\n \"媦\",\n \"媥\",\n \"媬\",\n \"媕\",\n \"娷\",\n \"嫇\",\n \"嫋\",\n \"媰\",\n \"媻\",\n \"嫮\",\n \"嫥\",\n \"嫢\",\n \"嫛\",\n \"嫿\",\n \"嫴\",\n \"嫷\",\n \"嫶\",\n \"嬎\",\n \"嬓\",\n \"嬐\",\n \"嬲\",\n \"嬽\",\n \"孈\",\n \"屘\",\n \"孲\",\n \"孷\",\n \"宎\",\n \"宨\",\n \"寪\",\n \"寍\",\n \"寋\",\n \"寑\",\n \"寙\",\n \"寠\",\n \"寱\",\n \"尌\",\n \"尒\",\n \"尟\",\n \"尰\",\n \"尳\",\n \"屖\",\n \"屔\",\n \"屝\",\n \"屧\",\n \"屩\",\n \"屮\",\n \"屴\",\n \"岏\",\n \"岋\",\n \"岉\",\n \"岒\",\n \"岮\",\n \"岤\",\n \"岯\",\n \"岟\",\n \"岝\",\n \"峐\",\n \"峌\",\n \"峞\",\n \"峉\",\n \"峊\",\n \"峬\",\n \"峮\",\n \"峷\",\n \"崝\",\n \"崨\",\n \"崥\",\n \"崏\",\n \"崰\",\n \"崣\",\n \"崷\",\n \"嵃\",\n \"嵑\",\n \"崳\",\n \"崺\",\n \"嵂\",\n \"嵱\",\n \"嵣\",\n \"嵥\",\n \"嵞\",\n \"嶀\",\n \"嵽\",\n \"嶆\",\n \"嵺\",\n \"嵷\",\n \"嶊\",\n \"嶉\",\n \"嶈\",\n \"嵾\",\n \"嶕\",\n \"嶜\",\n \"嶡\",\n \"嶚\",\n \"嶞\",\n \"嶱\",\n \"嶩\",\n \"嶵\",\n \"嶭\",\n \"巃\",\n \"巏\",\n \"巕\",\n \"巟\",\n \"巹\",\n \"帊\",\n \"帗\",\n \"帟\",\n \"帣\",\n \"帠\",\n \"帤\",\n \"帩\",\n \"帾\",\n \"帴\",\n \"幏\",\n \"幎\",\n \"幓\",\n \"幩\",\n \"幝\",\n \"幠\",\n \"幧\",\n \"幨\",\n \"幦\",\n \"幭\",\n \"幰\",\n \"庂\",\n \"庉\",\n \"庌\",\n \"庈\",\n \"庰\",\n \"庛\",\n \"庣\",\n \"庨\",\n \"庮\",\n \"庪\",\n \"庬\",\n \"庴\",\n \"廅\",\n \"廇\",\n \"廘\",\n \"廗\",\n \"廎\",\n \"廜\",\n \"緳\",\n \"廦\",\n \"廥\",\n \"廮\",\n \"廯\",\n \"蠯\",\n \"廾\",\n \"弚\",\n \"弝\",\n \"弣\",\n \"弤\",\n \"弮\",\n \"弳\",\n \"彃\",\n \"彉\",\n \"彋\",\n \"彏\",\n \"彯\",\n \"彴\",\n \"彸\",\n \"彾\",\n \"徦\",\n \"徥\",\n \"徯\",\n \"徲\",\n \"徾\",\n \"徿\",\n \"忀\",\n \"忁\",\n \"忔\",\n \"忕\",\n \"忨\",\n \"忣\",\n \"忷\",\n \"忥\",\n \"怭\",\n \"怲\",\n \"怋\",\n \"怴\",\n \"怗\",\n \"怚\",\n \"怞\",\n \"怬\",\n \"怢\",\n \"怐\",\n \"怮\",\n \"怓\",\n \"怷\",\n \"怹\",\n \"恲\",\n \"恞\",\n \"恅\",\n \"恇\",\n \"恉\",\n \"恛\",\n \"恌\",\n \"恀\",\n \"恟\",\n \"悀\",\n \"悁\",\n \"悕\",\n \"悗\",\n \"悇\",\n \"悊\",\n \"悐\",\n \"悾\",\n \"悺\",\n \"惓\",\n \"惤\",\n \"惈\",\n \"悷\",\n \"惉\",\n \"悹\",\n \"惌\",\n \"惢\",\n \"惄\",\n \"愊\",\n \"愖\",\n \"愅\",\n \"惵\",\n \"愓\",\n \"惸\",\n \"惼\",\n \"惾\",\n \"慉\",\n \"慅\",\n \"愶\",\n \"愲\",\n \"愮\",\n \"愯\",\n \"愬\",\n \"慁\",\n \"慞\",\n \"慱\",\n \"慒\",\n \"慓\",\n \"慲\",\n \"憀\",\n \"慴\",\n \"慔\",\n \"慺\",\n \"慛\",\n \"憃\",\n \"慹\",\n \"憱\",\n \"憰\",\n \"憢\",\n \"憉\",\n \"憛\",\n \"憯\",\n \"憟\",\n \"憪\",\n \"憡\",\n \"憝\",\n \"憖\",\n \"懅\",\n \"憴\",\n \"懆\",\n \"懁\",\n \"憿\",\n \"憸\",\n \"憵\",\n \"憼\",\n \"懧\",\n \"懠\",\n \"懥\",\n \"懤\",\n \"懘\",\n \"懭\",\n \"懱\",\n \"懪\",\n \"懰\",\n \"懫\",\n \"懻\",\n \"戁\",\n \"戃\",\n \"戄\",\n \"戉\",\n \"戠\",\n \"酨\",\n \"戺\",\n \"扐\",\n \"扜\",\n \"扤\",\n \"扡\",\n \"扢\",\n \"抆\",\n \"抌\",\n \"抎\",\n \"抏\",\n \"扻\",\n \"抭\",\n \"抴\",\n \"拑\",\n \"抾\",\n \"抪\",\n \"抶\",\n \"抮\",\n \"挍\",\n \"挋\",\n \"挃\",\n \"拫\",\n \"拹\",\n \"挏\",\n \"挌\",\n \"拸\",\n \"挀\",\n \"拲\",\n \"捖\",\n \"挬\",\n \"挶\",\n \"揤\",\n \"捊\",\n \"挼\",\n \"挩\",\n \"捁\",\n \"挴\",\n \"捘\",\n \"捔\",\n \"捥\",\n \"掝\",\n \"掗\",\n \"掫\",\n \"掯\",\n \"捵\",\n \"掜\",\n \"捼\",\n \"掤\",\n \"掔\",\n \"掱\",\n \"揎\",\n \"揥\",\n \"揨\",\n \"揯\",\n \"揊\",\n \"揲\",\n \"揵\",\n \"摡\",\n \"揟\",\n \"揝\",\n \"揜\",\n \"揘\",\n \"揅\",\n \"揱\",\n \"搆\",\n \"搟\",\n \"搕\",\n \"搘\",\n \"搹\",\n \"搷\",\n \"搣\",\n \"搰\",\n \"搊\",\n \"搚\",\n \"摀\",\n \"搧\",\n \"搫\",\n \"摍\",\n \"摝\",\n \"摲\",\n \"摦\",\n \"摎\",\n \"摋\",\n \"摓\",\n \"摐\",\n \"摿\",\n \"摮\",\n \"摰\",\n \"撢\",\n \"撠\",\n \"撗\",\n \"撜\",\n \"撋\",\n \"撊\",\n \"撌\",\n \"撟\",\n \"擗\",\n \"擖\",\n \"擏\",\n \"擉\",\n \"撽\",\n \"擩\",\n \"擣\",\n \"擫\",\n \"擭\",\n \"擨\",\n \"擽\",\n \"擸\",\n \"攇\",\n \"攐\",\n \"攍\",\n \"攌\",\n \"攗\",\n \"攕\",\n \"攓\",\n \"攡\",\n \"攠\",\n \"攦\",\n \"攩\",\n \"攭\",\n \"攲\",\n \"攳\",\n \"敁\",\n \"敊\",\n \"敆\",\n \"敓\",\n \"敧\",\n \"敪\",\n \"敤\",\n \"敜\",\n \"敯\",\n \"敳\",\n \"敶\",\n \"敺\",\n \"敹\",\n \"敿\",\n \"斁\",\n \"斀\",\n \"斄\",\n \"斒\",\n \"斔\",\n \"斞\",\n \"斨\",\n \"斪\",\n \"斻\",\n \"旍\",\n \"旓\",\n \"旚\",\n \"旝\",\n \"旟\",\n \"昲\",\n \"昦\",\n \"昢\",\n \"晇\",\n \"晥\",\n \"晜\",\n \"晼\",\n \"晬\",\n \"暀\",\n \"暆\",\n \"暍\",\n \"暋\",\n \"暡\",\n \"暰\",\n \"暩\",\n \"曀\",\n \"曊\",\n \"曋\",\n \"曏\",\n \"曒\",\n \"曚\",\n \"曣\",\n \"曭\",\n \"朁\",\n \"朅\",\n \"朄\",\n \"朒\",\n \"朘\",\n \"朣\",\n \"朾\",\n \"朹\",\n \"朻\",\n \"朼\",\n \"杅\",\n \"杇\",\n \"杝\",\n \"杗\",\n \"枎\",\n \"杶\",\n \"枆\",\n \"枌\",\n \"柲\",\n \"枺\",\n \"枻\",\n \"柸\",\n \"柀\",\n \"柅\",\n \"柫\",\n \"柤\",\n \"柍\",\n \"柮\",\n \"柣\",\n \"柂\",\n \"柧\",\n \"栚\",\n \"桋\",\n \"桏\",\n \"栱\",\n \"栵\",\n \"栫\",\n \"栭\",\n \"栯\",\n \"栘\",\n \"栔\",\n \"梡\",\n \"梇\",\n \"梐\",\n \"桭\",\n \"梮\",\n \"楖\",\n \"梬\",\n \"梩\",\n \"桵\",\n \"梒\",\n \"椌\",\n \"椄\",\n \"棜\",\n \"棷\",\n \"棳\",\n \"棌\",\n \"椈\",\n \"楰\",\n \"棯\",\n \"椔\",\n \"棸\",\n \"楟\",\n \"楎\",\n \"楱\",\n \"楅\",\n \"楺\",\n \"楈\",\n \"楛\",\n \"楉\",\n \"楬\",\n \"椳\",\n \"楀\",\n \"楄\",\n \"楶\",\n \"楘\",\n \"榶\",\n \"槉\",\n \"榠\",\n \"榬\",\n \"榼\",\n \"榙\",\n \"榩\",\n \"榾\",\n \"榯\",\n \"槄\",\n \"榽\",\n \"榹\",\n \"槥\",\n \"槸\",\n \"樕\",\n \"樠\",\n \"槬\",\n \"槢\",\n \"樛\",\n \"樝\",\n \"槾\",\n \"樧\",\n \"槮\",\n \"樔\",\n \"槷\",\n \"橀\",\n \"樴\",\n \"橉\",\n \"橧\",\n \"樲\",\n \"橨\",\n \"橝\",\n \"橭\",\n \"橶\",\n \"樿\",\n \"橁\",\n \"檍\",\n \"檖\",\n \"檁\",\n \"檟\",\n \"橾\",\n \"檛\",\n \"檓\",\n \"檕\",\n \"檃\",\n \"櫅\",\n \"檹\",\n \"櫡\",\n \"櫠\",\n \"櫌\",\n \"櫑\",\n \"櫙\",\n \"櫋\",\n \"櫜\",\n \"櫐\",\n \"櫫\",\n \"櫬\",\n \"櫰\",\n \"櫹\",\n \"櫺\",\n \"櫼\",\n \"欃\",\n \"欋\",\n \"欈\",\n \"欐\",\n \"欑\",\n \"欘\",\n \"欨\",\n \"欴\",\n \"欯\",\n \"欭\",\n \"欱\",\n \"欶\",\n \"欳\",\n \"欷\",\n \"欿\",\n \"歂\",\n \"歈\",\n \"歍\",\n \"歋\",\n \"歕\",\n \"歔\",\n \"歜\",\n \"歠\",\n \"歭\",\n \"歾\",\n \"肂\",\n \"殈\",\n \"殏\",\n \"殔\",\n \"殗\",\n \"殙\",\n \"殠\",\n \"殥\",\n \"殢\",\n \"殦\",\n \"殧\",\n \"殰\",\n \"殶\",\n \"毃\",\n \"毄\",\n \"毈\",\n \"毇\",\n \"毊\",\n \"毚\",\n \"毞\",\n \"毦\",\n \"毤\",\n \"毨\",\n \"毣\",\n \"毰\",\n \"毲\",\n \"毻\",\n \"毼\",\n \"毾\",\n \"氁\",\n \"氀\",\n \"氄\",\n \"氠\",\n \"氶\",\n \"汃\",\n \"汒\",\n \"汏\",\n \"汍\",\n \"汸\",\n \"沋\",\n \"汱\",\n \"汯\",\n \"沕\",\n \"汦\",\n \"汳\",\n \"泬\",\n \"沶\",\n \"沬\",\n \"泧\",\n \"沷\",\n \"泭\",\n \"泲\",\n \"泒\",\n \"沴\",\n \"洟\",\n \"洊\",\n \"洀\",\n \"浺\",\n \"浶\",\n \"洍\",\n \"涒\",\n \"浘\",\n \"浢\",\n \"涊\",\n \"涆\",\n \"浧\",\n \"涗\",\n \"涳\",\n \"涬\",\n \"淢\",\n \"涷\",\n \"淔\",\n \"渀\",\n \"淈\",\n \"涾\",\n \"淊\",\n \"涽\",\n \"淭\",\n \"湆\",\n \"湇\",\n \"湅\",\n \"湢\",\n \"渿\",\n \"湁\",\n \"渜\",\n \"渳\",\n \"湀\",\n \"渻\",\n \"渮\",\n \"湨\",\n \"湡\",\n \"渱\",\n \"渨\",\n \"湠\",\n \"湱\",\n \"湩\",\n \"渹\",\n \"溛\",\n \"滖\",\n \"溓\",\n \"溔\",\n \"滒\",\n \"溰\",\n \"溾\",\n \"滜\",\n \"滵\",\n \"滱\",\n \"漃\",\n \"漥\",\n \"漮\",\n \"潎\",\n \"漙\",\n \"漧\",\n \"漘\",\n \"漒\",\n \"滭\",\n \"漊\",\n \"潳\",\n \"滮\",\n \"潀\",\n \"漰\",\n \"潃\",\n \"漅\",\n \"濆\",\n \"澒\",\n \"澅\",\n \"潚\",\n \"潠\",\n \"澖\",\n \"潶\",\n \"潬\",\n \"潒\",\n \"潐\",\n \"潗\",\n \"澓\",\n \"潝\",\n \"濇\",\n \"濎\",\n \"濈\",\n \"濄\",\n \"澞\",\n \"澨\",\n \"瀄\",\n \"濌\",\n \"澩\",\n \"濴\",\n \"濔\",\n \"濣\",\n \"濭\",\n \"濧\",\n \"濦\",\n \"瀇\",\n \"瀎\",\n \"濿\",\n \"瀀\",\n \"濻\",\n \"瀙\",\n \"瀖\",\n \"瀫\",\n \"瀡\",\n \"瀢\",\n \"瀩\",\n \"瀯\",\n \"瀷\",\n \"灂\",\n \"瀸\",\n \"瀿\",\n \"瀺\",\n \"灄\",\n \"灉\",\n \"灖\",\n \"灗\",\n \"灛\",\n \"灟\",\n \"灨\",\n \"灩\",\n \"灪\",\n \"炾\",\n \"炰\",\n \"烓\",\n \"烑\",\n \"缹\",\n \"焍\",\n \"烰\",\n \"焠\",\n \"焮\",\n \"焣\",\n \"煆\",\n \"煣\",\n \"煝\",\n \"熐\",\n \"熉\",\n \"熀\",\n \"熂\",\n \"熚\",\n \"燅\",\n \"燂\",\n \"熸\",\n \"燀\",\n \"燡\",\n \"爁\",\n \"爊\",\n \"爂\",\n \"爓\",\n \"爞\",\n \"爢\",\n \"爣\",\n \"牄\",\n \"牉\",\n \"牋\",\n \"牏\",\n \"牣\",\n \"牬\",\n \"牰\",\n \"牸\",\n \"牷\",\n \"犈\",\n \"犉\",\n \"犆\",\n \"犅\",\n \"犌\",\n \"犑\",\n \"犐\",\n \"犗\",\n \"犕\",\n \"犓\",\n \"犘\",\n \"犚\",\n \"犝\",\n \"犞\",\n \"犥\",\n \"犦\",\n \"犤\",\n \"犣\",\n \"犩\",\n \"犪\",\n \"犮\",\n \"犵\",\n \"犿\",\n \"狆\",\n \"狖\",\n \"狋\",\n \"狘\",\n \"狜\",\n \"狔\",\n \"狚\",\n \"狌\",\n \"狑\",\n \"狊\",\n \"狤\",\n \"狫\",\n \"狪\",\n \"狣\",\n \"猀\",\n \"狾\",\n \"猑\",\n \"猘\",\n \"猈\",\n \"狿\",\n \"猏\",\n \"猋\",\n \"猒\",\n \"猧\",\n \"猲\",\n \"猭\",\n \"猦\",\n \"猣\",\n \"猵\",\n \"猼\",\n \"獂\",\n \"獀\",\n \"獊\",\n \"獑\",\n \"獌\",\n \"獘\",\n \"獞\",\n \"獟\",\n \"獝\",\n \"獛\",\n \"獡\",\n \"獩\",\n \"獦\",\n \"獥\",\n \"獳\",\n \"獶\",\n \"獽\",\n \"獿\",\n \"玂\",\n \"玁\",\n \"玈\",\n \"玊\",\n \"玔\",\n \"珓\",\n \"珶\",\n \"琖\",\n \"瑵\",\n \"璊\",\n \"瑽\",\n \"璅\",\n \"瑿\",\n \"璗\",\n \"瓁\",\n \"瓋\",\n \"瓝\",\n \"瓟\",\n \"瓡\",\n \"瓥\",\n \"瓨\",\n \"瓬\",\n \"瓵\",\n \"瓾\",\n \"瓽\",\n \"甀\",\n \"甃\",\n \"甈\",\n \"甋\",\n \"甐\",\n \"甒\",\n \"甔\",\n \"甖\",\n \"甝\",\n \"甮\",\n \"甿\",\n \"畟\",\n \"畣\",\n \"畽\",\n \"疀\",\n \"疧\",\n \"痁\",\n \"疻\",\n \"痀\",\n \"痎\",\n \"痏\",\n \"痋\",\n \"痌\",\n \"痑\",\n \"痚\",\n \"痡\",\n \"痝\",\n \"痗\",\n \"痯\",\n \"瘏\",\n \"痷\",\n \"痸\",\n \"痻\",\n \"瘈\",\n \"瘑\",\n \"瘝\",\n \"瘣\",\n \"瘯\",\n \"瘱\",\n \"瘽\",\n \"癈\",\n \"癉\",\n \"癙\",\n \"癐\",\n \"癓\",\n \"癠\",\n \"癵\",\n \"癹\",\n \"皊\",\n \"皏\",\n \"皫\",\n \"皯\",\n \"皵\",\n \"皻\",\n \"皽\",\n \"皾\",\n \"盄\",\n \"盓\",\n \"盝\",\n \"盬\",\n \"盭\",\n \"盳\",\n \"眃\",\n \"眅\",\n \"盻\",\n \"眝\",\n \"眐\",\n \"眓\",\n \"眒\",\n \"眣\",\n \"眑\",\n \"眕\",\n \"眹\",\n \"眱\",\n \"眲\",\n \"眴\",\n \"眳\",\n \"眽\",\n \"睆\",\n \"睅\",\n \"睊\",\n \"睋\",\n \"睌\",\n \"睕\",\n \"睟\",\n \"睒\",\n \"睖\",\n \"睩\",\n \"睧\",\n \"睔\",\n \"瞁\",\n \"睼\",\n \"瞂\",\n \"睮\",\n \"睯\",\n \"瞏\",\n \"瞉\",\n \"瞚\",\n \"瞝\",\n \"瞡\",\n \"瞛\",\n \"瞲\",\n \"瞷\",\n \"瞶\",\n \"瞴\",\n \"矂\",\n \"矉\",\n \"矊\",\n \"矌\",\n \"矎\",\n \"矏\",\n \"矐\",\n \"矔\",\n \"矕\",\n \"矘\",\n \"矠\",\n \"矱\",\n \"矲\",\n \"矹\",\n \"矺\",\n \"砅\",\n \"砐\",\n \"砏\",\n \"砎\",\n \"砨\",\n \"硈\",\n \"硉\",\n \"硠\",\n \"硥\",\n \"硱\",\n \"硰\",\n \"硩\",\n \"碔\",\n \"碄\",\n \"碅\",\n \"碆\",\n \"硾\",\n \"碫\",\n \"碞\",\n \"磍\",\n \"磌\",\n \"磎\",\n \"磈\",\n \"磃\",\n \"磝\",\n \"磩\",\n \"磥\",\n \"磞\",\n \"磛\",\n \"磳\",\n \"磼\",\n \"磿\",\n \"礔\",\n \"礉\",\n \"礝\",\n \"礛\",\n \"礜\",\n \"礥\",\n \"礣\",\n \"礧\",\n \"礨\",\n \"礭\",\n \"礿\",\n \"祌\",\n \"祅\",\n \"祔\",\n \"祒\",\n \"祑\",\n \"祤\",\n \"祩\",\n \"祪\",\n \"祣\",\n \"祫\",\n \"祡\",\n \"祴\",\n \"祳\",\n \"禂\",\n \"禗\",\n \"禜\",\n \"禫\",\n \"禭\",\n \"禬\",\n \"禴\",\n \"禷\",\n \"禸\",\n \"歶\",\n \"秅\",\n \"秏\",\n \"秖\",\n \"秎\",\n \"秮\",\n \"秪\",\n \"秺\",\n \"秶\",\n \"稊\",\n \"稒\",\n \"稫\",\n \"穊\",\n \"稰\",\n \"稯\",\n \"穋\",\n \"穛\",\n \"穖\",\n \"穧\",\n \"穨\",\n \"穮\",\n \"穬\",\n \"穭\",\n \"穱\",\n \"穾\",\n \"窆\",\n \"窉\",\n \"窌\",\n \"窏\",\n \"窔\",\n \"窐\",\n \"窙\",\n \"窢\",\n \"窞\",\n \"窫\",\n \"窲\",\n \"窴\",\n \"窱\",\n \"窾\",\n \"竀\",\n \"竁\",\n \"竷\",\n \"笐\",\n \"笓\",\n \"笅\",\n \"笵\",\n \"笻\",\n \"笴\",\n \"笰\",\n \"笢\",\n \"笝\",\n \"笲\",\n \"筄\",\n \"筡\",\n \"箈\",\n \"箊\",\n \"箌\",\n \"箛\",\n \"箎\",\n \"箘\",\n \"箄\",\n \"箷\",\n \"箾\",\n \"篎\",\n \"箯\",\n \"箹\",\n \"篞\",\n \"篣\",\n \"篧\",\n \"篕\",\n \"篨\",\n \"篹\",\n \"簅\",\n \"篲\",\n \"篿\",\n \"篻\",\n \"簎\",\n \"篴\",\n \"簂\",\n \"簁\",\n \"篸\",\n \"篽\",\n \"簜\",\n \"簩\",\n \"簙\",\n \"簭\",\n \"簦\",\n \"簨\",\n \"簢\",\n \"簥\",\n \"簳\",\n \"簼\",\n \"簬\",\n \"簻\",\n \"籉\",\n \"籈\",\n \"籊\",\n \"籔\",\n \"籗\",\n \"籧\",\n \"籦\",\n \"籯\",\n \"籺\",\n \"籸\",\n \"籹\",\n \"粊\",\n \"粔\",\n \"粻\",\n \"糔\",\n \"糪\",\n \"糱\",\n \"糷\",\n \"紎\",\n \"紟\",\n \"紒\",\n \"紽\",\n \"紸\",\n \"紶\",\n \"紩\",\n \"絇\",\n \"紾\",\n \"絘\",\n \"絯\",\n \"絓\",\n \"絧\",\n \"絏\",\n \"絭\",\n \"絫\",\n \"綀\",\n \"綍\",\n \"絿\",\n \"綅\",\n \"絻\",\n \"絼\",\n \"綔\",\n \"綷\",\n \"緂\",\n \"綪\",\n \"緀\",\n \"緅\",\n \"緎\",\n \"緆\",\n \"緌\",\n \"綯\",\n \"綼\",\n \"緷\",\n \"緛\",\n \"緪\",\n \"緧\",\n \"縃\",\n \"緺\",\n \"緶\",\n \"緰\",\n \"縗\",\n \"縌\",\n \"縓\",\n \"縎\",\n \"縜\",\n \"縚\",\n \"縏\",\n \"縼\",\n \"繂\",\n \"縳\",\n \"顈\",\n \"繈\",\n \"縸\",\n \"縪\",\n \"繉\",\n \"繀\",\n \"縩\",\n \"緵\",\n \"縰\",\n \"縿\",\n \"縶\",\n \"繜\",\n \"繐\",\n \"繣\",\n \"繘\",\n \"繢\",\n \"繟\",\n \"繑\",\n \"繠\",\n \"繶\",\n \"繵\",\n \"繸\",\n \"繷\",\n \"繺\",\n \"繲\",\n \"繴\",\n \"纀\",\n \"纇\",\n \"纋\",\n \"纆\",\n \"纑\",\n \"纗\",\n \"纚\",\n \"缿\",\n \"罊\",\n \"罏\",\n \"罜\",\n \"罞\",\n \"罝\",\n \"罛\",\n \"罣\",\n \"罥\",\n \"罦\",\n \"罭\",\n \"罫\",\n \"罬\",\n \"罻\",\n \"罼\",\n \"罺\",\n \"罿\",\n \"羃\",\n \"羉\",\n \"羍\",\n \"羒\",\n \"羜\",\n \"羛\",\n \"羢\",\n \"羠\",\n \"羦\",\n \"羬\",\n \"羭\",\n \"羵\",\n \"羳\",\n \"羷\",\n \"羺\",\n \"羾\",\n \"翋\",\n \"翍\",\n \"翐\",\n \"翑\",\n \"翇\",\n \"翢\",\n \"翣\",\n \"翭\",\n \"翪\",\n \"翨\",\n \"翴\",\n \"翲\",\n \"翽\",\n \"翿\",\n \"耟\",\n \"耞\",\n \"耡\",\n \"耴\",\n \"耾\",\n \"耹\",\n \"聇\",\n \"聈\",\n \"聑\",\n \"聏\",\n \"聝\",\n \"肕\",\n \"肙\",\n \"肒\",\n \"肣\",\n \"肵\",\n \"胘\",\n \"胑\",\n \"胐\",\n \"胕\",\n \"胉\",\n \"胏\",\n \"胹\",\n \"胵\",\n \"脁\",\n \"胻\",\n \"脀\",\n \"胾\",\n \"胔\",\n \"脰\",\n \"脥\",\n \"脤\",\n \"脙\",\n \"脡\",\n \"脕\",\n \"脧\",\n \"腃\",\n \"腏\",\n \"腄\",\n \"腇\",\n \"脽\",\n \"腍\",\n \"腤\",\n \"腷\",\n \"腜\",\n \"腛\",\n \"腢\",\n \"腲\",\n \"朡\",\n \"腞\",\n \"腶\",\n \"膉\",\n \"膆\",\n \"膃\",\n \"膇\",\n \"膍\",\n \"膌\",\n \"膋\",\n \"膟\",\n \"膕\",\n \"膢\",\n \"膱\",\n \"膹\",\n \"膫\",\n \"膰\",\n \"膬\",\n \"膴\",\n \"膲\",\n \"臇\",\n \"膷\",\n \"臄\",\n \"臅\",\n \"臒\",\n \"臐\",\n \"臗\",\n \"臛\",\n \"臡\",\n \"臦\",\n \"臩\",\n \"臮\",\n \"臲\",\n \"臷\",\n \"臸\",\n \"臿\",\n \"舋\",\n \"舑\",\n \"舕\",\n \"舝\",\n \"舡\",\n \"舼\",\n \"舽\",\n \"艀\",\n \"艂\",\n \"艓\",\n \"艒\",\n \"艐\",\n \"艑\",\n \"艕\",\n \"艛\",\n \"艵\",\n \"艼\",\n \"芀\",\n \"芐\",\n \"芅\",\n \"芓\",\n \"芔\",\n \"苀\",\n \"芚\",\n \"芵\",\n \"芧\",\n \"芞\",\n \"芺\",\n \"苙\",\n \"苨\",\n \"苖\",\n \"苬\",\n \"苲\",\n \"苵\",\n \"苶\",\n \"茙\",\n \"茥\",\n \"茿\",\n \"茦\",\n \"茢\",\n \"荂\",\n \"茪\",\n \"荍\",\n \"茖\",\n \"茤\",\n \"茠\",\n \"茩\",\n \"茻\",\n \"莐\",\n \"莣\",\n \"莍\",\n \"荺\",\n \"莤\",\n \"荴\",\n \"莏\",\n \"莁\",\n \"荵\",\n \"莔\",\n \"莃\",\n \"莌\",\n \"莋\",\n \"荾\",\n \"莥\",\n \"菨\",\n \"萒\",\n \"菧\",\n \"菤\",\n \"菆\",\n \"菣\",\n \"菿\",\n \"菋\",\n \"菎\",\n \"菵\",\n \"萉\",\n \"菞\",\n \"菳\",\n \"菕\",\n \"蓱\",\n \"萿\",\n \"葹\",\n \"葥\",\n \"葀\",\n \"葧\",\n \"萰\",\n \"葍\",\n \"葽\",\n \"蔇\",\n \"葞\",\n \"萷\",\n \"萺\",\n \"萴\",\n \"葅\",\n \"菙\",\n \"葋\",\n \"萯\",\n \"葂\",\n \"葟\",\n \"葌\",\n \"蓎\",\n \"蒬\",\n \"蒮\",\n \"蒫\",\n \"蒪\",\n \"蒚\",\n \"蒝\",\n \"蓌\",\n \"蒛\",\n \"蒩\",\n \"蒘\",\n \"蒶\",\n \"蒠\",\n \"蔤\",\n \"蔏\",\n \"蔩\",\n \"蔉\",\n \"蔍\",\n \"蔧\",\n \"蔜\",\n \"蓻\",\n \"蓺\",\n \"蓴\",\n \"蔪\",\n \"蓲\",\n \"蓷\",\n \"蓫\",\n \"蔒\",\n \"蓩\",\n \"蔖\",\n \"蓾\",\n \"蔨\",\n \"蔮\",\n \"蔂\",\n \"蓶\",\n \"蔱\",\n \"蓹\",\n \"蔠\",\n \"蔰\",\n \"蕫\",\n \"蕍\",\n \"蕀\",\n \"蕆\",\n \"蕄\",\n \"蕇\",\n \"蕣\",\n \"蕛\",\n \"蕱\",\n \"蕵\",\n \"蕮\",\n \"蕧\",\n \"蕠\",\n \"蕦\",\n \"蕝\",\n \"薃\",\n \"薧\",\n \"薕\",\n \"薠\",\n \"薋\",\n \"薣\",\n \"薚\",\n \"蕼\",\n \"薉\",\n \"蕸\",\n \"薎\",\n \"薖\",\n \"薍\",\n \"薝\",\n \"薂\",\n \"藆\",\n \"藀\",\n \"藃\",\n \"藂\",\n \"薵\",\n \"薽\",\n \"藇\",\n \"藄\",\n \"藋\",\n \"藈\",\n \"藅\",\n \"薱\",\n \"薶\",\n \"藒\",\n \"藫\",\n \"藱\",\n \"藙\",\n \"藡\",\n \"藚\",\n \"藗\",\n \"藲\",\n \"藬\",\n \"藘\",\n \"藣\",\n \"藑\",\n \"藰\",\n \"蘁\",\n \"藾\",\n \"蘛\",\n \"蘉\",\n \"蘌\",\n \"蘪\",\n \"蘦\",\n \"蘟\",\n \"蘣\",\n \"蘜\",\n \"蘙\",\n \"蘮\",\n \"蘡\",\n \"蘠\",\n \"蘥\",\n \"蘴\",\n \"蘳\",\n \"蘬\",\n \"虀\",\n \"蘹\",\n \"蘱\",\n \"蘻\",\n \"蘾\",\n \"虃\",\n \"虆\",\n \"虇\",\n \"虈\",\n \"虌\",\n \"虋\",\n \"虙\",\n \"虡\",\n \"虣\",\n \"虩\",\n \"虪\",\n \"虰\",\n \"虭\",\n \"虴\",\n \"蚑\",\n \"蚞\",\n \"蚇\",\n \"蚗\",\n \"蚚\",\n \"蚅\",\n \"蚥\",\n \"蚙\",\n \"蚿\",\n \"蚷\",\n \"蛂\",\n \"蛁\",\n \"蛅\",\n \"蛈\",\n \"蚹\",\n \"蚳\",\n \"蚸\",\n \"蛌\",\n \"蚻\",\n \"蛢\",\n \"蛦\",\n \"蛓\",\n \"蛣\",\n \"蛚\",\n \"蛪\",\n \"蛝\",\n \"蛫\",\n \"蛜\",\n \"蛬\",\n \"蛗\",\n \"蜄\",\n \"蛷\",\n \"蜌\",\n \"蛖\",\n \"蛵\",\n \"蜁\",\n \"蛶\",\n \"蜳\",\n \"蝫\",\n \"蜙\",\n \"蝃\",\n \"蜬\",\n \"蝁\",\n \"蝆\",\n \"蜠\",\n \"蜲\",\n \"蜪\",\n \"蜭\",\n \"蜼\",\n \"蜵\",\n \"蝂\",\n \"蜦\",\n \"蜧\",\n \"蜸\",\n \"蜤\",\n \"蜰\",\n \"蝖\",\n \"蝷\",\n \"蟡\",\n \"蝳\",\n \"蝔\",\n \"蝛\",\n \"蝒\",\n \"蝑\",\n \"蝞\",\n \"蝭\",\n \"蝪\",\n \"蝐\",\n \"蝝\",\n \"蝬\",\n \"蝺\",\n \"蝜\",\n \"螛\",\n \"螏\",\n \"螓\",\n \"螒\",\n \"螁\",\n \"螖\",\n \"螘\",\n \"蝹\",\n \"螇\",\n \"螑\",\n \"螝\",\n \"螜\",\n \"螚\",\n \"螪\",\n \"螰\",\n \"螹\",\n \"螼\",\n \"螮\",\n \"蟉\",\n \"蟃\",\n \"蟂\",\n \"螷\",\n \"螴\",\n \"螿\",\n \"螸\",\n \"蟞\",\n \"蟧\",\n \"蟦\",\n \"蟢\",\n \"蟟\",\n \"蟤\",\n \"蟔\",\n \"蟓\",\n \"蟭\",\n \"蟘\",\n \"螤\",\n \"蟗\",\n \"蟙\",\n \"蠁\",\n \"蟨\",\n \"蠀\",\n \"蟺\",\n \"蠉\",\n \"蠌\",\n \"蟼\",\n \"蠈\",\n \"蟿\",\n \"蠗\",\n \"蠩\",\n \"蠝\",\n \"蠛\",\n \"蠠\",\n \"蠤\",\n \"蠜\",\n \"蠫\",\n \"蠬\",\n \"蠨\",\n \"蠦\",\n \"蠪\",\n \"蠥\",\n \"蠰\",\n \"蠮\",\n \"蠳\",\n \"蠸\",\n \"蠾\",\n \"蠽\",\n \"蠿\",\n \"衁\",\n \"衈\",\n \"衋\",\n \"衧\",\n \"衪\",\n \"衭\",\n \"衶\",\n \"袀\",\n \"衱\",\n \"衯\",\n \"袃\",\n \"袉\",\n \"袕\",\n \"袨\",\n \"袚\",\n \"袑\",\n \"袡\",\n \"袘\",\n \"袧\",\n \"袬\",\n \"袌\",\n \"袺\",\n \"裗\",\n \"袹\",\n \"袸\",\n \"裀\",\n \"袶\",\n \"袽\",\n \"袲\",\n \"裋\",\n \"裍\",\n \"裞\",\n \"裚\",\n \"裷\",\n \"裧\",\n \"裺\",\n \"裮\",\n \"裶\",\n \"裯\",\n \"裻\",\n \"褁\",\n \"褅\",\n \"褋\",\n \"褗\",\n \"褆\",\n \"褖\",\n \"褑\",\n \"褦\",\n \"褮\",\n \"褱\",\n \"褢\",\n \"褩\",\n \"褵\",\n \"褼\",\n \"褾\",\n \"襒\",\n \"褷\",\n \"襂\",\n \"褽\",\n \"襓\",\n \"襋\",\n \"襆\",\n \"襐\",\n \"襛\",\n \"襗\",\n \"襡\",\n \"襘\",\n \"襝\",\n \"襣\",\n \"襭\",\n \"襩\",\n \"襮\",\n \"襳\",\n \"襹\",\n \"襺\",\n \"覂\",\n \"覅\",\n \"覕\",\n \"覛\",\n \"覝\",\n \"覢\",\n \"覤\",\n \"覣\",\n \"覭\",\n \"覮\",\n \"覶\",\n \"觓\",\n \"觤\",\n \"觡\",\n \"觠\",\n \"觢\",\n \"觩\",\n \"觰\",\n \"觬\",\n \"觲\",\n \"觷\",\n \"觺\",\n \"觻\",\n \"觼\",\n \"觾\",\n \"訑\",\n \"訰\",\n \"訧\",\n \"訬\",\n \"訞\",\n \"詍\",\n \"訹\",\n \"詙\",\n \"詀\",\n \"詄\",\n \"詅\",\n \"訿\",\n \"誂\",\n \"詻\",\n \"誃\",\n \"誫\",\n \"誙\",\n \"誋\",\n \"諆\",\n \"誸\",\n \"諔\",\n \"諕\",\n \"誻\",\n \"諀\",\n \"諅\",\n \"諵\",\n \"諝\",\n \"諰\",\n \"諈\",\n \"謞\",\n \"謘\",\n \"謑\",\n \"謋\",\n \"謒\",\n \"謕\",\n \"謍\",\n \"謈\",\n \"謪\",\n \"謧\",\n \"謣\",\n \"謰\",\n \"謵\",\n \"譇\",\n \"謯\",\n \"謱\",\n \"謥\",\n \"謷\",\n \"謦\",\n \"譐\",\n \"譈\",\n \"譊\",\n \"譀\",\n \"譋\",\n \"譕\",\n \"譑\",\n \"譠\",\n \"譪\",\n \"譝\",\n \"譨\",\n \"譣\",\n \"譥\",\n \"譹\",\n \"譸\",\n \"譅\",\n \"譺\",\n \"譻\",\n \"譾\",\n \"讄\",\n \"讂\",\n \"讆\",\n \"讋\",\n \"讔\",\n \"讘\",\n \"讟\",\n \"谹\",\n \"谻\",\n \"谽\",\n \"谾\",\n \"豃\",\n \"豋\",\n \"豍\",\n \"豏\",\n \"豗\",\n \"豜\",\n \"豝\",\n \"豟\",\n \"豥\",\n \"豤\",\n \"豦\",\n \"豭\",\n \"豰\",\n \"豲\",\n \"豱\",\n \"豯\",\n \"豵\",\n \"豷\",\n \"豶\",\n \"豻\",\n \"豽\",\n \"貁\",\n \"貀\",\n \"貄\",\n \"貏\",\n \"貑\",\n \"貕\",\n \"貙\",\n \"貗\",\n \"貜\",\n \"貣\",\n \"貾\",\n \"賌\",\n \"賥\",\n \"賟\",\n \"賙\",\n \"賵\",\n \"賮\",\n \"贆\",\n \"贕\",\n \"贙\",\n \"赨\",\n \"赩\",\n \"赮\",\n \"赸\",\n \"趀\",\n \"趌\",\n \"趎\",\n \"趏\",\n \"趍\",\n \"趓\",\n \"趠\",\n \"趜\",\n \"趡\",\n \"趥\",\n \"趧\",\n \"趬\",\n \"趪\",\n \"趭\",\n \"趫\",\n \"趮\",\n \"趷\",\n \"趹\",\n \"跘\",\n \"跓\",\n \"跍\",\n \"跇\",\n \"跜\",\n \"跕\",\n \"跙\",\n \"跈\",\n \"跰\",\n \"跠\",\n \"跮\",\n \"跦\",\n \"跢\",\n \"跧\",\n \"跲\",\n \"跫\",\n \"踂\",\n \"跿\",\n \"踍\",\n \"踃\",\n \"踇\",\n \"踆\",\n \"跾\",\n \"踠\",\n \"踥\",\n \"踤\",\n \"踡\",\n \"踕\",\n \"踛\",\n \"踖\",\n \"踑\",\n \"踙\",\n \"踧\",\n \"踘\",\n \"踓\",\n \"踳\",\n \"踾\",\n \"踸\",\n \"踼\",\n \"蹎\",\n \"蹍\",\n \"蹓\",\n \"蹗\",\n \"蹖\",\n \"蹞\",\n \"蹥\",\n \"蹛\",\n \"蹡\",\n \"蹝\",\n \"蹔\",\n \"蹸\",\n \"蹳\",\n \"蹪\",\n \"躆\",\n \"躈\",\n \"躖\",\n \"躗\",\n \"躟\",\n \"躠\",\n \"躤\",\n \"躣\",\n \"躩\",\n \"躨\",\n \"躽\",\n \"軓\",\n \"軘\",\n \"軞\",\n \"軯\",\n \"軷\",\n \"軦\",\n \"軮\",\n \"軥\",\n \"軵\",\n \"軧\",\n \"軨\",\n \"軶\",\n \"軱\",\n \"軬\",\n \"輆\",\n \"軿\",\n \"輁\",\n \"輀\",\n \"輂\",\n \"輐\",\n \"輑\",\n \"輤\",\n \"輘\",\n \"輚\",\n \"輠\",\n \"輣\",\n \"輖\",\n \"輗\",\n \"輮\",\n \"輵\",\n \"輲\",\n \"輹\",\n \"輷\",\n \"輴\",\n \"轃\",\n \"轇\",\n \"轈\",\n \"轒\",\n \"轑\",\n \"轏\",\n \"轐\",\n \"轓\",\n \"轙\",\n \"轖\",\n \"轗\",\n \"轕\",\n \"轚\",\n \"轞\",\n \"轛\",\n \"轠\",\n \"辴\",\n \"迉\",\n \"迒\",\n \"迋\",\n \"迍\",\n \"迖\",\n \"迣\",\n \"迡\",\n \"迾\",\n \"迿\",\n \"逜\",\n \"逿\",\n \"遝\",\n \"遳\",\n \"遰\",\n \"遻\",\n \"邆\",\n \"邅\",\n \"遾\",\n \"邍\",\n \"邔\",\n \"邟\",\n \"邥\",\n \"邞\",\n \"邧\",\n \"郱\",\n \"郕\",\n \"郖\",\n \"郠\",\n \"郙\",\n \"郣\",\n \"郥\",\n \"郘\",\n \"郰\",\n \"郲\",\n \"郔\",\n \"鄬\",\n \"郼\",\n \"鄈\",\n \"郹\",\n \"郻\",\n \"鄁\",\n \"鄇\",\n \"郺\",\n \"鄐\",\n \"鄍\",\n \"鄏\",\n \"鄎\",\n \"鄟\",\n \"鄝\",\n \"鄡\",\n \"鄛\",\n \"鄨\",\n \"鄪\",\n \"鄦\",\n \"鄮\",\n \"鄵\",\n \"鄸\",\n \"鄻\",\n \"鄾\",\n \"酀\",\n \"酁\",\n \"酄\",\n \"酇\",\n \"酖\",\n \"酘\",\n \"酓\",\n \"酟\",\n \"酳\",\n \"醆\",\n \"醊\",\n \"醓\",\n \"醙\",\n \"醟\",\n \"醥\",\n \"醧\",\n \"醰\",\n \"醱\",\n \"醷\",\n \"醲\",\n \"醳\",\n \"醹\",\n \"醽\",\n \"釂\",\n \"釃\",\n \"釢\",\n \"釱\",\n \"釳\",\n \"釸\",\n \"鈚\",\n \"鈌\",\n \"鈒\",\n \"釽\",\n \"鈆\",\n \"鉒\",\n \"鉠\",\n \"鉯\",\n \"鈶\",\n \"鉼\",\n \"銤\",\n \"銛\",\n \"銔\",\n \"鉹\",\n \"銗\",\n \"鋄\",\n \"鋀\",\n \"鋟\",\n \"鋘\",\n \"鋩\",\n \"鋝\",\n \"鋂\",\n \"鋊\",\n \"錧\",\n \"錼\",\n \"錭\",\n \"錎\",\n \"鋋\",\n \"鎡\",\n \"鎃\",\n \"鎯\",\n \"鍖\",\n \"鍜\",\n \"鍐\",\n \"鍭\",\n \"鍌\",\n \"鎒\",\n \"鎷\",\n \"鎝\",\n \"鎉\",\n \"鎎\",\n \"鎞\",\n \"鏏\",\n \"鏂\",\n \"鏚\",\n \"鏬\",\n \"鏙\",\n \"鐋\",\n \"鐏\",\n \"鏾\",\n \"鐕\",\n \"鐨\",\n \"鐍\",\n \"鐀\",\n \"鐎\",\n \"鐖\",\n \"鐻\",\n \"鐶\",\n \"鑐\",\n \"鑋\",\n \"鑕\",\n \"鑮\",\n \"鑯\",\n \"钂\",\n \"钀\",\n \"钁\",\n \"钃\",\n \"镺\",\n \"镻\",\n \"镼\",\n \"镽\",\n \"閈\",\n \"閍\",\n \"閺\",\n \"閵\",\n \"闀\",\n \"闉\",\n \"闅\",\n \"閷\",\n \"闒\",\n \"闑\",\n \"闚\",\n \"闛\",\n \"闠\",\n \"闟\",\n \"闤\",\n \"阞\",\n \"阢\",\n \"阤\",\n \"阠\",\n \"阰\",\n \"阹\",\n \"阸\",\n \"阺\",\n \"陏\",\n \"陓\",\n \"陊\",\n \"陼\",\n \"陭\",\n \"陫\",\n \"隇\",\n \"陾\",\n \"隉\",\n \"隒\",\n \"隓\",\n \"隞\",\n \"隤\",\n \"隿\",\n \"雂\",\n \"雈\",\n \"雓\",\n \"雔\",\n \"雗\",\n \"雚\",\n \"雟\",\n \"雘\",\n \"雺\",\n \"雽\",\n \"雿\",\n \"霂\",\n \"霋\",\n \"霒\",\n \"霐\",\n \"霠\",\n \"霣\",\n \"霢\",\n \"霩\",\n \"霫\",\n \"霬\",\n \"霮\",\n \"霵\",\n \"霿\",\n \"靆\",\n \"靃\",\n \"靪\",\n \"靮\",\n \"靷\",\n \"靲\",\n \"靾\",\n \"鞃\",\n \"鞀\",\n \"鞂\",\n \"靻\",\n \"鞊\",\n \"鞎\",\n \"鞈\",\n \"鞙\",\n \"鞗\",\n \"鞚\",\n \"鞜\",\n \"鞤\",\n \"鞪\",\n \"鞷\",\n \"鞶\",\n \"鞹\",\n \"鞻\",\n \"鞿\",\n \"韄\",\n \"韅\",\n \"韇\",\n \"韎\",\n \"韐\",\n \"韏\",\n \"韕\",\n \"韔\",\n \"韗\",\n \"韝\",\n \"韟\",\n \"韣\",\n \"韥\",\n \"韰\",\n \"韱\",\n \"韹\",\n \"韽\",\n \"頄\",\n \"頖\",\n \"頞\",\n \"頝\",\n \"頩\",\n \"頨\",\n \"頯\",\n \"頲\",\n \"顁\",\n \"顄\",\n \"顊\",\n \"顉\",\n \"顅\",\n \"顐\",\n \"顑\",\n \"顜\",\n \"顝\",\n \"顠\",\n \"顣\",\n \"顟\",\n \"顤\",\n \"顪\",\n \"顩\",\n \"顲\",\n \"颬\",\n \"颲\",\n \"颸\",\n \"颽\",\n \"颻\",\n \"颾\",\n \"飁\",\n \"飂\",\n \"飉\",\n \"飋\",\n \"飌\",\n \"飣\",\n \"飶\",\n \"餂\",\n \"餀\",\n \"飺\",\n \"餔\",\n \"餖\",\n \"餕\",\n \"餤\",\n \"餟\",\n \"餥\",\n \"餫\",\n \"餪\",\n \"餲\",\n \"餯\",\n \"餭\",\n \"餱\",\n \"餰\",\n \"饁\",\n \"饇\",\n \"饐\",\n \"饎\",\n \"饙\",\n \"饘\",\n \"饛\",\n \"饡\",\n \"馣\",\n \"馲\",\n \"馰\",\n \"馵\",\n \"馻\",\n \"馺\",\n \"駂\",\n \"馽\",\n \"駜\",\n \"駍\",\n \"駏\",\n \"駎\",\n \"駖\",\n \"駮\",\n \"駬\",\n \"駥\",\n \"駤\",\n \"駣\",\n \"駩\",\n \"駺\",\n \"駴\",\n \"駷\",\n \"駹\",\n \"駶\",\n \"駻\",\n \"駽\",\n \"駾\",\n \"騃\",\n \"騉\",\n \"騑\",\n \"騊\",\n \"騇\",\n \"騚\",\n \"騕\",\n \"騥\",\n \"騝\",\n \"騛\",\n \"騢\",\n \"騠\",\n \"騧\",\n \"騞\",\n \"騜\",\n \"騵\",\n \"騲\",\n \"騴\",\n \"騱\",\n \"騬\",\n \"騪\",\n \"騩\",\n \"騹\",\n \"騽\",\n \"驆\",\n \"騺\",\n \"驓\",\n \"驔\",\n \"驈\",\n \"驉\",\n \"驖\",\n \"驞\",\n \"驠\",\n \"驦\",\n \"驨\",\n \"骭\",\n \"骫\",\n \"骹\",\n \"骿\",\n \"骴\",\n \"骾\",\n \"髇\",\n \"髊\",\n \"髆\",\n \"髍\",\n \"髐\",\n \"髟\",\n \"髧\",\n \"髬\",\n \"髳\",\n \"髶\",\n \"髺\",\n \"髾\",\n \"鬁\",\n \"髼\",\n \"鬋\",\n \"鬊\",\n \"鬎\",\n \"鬌\",\n \"鬐\",\n \"鬕\",\n \"鬗\",\n \"鬖\",\n \"鬙\",\n \"鬞\",\n \"鬠\",\n \"鬤\",\n \"鬫\",\n \"鬳\",\n \"鬵\",\n \"鬺\",\n \"鬾\",\n \"鬿\",\n \"魊\",\n \"魌\",\n \"魖\",\n \"魠\",\n \"魡\",\n \"魧\",\n \"魱\",\n \"魦\",\n \"魶\",\n \"魵\",\n \"鮅\",\n \"鮇\",\n \"魼\",\n \"魾\",\n \"魻\",\n \"鮂\",\n \"鮚\",\n \"鮞\",\n \"鮛\",\n \"鮦\",\n \"鮥\",\n \"鮤\",\n \"鮆\",\n \"鯆\",\n \"鮿\",\n \"鮵\",\n \"鯈\",\n \"鯫\",\n \"鯠\",\n \"鯞\",\n \"鯦\",\n \"鯬\",\n \"鰌\",\n \"鰋\",\n \"鰅\",\n \"鯸\",\n \"鰫\",\n \"鰝\",\n \"鰬\",\n \"鱆\",\n \"鰿\",\n \"鱄\",\n \"鱁\",\n \"鰴\",\n \"鱐\",\n \"鱍\",\n \"鱋\",\n \"鱕\",\n \"鱦\",\n \"鱢\",\n \"鱞\",\n \"鱴\",\n \"鱳\",\n \"鱹\",\n \"鳦\",\n \"鳪\",\n \"鳭\",\n \"鳱\",\n \"鳵\",\n \"鳼\",\n \"鳺\",\n \"鳿\",\n \"鳷\",\n \"鴀\",\n \"鳹\",\n \"鳻\",\n \"鴅\",\n \"鴃\",\n \"鴥\",\n \"鴠\",\n \"鴔\",\n \"鴩\",\n \"鴘\",\n \"鴢\",\n \"鴐\",\n \"鴳\",\n \"鵁\",\n \"鵧\",\n \"鴶\",\n \"鴮\",\n \"鴱\",\n \"鴸\",\n \"鵅\",\n \"鵃\",\n \"鴾\",\n \"鵀\",\n \"鴽\",\n \"鵏\",\n \"鵊\",\n \"鵛\",\n \"鵋\",\n \"鵖\",\n \"鵌\",\n \"鵗\",\n \"鵔\",\n \"鵷\",\n \"鶁\",\n \"鶊\",\n \"鶄\",\n \"鶈\",\n \"鵱\",\n \"鶀\",\n \"鵸\",\n \"鶋\",\n \"鶌\",\n \"鵽\",\n \"鵫\",\n \"鵴\",\n \"鵩\",\n \"鶅\",\n \"鵳\",\n \"鵻\",\n \"鶂\",\n \"鵹\",\n \"鶟\",\n \"鶙\",\n \"鶤\",\n \"鶝\",\n \"鶐\",\n \"鶛\",\n \"鶠\",\n \"鶔\",\n \"鶜\",\n \"鶪\",\n \"鶗\",\n \"鶢\",\n \"鶨\",\n \"鶞\",\n \"鶣\",\n \"鶖\",\n \"鶷\",\n \"鶶\",\n \"鷁\",\n \"鷇\",\n \"鷊\",\n \"鷏\",\n \"鶾\",\n \"鷅\",\n \"鷃\",\n \"鶵\",\n \"鷈\",\n \"鶱\",\n \"鶭\",\n \"鷛\",\n \"鷒\",\n \"鷞\",\n \"鷋\",\n \"鷐\",\n \"鷜\",\n \"鷑\",\n \"鷩\",\n \"鷘\",\n \"鷖\",\n \"鷵\",\n \"鷕\",\n \"鷻\",\n \"鷷\",\n \"鷣\",\n \"鷤\",\n \"鷶\",\n \"鷡\",\n \"鷮\",\n \"鷢\",\n \"鸂\",\n \"鷾\",\n \"鸇\",\n \"鸃\",\n \"鸆\",\n \"鸅\",\n \"鸀\",\n \"鸁\",\n \"鸉\",\n \"鷿\",\n \"鷽\",\n \"鸄\",\n \"鸋\",\n \"鸍\",\n \"鸏\",\n \"鸒\",\n \"鸔\",\n \"鸓\",\n \"鸗\",\n \"鸙\",\n \"鹺\",\n \"麃\",\n \"麆\",\n \"麉\",\n \"麎\",\n \"麌\",\n \"麔\",\n \"麙\",\n \"麛\",\n \"麚\",\n \"麜\",\n \"麠\",\n \"麡\",\n \"麧\",\n \"麮\",\n \"麰\",\n \"麶\",\n \"麷\",\n \"黀\",\n \"黂\",\n \"黈\",\n \"黓\",\n \"黕\",\n \"黖\",\n \"黚\",\n \"黤\",\n \"黫\",\n \"黮\",\n \"黭\",\n \"黰\",\n \"黳\",\n \"黵\",\n \"黺\",\n \"鼁\",\n \"鼀\",\n \"鼆\",\n \"鼊\",\n \"鼏\",\n \"鼖\",\n \"鼛\",\n \"鼘\",\n \"鼜\",\n \"鼤\",\n \"鼣\",\n \"鼥\",\n \"鼪\",\n \"鼨\",\n \"鼭\",\n \"鼰\",\n \"鼮\",\n \"鼵\",\n \"鼳\",\n \"鼲\",\n \"鼸\",\n \"鼶\",\n \"齀\",\n \"齂\",\n \"齃\",\n \"齌\",\n \"齍\",\n \"齎\",\n \"齖\",\n \"齗\",\n \"齘\",\n \"齛\",\n \"齠\",\n \"齞\",\n \"齝\",\n \"齥\",\n \"齤\",\n \"齫\",\n \"齱\",\n \"齰\",\n \"齮\",\n \"齯\",\n \"齴\",\n \"齵\",\n \"齸\",\n \"齻\",\n \"齺\",\n \"齹\",\n \"齾\",\n \"龒\",\n \"龤\",\n \"堔\",\n \"礂\",\n \"蒏\",\n \"蒆\",\n \"兙\",\n \"兛\",\n \"兞\",\n \"兝\",\n \"兡\",\n \"兣\",\n \"嗧\",\n \"瓩\",\n \"忼\",\n \"擡\",\n \"氊\",\n \"穇\",\n \"擧\",\n \"譌\",\n \"!\",\n \"\\\"\",\n \"#\",\n \"$\",\n \"%\",\n \"&\",\n \"'\",\n \"(\",\n \")\",\n \"*\",\n \"+\",\n \",\",\n \"-\",\n \".\",\n \"/\",\n \"0\",\n \"1\",\n \"2\",\n \"3\",\n \"4\",\n \"5\",\n \"6\",\n \"7\",\n \"8\",\n \"9\",\n \":\",\n \";\",\n \"<\",\n \"=\",\n \">\",\n \"?\",\n \"A\",\n \"B\",\n \"C\",\n \"D\",\n \"E\",\n \"F\",\n \"G\",\n \"H\",\n \"I\",\n \"J\",\n \"K\",\n \"L\",\n \"M\",\n \"N\",\n \"O\",\n \"P\",\n \"Q\",\n \"R\",\n \"S\",\n \"T\",\n \"U\",\n \"V\",\n \"W\",\n \"X\",\n \"Y\",\n \"Z\",\n \"[\",\n \"]\",\n \"_\",\n \"`\",\n \"a\",\n \"b\",\n \"c\",\n \"d\",\n \"e\",\n \"f\",\n \"g\",\n \"h\",\n \"i\",\n \"j\",\n \"k\",\n \"l\",\n \"m\",\n \"n\",\n \"o\",\n \"p\",\n \"q\",\n \"r\",\n \"s\",\n \"t\",\n \"u\",\n \"v\",\n \"w\",\n \"x\",\n \"y\",\n \"z\",\n \"©\",\n \"°\",\n \"²\",\n \"´\",\n \"½\",\n \"Á\",\n \"Ä\",\n \"Å\",\n \"Ç\",\n \"È\",\n \"É\",\n \"Í\",\n \"Ó\",\n \"Ö\",\n \"×\",\n \"Ü\",\n \"ß\",\n \"à\",\n \"á\",\n \"â\",\n \"ã\",\n \"ä\",\n \"å\",\n \"æ\",\n \"ç\",\n \"è\",\n \"é\",\n \"ê\",\n \"ë\",\n \"í\",\n \"ð\",\n \"ñ\",\n \"ò\",\n \"ó\",\n \"ô\",\n \"õ\",\n \"ö\",\n \"ø\",\n \"ú\",\n \"û\",\n \"ü\",\n \"ý\",\n \"ā\",\n \"ă\",\n \"ą\",\n \"ć\",\n \"Č\",\n \"č\",\n \"đ\",\n \"ē\",\n \"ė\",\n \"ę\",\n \"ğ\",\n \"ī\",\n \"ı\",\n \"Ł\",\n \"ł\",\n \"ń\",\n \"ň\",\n \"ō\",\n \"ř\",\n \"Ş\",\n \"ş\",\n \"Š\",\n \"š\",\n \"ţ\",\n \"ū\",\n \"ż\",\n \"Ž\",\n \"ž\",\n \"Ș\",\n \"ș\",\n \"ț\",\n \"Δ\",\n \"α\",\n \"λ\",\n \"μ\",\n \"φ\",\n \"Г\",\n \"О\",\n \"а\",\n \"в\",\n \"л\",\n \"о\",\n \"р\",\n \"с\",\n \"т\",\n \"я\",\n \"ồ\",\n \"—\",\n \"―\",\n \"’\",\n \"“\",\n \"”\",\n \"…\",\n \"℃\",\n \"→\",\n \"∇\",\n \"−\",\n \"■\",\n \"☆\",\n \"、\",\n \"。\",\n \"々\",\n \"〆\",\n \"〈\",\n \"〉\",\n \"「\",\n \"」\",\n \"『\",\n \"』\",\n \"〔\",\n \"〕\",\n \"〜\",\n \"！\",\n \"＃\",\n \"％\",\n \"＆\",\n \"（\",\n \"）\",\n \"＋\",\n \"，\",\n \"－\",\n \"．\",\n \"／\",\n \"０\",\n \"１\",\n \"２\",\n \"３\",\n \"４\",\n \"５\",\n \"６\",\n \"７\",\n \"８\",\n \"９\",\n \"：\",\n \"；\",\n \"＝\",\n \"？\",\n \"＠\",\n \"Ａ\",\n \"Ｂ\",\n \"Ｃ\",\n \"Ｄ\",\n \"Ｅ\",\n \"Ｆ\",\n \"Ｇ\",\n \"Ｈ\",\n \"Ｉ\",\n \"Ｊ\",\n \"Ｋ\",\n \"Ｌ\",\n \"Ｍ\",\n \"Ｎ\",\n \"Ｏ\",\n \"Ｐ\",\n \"Ｒ\",\n \"Ｓ\",\n \"Ｔ\",\n \"Ｕ\",\n \"Ｖ\",\n \"Ｗ\",\n \"Ｘ\",\n \"Ｚ\",\n \"ａ\",\n \"ｂ\",\n \"ｃ\",\n \"ｄ\",\n \"ｅ\",\n \"ｆ\",\n \"ｇ\",\n \"ｈ\",\n \"ｉ\",\n \"ｊ\",\n \"ｋ\",\n \"ｌ\",\n \"ｍ\",\n \"ｎ\",\n \"ｏ\",\n \"ｐ\",\n \"ｑ\",\n \"ｒ\",\n \"ｓ\",\n \"ｔ\",\n \"ｕ\",\n \"ｖ\",\n \"ｗ\",\n \"ｘ\",\n \"ｙ\",\n \"ｚ\",\n \"～\",\n \"･\",\n \"ǎ\",\n \"ǒ\",\n \"ě\",\n \"ǐ\",\n \"ì\",\n \"ǔ\",\n \"ù\",\n \"ǖ\",\n \"ǘ\",\n \"ǚ\",\n \"ǜ\",\n \"【\",\n \"】\",\n \"《\",\n \"》\",\n \"‥\",\n \"{\",\n \"}\",\n \"\\\\\",\n \"|\",\n \"@\",\n \"^\",\n \"~\",\n \"÷\",\n \"∕\",\n \"∙\",\n \"⋅\",\n \"·\",\n \"⊕\",\n \"⊖\",\n \"⊗\",\n \"⊘\",\n \"⊙\",\n \"±\",\n \"∓\",\n \"∩\",\n \"∪\",\n \"□\",\n \"⊎\",\n \"⊓\",\n \"⊔\",\n \"≠\",\n \"≈\",\n \"≡\",\n \"≤\",\n \"≥\",\n \"≪\",\n \"≫\",\n \"≲\",\n \"≳\",\n \"≶\",\n \"≷\",\n \"≺\",\n \"≻\",\n \"≼\",\n \"≽\",\n \"∈\",\n \"∉\",\n \"⊂\",\n \"⊃\",\n \"⊆\",\n \"⊇\",\n \"⊄\",\n \"⊅\",\n \"∅\",\n \"∖\",\n \"∁\",\n \"∆\",\n \"∧\",\n \"∨\",\n \"¬\",\n \"⊻\",\n \"⊼\",\n \"⊽\",\n \"←\",\n \"↔\",\n \"⇒\",\n \"⇐\",\n \"⇔\",\n \"∀\",\n \"∃\",\n \"∄\",\n \"∴\",\n \"∵\",\n \"∝\",\n \"∞\",\n \"⊥\",\n \"∟\",\n \"∠\",\n \"∡\",\n \"∢\",\n \"′\",\n \"″\",\n \"∥\",\n \"⊾\",\n \"⊿\",\n \"∂\",\n \"∫\",\n \"∬\",\n \"∭\",\n \"∮\",\n \"∯\",\n \"∰\",\n \"∑\",\n \"∏\",\n \"√\",\n \"∛\",\n \"∜\",\n \"∱\",\n \"∲\",\n \"∳\",\n \"∶\",\n \"∷\",\n \"∼\",\n \"®\",\n \"≄\",\n \"≅\",\n \"≃\",\n \"≦\",\n \"≧\",\n \"⊈\",\n \"⊉\",\n \"⊢\",\n \"⊤\",\n \"⊨\",\n \"⊧\",\n \"℉\",\n \"Ω\",\n \"℧\",\n \"Å\",\n \"⌀\",\n \"ℏ\",\n \"⅀\",\n \"⍺\",\n \"⍵\",\n \"¢\",\n \"€\",\n \"£\",\n \"¥\",\n \"￥\",\n \"₿\",\n \"↑\",\n \"↓\",\n \"↕\",\n \"↖\",\n \"↗\",\n \"↘\",\n \"↙\",\n \"↺\",\n \"↻\",\n \"↼\",\n \"↽\",\n \"↾\",\n \"↿\",\n \"⇀\",\n \"⇁\",\n \"⇂\",\n \"⇃\",\n \"⇋\",\n \"⇌\",\n \"ª\",\n \"º\",\n \"⁰\",\n \"¹\",\n \"³\",\n \"⁴\",\n \"⁵\",\n \"⁶\",\n \"⁷\",\n \"⁸\",\n \"⁹\",\n \"⁺\",\n \"⁻\",\n \"⁼\",\n \"⁽\",\n \"⁾\",\n \"ⁿ\",\n \"₀\",\n \"₁\",\n \"₂\",\n \"₃\",\n \"₄\",\n \"₅\",\n \"₆\",\n \"₇\",\n \"₈\",\n \"₉\",\n \"₊\",\n \"₋\",\n \"₌\",\n \"₍\",\n \"₎\",\n \"Ⅰ\",\n \"Ⅱ\",\n \"Ⅲ\",\n \"Ⅳ\",\n \"Ⅴ\",\n \"Ⅵ\",\n \"Ⅶ\",\n \"Ⅷ\",\n \"Ⅸ\",\n \"Ⅹ\",\n \"Ⅺ\",\n \"Ⅻ\",\n \"ⅰ\",\n \"ⅱ\",\n \"ⅲ\",\n \"ⅳ\",\n \"ⅴ\",\n \"ⅵ\",\n \"ⅶ\",\n \"ⅷ\",\n \"ⅸ\",\n \"ⅹ\",\n \"ⅺ\",\n \"ⅻ\",\n \"☰\",\n \"☱\",\n \"☲\",\n \"☳\",\n \"☴\",\n \"☵\",\n \"☶\",\n \"☷\",\n \"♀\",\n \"♂\",\n \"♳\",\n \"♴\",\n \"♵\",\n \"♶\",\n \"♷\",\n \"♸\",\n \"♹\",\n \"♺\",\n \"♩\",\n \"♪\",\n \"♫\",\n \"♬\",\n \"⚪\",\n \"⚫\",\n \"⚬\",\n \"✶\",\n \"✷\",\n \"✸\",\n \"➀\",\n \"➁\",\n \"➂\",\n \"➃\",\n \"➄\",\n \"➅\",\n \"➆\",\n \"➇\",\n \"➈\",\n \"➉\",\n \"➊\",\n \"➋\",\n \"➌\",\n \"➍\",\n \"➎\",\n \"➏\",\n \"➐\",\n \"➑\",\n \"➒\",\n \"➓\",\n \"⏀\",\n \"⏁\",\n \"⏂\",\n \"⏃\",\n \"⏄\",\n \"⏅\",\n \"⏆\",\n \"⏇\",\n \"⏈\",\n \"⏉\",\n \"⏊\",\n \"⏋\",\n \"⏌\",\n \"⏚\",\n \"⏴\",\n \"⏵\",\n \"⏶\",\n \"⏷\",\n \"⏸\",\n \"⏹\",\n \"⏺\",\n \"⏻\",\n \"⏼\",\n \"Α\",\n \"Β\",\n \"Γ\",\n \"Ε\",\n \"Ζ\",\n \"Η\",\n \"Θ\",\n \"Ι\",\n \"Κ\",\n \"Λ\",\n \"Μ\",\n \"Ν\",\n \"Ξ\",\n \"Ο\",\n \"Π\",\n \"Ρ\",\n \"Σ\",\n \"Τ\",\n \"Υ\",\n \"Φ\",\n \"Χ\",\n \"Ψ\",\n \"β\",\n \"γ\",\n \"δ\",\n \"ε\",\n \"ζ\",\n \"η\",\n \"θ\",\n \"ι\",\n \"κ\",\n \"ν\",\n \"ξ\",\n \"ο\",\n \"π\",\n \"ρ\",\n \"σ\",\n \"τ\",\n \"υ\",\n \"χ\",\n \"ψ\",\n \"ω\",\n \"ϐ\",\n \"ϑ\",\n \"ϒ\",\n \"ϕ\",\n \"█\",\n \"ϖ\",\n \"ϰ\",\n \"ϱ\",\n \"ϴ\",\n \"ϵ\",\n \"ϝ\",\n \"Ϟ\",\n \"ϟ\",\n \"Ϡ\",\n \"ϡ\",\n \"Ϣ\",\n \"ϣ\",\n \"Ϥ\",\n \"ϥ\",\n \"Ϧ\",\n \"ϧ\",\n \"Ϩ\",\n \"ϩ\",\n \"Ϫ\",\n \"ϫ\",\n \"Ϭ\",\n \"ϭ\",\n \"Ϯ\",\n \"ϯ\",\n \"∸\",\n \"∹\",\n \"∺\",\n \"∻\",\n \"∽\",\n \"∾\",\n \"∿\",\n \"≀\",\n \"≁\",\n \"≂\",\n \"≆\",\n \"≇\",\n \"≉\",\n \"≊\",\n \"≋\",\n \"≌\",\n \"≍\",\n \"≎\",\n \"≏\",\n \"≐\",\n \"≑\",\n \"≒\",\n \"≓\",\n \"≔\",\n \"≕\",\n \"≖\",\n \"≗\",\n \"≘\",\n \"≙\",\n \"≚\",\n \"≛\",\n \"≜\",\n \"≝\",\n \"≞\",\n \"≟\",\n \"≢\",\n \"≣\",\n \"≨\",\n \"≩\",\n \"≬\",\n \"≭\",\n \"≮\",\n \"≯\",\n \"≰\",\n \"≱\",\n \"≴\",\n \"≵\",\n \"≸\",\n \"≹\",\n \"≾\",\n \"≿\",\n \"⊀\",\n \"⊁\",\n \"⊊\",\n \"⊋\",\n \"⊌\",\n \"⊍\",\n \"⊏\",\n \"⊐\",\n \"⊑\",\n \"⊒\",\n \"⊚\",\n \"⊛\",\n \"⊜\",\n \"⊝\",\n \"⊞\",\n \"⊟\",\n \"⊠\",\n \"⊡\",\n \"⊣\",\n \"⊦\",\n \"⊩\",\n \"⊪\",\n \"⊫\",\n \"⊬\",\n \"⊭\",\n \"⊮\",\n \"⊯\",\n \"⊰\",\n \"⊱\",\n \"⊲\",\n \"⊳\",\n \"⊴\",\n \"⊵\",\n \"⊶\",\n \"⊷\",\n \"⊸\",\n \"⊹\",\n \"⊺\",\n \"ℎ\",\n \"℘\",\n \"ℜ\",\n \"ℑ\",\n \"ℵ\",\n \"ℶ\",\n \"ℷ\",\n \"ℸ\",\n \"⌬\",\n \"⌭\",\n \"⌮\",\n \"⌯\",\n \"⎔\",\n \"¤\",\n \"₠\",\n \"₡\",\n \"₢\",\n \"₣\",\n \"₤\",\n \"₥\",\n \"₦\",\n \"₧\",\n \"₨\",\n \"₩\",\n \"₪\",\n \"₫\",\n \"₭\",\n \"₮\",\n \"₯\",\n \"₰\",\n \"₱\",\n \"₲\",\n \"₳\",\n \"₴\",\n \"₵\",\n \"₶\",\n \"₷\",\n \"₸\",\n \"₹\",\n \"₺\",\n \"₻\",\n \"₼\",\n \"₽\",\n \"₾\",\n \"↚\",\n \"↛\",\n \"↜\",\n \"↝\",\n \"↞\",\n \"↟\",\n \"↠\",\n \"↡\",\n \"↢\",\n \"↣\",\n \"↤\",\n \"↥\",\n \"↦\",\n \"↧\",\n \"↨\",\n \"↩\",\n \"↪\",\n \"↫\",\n \"↬\",\n \"↭\",\n \"↮\",\n \"↯\",\n \"↰\",\n \"↱\",\n \"↲\",\n \"↳\",\n \"↴\",\n \"↵\",\n \"↶\",\n \"↷\",\n \"↸\",\n \"↹\",\n \"⇄\",\n \"⇅\",\n \"⇆\",\n \"⇇\",\n \"⇈\",\n \"⇉\",\n \"⇊\",\n \"⇍\",\n \"⇎\",\n \"⇏\",\n \"⇑\",\n \"⇓\",\n \"⇕\",\n \"⇖\",\n \"⇗\",\n \"⇘\",\n \"⇙\",\n \"⇚\",\n \"⇛\",\n \"⇜\",\n \"⇝\",\n \"⇞\",\n \"⇟\",\n \"⇠\",\n \"⇡\",\n \"⇢\",\n \"⇣\",\n \"⇤\",\n \"⇥\",\n \"⇦\",\n \"⇧\",\n \"⇨\",\n \"⇩\",\n \"⇪\",\n \"⇫\",\n \"⇬\",\n \"⇭\",\n \"⇮\",\n \"⇯\",\n \"⇰\",\n \"⇱\",\n \"⇲\",\n \"⇳\",\n \"⇴\",\n \"⇵\",\n \"⇶\",\n \"⇷\",\n \"⇸\",\n \"⇹\",\n \"⇺\",\n \"⇻\",\n \"⇼\",\n \"⇽\",\n \"⇾\",\n \"⇿\",\n \"ↀ\",\n \"ↁ\",\n \"ↂ\",\n \"☀\",\n \"☁\",\n \"☂\",\n \"☃\",\n \"☄\",\n \"★\",\n \"☇\",\n \"☈\",\n \"☉\",\n \"☊\",\n \"☋\",\n \"☌\",\n \"☍\",\n \"☎\",\n \"☏\",\n \"☐\",\n \"☑\",\n \"☒\",\n \"☓\",\n \"☔\",\n \"☕\",\n \"☖\",\n \"☗\",\n \"☘\",\n \"☙\",\n \"☚\",\n \"☛\",\n \"☜\",\n \"☝\",\n \"☞\",\n \"☟\",\n \"☠\",\n \"☡\",\n \"☢\",\n \"☣\",\n \"☤\",\n \"☥\",\n \"☦\",\n \"☧\",\n \"☨\",\n \"☩\",\n \"☪\",\n \"☫\",\n \"☬\",\n \"☭\",\n \"☮\",\n \"☯\",\n \"☸\",\n \"☹\",\n \"☺\",\n \"☻\",\n \"☼\",\n \"☽\",\n \"☾\",\n \"☿\",\n \"♁\",\n \"♃\",\n \"♄\",\n \"♅\",\n \"♆\",\n \"♇\",\n \"♔\",\n \"♕\",\n \"♖\",\n \"♗\",\n \"♘\",\n \"♙\",\n \"♚\",\n \"♛\",\n \"♜\",\n \"♝\",\n \"♞\",\n \"♟\",\n \"♠\",\n \"♡\",\n \"♢\",\n \"♣\",\n \"♤\",\n \"♥\",\n \"♦\",\n \"♧\",\n \"♨\",\n \"♭\",\n \"♮\",\n \"♯\",\n \"♰\",\n \"♱\",\n \"♲\",\n \"♻\",\n \"♼\",\n \"♽\",\n \"♾\",\n \"⚀\",\n \"⚁\",\n \"⚂\",\n \"⚃\",\n \"⚄\",\n \"⚅\",\n \"⚆\",\n \"⚇\",\n \"⚈\",\n \"⚉\",\n \"⚊\",\n \"⚋\",\n \"⚌\",\n \"⚍\",\n \"⚎\",\n \"⚏\",\n \"⚐\",\n \"⚑\",\n \"⚒\",\n \"⚓\",\n \"⚔\",\n \"⚕\",\n \"⚖\",\n \"⚗\",\n \"⚘\",\n \"⚙\",\n \"⚚\",\n \"⚛\",\n \"⚜\",\n \"⚝\",\n \"⚞\",\n \"⚟\",\n \"⚠\",\n \"⚡\",\n \"⚢\",\n \"⚣\",\n \"⚤\",\n \"⚥\",\n \"⚦\",\n \"⚧\",\n \"⚨\",\n \"⚩\",\n \"⚭\",\n \"⚮\",\n \"⚯\",\n \"⚰\",\n \"⚱\",\n \"⚲\",\n \"⚳\",\n \"⚴\",\n \"⚵\",\n \"⚶\",\n \"⚷\",\n \"⚸\",\n \"⚹\",\n \"⚺\",\n \"⚻\",\n \"⚼\",\n \"⚿\",\n \"⛀\",\n \"⛁\",\n \"⛂\",\n \"⛃\",\n \"⛆\",\n \"⛇\",\n \"⛈\",\n \"⛉\",\n \"⛊\",\n \"⛋\",\n \"⛌\",\n \"⛍\",\n \"⛏\",\n \"⛐\",\n \"⛑\",\n \"⛒\",\n \"⛓\",\n \"⛕\",\n \"⛖\",\n \"⛗\",\n \"⛘\",\n \"⛙\",\n \"⛚\",\n \"⛛\",\n \"⛜\",\n \"⛝\",\n \"⛞\",\n \"⛠\",\n \"⛡\",\n \"⛢\",\n \"⛣\",\n \"⛤\",\n \"⛥\",\n \"⛦\",\n \"⛧\",\n \"⛨\",\n \"⛩\",\n \"⛪\",\n \"⛫\",\n \"⛬\",\n \"⛭\",\n \"⛮\",\n \"⛯\",\n \"⛶\",\n \"⛾\",\n \"⛿\",\n \"✆\",\n \"✇\",\n \"✈\",\n \"✉\",\n \"✌\",\n \"✍\",\n \"✎\",\n \"✏\",\n \"✐\",\n \"✑\",\n \"✒\",\n \"✓\",\n \"✔\",\n \"✕\",\n \"✙\",\n \"✚\",\n \"✛\",\n \"✜\",\n \"✝\",\n \"✞\",\n \"✟\",\n \"✠\",\n \"✡\",\n \"✢\",\n \"✣\",\n \"✤\",\n \"✥\",\n \"✦\",\n \"✧\",\n \"✩\",\n \"✪\",\n \"✫\",\n \"✬\",\n \"✭\",\n \"✮\",\n \"✯\",\n \"✰\",\n \"✱\",\n \"✲\",\n \"✳\",\n \"✴\",\n \"✵\",\n \"✹\",\n \"✺\",\n \"✻\",\n \"✼\",\n \"✽\",\n \"✾\",\n \"✿\",\n \"❀\",\n \"❁\",\n \"❂\",\n \"❃\",\n \"❄\",\n \"❅\",\n \"❆\",\n \"❇\",\n \"❈\",\n \"❉\",\n \"❊\",\n \"❋\",\n \"❍\",\n \"❏\",\n \"❐\",\n \"❑\",\n \"❒\",\n \"❖\",\n \"❘\",\n \"❙\",\n \"❚\",\n \"❛\",\n \"❜\",\n \"❝\",\n \"❞\",\n \"❡\",\n \"❢\",\n \"❣\",\n \"❤\",\n \"❥\",\n \"❦\",\n \"❧\",\n \"❨\",\n \"❩\",\n \"❪\",\n \"❫\",\n \"❬\",\n \"❭\",\n \"❮\",\n \"❯\",\n \"❰\",\n \"❱\",\n \"❲\",\n \"❳\",\n \"❴\",\n \"❵\",\n \"❶\",\n \"❷\",\n \"❸\",\n \"❹\",\n \"❺\",\n \"❻\",\n \"❼\",\n \"❽\",\n \"❾\",\n \"❿\",\n \"①\",\n \"②\",\n \"③\",\n \"④\",\n \"⑤\",\n \"⑥\",\n \"⑦\",\n \"⑧\",\n \"⑨\",\n \"⑩\",\n \"➔\",\n \"➕\",\n \"➖\",\n \"➗\",\n \"➘\",\n \"➙\",\n \"➚\",\n \"➛\",\n \"➜\",\n \"➝\",\n \"➞\",\n \"➟\",\n \"➠\",\n \"➡\",\n \"➢\",\n \"➣\",\n \"➤\",\n \"➥\",\n \"➦\",\n \"➧\",\n \"➨\",\n \"➩\",\n \"➪\",\n \"➫\",\n \"➬\",\n \"➭\",\n \"➮\",\n \"➯\",\n \"➰\",\n \"➱\",\n \"➲\",\n \"➳\",\n \"➴\",\n \"➵\",\n \"➶\",\n \"➷\",\n \"➸\",\n \"➹\",\n \"➺\",\n \"➻\",\n \"➼\",\n \"➽\",\n \"➾\",\n \"➿\",\n \"⌘\",\n \"⌥\",\n \"⌃\",\n \"⎋\",\n \"⌫\",\n \"⌦\",\n \"⏏\",\n \"⌤\",\n \"⌧\",\n \"⌨\",\n \"⎆\",\n \"⎇\",\n \"⎈\",\n \"⎉\",\n \"⎊\",\n \"⎌\",\n \"⎍\",\n \"⎎\",\n \"⎏\",\n \"⎐\",\n \"⎑\",\n \"⎒\",\n \"⎓\",\n \"⎕\",\n \"⎖\",\n \"⎗\",\n \"⎘\",\n \"⎙\",\n \"⎚\",\n \"⎛\",\n \"⎜\",\n \"⎝\",\n \"⎞\",\n \"⎟\",\n \"⎠\",\n \"⎡\",\n \"⎢\",\n \"⎣\",\n \"⎤\",\n \"⎥\",\n \"⎦\",\n \"⎧\",\n \"⎨\",\n \"⎩\",\n \"⎪\",\n \"⎫\",\n \"⎬\",\n \"⎭\",\n \"⎮\",\n \"⎯\",\n \"⎰\",\n \"⎱\",\n \"⎲\",\n \"⎳\",\n \"⎴\",\n \"⎵\",\n \"⎶\",\n \"⎷\",\n \"⎸\",\n \"⎹\",\n \"⎺\",\n \"⎻\",\n \"⎼\",\n \"⎽\",\n \"⎾\",\n \"⎿\",\n \"⏍\",\n \"⏎\",\n \"⏐\",\n \"⏑\",\n \"⏒\",\n \"⏓\",\n \"⏔\",\n \"⏕\",\n \"⏖\",\n \"⏗\",\n \"⏘\",\n \"⏙\",\n \"⏛\",\n \"⏜\",\n \"⏝\",\n \"⏞\",\n \"⏟\",\n \"⏠\",\n \"⏡\",\n \"⏢\",\n \"⏣\",\n \"⏤\",\n \"⏥\",\n \"⏦\",\n \"⏧\",\n \"⏨\",\n \"⏭\",\n \"⏮\",\n \"⏯\",\n \"⏱\",\n \"⏲\",\n \"▲\",\n \"▽\",\n \"◐\",\n \"⏽\",\n \"⏾\",\n \"⏿\",\n \"ɐ\",\n \"ɑ\",\n \"ɒ\",\n \"ɓ\",\n \"ɔ\",\n \"ɕ\",\n \"ɖ\",\n \"ɗ\",\n \"ɘ\",\n \"ə\",\n \"ɚ\",\n \"ɛ\",\n \"ɜ\",\n \"ɝ\",\n \"ɞ\",\n \"ɟ\",\n \"ɠ\",\n \"ɡ\",\n \"ɢ\",\n \"ɣ\",\n \"ɤ\",\n \"ɥ\",\n \"ɦ\",\n \"ɧ\",\n \"ɨ\",\n \"ɩ\",\n \"ɪ\",\n \"ɫ\",\n \"ɬ\",\n \"ɭ\",\n \"ɮ\",\n \"ɯ\",\n \"ɰ\",\n \"ɱ\",\n \"ɲ\",\n \"ɳ\",\n \"ɴ\",\n \"ɵ\",\n \"ɶ\",\n \"ɷ\",\n \"ɸ\",\n \"ɹ\",\n \"ɺ\",\n \"ɻ\",\n \"ɼ\",\n \"ɽ\",\n \"ɾ\",\n \"ɿ\",\n \"ʀ\",\n \"ʁ\",\n \"ʂ\",\n \"ʃ\",\n \"ʄ\",\n \"ʅ\",\n \"ʆ\",\n \"ʇ\",\n \"ʈ\",\n \"ʉ\",\n \"ʊ\",\n \"ʋ\",\n \"ʌ\",\n \"ʍ\",\n \"ʎ\",\n \"ʏ\",\n \"ʐ\",\n \"ʑ\",\n \"ʒ\",\n \"ʓ\",\n \"ʔ\",\n \"ʕ\",\n \"ʖ\",\n \"ʗ\",\n \"ʘ\",\n \"ʙ\",\n \"ʚ\",\n \"ʛ\",\n \"ʜ\",\n \"ʝ\",\n \"ʞ\",\n \"ʟ\",\n \"ʠ\",\n \"ʡ\",\n \"ʢ\",\n \"ʣ\",\n \"ʤ\",\n \"ʥ\",\n \"ʦ\",\n \"ʧ\",\n \"ʨ\",\n \"ʩ\",\n \"ʪ\",\n \"ʫ\",\n \"ʬ\",\n \"ʭ\",\n \"ʮ\",\n \"ʯ\",\n \"━\",\n \"Ǝ\",\n \"Ã\",\n \"●\",\n \"▶\",\n \"｜\",\n \"𝑢\",\n \"〖\",\n \"〗\",\n \"︽\",\n \"–\",\n \"﹥\",\n \"𝜓\",\n \"•\",\n \"∋\",\n \"ƒ\",\n \"०\",\n \"✘\",\n \"Е\",\n \"◉\",\n \"〒\",\n \"𝒱\",\n \"𝜆\",\n \"⟹\",\n \"﹪\",\n \"◊\",\n \"╆\",\n \"오\",\n \"˂\",\n \"〉\",\n \"𝝎\",\n \"▪\",\n \"△\",\n \"▁\",\n \"◼\",\n \"〇\",\n \"▷\",\n \"▬\",\n \"𝒮\",\n \"†\",\n \"ₒ\",\n \"⼁\",\n \"〵\",\n \"⭐\",\n \"╳\",\n \"⟶\",\n \"으\",\n \"⬆\",\n \"Ạ\",\n \"◀\",\n \"\",\n \"▫\",\n \"丄\",\n \"︾\",\n \"◥\",\n \"‖\",\n \"𝜌\",\n \"ⅼ\",\n \"▼\",\n \"⁎\",\n \"﹏\",\n \"😁\",\n \"😂\",\n \"😃\",\n \"😄\",\n \"😅\",\n \"😆\",\n \"😉\",\n \"😊\",\n \"😋\",\n \"😌\",\n \"😍\",\n \"😏\",\n \"😒\",\n \"😓\",\n \"😔\",\n \"😖\",\n \"😘\",\n \"😚\",\n \"😜\",\n \"😝\",\n \"😞\",\n \"😠\",\n \"😡\",\n \"😢\",\n \"😣\",\n \"😤\",\n \"😥\",\n \"😨\",\n \"😩\",\n \"😪\",\n \"😫\",\n \"😭\",\n \"😰\",\n \"😱\",\n \"😲\",\n \"😳\",\n \"😵\",\n \"😷\",\n \"😸\",\n \"😹\",\n \"😺\",\n \"😻\",\n \"😼\",\n \"😽\",\n \"😾\",\n \"😿\",\n \"🙀\",\n \"🙅\",\n \"🙆\",\n \"🙇\",\n \"🙈\",\n \"🙉\",\n \"🙊\",\n \"🙋\",\n \"🙌\",\n \"🙍\",\n \"🙎\",\n \"🙏\",\n \"✂\",\n \"✅\",\n \"✊\",\n \"✋\",\n \"✖\",\n \"✨\",\n \"❌\",\n \"❎\",\n \"❓\",\n \"❔\",\n \"❕\",\n \"❗\",\n \"🚀\",\n \"🚃\",\n \"🚄\",\n \"🚅\",\n \"🚇\",\n \"🚉\",\n \"🚌\",\n \"🚏\",\n \"🚑\",\n \"🚒\",\n \"🚓\",\n \"🚕\",\n \"🚗\",\n \"🚙\",\n \"🚚\",\n \"🚢\",\n \"🚤\",\n \"🚥\",\n \"🚧\",\n \"🚨\",\n \"🚩\",\n \"🚪\",\n \"🚫\",\n \"🚬\",\n \"🚭\",\n \"🚲\",\n \"🚶\",\n \"🚹\",\n \"🚺\",\n \"🚻\",\n \"🚼\",\n \"🚽\",\n \"🚾\",\n \"🛀\",\n \"Ⓜ\",\n \"🅰\",\n \"🅱\",\n \"🅾\",\n \"🅿\",\n \"🆎\",\n \"🆑\",\n \"🆒\",\n \"🆓\",\n \"🆔\",\n \"🆕\",\n \"🆖\",\n \"🆗\",\n \"🆘\",\n \"🆙\",\n \"🆚\",\n \"🇩🇪\",\n \"🇬🇧\",\n \"🇨🇳\",\n \"🇯🇵\",\n \"🇫🇷\",\n \"🇰🇷\",\n \"🇪🇸\",\n \"🇮🇹\",\n \"🇷🇺\",\n \"🇺🇸\",\n \"🈁\",\n \"ℹ\",\n \"⌚\",\n \"⌛\",\n \"⏩\",\n \"⏪\",\n \"⏫\",\n \"⏬\",\n \"⏰\",\n \"⏳\",\n \"◻\",\n \"◽\",\n \"◾\",\n \"♈\",\n \"♉\",\n \"♊\",\n \"♋\",\n \"♌\",\n \"♍\",\n \"♎\",\n \"♏\",\n \"♐\",\n \"♑\",\n \"♒\",\n \"♓\",\n \"♿\",\n \"⚽\",\n \"⚾\",\n \"⛄\",\n \"⛅\",\n \"⛎\",\n \"⛔\",\n \"⛲\",\n \"⛳\",\n \"⛵\",\n \"⛺\",\n \"⛽\",\n \"⤴\",\n \"⤵\",\n \"⬅\",\n \"⬇\",\n \"⬛\",\n \"⬜\",\n \"⭕\",\n \"〰\",\n \"〽\",\n \"㊗\",\n \"㊙\",\n \"🀄\",\n \"🃏\",\n \"🌀\",\n \"🌁\",\n \"🌂\",\n \"🌃\",\n \"🌄\",\n \"🌅\",\n \"🌆\",\n \"🌇\",\n \"🌈\",\n \"🌉\",\n \"🌊\",\n \"🌋\",\n \"🌌\",\n \"🌏\",\n \"🌑\",\n \"🌓\",\n \"🌔\",\n \"🌕\",\n \"🌙\",\n \"🌛\",\n \"🌟\",\n \"🌠\",\n \"🌰\",\n \"🌱\",\n \"🌴\",\n \"🌵\",\n \"🌷\",\n \"🌸\",\n \"🌹\",\n \"🌺\",\n \"🌻\",\n \"🌼\",\n \"🌽\",\n \"🌾\",\n \"🌿\",\n \"🍀\",\n \"🍁\",\n \"🍂\",\n \"🍃\",\n \"🍄\",\n \"🍅\",\n \"🍆\",\n \"🍇\",\n \"🍈\",\n \"🍉\",\n \"🍊\",\n \"🍌\",\n \"🍍\",\n \"🍎\",\n \"🍏\",\n \"🍑\",\n \"🍒\",\n \"🍓\",\n \"🍔\",\n \"🍕\",\n \"🍖\",\n \"🍗\",\n \"🍘\",\n \"🍙\",\n \"🍚\",\n \"🍛\",\n \"🍜\",\n \"🍝\",\n \"🍞\",\n \"🍟\",\n \"🍠\",\n \"🍡\",\n \"🍢\",\n \"🍣\",\n \"🍤\",\n \"🍥\",\n \"🍦\",\n \"🍧\",\n \"🍨\",\n \"🍩\",\n \"🍪\",\n \"🍫\",\n \"🍬\",\n \"🍭\",\n \"🍮\",\n \"🍯\",\n \"🍰\",\n \"🍱\",\n \"🍲\",\n \"🍳\",\n \"🍴\",\n \"🍵\",\n \"🍶\",\n \"🍷\",\n \"🍸\",\n \"🍹\",\n \"🍺\",\n \"🍻\",\n \"🎀\",\n \"🎁\",\n \"🎂\",\n \"🎃\",\n \"🎄\",\n \"🎅\",\n \"🎆\",\n \"🎇\",\n \"🎈\",\n \"🎉\",\n \"🎊\",\n \"🎋\",\n \"🎌\",\n \"🎍\",\n \"🎎\",\n \"🎏\",\n \"🎐\",\n \"🎑\",\n \"🎒\",\n \"🎓\",\n \"🎠\",\n \"🎡\",\n \"🎢\",\n \"🎣\",\n \"🎤\",\n \"🎥\",\n \"🎦\",\n \"🎧\",\n \"🎨\",\n \"🎩\",\n \"🎪\",\n \"🎫\",\n \"🎬\",\n \"🎭\",\n \"🎮\",\n \"🎯\",\n \"🎰\",\n \"🎱\",\n \"🎲\",\n \"🎳\",\n \"🎴\",\n \"🎵\",\n \"🎶\",\n \"🎷\",\n \"🎸\",\n \"🎹\",\n \"🎺\",\n \"🎻\",\n \"🎼\",\n \"🎽\",\n \"🎾\",\n \"🎿\",\n \"🏀\",\n \"🏁\",\n \"🏂\",\n \"🏃\",\n \"🏄\",\n \"🏆\",\n \"🏈\",\n \"🏊\",\n \"🏠\",\n \"🏡\",\n \"🏢\",\n \"🏣\",\n \"🏥\",\n \"🏦\",\n \"🏧\",\n \"🏨\",\n \"🏩\",\n \"🏪\",\n \"🏫\",\n \"🏬\",\n \"🏭\",\n \"🏮\",\n \"🏯\",\n \"🏰\",\n \"🐌\",\n \"🐍\",\n \"🐎\",\n \"🐑\",\n \"🐒\",\n \"🐔\",\n \"🐗\",\n \"🐘\",\n \"🐙\",\n \"🐚\",\n \"🐛\",\n \"🐜\",\n \"🐝\",\n \"🐞\",\n \"🐟\",\n \"🐠\",\n \"🐡\",\n \"🐢\",\n \"🐣\",\n \"🐤\",\n \"🐥\",\n \"🐦\",\n \"🐧\",\n \"🐨\",\n \"🐩\",\n \"🐫\",\n \"🐬\",\n \"🐭\",\n \"🐮\",\n \"🐯\",\n \"🐰\",\n \"🐱\",\n \"🐲\",\n \"🐳\",\n \"🐴\",\n \"🐵\",\n \"🐶\",\n \"🐷\",\n \"🐸\",\n \"🐹\",\n \"🐺\",\n \"🐻\",\n \"🐼\",\n \"🐽\",\n \"🐾\",\n \"👀\",\n \"👂\",\n \"👃\",\n \"👄\",\n \"👅\",\n \"👆\",\n \"👇\",\n \"👈\",\n \"👉\",\n \"👊\",\n \"👋\",\n \"👌\",\n \"👍\",\n \"👎\",\n \"👏\",\n \"👐\",\n \"👑\",\n \"👒\",\n \"👓\",\n \"👔\",\n \"👕\",\n \"👖\",\n \"👗\",\n \"👘\",\n \"👙\",\n \"👚\",\n \"👛\",\n \"👜\",\n \"👝\",\n \"👞\",\n \"👟\",\n \"👠\",\n \"👡\",\n \"👢\",\n \"👣\",\n \"👤\",\n \"👦\",\n \"👧\",\n \"👨\",\n \"👩\",\n \"👪\",\n \"👫\",\n \"👮\",\n \"👯\",\n \"👰\",\n \"👱\",\n \"👲\",\n \"👳\",\n \"👴\",\n \"👵\",\n \"👶\",\n \"👷\",\n \"👸\",\n \"👹\",\n \"👺\",\n \"👻\",\n \"👼\",\n \"👽\",\n \"👾\",\n \"👿\",\n \"💀\",\n \"💁\",\n \"💂\",\n \"💃\",\n \"💄\",\n \"💅\",\n \"💆\",\n \"💇\",\n \"💈\",\n \"💉\",\n \"💊\",\n \"💋\",\n \"💌\",\n \"💍\",\n \"💎\",\n \"💏\",\n \"💐\",\n \"💑\",\n \"💒\",\n \"💓\",\n \"💔\",\n \"💕\",\n \"💖\",\n \"💗\",\n \"💘\",\n \"💙\",\n \"💚\",\n \"💛\",\n \"💜\",\n \"💝\",\n \"💞\",\n \"💟\",\n \"💠\",\n \"💡\",\n \"💢\",\n \"💣\",\n \"💤\",\n \"💥\",\n \"💦\",\n \"💧\",\n \"💨\",\n \"💩\",\n \"💪\",\n \"💫\",\n \"💬\",\n \"💮\",\n \"💯\",\n \"💰\",\n \"💲\",\n \"💳\",\n \"💴\",\n \"💵\",\n \"💸\",\n \"💹\",\n \"💺\",\n \"💻\",\n \"💼\",\n \"💽\",\n \"💾\",\n \"💿\",\n \"📀\",\n \"📁\",\n \"📂\",\n \"📃\",\n \"📄\",\n \"📅\",\n \"📆\",\n \"📇\",\n \"📈\",\n \"📉\",\n \"📊\",\n \"📋\",\n \"📌\",\n \"📍\",\n \"📎\",\n \"📏\",\n \"📐\",\n \"📑\",\n \"📒\",\n \"📓\",\n \"📔\",\n \"📕\",\n \"📖\",\n \"📗\",\n \"📘\",\n \"📙\",\n \"📚\",\n \"📛\",\n \"📜\",\n \"📝\",\n \"📞\",\n \"📟\",\n \"📠\",\n \"📡\",\n \"📢\",\n \"📣\",\n \"📤\",\n \"📥\",\n \"📦\",\n \"📧\",\n \"📨\",\n \"📩\",\n \"📪\",\n \"📫\",\n \"📮\",\n \"📰\",\n \"📱\",\n \"📲\",\n \"📳\",\n \"📴\",\n \"📶\",\n \"📷\",\n \"📹\",\n \"📺\",\n \"📻\",\n \"📼\",\n \"🔃\",\n \"🔊\",\n \"🔋\",\n \"🔌\",\n \"🔍\",\n \"🔎\",\n \"🔏\",\n \"🔐\",\n \"🔑\",\n \"🔒\",\n \"🔓\",\n \"🔔\",\n \"🔖\",\n \"🔗\",\n \"🔘\",\n \"🔙\",\n \"🔚\",\n \"🔛\",\n \"🔜\",\n \"🔝\",\n \"🔞\",\n \"🔟\",\n \"🔠\",\n \"🔡\",\n \"🔢\",\n \"🔣\",\n \"🔤\",\n \"🔥\",\n \"🔦\",\n \"🔧\",\n \"🔨\",\n \"🔩\",\n \"🔪\",\n \"🔫\",\n \"🔮\",\n \"🔯\",\n \"🔰\",\n \"🔱\",\n \"🔲\",\n \"🔳\",\n \"🔴\",\n \"🔵\",\n \"🔶\",\n \"🔷\",\n \"🔸\",\n \"🔹\",\n \"🔺\",\n \"🔻\",\n \"🔼\",\n \"🔽\",\n \"🕐\",\n \"🕑\",\n \"🕒\",\n \"🕓\",\n \"🕔\",\n \"🕕\",\n \"🕖\",\n \"🕗\",\n \"🕘\",\n \"🕙\",\n \"🕚\",\n \"🕛\",\n \"🗻\",\n \"🗼\",\n \"🗽\",\n \"🗾\",\n \"🗿\",\n \"😀\",\n \"😇\",\n \"😈\",\n \"😎\",\n \"😐\",\n \"😑\",\n \"😕\",\n \"😗\",\n \"😙\",\n \"😛\",\n \"😟\",\n \"😦\",\n \"😧\",\n \"😬\",\n \"😮\",\n \"😯\",\n \"😴\",\n \"😶\",\n \"🚁\",\n \"🚂\",\n \"🚆\",\n \"🚈\",\n \"🚊\",\n \"🚍\",\n \"🚎\",\n \"🚐\",\n \"🚔\",\n \"🚖\",\n \"🚘\",\n \"🚛\",\n \"🚜\",\n \"🚝\",\n \"🚞\",\n \"🚟\",\n \"🚠\",\n \"🚡\",\n \"🚣\",\n \"🚦\",\n \"🚮\",\n \"🚯\",\n \"🚰\",\n \"🚱\",\n \"🚳\",\n \"🚴\",\n \"🚵\",\n \"🚷\",\n \"🚸\",\n \"🚿\",\n \"🛁\",\n \"🛂\",\n \"🛃\",\n \"🛄\",\n \"🛅\",\n \"🌍\",\n \"🌎\",\n \"🌐\",\n \"🌒\",\n \"🌖\",\n \"🌗\",\n \"🌘\",\n \"🌚\",\n \"🌜\",\n \"🌝\",\n \"🌞\",\n \"🌲\",\n \"🌳\",\n \"🍋\",\n \"🍐\",\n \"🍼\",\n \"🏇\",\n \"🏉\",\n \"🏤\",\n \"🐀\",\n \"🐁\",\n \"🐂\",\n \"🐃\",\n \"🐄\",\n \"🐅\",\n \"🐆\",\n \"🐇\",\n \"🐈\",\n \"🐉\",\n \"🐊\",\n \"🐋\",\n \"🐏\",\n \"🐐\",\n \"🐓\",\n \"🐕\",\n \"🐖\",\n \"🐪\",\n \"👥\",\n \"👬\",\n \"👭\",\n \"💭\",\n \"💶\",\n \"💷\",\n \"📬\",\n \"📭\",\n \"📯\",\n \"📵\",\n \"🔀\",\n \"🔁\",\n \"🔂\",\n \"🔄\",\n \"🔅\",\n \"🔆\",\n \"🔇\",\n \"🔉\",\n \"🔕\",\n \"🔬\",\n \"🔭\",\n \"🕜\",\n \"🕝\",\n \"🕞\",\n \"🕟\",\n \"🕠\",\n \"🕡\",\n \"🕢\",\n \"🕣\",\n \"🕤\",\n \"🕥\",\n \"🕦\",\n \"🕧\"\n};\n\nstatic const int character_dict_size = sizeof(character_dict) / sizeof(const char*);\n" }, { "path": "examples/retinaface.cpp", "content": "// Copyright 2019 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct FaceObject\n{\n cv::Rect_ rect;\n cv::Point2f landmark[5];\n float prob;\n};\n\nstatic inline float intersection_area(const FaceObject& a, const FaceObject& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const FaceObject& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const FaceObject& b = faceobjects[picked[j]];\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\n// copy from src/layer/proposal.cpp\nstatic ncnn::Mat generate_anchors(int base_size, const ncnn::Mat& ratios, const ncnn::Mat& scales)\n{\n int num_ratio = ratios.w;\n int num_scale = scales.w;\n\n ncnn::Mat anchors;\n anchors.create(4, num_ratio * num_scale);\n\n const float cx = base_size * 0.5f;\n const float cy = base_size * 0.5f;\n\n for (int i = 0; i < num_ratio; i++)\n {\n float ar = ratios[i];\n\n int r_w = round(base_size / sqrt(ar));\n int r_h = round(r_w * ar); //round(base_size * sqrt(ar));\n\n for (int j = 0; j < num_scale; j++)\n {\n float scale = scales[j];\n\n float rs_w = r_w * scale;\n float rs_h = r_h * scale;\n\n float* anchor = anchors.row(i * num_scale + j);\n\n anchor[0] = cx - rs_w * 0.5f;\n anchor[1] = cy - rs_h * 0.5f;\n anchor[2] = cx + rs_w * 0.5f;\n anchor[3] = cy + rs_h * 0.5f;\n }\n }\n\n return anchors;\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int feat_stride, const ncnn::Mat& score_blob, const ncnn::Mat& bbox_blob, const ncnn::Mat& landmark_blob, float prob_threshold, std::vector& faceobjects)\n{\n int w = score_blob.w;\n int h = score_blob.h;\n\n // generate face proposal from bbox deltas and shifted anchors\n const int num_anchors = anchors.h;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float* anchor = anchors.row(q);\n\n const ncnn::Mat score = score_blob.channel(q + num_anchors);\n const ncnn::Mat bbox = bbox_blob.channel_range(q * 4, 4);\n const ncnn::Mat landmark = landmark_blob.channel_range(q * 10, 10);\n\n // shifted anchor\n float anchor_y = anchor[1];\n\n float anchor_w = anchor[2] - anchor[0];\n float anchor_h = anchor[3] - anchor[1];\n\n for (int i = 0; i < h; i++)\n {\n float anchor_x = anchor[0];\n\n for (int j = 0; j < w; j++)\n {\n int index = i * w + j;\n\n float prob = score[index];\n\n if (prob >= prob_threshold)\n {\n // apply center size\n float dx = bbox.channel(0)[index];\n float dy = bbox.channel(1)[index];\n float dw = bbox.channel(2)[index];\n float dh = bbox.channel(3)[index];\n\n float cx = anchor_x + anchor_w * 0.5f;\n float cy = anchor_y + anchor_h * 0.5f;\n\n float pb_cx = cx + anchor_w * dx;\n float pb_cy = cy + anchor_h * dy;\n\n float pb_w = anchor_w * exp(dw);\n float pb_h = anchor_h * exp(dh);\n\n float x0 = pb_cx - pb_w * 0.5f;\n float y0 = pb_cy - pb_h * 0.5f;\n float x1 = pb_cx + pb_w * 0.5f;\n float y1 = pb_cy + pb_h * 0.5f;\n\n FaceObject obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0 + 1;\n obj.rect.height = y1 - y0 + 1;\n obj.landmark[0].x = cx + (anchor_w + 1) * landmark.channel(0)[index];\n obj.landmark[0].y = cy + (anchor_h + 1) * landmark.channel(1)[index];\n obj.landmark[1].x = cx + (anchor_w + 1) * landmark.channel(2)[index];\n obj.landmark[1].y = cy + (anchor_h + 1) * landmark.channel(3)[index];\n obj.landmark[2].x = cx + (anchor_w + 1) * landmark.channel(4)[index];\n obj.landmark[2].y = cy + (anchor_h + 1) * landmark.channel(5)[index];\n obj.landmark[3].x = cx + (anchor_w + 1) * landmark.channel(6)[index];\n obj.landmark[3].y = cy + (anchor_h + 1) * landmark.channel(7)[index];\n obj.landmark[4].x = cx + (anchor_w + 1) * landmark.channel(8)[index];\n obj.landmark[4].y = cy + (anchor_h + 1) * landmark.channel(9)[index];\n obj.prob = prob;\n\n faceobjects.push_back(obj);\n }\n\n anchor_x += feat_stride;\n }\n\n anchor_y += feat_stride;\n }\n }\n}\n\nstatic int detect_retinaface(const cv::Mat& bgr, std::vector& faceobjects)\n{\n ncnn::Net retinaface;\n\n retinaface.opt.use_vulkan_compute = true;\n\n // model is converted from\n // https://github.com/deepinsight/insightface/tree/master/RetinaFace#retinaface-pretrained-models\n // https://github.com/deepinsight/insightface/issues/669\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n // retinaface.load_param(\"retinaface-R50.param\");\n // retinaface.load_model(\"retinaface-R50.bin\");\n if (retinaface.load_param(\"mnet.25-opt.param\"))\n exit(-1);\n if (retinaface.load_model(\"mnet.25-opt.bin\"))\n exit(-1);\n\n const float prob_threshold = 0.8f;\n const float nms_threshold = 0.4f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h);\n\n ncnn::Extractor ex = retinaface.create_extractor();\n\n ex.input(\"data\", in);\n\n std::vector faceproposals;\n\n // stride 32\n {\n ncnn::Mat score_blob, bbox_blob, landmark_blob;\n ex.extract(\"face_rpn_cls_prob_reshape_stride32\", score_blob);\n ex.extract(\"face_rpn_bbox_pred_stride32\", bbox_blob);\n ex.extract(\"face_rpn_landmark_pred_stride32\", landmark_blob);\n\n const int base_size = 16;\n const int feat_stride = 32;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 32.f;\n scales[1] = 16.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects32;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects32);\n\n faceproposals.insert(faceproposals.end(), faceobjects32.begin(), faceobjects32.end());\n }\n\n // stride 16\n {\n ncnn::Mat score_blob, bbox_blob, landmark_blob;\n ex.extract(\"face_rpn_cls_prob_reshape_stride16\", score_blob);\n ex.extract(\"face_rpn_bbox_pred_stride16\", bbox_blob);\n ex.extract(\"face_rpn_landmark_pred_stride16\", landmark_blob);\n\n const int base_size = 16;\n const int feat_stride = 16;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 8.f;\n scales[1] = 4.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects16;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects16);\n\n faceproposals.insert(faceproposals.end(), faceobjects16.begin(), faceobjects16.end());\n }\n\n // stride 8\n {\n ncnn::Mat score_blob, bbox_blob, landmark_blob;\n ex.extract(\"face_rpn_cls_prob_reshape_stride8\", score_blob);\n ex.extract(\"face_rpn_bbox_pred_stride8\", bbox_blob);\n ex.extract(\"face_rpn_landmark_pred_stride8\", landmark_blob);\n\n const int base_size = 16;\n const int feat_stride = 8;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 2.f;\n scales[1] = 1.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects8;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold, faceobjects8);\n\n faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(faceproposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(faceproposals, picked, nms_threshold);\n\n int face_count = picked.size();\n\n faceobjects.resize(face_count);\n for (int i = 0; i < face_count; i++)\n {\n faceobjects[i] = faceproposals[picked[i]];\n\n // clip to image size\n float x0 = faceobjects[i].rect.x;\n float y0 = faceobjects[i].rect.y;\n float x1 = x0 + faceobjects[i].rect.width;\n float y1 = y0 + faceobjects[i].rect.height;\n\n x0 = std::max(std::min(x0, (float)img_w - 1), 0.f);\n y0 = std::max(std::min(y0, (float)img_h - 1), 0.f);\n x1 = std::max(std::min(x1, (float)img_w - 1), 0.f);\n y1 = std::max(std::min(y1, (float)img_h - 1), 0.f);\n\n faceobjects[i].rect.x = x0;\n faceobjects[i].rect.y = y0;\n faceobjects[i].rect.width = x1 - x0;\n faceobjects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_faceobjects(const cv::Mat& bgr, const std::vector& faceobjects)\n{\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < faceobjects.size(); i++)\n {\n const FaceObject& obj = faceobjects[i];\n\n fprintf(stderr, \"%.5f at %.2f %.2f %.2f x %.2f\\n\", obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));\n\n cv::circle(image, obj.landmark[0], 2, cv::Scalar(0, 255, 255), -1);\n cv::circle(image, obj.landmark[1], 2, cv::Scalar(0, 255, 255), -1);\n cv::circle(image, obj.landmark[2], 2, cv::Scalar(0, 255, 255), -1);\n cv::circle(image, obj.landmark[3], 2, cv::Scalar(0, 255, 255), -1);\n cv::circle(image, obj.landmark[4], 2, cv::Scalar(0, 255, 255), -1);\n\n char text[256];\n sprintf(text, \"%.1f%%\", obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector faceobjects;\n detect_retinaface(m, faceobjects);\n\n draw_faceobjects(m, faceobjects);\n\n return 0;\n}\n" }, { "path": "examples/rfcn.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic int detect_rfcn(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net rfcn;\n\n rfcn.opt.use_vulkan_compute = true;\n\n // original pretrained model from https://github.com/YuwenXiong/py-R-FCN\n // https://github.com/YuwenXiong/py-R-FCN/blob/master/models/pascal_voc/ResNet-50/rfcn_end2end/test_agnostic.prototxt\n // https://1drv.ms/u/s!AoN7vygOjLIQqUWHpY67oaC7mopf\n // resnet50_rfcn_final.caffemodel\n if (rfcn.load_param(\"rfcn_end2end.param\"))\n exit(-1);\n if (rfcn.load_model(\"rfcn_end2end.bin\"))\n exit(-1);\n\n const int target_size = 224;\n\n const int max_per_image = 100;\n const float confidence_thresh = 0.6f; // CONF_THRESH\n\n const float nms_threshold = 0.3f; // NMS_THRESH\n\n // scale to target detect size\n int w = bgr.cols;\n int h = bgr.rows;\n float scale = 1.f;\n if (w < h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, w, h);\n\n const float mean_vals[3] = {102.9801f, 115.9465f, 122.7717f};\n in.substract_mean_normalize(mean_vals, 0);\n\n ncnn::Mat im_info(3);\n im_info[0] = h;\n im_info[1] = w;\n im_info[2] = scale;\n\n // step1, extract feature and all rois\n ncnn::Extractor ex1 = rfcn.create_extractor();\n\n ex1.input(\"data\", in);\n ex1.input(\"im_info\", im_info);\n\n ncnn::Mat rfcn_cls;\n ncnn::Mat rfcn_bbox;\n ncnn::Mat rois; // all rois\n ex1.extract(\"rfcn_cls\", rfcn_cls);\n ex1.extract(\"rfcn_bbox\", rfcn_bbox);\n ex1.extract(\"rois\", rois);\n\n // step2, extract bbox and score for each roi\n std::vector > class_candidates;\n for (int i = 0; i < rois.c; i++)\n {\n ncnn::Extractor ex2 = rfcn.create_extractor();\n\n ncnn::Mat roi = rois.channel(i); // get single roi\n ex2.input(\"rfcn_cls\", rfcn_cls);\n ex2.input(\"rfcn_bbox\", rfcn_bbox);\n ex2.input(\"rois\", roi);\n\n ncnn::Mat bbox_pred;\n ncnn::Mat cls_prob;\n ex2.extract(\"bbox_pred\", bbox_pred);\n ex2.extract(\"cls_prob\", cls_prob);\n\n int num_class = cls_prob.w;\n class_candidates.resize(num_class);\n\n // find class id with highest score\n int label = 0;\n float score = 0.f;\n for (int i = 0; i < num_class; i++)\n {\n float class_score = cls_prob[i];\n if (class_score > score)\n {\n label = i;\n score = class_score;\n }\n }\n\n // ignore background or low score\n if (label == 0 || score <= confidence_thresh)\n continue;\n\n // fprintf(stderr, \"%d = %f\\n\", label, score);\n\n // unscale to image size\n float x1 = roi[0] / scale;\n float y1 = roi[1] / scale;\n float x2 = roi[2] / scale;\n float y2 = roi[3] / scale;\n\n float pb_w = x2 - x1 + 1;\n float pb_h = y2 - y1 + 1;\n\n // apply bbox regression\n float dx = bbox_pred[4];\n float dy = bbox_pred[4 + 1];\n float dw = bbox_pred[4 + 2];\n float dh = bbox_pred[4 + 3];\n\n float cx = x1 + pb_w * 0.5f;\n float cy = y1 + pb_h * 0.5f;\n\n float obj_cx = cx + pb_w * dx;\n float obj_cy = cy + pb_h * dy;\n\n float obj_w = pb_w * exp(dw);\n float obj_h = pb_h * exp(dh);\n\n float obj_x1 = obj_cx - obj_w * 0.5f;\n float obj_y1 = obj_cy - obj_h * 0.5f;\n float obj_x2 = obj_cx + obj_w * 0.5f;\n float obj_y2 = obj_cy + obj_h * 0.5f;\n\n // clip\n obj_x1 = std::max(std::min(obj_x1, (float)(bgr.cols - 1)), 0.f);\n obj_y1 = std::max(std::min(obj_y1, (float)(bgr.rows - 1)), 0.f);\n obj_x2 = std::max(std::min(obj_x2, (float)(bgr.cols - 1)), 0.f);\n obj_y2 = std::max(std::min(obj_y2, (float)(bgr.rows - 1)), 0.f);\n\n // append object\n Object obj;\n obj.rect = cv::Rect_(obj_x1, obj_y1, obj_x2 - obj_x1 + 1, obj_y2 - obj_y1 + 1);\n obj.label = label;\n obj.prob = score;\n\n class_candidates[label].push_back(obj);\n }\n\n // post process\n objects.clear();\n for (int i = 0; i < (int)class_candidates.size(); i++)\n {\n std::vector& candidates = class_candidates[i];\n\n qsort_descent_inplace(candidates);\n\n std::vector picked;\n nms_sorted_bboxes(candidates, picked, nms_threshold);\n\n for (int j = 0; j < (int)picked.size(); j++)\n {\n int z = picked[j];\n objects.push_back(candidates[z]);\n }\n }\n\n qsort_descent_inplace(objects);\n\n if (max_per_image > 0 && max_per_image < objects.size())\n {\n objects.resize(max_per_image);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_rfcn(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/rvm.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// ncnn model exported from https://github.com/PeterL1n/RobustVideoMatting\n//\n// import torch\n// from torch import nn\n// from model import MattingNetwork\n// from model.fast_guided_filter import FastGuidedFilterRefiner\n// from model.deep_guided_filter import DeepGuidedFilterRefiner\n//\n// class Model(nn.Module):\n// def __init__(self):\n// super().__init__()\n//\n// self.rvm = MattingNetwork('mobilenetv3').eval()\n// self.rvm.load_state_dict(torch.load('rvm_mobilenetv3.pth'))\n//\n// self.refiner_deep = DeepGuidedFilterRefiner()\n// self.refiner_fast = FastGuidedFilterRefiner()\n//\n// def forward_first_frame(self, src):\n// return self.rvm(src)\n//\n// def forward(self, src, src_sm, r1, r2, r3, r4):\n//\n// f1, f2, f3, f4 = self.rvm.backbone(src_sm)\n// f4 = self.rvm.aspp(f4)\n// hid, *rec = self.rvm.decoder(src_sm, f1, f2, f3, f4, r1, r2, r3, r4)\n//\n// # downsample\n// fgr_residual, pha = self.rvm.project_mat(hid).split([3, 1], dim=-3)\n// fgr = fgr_residual + src_sm\n//\n// # downsample + refiner_deep\n// fgr_residual_deep, pha_deep = self.refiner_deep(src, src_sm, fgr_residual, pha, hid)\n// fgr_deep = fgr_residual_deep + src\n//\n// # downsample + refiner_fast\n// fgr_residual_fast, pha_fast = self.refiner_fast(src, src_sm, fgr_residual, pha, hid)\n// fgr_fast = fgr_residual_fast + src\n//\n// # downsample + segmentation\n// seg = self.rvm.project_seg(hid)\n//\n// return fgr, pha, fgr_deep, pha_deep, fgr_fast, pha_fast, seg, *rec\n//\n// import pnnx\n//\n// model = Model().eval()\n//\n// x = torch.rand(1, 3, 512, 512)\n// x2 = torch.rand(1, 3, 256, 256)\n// x2_hr = torch.rand(1, 3, 1024, 1024)\n//\n// # generate feats via forward_first_frame, with different shapes\n// fgr, pha, r1, r2, r3, r4 = model.forward_first_frame(x)\n// fgr2, pha2, r12, r22, r32, r42 = model.forward_first_frame(x2)\n//\n// # export with dynamic shape\n// pnnx.export(model, \"rvm_mobilenetv3.pt\", (x, x, r1, r2, r3, r4), (x2_hr, x2, r12, r22, r32, r42))\n//\n// and then fix refiner_fast fp16 overflow issue in ncnn.param via appending 31=1 layer feat mask\n//\n// BinaryOp div_58 2 1 401 399 402 0=3 31=1\n//\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n\nstatic int detect_rvm(const cv::Mat& bgr, cv::Mat& fgr, cv::Mat& pha, cv::Mat& seg)\n{\n ncnn::Net rvm;\n\n rvm.opt.use_vulkan_compute = true;\n\n // https://github.com/nihui/ncnn-android-rvm/tree/master/app/src/main/assets\n // you shall also change r1,r2,r3,r4 shape below when model changed\n if (rvm.load_param(\"rvm_mobilenetv3.ncnn.param\"))\n exit(-1);\n if (rvm.load_model(\"rvm_mobilenetv3.ncnn.bin\"))\n exit(-1);\n // if (rvm.load_param(\"rvm_resnet50.ncnn.param\"))\n // exit(-1);\n // if (rvm.load_model(\"rvm_resnet50.ncnn.bin\"))\n // exit(-1);\n\n const int w = bgr.cols;\n const int h = bgr.rows;\n\n const int target_size = 512;\n const int max_stride = 16;\n\n bool refine_deep = true;\n // bool refine_fast = true;\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n\n ncnn::Mat in_pad;\n ncnn::Mat in_small_pad;\n\n int wpad = 0;\n int hpad = 0;\n\n bool downsample = std::max(w, h) > target_size;\n if (downsample)\n {\n // letterbox pad to multiple of max_stride\n int w2 = w;\n int h2 = h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w2 = target_size;\n h2 = h2 * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h2 = target_size;\n w2 = w2 * scale;\n }\n\n ncnn::Mat in_small = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, w, h, w2, h2);\n\n // letterbox pad to target_size rectangle\n int w2pad = (w2 + max_stride - 1) / max_stride * max_stride - w2;\n int h2pad = (h2 + max_stride - 1) / max_stride * max_stride - h2;\n ncnn::copy_make_border(in_small, in_small_pad, h2pad / 2, h2pad - h2pad / 2, w2pad / 2, w2pad - w2pad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n in_small_pad.substract_mean_normalize(0, norm_vals);\n\n int w3 = w;\n int h3 = h;\n if (w > h)\n {\n w3 = w;\n h3 = in_small_pad.h / scale;\n wpad = 0;\n hpad = h3 - h;\n }\n else\n {\n h3 = h;\n w3 = in_small_pad.w / scale;\n wpad = w3 - w;\n hpad = 0;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, w, h);\n\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n in_pad.substract_mean_normalize(0, norm_vals);\n }\n else\n {\n ncnn::Mat in = ncnn::Mat::from_pixels(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, w, h);\n\n // letterbox pad to target_size rectangle\n wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n in_pad.substract_mean_normalize(0, norm_vals);\n\n in_small_pad = in_pad;\n }\n\n // rvm_mobilenetv3\n ncnn::Mat r1(in_small_pad.w / 2, in_small_pad.h / 2, 16);\n ncnn::Mat r2(in_small_pad.w / 4, in_small_pad.h / 4, 20);\n ncnn::Mat r3(in_small_pad.w / 8, in_small_pad.h / 8, 40);\n ncnn::Mat r4(in_small_pad.w / 16, in_small_pad.h / 16, 64);\n\n // rvm_resnet50\n // ncnn::Mat r1(in_small_pad.w / 2, in_small_pad.h / 2, 16);\n // ncnn::Mat r2(in_small_pad.w / 4, in_small_pad.h / 4, 32);\n // ncnn::Mat r3(in_small_pad.w / 8, in_small_pad.h / 8, 64);\n // ncnn::Mat r4(in_small_pad.w / 16, in_small_pad.h / 16, 128);\n\n r1.fill(0.f);\n r2.fill(0.f);\n r3.fill(0.f);\n r4.fill(0.f);\n\n ncnn::Extractor ex = rvm.create_extractor();\n\n ex.input(\"in0\", in_pad);\n ex.input(\"in1\", in_small_pad);\n\n ex.input(\"in2\", r1);\n ex.input(\"in3\", r2);\n ex.input(\"in4\", r3);\n ex.input(\"in5\", r4);\n\n ncnn::Mat out_fgr;\n ncnn::Mat out_pha;\n\n if (downsample)\n {\n if (refine_deep)\n {\n // downsample + refine deep\n ex.extract(\"out2\", out_fgr);\n ex.extract(\"out3\", out_pha);\n }\n else // if (refine_fast)\n {\n // downsample + refine fast\n ex.extract(\"out4\", out_fgr);\n ex.extract(\"out5\", out_pha);\n }\n }\n else\n {\n // no downsample\n ex.extract(\"out0\", out_fgr);\n ex.extract(\"out1\", out_pha);\n }\n\n ncnn::Mat out_seg;\n\n // segmentation\n ex.extract(\"out6\", out_seg);\n\n // feats\n ex.extract(\"out7\", r1);\n ex.extract(\"out8\", r2);\n ex.extract(\"out9\", r3);\n ex.extract(\"out10\", r4);\n\n const float denorm_vals[3] = {255.f, 255.f, 255.f};\n\n out_fgr.substract_mean_normalize(0, denorm_vals);\n fgr.create(out_fgr.h, out_fgr.w, CV_8UC3);\n out_fgr.to_pixels(fgr.data, ncnn::Mat::PIXEL_RGB2BGR);\n\n out_pha.substract_mean_normalize(0, denorm_vals);\n pha.create(out_pha.h, out_pha.w, CV_8UC1);\n out_pha.to_pixels(pha.data, ncnn::Mat::PIXEL_GRAY);\n\n out_seg.substract_mean_normalize(0, denorm_vals);\n seg.create(in_pad.h, in_pad.w, CV_8UC1);\n out_seg.to_pixels_resize(seg.data, ncnn::Mat::PIXEL_GRAY, in_pad.w, in_pad.h);\n\n // cut letterbox pad\n fgr = fgr(cv::Rect(wpad / 2, hpad / 2, w, h));\n pha = pha(cv::Rect(wpad / 2, hpad / 2, w, h));\n seg = seg(cv::Rect(wpad / 2, hpad / 2, w, h));\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const cv::Mat& fgr, const cv::Mat& pha, const cv::Mat& seg)\n{\n const int w = bgr.cols;\n const int h = bgr.rows;\n\n // composite\n cv::Mat comp(h, w, CV_8UC3);\n for (int y = 0; y < h; y++)\n {\n const uchar* pf = fgr.ptr(y);\n const uchar* pa = pha.ptr(y);\n uchar* p = comp.ptr(y);\n for (int x = 0; x < w; x++)\n {\n const float alpha = pa[0] / 255.f;\n p[0] = cv::saturate_cast(pf[0] * alpha + (1 - alpha) * 155);\n p[1] = cv::saturate_cast(pf[1] * alpha + (1 - alpha) * 255);\n p[2] = cv::saturate_cast(pf[2] * alpha + (1 - alpha) * 120);\n pf += 3;\n pa += 1;\n p += 3;\n }\n }\n\n // composite seg\n cv::Mat comp_seg(h, w, CV_8UC3);\n for (int y = 0; y < h; y++)\n {\n const uchar* pb = bgr.ptr(y);\n const uchar* ps = seg.ptr(y);\n uchar* p = comp_seg.ptr(y);\n for (int x = 0; x < w; x++)\n {\n const float alpha = ps[0] / 255.f;\n p[0] = cv::saturate_cast(pb[0] * alpha + (1 - alpha) * 155);\n p[1] = cv::saturate_cast(pb[1] * alpha + (1 - alpha) * 255);\n p[2] = cv::saturate_cast(pb[2] * alpha + (1 - alpha) * 120);\n pb += 3;\n ps += 1;\n p += 3;\n }\n }\n\n cv::imshow(\"comp\", comp);\n cv::imshow(\"comp_seg\", comp_seg);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n cv::Mat fgr;\n cv::Mat pha;\n cv::Mat seg;\n detect_rvm(m, fgr, pha, seg);\n\n draw_objects(m, fgr, pha, seg);\n\n return 0;\n}\n" }, { "path": "examples/scrfd.cpp", "content": "// Copyright 2021 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct FaceObject\n{\n cv::Rect_ rect;\n float prob;\n};\n\nstatic inline float intersection_area(const FaceObject& a, const FaceObject& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const FaceObject& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const FaceObject& b = faceobjects[picked[j]];\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\n// insightface/detection/scrfd/mmdet/core/anchor/anchor_generator.py gen_single_level_base_anchors()\nstatic ncnn::Mat generate_anchors(int base_size, const ncnn::Mat& ratios, const ncnn::Mat& scales)\n{\n int num_ratio = ratios.w;\n int num_scale = scales.w;\n\n ncnn::Mat anchors;\n anchors.create(4, num_ratio * num_scale);\n\n const float cx = 0;\n const float cy = 0;\n\n for (int i = 0; i < num_ratio; i++)\n {\n float ar = ratios[i];\n\n int r_w = round(base_size / sqrt(ar));\n int r_h = round(r_w * ar); //round(base_size * sqrt(ar));\n\n for (int j = 0; j < num_scale; j++)\n {\n float scale = scales[j];\n\n float rs_w = r_w * scale;\n float rs_h = r_h * scale;\n\n float* anchor = anchors.row(i * num_scale + j);\n\n anchor[0] = cx - rs_w * 0.5f;\n anchor[1] = cy - rs_h * 0.5f;\n anchor[2] = cx + rs_w * 0.5f;\n anchor[3] = cy + rs_h * 0.5f;\n }\n }\n\n return anchors;\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int feat_stride, const ncnn::Mat& score_blob, const ncnn::Mat& bbox_blob, float prob_threshold, std::vector& faceobjects)\n{\n int w = score_blob.w;\n int h = score_blob.h;\n\n // generate face proposal from bbox deltas and shifted anchors\n const int num_anchors = anchors.h;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float* anchor = anchors.row(q);\n\n const ncnn::Mat score = score_blob.channel(q);\n const ncnn::Mat bbox = bbox_blob.channel_range(q * 4, 4);\n\n // shifted anchor\n float anchor_y = anchor[1];\n\n float anchor_w = anchor[2] - anchor[0];\n float anchor_h = anchor[3] - anchor[1];\n\n for (int i = 0; i < h; i++)\n {\n float anchor_x = anchor[0];\n\n for (int j = 0; j < w; j++)\n {\n int index = i * w + j;\n\n float prob = score[index];\n\n if (prob >= prob_threshold)\n {\n // insightface/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py _get_bboxes_single()\n float dx = bbox.channel(0)[index] * feat_stride;\n float dy = bbox.channel(1)[index] * feat_stride;\n float dw = bbox.channel(2)[index] * feat_stride;\n float dh = bbox.channel(3)[index] * feat_stride;\n\n // insightface/detection/scrfd/mmdet/core/bbox/transforms.py distance2bbox()\n float cx = anchor_x + anchor_w * 0.5f;\n float cy = anchor_y + anchor_h * 0.5f;\n\n float x0 = cx - dx;\n float y0 = cy - dy;\n float x1 = cx + dw;\n float y1 = cy + dh;\n\n FaceObject obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0 + 1;\n obj.rect.height = y1 - y0 + 1;\n obj.prob = prob;\n\n faceobjects.push_back(obj);\n }\n\n anchor_x += feat_stride;\n }\n\n anchor_y += feat_stride;\n }\n }\n}\n\nstatic int detect_scrfd(const cv::Mat& bgr, std::vector& faceobjects)\n{\n ncnn::Net scrfd;\n\n scrfd.opt.use_vulkan_compute = true;\n\n // model is converted from\n // https://github.com/deepinsight/insightface/tree/master/detection/scrfd\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (scrfd.load_param(\"scrfd_500m-opt2.param\"))\n exit(-1);\n if (scrfd.load_model(\"scrfd_500m-opt2.bin\"))\n exit(-1);\n\n int width = bgr.cols;\n int height = bgr.rows;\n\n // insightface/detection/scrfd/configs/scrfd/scrfd_500m.py\n const int target_size = 640;\n const float prob_threshold = 0.3f;\n const float nms_threshold = 0.45f;\n\n // pad to multiple of 32\n int w = width;\n int h = height;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, width, height, w, h);\n\n // pad to target_size rectangle\n int wpad = (w + 31) / 32 * 32 - w;\n int hpad = (h + 31) / 32 * 32 - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 0.f);\n\n const float mean_vals[3] = {127.5f, 127.5f, 127.5f};\n const float norm_vals[3] = {1 / 128.f, 1 / 128.f, 1 / 128.f};\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = scrfd.create_extractor();\n\n ex.input(\"input.1\", in_pad);\n\n std::vector faceproposals;\n\n // stride 32\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"412\", score_blob);\n ex.extract(\"415\", bbox_blob);\n\n const int base_size = 16;\n const int feat_stride = 8;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 1.f;\n scales[1] = 2.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects32;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects32);\n\n faceproposals.insert(faceproposals.end(), faceobjects32.begin(), faceobjects32.end());\n }\n\n // stride 16\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"474\", score_blob);\n ex.extract(\"477\", bbox_blob);\n\n const int base_size = 64;\n const int feat_stride = 16;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 1.f;\n scales[1] = 2.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects16;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects16);\n\n faceproposals.insert(faceproposals.end(), faceobjects16.begin(), faceobjects16.end());\n }\n\n // stride 8\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"536\", score_blob);\n ex.extract(\"539\", bbox_blob);\n\n const int base_size = 256;\n const int feat_stride = 32;\n ncnn::Mat ratios(1);\n ratios[0] = 1.f;\n ncnn::Mat scales(2);\n scales[0] = 1.f;\n scales[1] = 2.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects8;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);\n\n faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(faceproposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(faceproposals, picked, nms_threshold);\n\n int face_count = picked.size();\n\n faceobjects.resize(face_count);\n for (int i = 0; i < face_count; i++)\n {\n faceobjects[i] = faceproposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (faceobjects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (faceobjects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (faceobjects[i].rect.x + faceobjects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (faceobjects[i].rect.y + faceobjects[i].rect.height - (hpad / 2)) / scale;\n\n x0 = std::max(std::min(x0, (float)width - 1), 0.f);\n y0 = std::max(std::min(y0, (float)height - 1), 0.f);\n x1 = std::max(std::min(x1, (float)width - 1), 0.f);\n y1 = std::max(std::min(y1, (float)height - 1), 0.f);\n\n faceobjects[i].rect.x = x0;\n faceobjects[i].rect.y = y0;\n faceobjects[i].rect.width = x1 - x0;\n faceobjects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_faceobjects(const cv::Mat& bgr, const std::vector& faceobjects)\n{\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < faceobjects.size(); i++)\n {\n const FaceObject& obj = faceobjects[i];\n\n fprintf(stderr, \"%.5f at %.2f %.2f %.2f x %.2f\\n\", obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));\n\n char text[256];\n sprintf(text, \"%.1f%%\", obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector faceobjects;\n detect_scrfd(m, faceobjects);\n\n draw_faceobjects(m, faceobjects);\n\n return 0;\n}\n" }, { "path": "examples/scrfd_crowdhuman.cpp", "content": "// Copyright 2021 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct FaceObject\n{\n cv::Rect_ rect;\n float prob;\n};\n\nstatic inline float intersection_area(const FaceObject& a, const FaceObject& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const FaceObject& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const FaceObject& b = faceobjects[picked[j]];\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\n// insightface/detection/scrfd/mmdet/core/anchor/anchor_generator.py gen_single_level_base_anchors()\nstatic ncnn::Mat generate_anchors(int base_size, const ncnn::Mat& ratios, const ncnn::Mat& scales)\n{\n int num_ratio = ratios.w;\n int num_scale = scales.w;\n\n ncnn::Mat anchors;\n anchors.create(4, num_ratio * num_scale);\n\n const float cx = 0;\n const float cy = 0;\n\n for (int i = 0; i < num_ratio; i++)\n {\n float ar = ratios[i];\n\n int r_w = round(base_size / sqrt(ar));\n int r_h = round(r_w * ar); //round(base_size * sqrt(ar));\n\n for (int j = 0; j < num_scale; j++)\n {\n float scale = scales[j];\n\n float rs_w = r_w * scale;\n float rs_h = r_h * scale;\n\n float* anchor = anchors.row(i * num_scale + j);\n\n anchor[0] = cx - rs_w * 0.5f;\n anchor[1] = cy - rs_h * 0.5f;\n anchor[2] = cx + rs_w * 0.5f;\n anchor[3] = cy + rs_h * 0.5f;\n }\n }\n\n return anchors;\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int feat_stride, const ncnn::Mat& score_blob, const ncnn::Mat& bbox_blob, float prob_threshold, std::vector& faceobjects)\n{\n int w = score_blob.w;\n int h = score_blob.h;\n\n // generate face proposal from bbox deltas and shifted anchors\n const int num_anchors = anchors.h;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float* anchor = anchors.row(q);\n\n const ncnn::Mat score = score_blob.channel(q);\n const ncnn::Mat bbox = bbox_blob.channel_range(q * 4, 4);\n\n // shifted anchor\n float anchor_y = anchor[1];\n\n float anchor_w = anchor[2] - anchor[0];\n float anchor_h = anchor[3] - anchor[1];\n\n for (int i = 0; i < h; i++)\n {\n float anchor_x = anchor[0];\n\n for (int j = 0; j < w; j++)\n {\n int index = i * w + j;\n\n float prob = score[index];\n\n if (prob >= prob_threshold)\n {\n // insightface/detection/scrfd/mmdet/models/dense_heads/scrfd_head.py _get_bboxes_single()\n float dx = bbox.channel(0)[index] * feat_stride;\n float dy = bbox.channel(1)[index] * feat_stride;\n float dw = bbox.channel(2)[index] * feat_stride;\n float dh = bbox.channel(3)[index] * feat_stride;\n\n // insightface/detection/scrfd/mmdet/core/bbox/transforms.py distance2bbox()\n float cx = anchor_x + anchor_w * 0.5f;\n float cy = anchor_y + anchor_h * 0.5f;\n\n float x0 = cx - dx;\n float y0 = cy - dy;\n float x1 = cx + dw;\n float y1 = cy + dh;\n\n FaceObject obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0 + 1;\n obj.rect.height = y1 - y0 + 1;\n obj.prob = prob;\n\n faceobjects.push_back(obj);\n }\n\n anchor_x += feat_stride;\n }\n\n anchor_y += feat_stride;\n }\n }\n}\n\nstatic int detect_scrfd(const cv::Mat& bgr, std::vector& faceobjects)\n{\n ncnn::Net scrfd;\n\n scrfd.opt.use_vulkan_compute = true;\n\n // Insight face does not provided a trained scrfd_crowdhuman model\n // but I have one for detecing cat face, you can have a try here:\n // https://drive.google.com/file/d/1JogkKa0f_09HkENbCnXy9hRYxm35wKTn\n\n if (scrfd.load_param(\"scrfd_crowdhuman.param\"))\n exit(-1);\n if (scrfd.load_model(\"scrfd_crowdhuman.bin\"))\n exit(-1);\n\n int width = bgr.cols;\n int height = bgr.rows;\n\n const int target_size = 640;\n const float prob_threshold = 0.3f;\n const float nms_threshold = 0.45f;\n\n // pad to multiple of 32\n int w = width;\n int h = height;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, width, height, w, h);\n\n // pad to target_size rectangle\n int wpad = (w + 31) / 32 * 32 - w;\n int hpad = (h + 31) / 32 * 32 - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 0.f);\n\n const float mean_vals[3] = {127.5f, 127.5f, 127.5f};\n const float norm_vals[3] = {1 / 128.f, 1 / 128.f, 1 / 128.f};\n in_pad.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = scrfd.create_extractor();\n\n ex.input(\"input.1\", in_pad);\n\n std::vector faceproposals;\n\n // stride 8\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"490\", score_blob);\n ex.extract(\"493\", bbox_blob);\n\n const int base_size = 8;\n const int feat_stride = 8;\n ncnn::Mat ratios(1);\n ratios[0] = 2.f;\n ncnn::Mat scales(1);\n scales[0] = 3.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects32;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects32);\n\n faceproposals.insert(faceproposals.end(), faceobjects32.begin(), faceobjects32.end());\n }\n\n // stride 16\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"510\", score_blob);\n ex.extract(\"513\", bbox_blob);\n\n const int base_size = 16;\n const int feat_stride = 16;\n ncnn::Mat ratios(1);\n ratios[0] = 2.f;\n ncnn::Mat scales(1);\n scales[0] = 3.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects16;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects16);\n\n faceproposals.insert(faceproposals.end(), faceobjects16.begin(), faceobjects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat score_blob, bbox_blob;\n ex.extract(\"530\", score_blob);\n ex.extract(\"533\", bbox_blob);\n\n const int base_size = 32;\n const int feat_stride = 32;\n ncnn::Mat ratios(1);\n ratios[0] = 2.f;\n ncnn::Mat scales(1);\n scales[0] = 3.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects8;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);\n\n faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());\n }\n\n // stride 64\n {\n ncnn::Mat score_blob, bbox_blob, kps_blob;\n ex.extract(\"550\", score_blob);\n ex.extract(\"553\", bbox_blob);\n\n const int base_size = 64;\n const int feat_stride = 64;\n ncnn::Mat ratios(1);\n ratios[0] = 2.f;\n ncnn::Mat scales(1);\n scales[0] = 3.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects8;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);\n\n faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());\n }\n\n // stride 128\n {\n ncnn::Mat score_blob, bbox_blob, kps_blob;\n ex.extract(\"570\", score_blob);\n ex.extract(\"573\", bbox_blob);\n\n const int base_size = 128;\n const int feat_stride = 128;\n ncnn::Mat ratios(1);\n ratios[0] = 2.f;\n ncnn::Mat scales(1);\n scales[0] = 3.f;\n ncnn::Mat anchors = generate_anchors(base_size, ratios, scales);\n\n std::vector faceobjects8;\n generate_proposals(anchors, feat_stride, score_blob, bbox_blob, prob_threshold, faceobjects8);\n\n faceproposals.insert(faceproposals.end(), faceobjects8.begin(), faceobjects8.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(faceproposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(faceproposals, picked, nms_threshold);\n\n int face_count = picked.size();\n\n faceobjects.resize(face_count);\n for (int i = 0; i < face_count; i++)\n {\n faceobjects[i] = faceproposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (faceobjects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (faceobjects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (faceobjects[i].rect.x + faceobjects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (faceobjects[i].rect.y + faceobjects[i].rect.height - (hpad / 2)) / scale;\n\n x0 = std::max(std::min(x0, (float)width - 1), 0.f);\n y0 = std::max(std::min(y0, (float)height - 1), 0.f);\n x1 = std::max(std::min(x1, (float)width - 1), 0.f);\n y1 = std::max(std::min(y1, (float)height - 1), 0.f);\n\n faceobjects[i].rect.x = x0;\n faceobjects[i].rect.y = y0;\n faceobjects[i].rect.width = x1 - x0;\n faceobjects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_faceobjects(const cv::Mat& bgr, const std::vector& faceobjects)\n{\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < faceobjects.size(); i++)\n {\n const FaceObject& obj = faceobjects[i];\n\n fprintf(stderr, \"%.5f at %.2f %.2f %.2f x %.2f\\n\", obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(0, 255, 0));\n\n char text[256];\n sprintf(text, \"%.1f%%\", obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector faceobjects;\n detect_scrfd(m, faceobjects);\n\n draw_faceobjects(m, faceobjects);\n\n return 0;\n}\n" }, { "path": "examples/shufflenetv2.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#endif\n#include \n#include \n\nstatic int detect_shufflenetv2(const cv::Mat& bgr, std::vector& cls_scores)\n{\n ncnn::Net shufflenetv2;\n\n shufflenetv2.opt.use_vulkan_compute = true;\n\n // https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe\n // models can be downloaded from https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe/releases\n if (shufflenetv2.load_param(\"shufflenet_v2_x0.5.param\"))\n exit(-1);\n if (shufflenetv2.load_model(\"shufflenet_v2_x0.5.bin\"))\n exit(-1);\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, 224, 224);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = shufflenetv2.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"fc\", out);\n\n // manually call softmax on the fc output\n // convert result into probability\n // skip if your model already has softmax operation\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n softmax->load_param(pd);\n\n softmax->forward_inplace(out, shufflenetv2.opt);\n\n delete softmax;\n }\n\n out = out.reshape(out.w * out.h * out.c);\n\n cls_scores.resize(out.w);\n for (int j = 0; j < out.w; j++)\n {\n cls_scores[j] = out[j];\n }\n\n return 0;\n}\n\nstatic int print_topk(const std::vector& cls_scores, int topk)\n{\n // partial sort topk with index\n int size = cls_scores.size();\n std::vector > vec;\n vec.resize(size);\n for (int i = 0; i < size; i++)\n {\n vec[i] = std::make_pair(cls_scores[i], i);\n }\n\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(),\n std::greater >());\n\n // print topk and score\n for (int i = 0; i < topk; i++)\n {\n float score = vec[i].first;\n int index = vec[i].second;\n fprintf(stderr, \"%d = %f\\n\", index, score);\n }\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector cls_scores;\n detect_shufflenetv2(m, cls_scores);\n\n print_topk(cls_scores, 3);\n\n return 0;\n}\n" }, { "path": "examples/simplepose.cpp", "content": "// Copyright 2019 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct KeyPoint\n{\n cv::Point2f p;\n float prob;\n};\n\nstatic int detect_posenet(const cv::Mat& bgr, std::vector& keypoints)\n{\n ncnn::Net posenet;\n\n posenet.opt.use_vulkan_compute = true;\n\n // the simple baseline human pose estimation from gluon-cv\n // https://gluon-cv.mxnet.io/build/examples_pose/demo_simple_pose.html\n // mxnet model exported via\n // pose_net.hybridize()\n // pose_net.export('pose')\n // then mxnet2ncnn\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (posenet.load_param(\"pose.param\"))\n exit(-1);\n if (posenet.load_model(\"pose.bin\"))\n exit(-1);\n\n int w = bgr.cols;\n int h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, w, h, 192, 256);\n\n // transforms.ToTensor(),\n // transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),\n // R' = (R / 255 - 0.485) / 0.229 = (R - 0.485 * 255) / 0.229 / 255\n // G' = (G / 255 - 0.456) / 0.224 = (G - 0.456 * 255) / 0.224 / 255\n // B' = (B / 255 - 0.406) / 0.225 = (B - 0.406 * 255) / 0.225 / 255\n const float mean_vals[3] = {0.485f * 255.f, 0.456f * 255.f, 0.406f * 255.f};\n const float norm_vals[3] = {1 / 0.229f / 255.f, 1 / 0.224f / 255.f, 1 / 0.225f / 255.f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = posenet.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"conv3_fwd\", out);\n\n // resolve point from heatmap\n keypoints.clear();\n for (int p = 0; p < out.c; p++)\n {\n const ncnn::Mat m = out.channel(p);\n\n float max_prob = 0.f;\n int max_x = 0;\n int max_y = 0;\n for (int y = 0; y < out.h; y++)\n {\n const float* ptr = m.row(y);\n for (int x = 0; x < out.w; x++)\n {\n float prob = ptr[x];\n if (prob > max_prob)\n {\n max_prob = prob;\n max_x = x;\n max_y = y;\n }\n }\n }\n\n KeyPoint keypoint;\n keypoint.p = cv::Point2f(max_x * w / (float)out.w, max_y * h / (float)out.h);\n keypoint.prob = max_prob;\n\n keypoints.push_back(keypoint);\n }\n\n return 0;\n}\n\nstatic void draw_pose(const cv::Mat& bgr, const std::vector& keypoints)\n{\n cv::Mat image = bgr.clone();\n\n // draw bone\n static const int joint_pairs[16][2] = {\n {0, 1}, {1, 3}, {0, 2}, {2, 4}, {5, 6}, {5, 7}, {7, 9}, {6, 8}, {8, 10}, {5, 11}, {6, 12}, {11, 12}, {11, 13}, {12, 14}, {13, 15}, {14, 16}\n };\n\n for (int i = 0; i < 16; i++)\n {\n const KeyPoint& p1 = keypoints[joint_pairs[i][0]];\n const KeyPoint& p2 = keypoints[joint_pairs[i][1]];\n\n if (p1.prob < 0.2f || p2.prob < 0.2f)\n continue;\n\n cv::line(image, p1.p, p2.p, cv::Scalar(255, 0, 0), 2);\n }\n\n // draw joint\n for (size_t i = 0; i < keypoints.size(); i++)\n {\n const KeyPoint& keypoint = keypoints[i];\n\n fprintf(stderr, \"%.2f %.2f = %.5f\\n\", keypoint.p.x, keypoint.p.y, keypoint.prob);\n\n if (keypoint.prob < 0.2f)\n continue;\n\n cv::circle(image, keypoint.p, 3, cv::Scalar(0, 255, 0), -1);\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector keypoints;\n detect_posenet(m, keypoints);\n\n draw_pose(m, keypoints);\n\n return 0;\n}\n" }, { "path": "examples/squeezencnn/README.md", "content": "The squeezenet android example project has been moved to https://github.com/nihui/ncnn-android-squeezenet\n" }, { "path": "examples/squeezenet.cpp", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#endif\n#include \n#include \n\nstatic int detect_squeezenet(const cv::Mat& bgr, std::vector& cls_scores)\n{\n ncnn::Net squeezenet;\n\n squeezenet.opt.use_vulkan_compute = true;\n\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (squeezenet.load_param(\"squeezenet_v1.1.param\"))\n exit(-1);\n if (squeezenet.load_model(\"squeezenet_v1.1.bin\"))\n exit(-1);\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, 227, 227);\n\n const float mean_vals[3] = {104.f, 117.f, 123.f};\n in.substract_mean_normalize(mean_vals, 0);\n\n ncnn::Extractor ex = squeezenet.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"prob\", out);\n\n cls_scores.resize(out.w);\n for (int j = 0; j < out.w; j++)\n {\n cls_scores[j] = out[j];\n }\n\n return 0;\n}\n\nstatic int print_topk(const std::vector& cls_scores, int topk)\n{\n // partial sort topk with index\n int size = cls_scores.size();\n std::vector > vec;\n vec.resize(size);\n for (int i = 0; i < size; i++)\n {\n vec[i] = std::make_pair(cls_scores[i], i);\n }\n\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(),\n std::greater >());\n\n // print topk and score\n for (int i = 0; i < topk; i++)\n {\n float score = vec[i].first;\n int index = vec[i].second;\n fprintf(stderr, \"%d = %f\\n\", index, score);\n }\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector cls_scores;\n detect_squeezenet(m, cls_scores);\n\n print_topk(cls_scores, 3);\n\n return 0;\n}\n" }, { "path": "examples/squeezenet_c_api.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"c_api.h\"\n\n#include \n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#endif\n#include \n#include \n\nstatic int detect_squeezenet(const cv::Mat& bgr, std::vector& cls_scores)\n{\n ncnn_net_t squeezenet = ncnn_net_create();\n\n ncnn_option_t opt = ncnn_option_create();\n ncnn_option_set_use_vulkan_compute(opt, 1);\n\n ncnn_net_set_option(squeezenet, opt);\n\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (ncnn_net_load_param(squeezenet, \"squeezenet_v1.1.param\"))\n exit(-1);\n if (ncnn_net_load_model(squeezenet, \"squeezenet_v1.1.bin\"))\n exit(-1);\n\n ncnn_mat_t in = ncnn_mat_from_pixels_resize(bgr.data, NCNN_MAT_PIXEL_BGR, bgr.cols, bgr.rows, bgr.cols * 3, 227, 227, NULL);\n\n const float mean_vals[3] = {104.f, 117.f, 123.f};\n ncnn_mat_substract_mean_normalize(in, mean_vals, 0);\n\n ncnn_extractor_t ex = ncnn_extractor_create(squeezenet);\n\n ncnn_extractor_input(ex, \"data\", in);\n\n ncnn_mat_t out;\n ncnn_extractor_extract(ex, \"prob\", &out);\n\n const int out_w = ncnn_mat_get_w(out);\n const float* out_data = (const float*)ncnn_mat_get_data(out);\n\n cls_scores.resize(out_w);\n for (int j = 0; j < out_w; j++)\n {\n cls_scores[j] = out_data[j];\n }\n\n ncnn_mat_destroy(in);\n ncnn_mat_destroy(out);\n\n ncnn_extractor_destroy(ex);\n\n ncnn_option_destroy(opt);\n\n ncnn_net_destroy(squeezenet);\n\n return 0;\n}\n\nstatic int print_topk(const std::vector& cls_scores, int topk)\n{\n // partial sort topk with index\n int size = cls_scores.size();\n std::vector > vec;\n vec.resize(size);\n for (int i = 0; i < size; i++)\n {\n vec[i] = std::make_pair(cls_scores[i], i);\n }\n\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(),\n std::greater >());\n\n // print topk and score\n for (int i = 0; i < topk; i++)\n {\n float score = vec[i].first;\n int index = vec[i].second;\n fprintf(stderr, \"%d = %f\\n\", index, score);\n }\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector cls_scores;\n detect_squeezenet(m, cls_scores);\n\n print_topk(cls_scores, 3);\n\n return 0;\n}\n" }, { "path": "examples/squeezenet_v1.1.param", "content": "7767517\n75 83\nInput data 0 1 data 0=227 1=227 2=3\nConvolution conv1 1 1 data conv1 0=64 1=3 2=1 3=2 4=0 5=1 6=1728\nReLU relu_conv1 1 1 conv1 conv1_relu_conv1 0=0.000000\nPooling pool1 1 1 conv1_relu_conv1 pool1 0=0 1=3 2=2 3=0 4=0\nConvolution fire2/squeeze1x1 1 1 pool1 fire2/squeeze1x1 0=16 1=1 2=1 3=1 4=0 5=1 6=1024\nReLU fire2/relu_squeeze1x1 1 1 fire2/squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1 0=0.000000\nSplit splitncnn_0 1 2 fire2/squeeze1x1_fire2/relu_squeeze1x1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1\nConvolution fire2/expand1x1 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_1 fire2/expand1x1 0=64 1=1 2=1 3=1 4=0 5=1 6=1024\nReLU fire2/relu_expand1x1 1 1 fire2/expand1x1 fire2/expand1x1_fire2/relu_expand1x1 0=0.000000\nConvolution fire2/expand3x3 1 1 fire2/squeeze1x1_fire2/relu_squeeze1x1_splitncnn_0 fire2/expand3x3 0=64 1=3 2=1 3=1 4=1 5=1 6=9216\nReLU fire2/relu_expand3x3 1 1 fire2/expand3x3 fire2/expand3x3_fire2/relu_expand3x3 0=0.000000\nConcat fire2/concat 2 1 fire2/expand1x1_fire2/relu_expand1x1 fire2/expand3x3_fire2/relu_expand3x3 fire2/concat 0=0\nConvolution fire3/squeeze1x1 1 1 fire2/concat fire3/squeeze1x1 0=16 1=1 2=1 3=1 4=0 5=1 6=2048\nReLU fire3/relu_squeeze1x1 1 1 fire3/squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1 0=0.000000\nSplit splitncnn_1 1 2 fire3/squeeze1x1_fire3/relu_squeeze1x1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1\nConvolution fire3/expand1x1 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_1 fire3/expand1x1 0=64 1=1 2=1 3=1 4=0 5=1 6=1024\nReLU fire3/relu_expand1x1 1 1 fire3/expand1x1 fire3/expand1x1_fire3/relu_expand1x1 0=0.000000\nConvolution fire3/expand3x3 1 1 fire3/squeeze1x1_fire3/relu_squeeze1x1_splitncnn_0 fire3/expand3x3 0=64 1=3 2=1 3=1 4=1 5=1 6=9216\nReLU fire3/relu_expand3x3 1 1 fire3/expand3x3 fire3/expand3x3_fire3/relu_expand3x3 0=0.000000\nConcat fire3/concat 2 1 fire3/expand1x1_fire3/relu_expand1x1 fire3/expand3x3_fire3/relu_expand3x3 fire3/concat 0=0\nPooling pool3 1 1 fire3/concat pool3 0=0 1=3 2=2 3=0 4=0\nConvolution fire4/squeeze1x1 1 1 pool3 fire4/squeeze1x1 0=32 1=1 2=1 3=1 4=0 5=1 6=4096\nReLU fire4/relu_squeeze1x1 1 1 fire4/squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1 0=0.000000\nSplit splitncnn_2 1 2 fire4/squeeze1x1_fire4/relu_squeeze1x1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1\nConvolution fire4/expand1x1 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_1 fire4/expand1x1 0=128 1=1 2=1 3=1 4=0 5=1 6=4096\nReLU fire4/relu_expand1x1 1 1 fire4/expand1x1 fire4/expand1x1_fire4/relu_expand1x1 0=0.000000\nConvolution fire4/expand3x3 1 1 fire4/squeeze1x1_fire4/relu_squeeze1x1_splitncnn_0 fire4/expand3x3 0=128 1=3 2=1 3=1 4=1 5=1 6=36864\nReLU fire4/relu_expand3x3 1 1 fire4/expand3x3 fire4/expand3x3_fire4/relu_expand3x3 0=0.000000\nConcat fire4/concat 2 1 fire4/expand1x1_fire4/relu_expand1x1 fire4/expand3x3_fire4/relu_expand3x3 fire4/concat 0=0\nConvolution fire5/squeeze1x1 1 1 fire4/concat fire5/squeeze1x1 0=32 1=1 2=1 3=1 4=0 5=1 6=8192\nReLU fire5/relu_squeeze1x1 1 1 fire5/squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1 0=0.000000\nSplit splitncnn_3 1 2 fire5/squeeze1x1_fire5/relu_squeeze1x1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1\nConvolution fire5/expand1x1 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_1 fire5/expand1x1 0=128 1=1 2=1 3=1 4=0 5=1 6=4096\nReLU fire5/relu_expand1x1 1 1 fire5/expand1x1 fire5/expand1x1_fire5/relu_expand1x1 0=0.000000\nConvolution fire5/expand3x3 1 1 fire5/squeeze1x1_fire5/relu_squeeze1x1_splitncnn_0 fire5/expand3x3 0=128 1=3 2=1 3=1 4=1 5=1 6=36864\nReLU fire5/relu_expand3x3 1 1 fire5/expand3x3 fire5/expand3x3_fire5/relu_expand3x3 0=0.000000\nConcat fire5/concat 2 1 fire5/expand1x1_fire5/relu_expand1x1 fire5/expand3x3_fire5/relu_expand3x3 fire5/concat 0=0\nPooling pool5 1 1 fire5/concat pool5 0=0 1=3 2=2 3=0 4=0\nConvolution fire6/squeeze1x1 1 1 pool5 fire6/squeeze1x1 0=48 1=1 2=1 3=1 4=0 5=1 6=12288\nReLU fire6/relu_squeeze1x1 1 1 fire6/squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1 0=0.000000\nSplit splitncnn_4 1 2 fire6/squeeze1x1_fire6/relu_squeeze1x1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1\nConvolution fire6/expand1x1 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_1 fire6/expand1x1 0=192 1=1 2=1 3=1 4=0 5=1 6=9216\nReLU fire6/relu_expand1x1 1 1 fire6/expand1x1 fire6/expand1x1_fire6/relu_expand1x1 0=0.000000\nConvolution fire6/expand3x3 1 1 fire6/squeeze1x1_fire6/relu_squeeze1x1_splitncnn_0 fire6/expand3x3 0=192 1=3 2=1 3=1 4=1 5=1 6=82944\nReLU fire6/relu_expand3x3 1 1 fire6/expand3x3 fire6/expand3x3_fire6/relu_expand3x3 0=0.000000\nConcat fire6/concat 2 1 fire6/expand1x1_fire6/relu_expand1x1 fire6/expand3x3_fire6/relu_expand3x3 fire6/concat 0=0\nConvolution fire7/squeeze1x1 1 1 fire6/concat fire7/squeeze1x1 0=48 1=1 2=1 3=1 4=0 5=1 6=18432\nReLU fire7/relu_squeeze1x1 1 1 fire7/squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1 0=0.000000\nSplit splitncnn_5 1 2 fire7/squeeze1x1_fire7/relu_squeeze1x1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1\nConvolution fire7/expand1x1 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_1 fire7/expand1x1 0=192 1=1 2=1 3=1 4=0 5=1 6=9216\nReLU fire7/relu_expand1x1 1 1 fire7/expand1x1 fire7/expand1x1_fire7/relu_expand1x1 0=0.000000\nConvolution fire7/expand3x3 1 1 fire7/squeeze1x1_fire7/relu_squeeze1x1_splitncnn_0 fire7/expand3x3 0=192 1=3 2=1 3=1 4=1 5=1 6=82944\nReLU fire7/relu_expand3x3 1 1 fire7/expand3x3 fire7/expand3x3_fire7/relu_expand3x3 0=0.000000\nConcat fire7/concat 2 1 fire7/expand1x1_fire7/relu_expand1x1 fire7/expand3x3_fire7/relu_expand3x3 fire7/concat 0=0\nConvolution fire8/squeeze1x1 1 1 fire7/concat fire8/squeeze1x1 0=64 1=1 2=1 3=1 4=0 5=1 6=24576\nReLU fire8/relu_squeeze1x1 1 1 fire8/squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1 0=0.000000\nSplit splitncnn_6 1 2 fire8/squeeze1x1_fire8/relu_squeeze1x1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1\nConvolution fire8/expand1x1 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_1 fire8/expand1x1 0=256 1=1 2=1 3=1 4=0 5=1 6=16384\nReLU fire8/relu_expand1x1 1 1 fire8/expand1x1 fire8/expand1x1_fire8/relu_expand1x1 0=0.000000\nConvolution fire8/expand3x3 1 1 fire8/squeeze1x1_fire8/relu_squeeze1x1_splitncnn_0 fire8/expand3x3 0=256 1=3 2=1 3=1 4=1 5=1 6=147456\nReLU fire8/relu_expand3x3 1 1 fire8/expand3x3 fire8/expand3x3_fire8/relu_expand3x3 0=0.000000\nConcat fire8/concat 2 1 fire8/expand1x1_fire8/relu_expand1x1 fire8/expand3x3_fire8/relu_expand3x3 fire8/concat 0=0\nConvolution fire9/squeeze1x1 1 1 fire8/concat fire9/squeeze1x1 0=64 1=1 2=1 3=1 4=0 5=1 6=32768\nReLU fire9/relu_squeeze1x1 1 1 fire9/squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1 0=0.000000\nSplit splitncnn_7 1 2 fire9/squeeze1x1_fire9/relu_squeeze1x1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1\nConvolution fire9/expand1x1 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_1 fire9/expand1x1 0=256 1=1 2=1 3=1 4=0 5=1 6=16384\nReLU fire9/relu_expand1x1 1 1 fire9/expand1x1 fire9/expand1x1_fire9/relu_expand1x1 0=0.000000\nConvolution fire9/expand3x3 1 1 fire9/squeeze1x1_fire9/relu_squeeze1x1_splitncnn_0 fire9/expand3x3 0=256 1=3 2=1 3=1 4=1 5=1 6=147456\nReLU fire9/relu_expand3x3 1 1 fire9/expand3x3 fire9/expand3x3_fire9/relu_expand3x3 0=0.000000\nConcat fire9/concat 2 1 fire9/expand1x1_fire9/relu_expand1x1 fire9/expand3x3_fire9/relu_expand3x3 fire9/concat 0=0\nDropout drop9 1 1 fire9/concat fire9/concat_drop9\nConvolution conv10 1 1 fire9/concat_drop9 conv10 0=1000 1=1 2=1 3=1 4=1 5=1 6=512000\nReLU relu_conv10 1 1 conv10 conv10_relu_conv10 0=0.000000\nPooling pool10 1 1 conv10_relu_conv10 pool10 0=1 1=0 2=1 3=0 4=1\nSoftmax prob 1 1 pool10 prob 0=0\n" }, { "path": "examples/squeezenet_v1.1.prototxt", "content": "name: \"squeezenet_v1.1_deploy\"\n\nlayer {\n name: \"data\"\n type: \"Input\"\n top: \"data\"\n input_param { shape: { dim: 1 dim: 3 dim: 227 dim: 227 } }\n}\nlayer {\n name: \"conv1\"\n type: \"Convolution\"\n bottom: \"data\"\n top: \"conv1\"\n convolution_param {\n num_output: 64\n kernel_size: 3\n stride: 2\n }\n}\nlayer {\n name: \"relu_conv1\"\n type: \"ReLU\"\n bottom: \"conv1\"\n top: \"conv1\"\n}\nlayer {\n name: \"pool1\"\n type: \"Pooling\"\n bottom: \"conv1\"\n top: \"pool1\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\nlayer {\n name: \"fire2/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"pool1\"\n top: \"fire2/squeeze1x1\"\n convolution_param {\n num_output: 16\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire2/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire2/squeeze1x1\"\n top: \"fire2/squeeze1x1\"\n}\nlayer {\n name: \"fire2/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire2/squeeze1x1\"\n top: \"fire2/expand1x1\"\n convolution_param {\n num_output: 64\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire2/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire2/expand1x1\"\n top: \"fire2/expand1x1\"\n}\nlayer {\n name: \"fire2/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire2/squeeze1x1\"\n top: \"fire2/expand3x3\"\n convolution_param {\n num_output: 64\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire2/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire2/expand3x3\"\n top: \"fire2/expand3x3\"\n}\nlayer {\n name: \"fire2/concat\"\n type: \"Concat\"\n bottom: \"fire2/expand1x1\"\n bottom: \"fire2/expand3x3\"\n top: \"fire2/concat\"\n}\nlayer {\n name: \"fire3/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"fire2/concat\"\n top: \"fire3/squeeze1x1\"\n convolution_param {\n num_output: 16\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire3/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire3/squeeze1x1\"\n top: \"fire3/squeeze1x1\"\n}\nlayer {\n name: \"fire3/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire3/squeeze1x1\"\n top: \"fire3/expand1x1\"\n convolution_param {\n num_output: 64\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire3/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire3/expand1x1\"\n top: \"fire3/expand1x1\"\n}\nlayer {\n name: \"fire3/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire3/squeeze1x1\"\n top: \"fire3/expand3x3\"\n convolution_param {\n num_output: 64\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire3/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire3/expand3x3\"\n top: \"fire3/expand3x3\"\n}\nlayer {\n name: \"fire3/concat\"\n type: \"Concat\"\n bottom: \"fire3/expand1x1\"\n bottom: \"fire3/expand3x3\"\n top: \"fire3/concat\"\n}\nlayer {\n name: \"pool3\"\n type: \"Pooling\"\n bottom: \"fire3/concat\"\n top: \"pool3\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\nlayer {\n name: \"fire4/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"pool3\"\n top: \"fire4/squeeze1x1\"\n convolution_param {\n num_output: 32\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire4/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire4/squeeze1x1\"\n top: \"fire4/squeeze1x1\"\n}\nlayer {\n name: \"fire4/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire4/squeeze1x1\"\n top: \"fire4/expand1x1\"\n convolution_param {\n num_output: 128\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire4/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire4/expand1x1\"\n top: \"fire4/expand1x1\"\n}\nlayer {\n name: \"fire4/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire4/squeeze1x1\"\n top: \"fire4/expand3x3\"\n convolution_param {\n num_output: 128\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire4/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire4/expand3x3\"\n top: \"fire4/expand3x3\"\n}\nlayer {\n name: \"fire4/concat\"\n type: \"Concat\"\n bottom: \"fire4/expand1x1\"\n bottom: \"fire4/expand3x3\"\n top: \"fire4/concat\"\n}\nlayer {\n name: \"fire5/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"fire4/concat\"\n top: \"fire5/squeeze1x1\"\n convolution_param {\n num_output: 32\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire5/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire5/squeeze1x1\"\n top: \"fire5/squeeze1x1\"\n}\nlayer {\n name: \"fire5/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire5/squeeze1x1\"\n top: \"fire5/expand1x1\"\n convolution_param {\n num_output: 128\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire5/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire5/expand1x1\"\n top: \"fire5/expand1x1\"\n}\nlayer {\n name: \"fire5/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire5/squeeze1x1\"\n top: \"fire5/expand3x3\"\n convolution_param {\n num_output: 128\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire5/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire5/expand3x3\"\n top: \"fire5/expand3x3\"\n}\nlayer {\n name: \"fire5/concat\"\n type: \"Concat\"\n bottom: \"fire5/expand1x1\"\n bottom: \"fire5/expand3x3\"\n top: \"fire5/concat\"\n}\nlayer {\n name: \"pool5\"\n type: \"Pooling\"\n bottom: \"fire5/concat\"\n top: \"pool5\"\n pooling_param {\n pool: MAX\n kernel_size: 3\n stride: 2\n }\n}\nlayer {\n name: \"fire6/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"pool5\"\n top: \"fire6/squeeze1x1\"\n convolution_param {\n num_output: 48\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire6/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire6/squeeze1x1\"\n top: \"fire6/squeeze1x1\"\n}\nlayer {\n name: \"fire6/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire6/squeeze1x1\"\n top: \"fire6/expand1x1\"\n convolution_param {\n num_output: 192\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire6/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire6/expand1x1\"\n top: \"fire6/expand1x1\"\n}\nlayer {\n name: \"fire6/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire6/squeeze1x1\"\n top: \"fire6/expand3x3\"\n convolution_param {\n num_output: 192\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire6/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire6/expand3x3\"\n top: \"fire6/expand3x3\"\n}\nlayer {\n name: \"fire6/concat\"\n type: \"Concat\"\n bottom: \"fire6/expand1x1\"\n bottom: \"fire6/expand3x3\"\n top: \"fire6/concat\"\n}\nlayer {\n name: \"fire7/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"fire6/concat\"\n top: \"fire7/squeeze1x1\"\n convolution_param {\n num_output: 48\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire7/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire7/squeeze1x1\"\n top: \"fire7/squeeze1x1\"\n}\nlayer {\n name: \"fire7/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire7/squeeze1x1\"\n top: \"fire7/expand1x1\"\n convolution_param {\n num_output: 192\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire7/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire7/expand1x1\"\n top: \"fire7/expand1x1\"\n}\nlayer {\n name: \"fire7/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire7/squeeze1x1\"\n top: \"fire7/expand3x3\"\n convolution_param {\n num_output: 192\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire7/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire7/expand3x3\"\n top: \"fire7/expand3x3\"\n}\nlayer {\n name: \"fire7/concat\"\n type: \"Concat\"\n bottom: \"fire7/expand1x1\"\n bottom: \"fire7/expand3x3\"\n top: \"fire7/concat\"\n}\nlayer {\n name: \"fire8/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"fire7/concat\"\n top: \"fire8/squeeze1x1\"\n convolution_param {\n num_output: 64\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire8/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire8/squeeze1x1\"\n top: \"fire8/squeeze1x1\"\n}\nlayer {\n name: \"fire8/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire8/squeeze1x1\"\n top: \"fire8/expand1x1\"\n convolution_param {\n num_output: 256\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire8/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire8/expand1x1\"\n top: \"fire8/expand1x1\"\n}\nlayer {\n name: \"fire8/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire8/squeeze1x1\"\n top: \"fire8/expand3x3\"\n convolution_param {\n num_output: 256\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire8/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire8/expand3x3\"\n top: \"fire8/expand3x3\"\n}\nlayer {\n name: \"fire8/concat\"\n type: \"Concat\"\n bottom: \"fire8/expand1x1\"\n bottom: \"fire8/expand3x3\"\n top: \"fire8/concat\"\n}\nlayer {\n name: \"fire9/squeeze1x1\"\n type: \"Convolution\"\n bottom: \"fire8/concat\"\n top: \"fire9/squeeze1x1\"\n convolution_param {\n num_output: 64\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire9/relu_squeeze1x1\"\n type: \"ReLU\"\n bottom: \"fire9/squeeze1x1\"\n top: \"fire9/squeeze1x1\"\n}\nlayer {\n name: \"fire9/expand1x1\"\n type: \"Convolution\"\n bottom: \"fire9/squeeze1x1\"\n top: \"fire9/expand1x1\"\n convolution_param {\n num_output: 256\n kernel_size: 1\n }\n}\nlayer {\n name: \"fire9/relu_expand1x1\"\n type: \"ReLU\"\n bottom: \"fire9/expand1x1\"\n top: \"fire9/expand1x1\"\n}\nlayer {\n name: \"fire9/expand3x3\"\n type: \"Convolution\"\n bottom: \"fire9/squeeze1x1\"\n top: \"fire9/expand3x3\"\n convolution_param {\n num_output: 256\n pad: 1\n kernel_size: 3\n }\n}\nlayer {\n name: \"fire9/relu_expand3x3\"\n type: \"ReLU\"\n bottom: \"fire9/expand3x3\"\n top: \"fire9/expand3x3\"\n}\nlayer {\n name: \"fire9/concat\"\n type: \"Concat\"\n bottom: \"fire9/expand1x1\"\n bottom: \"fire9/expand3x3\"\n top: \"fire9/concat\"\n}\nlayer {\n name: \"drop9\"\n type: \"Dropout\"\n bottom: \"fire9/concat\"\n top: \"fire9/concat\"\n dropout_param {\n dropout_ratio: 0.5\n }\n}\nlayer {\n name: \"conv10\"\n type: \"Convolution\"\n bottom: \"fire9/concat\"\n top: \"conv10\"\n convolution_param {\n num_output: 1000\n pad: 1\n kernel_size: 1\n }\n}\nlayer {\n name: \"relu_conv10\"\n type: \"ReLU\"\n bottom: \"conv10\"\n top: \"conv10\"\n}\nlayer {\n name: \"pool10\"\n type: \"Pooling\"\n bottom: \"conv10\"\n top: \"pool10\"\n pooling_param {\n pool: AVE\n global_pooling: true\n }\n}\nlayer {\n name: \"prob\"\n type: \"Softmax\"\n bottom: \"pool10\"\n top: \"prob\"\n}\n" }, { "path": "examples/squeezenetssd.cpp", "content": "// Copyright 2017 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_squeezenet(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net squeezenet;\n\n squeezenet.opt.use_vulkan_compute = true;\n\n // original pretrained model from https://github.com/chuanqi305/SqueezeNet-SSD\n // squeezenet_ssd_voc_deploy.prototxt\n // https://drive.google.com/open?id=0B3gersZ2cHIxdGpyZlZnbEQ5Snc\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (squeezenet.load_param(\"squeezenet_ssd_voc.param\"))\n exit(-1);\n if (squeezenet.load_model(\"squeezenet_ssd_voc.bin\"))\n exit(-1);\n\n const int target_size = 300;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {104.f, 117.f, 123.f};\n in.substract_mean_normalize(mean_vals, 0);\n\n ncnn::Extractor ex = squeezenet.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_squeezenet(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/synset_words.txt", "content": "n01440764 tench, Tinca tinca\nn01443537 goldfish, Carassius auratus\nn01484850 great white shark, white shark, man-eater, man-eating shark, Carcharodon carcharias\nn01491361 tiger shark, Galeocerdo cuvieri\nn01494475 hammerhead, hammerhead shark\nn01496331 electric ray, crampfish, numbfish, torpedo\nn01498041 stingray\nn01514668 cock\nn01514859 hen\nn01518878 ostrich, Struthio camelus\nn01530575 brambling, Fringilla montifringilla\nn01531178 goldfinch, Carduelis carduelis\nn01532829 house finch, linnet, Carpodacus mexicans\nn01534433 junco, snowbird\nn01537544 indigo bunting, indigo finch, indigo bird, Passerina cyanea\nn01558993 robin, American robin, Turdus migratorius\nn01560419 bulbul\nn01580077 jay\nn01582220 magpie\nn01592084 chickadee\nn01601694 water ouzel, dipper\nn01608432 kite\nn01614925 bald eagle, American eagle, Haliaeetus leucocephalus\nn01616318 vulture\nn01622779 great grey owl, great gray owl, Strix nebulosa\nn01629819 European fire salamander, Salamandra salamandra\nn01630670 common newt, Triturus vulgaris\nn01631663 eft\nn01632458 spotted salamander, Ambystoma maculatum\nn01632777 axolotl, mud puppy, Ambystoma mexicanum\nn01641577 bullfrog, Rana catesbeiana\nn01644373 tree frog, tree-frog\nn01644900 tailed frog, bell toad, ribbed toad, tailed toad, Ascaphus trui\nn01664065 loggerhead, loggerhead turtle, Caretta caretta\nn01665541 leatherback turtle, leatherback, leathery turtle, Dermochelys coriacea\nn01667114 mud turtle\nn01667778 terrapin\nn01669191 box turtle, box tortoise\nn01675722 banded gecko\nn01677366 common iguana, iguana, Iguana iguana\nn01682714 American chameleon, anole, Anolis carolinensis\nn01685808 whiptail, whiptail lizard\nn01687978 agama\nn01688243 frilled lizard, Chlamydosaurus kingi\nn01689811 alligator lizard\nn01692333 Gila monster, Heloderma suspectum\nn01693334 green lizard, Lacerta viridis\nn01694178 African chameleon, Chamaeleo chamaeleon\nn01695060 Komodo dragon, Komodo lizard, dragon lizard, giant lizard, Varanus komodoensis\nn01697457 African crocodile, Nile crocodile, Crocodylus niloticus\nn01698640 American alligator, Alligator mississipiensis\nn01704323 triceratops\nn01728572 thunder snake, worm snake, Carphophis amoenus\nn01728920 ringneck snake, ring-necked snake, ring snake\nn01729322 hognose snake, puff adder, sand viper\nn01729977 green snake, grass snake\nn01734418 king snake, kingsnake\nn01735189 garter snake, grass snake\nn01737021 water snake\nn01739381 vine snake\nn01740131 night snake, Hypsiglena torquata\nn01742172 boa constrictor, Constrictor constrictor\nn01744401 rock python, rock snake, Python sebae\nn01748264 Indian cobra, Naja naja\nn01749939 green mamba\nn01751748 sea snake\nn01753488 horned viper, cerastes, sand viper, horned asp, Cerastes cornutus\nn01755581 diamondback, diamondback rattlesnake, Crotalus adamanteus\nn01756291 sidewinder, horned rattlesnake, Crotalus cerastes\nn01768244 trilobite\nn01770081 harvestman, daddy longlegs, Phalangium opilio\nn01770393 scorpion\nn01773157 black and gold garden spider, Argiope aurantia\nn01773549 barn spider, Araneus cavaticus\nn01773797 garden spider, Aranea diademata\nn01774384 black widow, Latrodectus mactans\nn01774750 tarantula\nn01775062 wolf spider, hunting spider\nn01776313 tick\nn01784675 centipede\nn01795545 black grouse\nn01796340 ptarmigan\nn01797886 ruffed grouse, partridge, Bonasa umbellus\nn01798484 prairie chicken, prairie grouse, prairie fowl\nn01806143 peacock\nn01806567 quail\nn01807496 partridge\nn01817953 African grey, African gray, Psittacus erithacus\nn01818515 macaw\nn01819313 sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita\nn01820546 lorikeet\nn01824575 coucal\nn01828970 bee eater\nn01829413 hornbill\nn01833805 hummingbird\nn01843065 jacamar\nn01843383 toucan\nn01847000 drake\nn01855032 red-breasted merganser, Mergus serrator\nn01855672 goose\nn01860187 black swan, Cygnus atratus\nn01871265 tusker\nn01872401 echidna, spiny anteater, anteater\nn01873310 platypus, duckbill, duckbilled platypus, duck-billed platypus, Ornithorhynchus anatinus\nn01877812 wallaby, brush kangaroo\nn01882714 koala, koala bear, kangaroo bear, native bear, Phascolarctos cinereus\nn01883070 wombat\nn01910747 jellyfish\nn01914609 sea anemone, anemone\nn01917289 brain coral\nn01924916 flatworm, platyhelminth\nn01930112 nematode, nematode worm, roundworm\nn01943899 conch\nn01944390 snail\nn01945685 slug\nn01950731 sea slug, nudibranch\nn01955084 chiton, coat-of-mail shell, sea cradle, polyplacophore\nn01968897 chambered nautilus, pearly nautilus, nautilus\nn01978287 Dungeness crab, Cancer magister\nn01978455 rock crab, Cancer irroratus\nn01980166 fiddler crab\nn01981276 king crab, Alaska crab, Alaskan king crab, Alaska king crab, Paralithodes camtschatica\nn01983481 American lobster, Northern lobster, Maine lobster, Homarus americans\nn01984695 spiny lobster, langouste, rock lobster, crawfish, crayfish, sea crawfish\nn01985128 crayfish, crawfish, crawdad, crawdaddy\nn01986214 hermit crab\nn01990800 isopod\nn02002556 white stork, Ciconia ciconia\nn02002724 black stork, Ciconia nigra\nn02006656 spoonbill\nn02007558 flamingo\nn02009229 little blue heron, Egretta caerulea\nn02009912 American egret, great white heron, Egretta albus\nn02011460 bittern\nn02012849 crane\nn02013706 limpkin, Aramus pictus\nn02017213 European gallinule, Porphyrio porphyrio\nn02018207 American coot, marsh hen, mud hen, water hen, Fulica americana\nn02018795 bustard\nn02025239 ruddy turnstone, Arenaria interpres\nn02027492 red-backed sandpiper, dunlin, Erolia alpina\nn02028035 redshank, Tringa totanus\nn02033041 dowitcher\nn02037110 oystercatcher, oyster catcher\nn02051845 pelican\nn02056570 king penguin, Aptenodytes patagonica\nn02058221 albatross, mollymawk\nn02066245 grey whale, gray whale, devilfish, Eschrichtius gibbosus, Eschrichtius robustus\nn02071294 killer whale, killer, orca, grampus, sea wolf, Orcinus orca\nn02074367 dugong, Dugong dugon\nn02077923 sea lion\nn02085620 Chihuahua\nn02085782 Japanese spaniel\nn02085936 Maltese dog, Maltese terrier, Maltese\nn02086079 Pekinese, Pekingese, Peke\nn02086240 Shih-Tzu\nn02086646 Blenheim spaniel\nn02086910 papillon\nn02087046 toy terrier\nn02087394 Rhodesian ridgeback\nn02088094 Afghan hound, Afghan\nn02088238 basset, basset hound\nn02088364 beagle\nn02088466 bloodhound, sleuthhound\nn02088632 bluetick\nn02089078 black-and-tan coonhound\nn02089867 Walker hound, Walker foxhound\nn02089973 English foxhound\nn02090379 redbone\nn02090622 borzoi, Russian wolfhound\nn02090721 Irish wolfhound\nn02091032 Italian greyhound\nn02091134 whippet\nn02091244 Ibizan hound, Ibizan Podenco\nn02091467 Norwegian elkhound, elkhound\nn02091635 otterhound, otter hound\nn02091831 Saluki, gazelle hound\nn02092002 Scottish deerhound, deerhound\nn02092339 Weimaraner\nn02093256 Staffordshire bullterrier, Staffordshire bull terrier\nn02093428 American Staffordshire terrier, Staffordshire terrier, American pit bull terrier, pit bull terrier\nn02093647 Bedlington terrier\nn02093754 Border terrier\nn02093859 Kerry blue terrier\nn02093991 Irish terrier\nn02094114 Norfolk terrier\nn02094258 Norwich terrier\nn02094433 Yorkshire terrier\nn02095314 wire-haired fox terrier\nn02095570 Lakeland terrier\nn02095889 Sealyham terrier, Sealyham\nn02096051 Airedale, Airedale terrier\nn02096177 cairn, cairn terrier\nn02096294 Australian terrier\nn02096437 Dandie Dinmont, Dandie Dinmont terrier\nn02096585 Boston bull, Boston terrier\nn02097047 miniature schnauzer\nn02097130 giant schnauzer\nn02097209 standard schnauzer\nn02097298 Scotch terrier, Scottish terrier, Scottie\nn02097474 Tibetan terrier, chrysanthemum dog\nn02097658 silky terrier, Sydney silky\nn02098105 soft-coated wheaten terrier\nn02098286 West Highland white terrier\nn02098413 Lhasa, Lhasa apso\nn02099267 flat-coated retriever\nn02099429 curly-coated retriever\nn02099601 golden retriever\nn02099712 Labrador retriever\nn02099849 Chesapeake Bay retriever\nn02100236 German short-haired pointer\nn02100583 vizsla, Hungarian pointer\nn02100735 English setter\nn02100877 Irish setter, red setter\nn02101006 Gordon setter\nn02101388 Brittany spaniel\nn02101556 clumber, clumber spaniel\nn02102040 English springer, English springer spaniel\nn02102177 Welsh springer spaniel\nn02102318 cocker spaniel, English cocker spaniel, cocker\nn02102480 Sussex spaniel\nn02102973 Irish water spaniel\nn02104029 kuvasz\nn02104365 schipperke\nn02105056 groenendael\nn02105162 malinois\nn02105251 briard\nn02105412 kelpie\nn02105505 komondor\nn02105641 Old English sheepdog, bobtail\nn02105855 Shetland sheepdog, Shetland sheep dog, Shetland\nn02106030 collie\nn02106166 Border collie\nn02106382 Bouvier des Flandres, Bouviers des Flandres\nn02106550 Rottweiler\nn02106662 German shepherd, German shepherd dog, German police dog, alsatian\nn02107142 Doberman, Doberman pinscher\nn02107312 miniature pinscher\nn02107574 Greater Swiss Mountain dog\nn02107683 Bernese mountain dog\nn02107908 Appenzeller\nn02108000 EntleBucher\nn02108089 boxer\nn02108422 bull mastiff\nn02108551 Tibetan mastiff\nn02108915 French bulldog\nn02109047 Great Dane\nn02109525 Saint Bernard, St Bernard\nn02109961 Eskimo dog, husky\nn02110063 malamute, malemute, Alaskan malamute\nn02110185 Siberian husky\nn02110341 dalmatian, coach dog, carriage dog\nn02110627 affenpinscher, monkey pinscher, monkey dog\nn02110806 basenji\nn02110958 pug, pug-dog\nn02111129 Leonberg\nn02111277 Newfoundland, Newfoundland dog\nn02111500 Great Pyrenees\nn02111889 Samoyed, Samoyede\nn02112018 Pomeranian\nn02112137 chow, chow chow\nn02112350 keeshond\nn02112706 Brabancon griffon\nn02113023 Pembroke, Pembroke Welsh corgi\nn02113186 Cardigan, Cardigan Welsh corgi\nn02113624 toy poodle\nn02113712 miniature poodle\nn02113799 standard poodle\nn02113978 Mexican hairless\nn02114367 timber wolf, grey wolf, gray wolf, Canis lupus\nn02114548 white wolf, Arctic wolf, Canis lupus tundrarum\nn02114712 red wolf, maned wolf, Canis rufus, Canis niger\nn02114855 coyote, prairie wolf, brush wolf, Canis latrans\nn02115641 dingo, warrigal, warragal, Canis dingo\nn02115913 dhole, Cuon alpinus\nn02116738 African hunting dog, hyena dog, Cape hunting dog, Lycaon pictus\nn02117135 hyena, hyaena\nn02119022 red fox, Vulpes vulpes\nn02119789 kit fox, Vulpes macrotis\nn02120079 Arctic fox, white fox, Alopex lagopus\nn02120505 grey fox, gray fox, Urocyon cinereoargenteus\nn02123045 tabby, tabby cat\nn02123159 tiger cat\nn02123394 Persian cat\nn02123597 Siamese cat, Siamese\nn02124075 Egyptian cat\nn02125311 cougar, puma, catamount, mountain lion, painter, panther, Felis concolor\nn02127052 lynx, catamount\nn02128385 leopard, Panthera pardus\nn02128757 snow leopard, ounce, Panthera uncia\nn02128925 jaguar, panther, Panthera onca, Felis onca\nn02129165 lion, king of beasts, Panthera leo\nn02129604 tiger, Panthera tigris\nn02130308 cheetah, chetah, Acinonyx jubatus\nn02132136 brown bear, bruin, Ursus arctos\nn02133161 American black bear, black bear, Ursus americans, Euarctos americans\nn02134084 ice bear, polar bear, Ursus Maritimus, Thalarctos maritimus\nn02134418 sloth bear, Melursus ursinus, Ursus ursinus\nn02137549 mongoose\nn02138441 meerkat, mierkat\nn02165105 tiger beetle\nn02165456 ladybug, ladybeetle, lady beetle, ladybird, ladybird beetle\nn02167151 ground beetle, carabid beetle\nn02168699 long-horned beetle, longicorn, longicorn beetle\nn02169497 leaf beetle, chrysomelid\nn02172182 dung beetle\nn02174001 rhinoceros beetle\nn02177972 weevil\nn02190166 fly\nn02206856 bee\nn02219486 ant, emmet, pismire\nn02226429 grasshopper, hopper\nn02229544 cricket\nn02231487 walking stick, walkingstick, stick insect\nn02233338 cockroach, roach\nn02236044 mantis, mantid\nn02256656 cicada, cicala\nn02259212 leafhopper\nn02264363 lacewing, lacewing fly\nn02268443 dragonfly, darning needle, devil's darning needle, sewing needle, snake feeder, snake doctor, mosquito hawk, skeeter hawk\nn02268853 damselfly\nn02276258 admiral\nn02277742 ringlet, ringlet butterfly\nn02279972 monarch, monarch butterfly, milkweed butterfly, Danaus plexippus\nn02280649 cabbage butterfly\nn02281406 sulphur butterfly, sulfur butterfly\nn02281787 lycaenid, lycaenid butterfly\nn02317335 starfish, sea star\nn02319095 sea urchin\nn02321529 sea cucumber, holothurian\nn02325366 wood rabbit, cottontail, cottontail rabbit\nn02326432 hare\nn02328150 Angora, Angora rabbit\nn02342885 hamster\nn02346627 porcupine, hedgehog\nn02356798 fox squirrel, eastern fox squirrel, Sciurus niger\nn02361337 marmot\nn02363005 beaver\nn02364673 guinea pig, Cavia cobaya\nn02389026 sorrel\nn02391049 zebra\nn02395406 hog, pig, grunter, squealer, Sus scrofa\nn02396427 wild boar, boar, Sus scrofa\nn02397096 warthog\nn02398521 hippopotamus, hippo, river horse, Hippopotamus amphibius\nn02403003 ox\nn02408429 water buffalo, water ox, Asiatic buffalo, Bubalus bubalis\nn02410509 bison\nn02412080 ram, tup\nn02415577 bighorn, bighorn sheep, cimarron, Rocky Mountain bighorn, Rocky Mountain sheep, Ovis canadensis\nn02417914 ibex, Capra ibex\nn02422106 hartebeest\nn02422699 impala, Aepyceros melampus\nn02423022 gazelle\nn02437312 Arabian camel, dromedary, Camelus dromedarius\nn02437616 llama\nn02441942 weasel\nn02442845 mink\nn02443114 polecat, fitch, foulmart, foumart, Mustela putorius\nn02443484 black-footed ferret, ferret, Mustela nigripes\nn02444819 otter\nn02445715 skunk, polecat, wood pussy\nn02447366 badger\nn02454379 armadillo\nn02457408 three-toed sloth, ai, Bradypus tridactylus\nn02480495 orangutan, orang, orangutang, Pongo pygmaeus\nn02480855 gorilla, Gorilla gorilla\nn02481823 chimpanzee, chimp, Pan troglodytes\nn02483362 gibbon, Hylobates lar\nn02483708 siamang, Hylobates syndactylus, Symphalangus syndactylus\nn02484975 guenon, guenon monkey\nn02486261 patas, hussar monkey, Erythrocebus patas\nn02486410 baboon\nn02487347 macaque\nn02488291 langur\nn02488702 colobus, colobus monkey\nn02489166 proboscis monkey, Nasalis larvatus\nn02490219 marmoset\nn02492035 capuchin, ringtail, Cebus capucinus\nn02492660 howler monkey, howler\nn02493509 titi, titi monkey\nn02493793 spider monkey, Ateles geoffroyi\nn02494079 squirrel monkey, Saimiri sciureus\nn02497673 Madagascar cat, ring-tailed lemur, Lemur catta\nn02500267 indri, indris, Indri indri, Indri brevicaudatus\nn02504013 Indian elephant, Elephas maximus\nn02504458 African elephant, Loxodonta africana\nn02509815 lesser panda, red panda, panda, bear cat, cat bear, Ailurus fulgens\nn02510455 giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca\nn02514041 barracouta, snoek\nn02526121 eel\nn02536864 coho, cohoe, coho salmon, blue jack, silver salmon, Oncorhynchus kisutch\nn02606052 rock beauty, Holocanthus tricolor\nn02607072 anemone fish\nn02640242 sturgeon\nn02641379 gar, garfish, garpike, billfish, Lepisosteus osseus\nn02643566 lionfish\nn02655020 puffer, pufferfish, blowfish, globefish\nn02666196 abacus\nn02667093 abaya\nn02669723 academic gown, academic robe, judge's robe\nn02672831 accordion, piano accordion, squeeze box\nn02676566 acoustic guitar\nn02687172 aircraft carrier, carrier, flattop, attack aircraft carrier\nn02690373 airliner\nn02692877 airship, dirigible\nn02699494 altar\nn02701002 ambulance\nn02704792 amphibian, amphibious vehicle\nn02708093 analog clock\nn02727426 apiary, bee house\nn02730930 apron\nn02747177 ashcan, trash can, garbage can, wastebin, ash bin, ash-bin, ashbin, dustbin, trash barrel, trash bin\nn02749479 assault rifle, assault gun\nn02769748 backpack, back pack, knapsack, packsack, rucksack, haversack\nn02776631 bakery, bakeshop, bakehouse\nn02777292 balance beam, beam\nn02782093 balloon\nn02783161 ballpoint, ballpoint pen, ballpen, Biro\nn02786058 Band Aid\nn02787622 banjo\nn02788148 bannister, banister, balustrade, balusters, handrail\nn02790996 barbell\nn02791124 barber chair\nn02791270 barbershop\nn02793495 barn\nn02794156 barometer\nn02795169 barrel, cask\nn02797295 barrow, garden cart, lawn cart, wheelbarrow\nn02799071 baseball\nn02802426 basketball\nn02804414 bassinet\nn02804610 bassoon\nn02807133 bathing cap, swimming cap\nn02808304 bath towel\nn02808440 bathtub, bathing tub, bath, tub\nn02814533 beach wagon, station wagon, wagon, estate car, beach waggon, station waggon, waggon\nn02814860 beacon, lighthouse, beacon light, pharos\nn02815834 beaker\nn02817516 bearskin, busby, shako\nn02823428 beer bottle\nn02823750 beer glass\nn02825657 bell cote, bell cot\nn02834397 bib\nn02835271 bicycle-built-for-two, tandem bicycle, tandem\nn02837789 bikini, two-piece\nn02840245 binder, ring-binder\nn02841315 binoculars, field glasses, opera glasses\nn02843684 birdhouse\nn02859443 boathouse\nn02860847 bobsled, bobsleigh, bob\nn02865351 bolo tie, bolo, bola tie, bola\nn02869837 bonnet, poke bonnet\nn02870880 bookcase\nn02871525 bookshop, bookstore, bookstall\nn02877765 bottlecap\nn02879718 bow\nn02883205 bow tie, bow-tie, bowtie\nn02892201 brass, memorial tablet, plaque\nn02892767 brassiere, bra, bandeau\nn02894605 breakwater, groin, groyne, mole, bulwark, seawall, jetty\nn02895154 breastplate, aegis, egis\nn02906734 broom\nn02909870 bucket, pail\nn02910353 buckle\nn02916936 bulletproof vest\nn02917067 bullet train, bullet\nn02927161 butcher shop, meat market\nn02930766 cab, hack, taxi, taxicab\nn02939185 caldron, cauldron\nn02948072 candle, taper, wax light\nn02950826 cannon\nn02951358 canoe\nn02951585 can opener, tin opener\nn02963159 cardigan\nn02965783 car mirror\nn02966193 carousel, carrousel, merry-go-round, roundabout, whirligig\nn02966687 carpenter's kit, tool kit\nn02971356 carton\nn02974003 car wheel\nn02977058 cash machine, cash dispenser, automated teller machine, automatic teller machine, automated teller, automatic teller, ATM\nn02978881 cassette\nn02979186 cassette player\nn02980441 castle\nn02981792 catamaran\nn02988304 CD player\nn02992211 cello, violoncello\nn02992529 cellular telephone, cellular phone, cellphone, cell, mobile phone\nn02999410 chain\nn03000134 chainlink fence\nn03000247 chain mail, ring mail, mail, chain armor, chain armour, ring armor, ring armour\nn03000684 chain saw, chainsaw\nn03014705 chest\nn03016953 chiffonier, commode\nn03017168 chime, bell, gong\nn03018349 china cabinet, china closet\nn03026506 Christmas stocking\nn03028079 church, church building\nn03032252 cinema, movie theater, movie theatre, movie house, picture palace\nn03041632 cleaver, meat cleaver, chopper\nn03042490 cliff dwelling\nn03045698 cloak\nn03047690 clog, geta, patten, sabot\nn03062245 cocktail shaker\nn03063599 coffee mug\nn03063689 coffeepot\nn03065424 coil, spiral, volute, whorl, helix\nn03075370 combination lock\nn03085013 computer keyboard, keypad\nn03089624 confectionery, confectionary, candy store\nn03095699 container ship, containership, container vessel\nn03100240 convertible\nn03109150 corkscrew, bottle screw\nn03110669 cornet, horn, trumpet, trump\nn03124043 cowboy boot\nn03124170 cowboy hat, ten-gallon hat\nn03125729 cradle\nn03126707 crane\nn03127747 crash helmet\nn03127925 crate\nn03131574 crib, cot\nn03133878 Crock Pot\nn03134739 croquet ball\nn03141823 crutch\nn03146219 cuirass\nn03160309 dam, dike, dyke\nn03179701 desk\nn03180011 desktop computer\nn03187595 dial telephone, dial phone\nn03188531 diaper, nappy, napkin\nn03196217 digital clock\nn03197337 digital watch\nn03201208 dining table, board\nn03207743 dishrag, dishcloth\nn03207941 dishwasher, dish washer, dishwashing machine\nn03208938 disk brake, disc brake\nn03216828 dock, dockage, docking facility\nn03218198 dogsled, dog sled, dog sleigh\nn03220513 dome\nn03223299 doormat, welcome mat\nn03240683 drilling platform, offshore rig\nn03249569 drum, membranophone, tympan\nn03250847 drumstick\nn03255030 dumbbell\nn03259280 Dutch oven\nn03271574 electric fan, blower\nn03272010 electric guitar\nn03272562 electric locomotive\nn03290653 entertainment center\nn03291819 envelope\nn03297495 espresso maker\nn03314780 face powder\nn03325584 feather boa, boa\nn03337140 file, file cabinet, filing cabinet\nn03344393 fireboat\nn03345487 fire engine, fire truck\nn03347037 fire screen, fireguard\nn03355925 flagpole, flagstaff\nn03372029 flute, transverse flute\nn03376595 folding chair\nn03379051 football helmet\nn03384352 forklift\nn03388043 fountain\nn03388183 fountain pen\nn03388549 four-poster\nn03393912 freight car\nn03394916 French horn, horn\nn03400231 frying pan, frypan, skillet\nn03404251 fur coat\nn03417042 garbage truck, dustcart\nn03424325 gasmask, respirator, gas helmet\nn03425413 gas pump, gasoline pump, petrol pump, island dispenser\nn03443371 goblet\nn03444034 go-kart\nn03445777 golf ball\nn03445924 golfcart, golf cart\nn03447447 gondola\nn03447721 gong, tam-tam\nn03450230 gown\nn03452741 grand piano, grand\nn03457902 greenhouse, nursery, glasshouse\nn03459775 grille, radiator grille\nn03461385 grocery store, grocery, food market, market\nn03467068 guillotine\nn03476684 hair slide\nn03476991 hair spray\nn03478589 half track\nn03481172 hammer\nn03482405 hamper\nn03483316 hand blower, blow dryer, blow drier, hair dryer, hair drier\nn03485407 hand-held computer, hand-held microcomputer\nn03485794 handkerchief, hankie, hanky, hankey\nn03492542 hard disc, hard disk, fixed disk\nn03494278 harmonica, mouth organ, harp, mouth harp\nn03495258 harp\nn03496892 harvester, reaper\nn03498962 hatchet\nn03527444 holster\nn03529860 home theater, home theatre\nn03530642 honeycomb\nn03532672 hook, claw\nn03534580 hoopskirt, crinoline\nn03535780 horizontal bar, high bar\nn03538406 horse cart, horse-cart\nn03544143 hourglass\nn03584254 iPod\nn03584829 iron, smoothing iron\nn03590841 jack-o'-lantern\nn03594734 jean, blue jean, denim\nn03594945 jeep, landrover\nn03595614 jersey, T-shirt, tee shirt\nn03598930 jigsaw puzzle\nn03599486 jinrikisha, ricksha, rickshaw\nn03602883 joystick\nn03617480 kimono\nn03623198 knee pad\nn03627232 knot\nn03630383 lab coat, laboratory coat\nn03633091 ladle\nn03637318 lampshade, lamp shade\nn03642806 laptop, laptop computer\nn03649909 lawn mower, mower\nn03657121 lens cap, lens cover\nn03658185 letter opener, paper knife, paperknife\nn03661043 library\nn03662601 lifeboat\nn03666591 lighter, light, igniter, ignitor\nn03670208 limousine, limo\nn03673027 liner, ocean liner\nn03676483 lipstick, lip rouge\nn03680355 Loafer\nn03690938 lotion\nn03691459 loudspeaker, speaker, speaker unit, loudspeaker system, speaker system\nn03692522 loupe, jeweler's loupe\nn03697007 lumbermill, sawmill\nn03706229 magnetic compass\nn03709823 mailbag, postbag\nn03710193 mailbox, letter box\nn03710637 maillot\nn03710721 maillot, tank suit\nn03717622 manhole cover\nn03720891 maraca\nn03721384 marimba, xylophone\nn03724870 mask\nn03729826 matchstick\nn03733131 maypole\nn03733281 maze, labyrinth\nn03733805 measuring cup\nn03742115 medicine chest, medicine cabinet\nn03743016 megalith, megalithic structure\nn03759954 microphone, mike\nn03761084 microwave, microwave oven\nn03763968 military uniform\nn03764736 milk can\nn03769881 minibus\nn03770439 miniskirt, mini\nn03770679 minivan\nn03773504 missile\nn03775071 mitten\nn03775546 mixing bowl\nn03776460 mobile home, manufactured home\nn03777568 Model T\nn03777754 modem\nn03781244 monastery\nn03782006 monitor\nn03785016 moped\nn03786901 mortar\nn03787032 mortarboard\nn03788195 mosque\nn03788365 mosquito net\nn03791053 motor scooter, scooter\nn03792782 mountain bike, all-terrain bike, off-roader\nn03792972 mountain tent\nn03793489 mouse, computer mouse\nn03794056 mousetrap\nn03796401 moving van\nn03803284 muzzle\nn03804744 nail\nn03814639 neck brace\nn03814906 necklace\nn03825788 nipple\nn03832673 notebook, notebook computer\nn03837869 obelisk\nn03838899 oboe, hautboy, hautbois\nn03840681 ocarina, sweet potato\nn03841143 odometer, hodometer, mileometer, milometer\nn03843555 oil filter\nn03854065 organ, pipe organ\nn03857828 oscilloscope, scope, cathode-ray oscilloscope, CRO\nn03866082 overskirt\nn03868242 oxcart\nn03868863 oxygen mask\nn03871628 packet\nn03873416 paddle, boat paddle\nn03874293 paddlewheel, paddle wheel\nn03874599 padlock\nn03876231 paintbrush\nn03877472 pajama, pyjama, pj's, jammies\nn03877845 palace\nn03884397 panpipe, pandean pipe, syrinx\nn03887697 paper towel\nn03888257 parachute, chute\nn03888605 parallel bars, bars\nn03891251 park bench\nn03891332 parking meter\nn03895866 passenger car, coach, carriage\nn03899768 patio, terrace\nn03902125 pay-phone, pay-station\nn03903868 pedestal, plinth, footstall\nn03908618 pencil box, pencil case\nn03908714 pencil sharpener\nn03916031 perfume, essence\nn03920288 Petri dish\nn03924679 photocopier\nn03929660 pick, plectrum, plectron\nn03929855 pickelhaube\nn03930313 picket fence, paling\nn03930630 pickup, pickup truck\nn03933933 pier\nn03935335 piggy bank, penny bank\nn03937543 pill bottle\nn03938244 pillow\nn03942813 ping-pong ball\nn03944341 pinwheel\nn03947888 pirate, pirate ship\nn03950228 pitcher, ewer\nn03954731 plane, carpenter's plane, woodworking plane\nn03956157 planetarium\nn03958227 plastic bag\nn03961711 plate rack\nn03967562 plow, plough\nn03970156 plunger, plumber's helper\nn03976467 Polaroid camera, Polaroid Land camera\nn03976657 pole\nn03977966 police van, police wagon, paddy wagon, patrol wagon, wagon, black Maria\nn03980874 poncho\nn03982430 pool table, billiard table, snooker table\nn03983396 pop bottle, soda bottle\nn03991062 pot, flowerpot\nn03992509 potter's wheel\nn03995372 power drill\nn03998194 prayer rug, prayer mat\nn04004767 printer\nn04005630 prison, prison house\nn04008634 projectile, missile\nn04009552 projector\nn04019541 puck, hockey puck\nn04023962 punching bag, punch bag, punching ball, punchball\nn04026417 purse\nn04033901 quill, quill pen\nn04033995 quilt, comforter, comfort, puff\nn04037443 racer, race car, racing car\nn04039381 racket, racquet\nn04040759 radiator\nn04041544 radio, wireless\nn04044716 radio telescope, radio reflector\nn04049303 rain barrel\nn04065272 recreational vehicle, RV, R.V.\nn04067472 reel\nn04069434 reflex camera\nn04070727 refrigerator, icebox\nn04074963 remote control, remote\nn04081281 restaurant, eating house, eating place, eatery\nn04086273 revolver, six-gun, six-shooter\nn04090263 rifle\nn04099969 rocking chair, rocker\nn04111531 rotisserie\nn04116512 rubber eraser, rubber, pencil eraser\nn04118538 rugby ball\nn04118776 rule, ruler\nn04120489 running shoe\nn04125021 safe\nn04127249 safety pin\nn04131690 saltshaker, salt shaker\nn04133789 sandal\nn04136333 sarong\nn04141076 sax, saxophone\nn04141327 scabbard\nn04141975 scale, weighing machine\nn04146614 school bus\nn04147183 schooner\nn04149813 scoreboard\nn04152593 screen, CRT screen\nn04153751 screw\nn04154565 screwdriver\nn04162706 seat belt, seatbelt\nn04179913 sewing machine\nn04192698 shield, buckler\nn04200800 shoe shop, shoe-shop, shoe store\nn04201297 shoji\nn04204238 shopping basket\nn04204347 shopping cart\nn04208210 shovel\nn04209133 shower cap\nn04209239 shower curtain\nn04228054 ski\nn04229816 ski mask\nn04235860 sleeping bag\nn04238763 slide rule, slipstick\nn04239074 sliding door\nn04243546 slot, one-armed bandit\nn04251144 snorkel\nn04252077 snowmobile\nn04252225 snowplow, snowplough\nn04254120 soap dispenser\nn04254680 soccer ball\nn04254777 sock\nn04258138 solar dish, solar collector, solar furnace\nn04259630 sombrero\nn04263257 soup bowl\nn04264628 space bar\nn04265275 space heater\nn04266014 space shuttle\nn04270147 spatula\nn04273569 speedboat\nn04275548 spider web, spider's web\nn04277352 spindle\nn04285008 sports car, sport car\nn04286575 spotlight, spot\nn04296562 stage\nn04310018 steam locomotive\nn04311004 steel arch bridge\nn04311174 steel drum\nn04317175 stethoscope\nn04325704 stole\nn04326547 stone wall\nn04328186 stopwatch, stop watch\nn04330267 stove\nn04332243 strainer\nn04335435 streetcar, tram, tramcar, trolley, trolley car\nn04336792 stretcher\nn04344873 studio couch, day bed\nn04346328 stupa, tope\nn04347754 submarine, pigboat, sub, U-boat\nn04350905 suit, suit of clothes\nn04355338 sundial\nn04355933 sunglass\nn04356056 sunglasses, dark glasses, shades\nn04357314 sunscreen, sunblock, sun blocker\nn04366367 suspension bridge\nn04367480 swab, swob, mop\nn04370456 sweatshirt\nn04371430 swimming trunks, bathing trunks\nn04371774 swing\nn04372370 switch, electric switch, electrical switch\nn04376876 syringe\nn04380533 table lamp\nn04389033 tank, army tank, armored combat vehicle, armoured combat vehicle\nn04392985 tape player\nn04398044 teapot\nn04399382 teddy, teddy bear\nn04404412 television, television system\nn04409515 tennis ball\nn04417672 thatch, thatched roof\nn04418357 theater curtain, theatre curtain\nn04423845 thimble\nn04428191 thresher, thrasher, threshing machine\nn04429376 throne\nn04435653 tile roof\nn04442312 toaster\nn04443257 tobacco shop, tobacconist shop, tobacconist\nn04447861 toilet seat\nn04456115 torch\nn04458633 totem pole\nn04461696 tow truck, tow car, wrecker\nn04462240 toyshop\nn04465501 tractor\nn04467665 trailer truck, tractor trailer, trucking rig, rig, articulated lorry, semi\nn04476259 tray\nn04479046 trench coat\nn04482393 tricycle, trike, velocipede\nn04483307 trimaran\nn04485082 tripod\nn04486054 triumphal arch\nn04487081 trolleybus, trolley coach, trackless trolley\nn04487394 trombone\nn04493381 tub, vat\nn04501370 turnstile\nn04505470 typewriter keyboard\nn04507155 umbrella\nn04509417 unicycle, monocycle\nn04515003 upright, upright piano\nn04517823 vacuum, vacuum cleaner\nn04522168 vase\nn04523525 vault\nn04525038 velvet\nn04525305 vending machine\nn04532106 vestment\nn04532670 viaduct\nn04536866 violin, fiddle\nn04540053 volleyball\nn04542943 waffle iron\nn04548280 wall clock\nn04548362 wallet, billfold, notecase, pocketbook\nn04550184 wardrobe, closet, press\nn04552348 warplane, military plane\nn04553703 washbasin, handbasin, washbowl, lavabo, wash-hand basin\nn04554684 washer, automatic washer, washing machine\nn04557648 water bottle\nn04560804 water jug\nn04562935 water tower\nn04579145 whiskey jug\nn04579432 whistle\nn04584207 wig\nn04589890 window screen\nn04590129 window shade\nn04591157 Windsor tie\nn04591713 wine bottle\nn04592741 wing\nn04596742 wok\nn04597913 wooden spoon\nn04599235 wool, woolen, woollen\nn04604644 worm fence, snake fence, snake-rail fence, Virginia fence\nn04606251 wreck\nn04612504 yawl\nn04613696 yurt\nn06359193 web site, website, internet site, site\nn06596364 comic book\nn06785654 crossword puzzle, crossword\nn06794110 street sign\nn06874185 traffic light, traffic signal, stoplight\nn07248320 book jacket, dust cover, dust jacket, dust wrapper\nn07565083 menu\nn07579787 plate\nn07583066 guacamole\nn07584110 consomme\nn07590611 hot pot, hotpot\nn07613480 trifle\nn07614500 ice cream, icecream\nn07615774 ice lolly, lolly, lollipop, popsicle\nn07684084 French loaf\nn07693725 bagel, beigel\nn07695742 pretzel\nn07697313 cheeseburger\nn07697537 hotdog, hot dog, red hot\nn07711569 mashed potato\nn07714571 head cabbage\nn07714990 broccoli\nn07715103 cauliflower\nn07716358 zucchini, courgette\nn07716906 spaghetti squash\nn07717410 acorn squash\nn07717556 butternut squash\nn07718472 cucumber, cuke\nn07718747 artichoke, globe artichoke\nn07720875 bell pepper\nn07730033 cardoon\nn07734744 mushroom\nn07742313 Granny Smith\nn07745940 strawberry\nn07747607 orange\nn07749582 lemon\nn07753113 fig\nn07753275 pineapple, ananas\nn07753592 banana\nn07754684 jackfruit, jak, jack\nn07760859 custard apple\nn07768694 pomegranate\nn07802026 hay\nn07831146 carbonara\nn07836838 chocolate sauce, chocolate syrup\nn07860988 dough\nn07871810 meat loaf, meatloaf\nn07873807 pizza, pizza pie\nn07875152 potpie\nn07880968 burrito\nn07892512 red wine\nn07920052 espresso\nn07930864 cup\nn07932039 eggnog\nn09193705 alp\nn09229709 bubble\nn09246464 cliff, drop, drop-off\nn09256479 coral reef\nn09288635 geyser\nn09332890 lakeside, lakeshore\nn09399592 promontory, headland, head, foreland\nn09421951 sandbar, sand bar\nn09428293 seashore, coast, seacoast, sea-coast\nn09468604 valley, vale\nn09472597 volcano\nn09835506 ballplayer, baseball player\nn10148035 groom, bridegroom\nn10565667 scuba diver\nn11879895 rapeseed\nn11939491 daisy\nn12057211 yellow lady's slipper, yellow lady-slipper, Cypripedium calceolus, Cypripedium parviflorum\nn12144580 corn\nn12267677 acorn\nn12620546 hip, rose hip, rosehip\nn12768682 buckeye, horse chestnut, conker\nn12985857 coral fungus\nn12998815 agaric\nn13037406 gyromitra\nn13040303 stinkhorn, carrion fungus\nn13044778 earthstar\nn13052670 hen-of-the-woods, hen of the woods, Polyporus frondosus, Grifola frondosa\nn13054560 bolete\nn13133613 ear, spike, capitulum\nn15075141 toilet tissue, toilet paper, bathroom tissue\n" }, { "path": "examples/whisper.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// whisper speech recognition implemented with ncnn library\n\n// convert openai-whisper checkpoints to ncnn models\n// 1. install pnnx via pip install -U pnnx\n// 2. obtain export_ncnn.py script from https://github.com/nihui/ncnn-android-whisper\n// 3. edit export_ncnn.py for changing the models among tiny/base/small/medium/large-v3-turbo\n// 4. make sure you have good internet connection\n// python export_ncnn.py\n\n// convert vocab.json to simple whisper_vocab.txt\n// 1. obtain vocab.json file from https://huggingface.co/openai/whisper-tiny/blob/main/vocab.json\n// 2. convert json dict into plain list, save to whisper_vocab.txt\n\n// NOTE large-v3-turbo has special token ids from others, one more language(yue) and does not support translation\n\n#include \"net.h\"\n#include \"layer.h\"\n#include \"layer_type.h\"\n\n#include \n#include \n#include \n#include \n#include \n#include \n#include \n\n// https://huggingface.co/openai/whisper-tiny/blob/main/tokenizer_config.json\nstatic const int token_endoftext = 50257;\nstatic const int token_startoftranscript = 50258;\nstatic const int token_lang_first = 50259;\nstatic const int token_lang_last = 50357;\nstatic const int token_lang_count = token_lang_last - token_lang_first + 1;\n// clang-format off\n// *INDENT-OFF*\nstatic const char* token_langs[] = {\n \"en\", \"zh\", \"de\", \"es\", \"ru\", \"ko\", \"fr\", \"ja\", \"pt\", \"tr\", \"pl\", \"ca\", \"nl\", \"ar\", \"sv\",\n \"it\", \"id\", \"hi\", \"fi\", \"vi\", \"he\", \"uk\", \"el\", \"ms\", \"cs\", \"ro\", \"da\", \"hu\", \"ta\", \"no\",\n \"th\", \"ur\", \"hr\", \"bg\", \"lt\", \"la\", \"mi\", \"ml\", \"cy\", \"sk\", \"te\", \"fa\", \"lv\", \"bn\", \"sr\",\n \"az\", \"sl\", \"kn\", \"et\", \"mk\", \"br\", \"eu\", \"is\", \"hy\", \"ne\", \"mn\", \"bs\", \"kk\", \"sq\", \"sw\",\n \"gl\", \"mr\", \"pa\", \"si\", \"km\", \"sn\", \"yo\", \"so\", \"af\", \"oc\", \"ka\", \"be\", \"tg\", \"sd\", \"gu\",\n \"am\", \"yi\", \"lo\", \"uz\", \"fo\", \"ht\", \"ps\", \"tk\", \"nn\", \"mt\", \"sa\", \"lb\", \"my\", \"bo\", \"tl\",\n \"mg\", \"as\", \"tt\", \"haw\", \"ln\", \"ha\", \"ba\", \"jw\", \"su\"\n};\n// *INDENT-ON*\n// clang-format on\nstatic const int token_translate = 50358;\nstatic const int token_transcribe = 50359;\nstatic const int token_startoflm = 50360;\nstatic const int token_startofprev = 50361;\nstatic const int token_nocaptions = 50362;\nstatic const int token_notimestamps = 50363;\nstatic const int token_timestamp_first = 50364;\nstatic const int token_timestamp_last = 51864;\n\n// https://huggingface.co/openai/whisper-large-v3-turbo/blob/main/tokenizer_config.json\n// static const int token_endoftext = 50257;\n// static const int token_startoftranscript = 50258;\n// static const int token_lang_first = 50259;\n// static const int token_lang_last = 50357;\n// static const int token_lang_count = token_lang_last - token_lang_first + 1;\n// // clang-format off\n// // *INDENT-OFF*\n// static const char* token_langs[] = {\n// \"en\", \"zh\", \"de\", \"es\", \"ru\", \"ko\", \"fr\", \"ja\", \"pt\", \"tr\", \"pl\", \"ca\", \"nl\", \"ar\", \"sv\",\n// \"it\", \"id\", \"hi\", \"fi\", \"vi\", \"he\", \"uk\", \"el\", \"ms\", \"cs\", \"ro\", \"da\", \"hu\", \"ta\", \"no\",\n// \"th\", \"ur\", \"hr\", \"bg\", \"lt\", \"la\", \"mi\", \"ml\", \"cy\", \"sk\", \"te\", \"fa\", \"lv\", \"bn\", \"sr\",\n// \"az\", \"sl\", \"kn\", \"et\", \"mk\", \"br\", \"eu\", \"is\", \"hy\", \"ne\", \"mn\", \"bs\", \"kk\", \"sq\", \"sw\",\n// \"gl\", \"mr\", \"pa\", \"si\", \"km\", \"sn\", \"yo\", \"so\", \"af\", \"oc\", \"ka\", \"be\", \"tg\", \"sd\", \"gu\",\n// \"am\", \"yi\", \"lo\", \"uz\", \"fo\", \"ht\", \"ps\", \"tk\", \"nn\", \"mt\", \"sa\", \"lb\", \"my\", \"bo\", \"tl\",\n// \"mg\", \"as\", \"tt\", \"haw\", \"ln\", \"ha\", \"ba\", \"jw\", \"su\", \"yue\"\n// };\n// // *INDENT-ON*\n// // clang-format on\n// static const int token_translate = 50359;\n// static const int token_transcribe = 50360;\n// static const int token_startoflm = 50361;\n// static const int token_startofprev = 50362;\n// static const int token_nospeech = 50363;\n// static const int token_notimestamps = 50364;\n// static const int token_timestamp_first = 50365;\n// static const int token_timestamp_last = 51865;\n\n// tokenizer for handling text tokens\nclass Tokenizer\n{\npublic:\n std::vector reverse_vocab;\n\n uint8_t byte_decoder[512]; // unicode code point to byte value\n\n // generate byte decoder for tokenization\n void generate_byte_decoder()\n {\n // initialize array to 0\n memset(byte_decoder, 0, 512 * sizeof(uint8_t));\n\n // define function to check if char is in \"printable\" range\n auto is_printable = [](int b) {\n return (b >= '!' && b <= '~') // '!' to '~'\n || (b >= 161 && b <= 172) // '¡' to '¬'\n || (b >= 174 && b <= 255); // '®' to 'ÿ'\n };\n\n // handle \"printable\" characters\n // for these chars, key and value are the same\n for (int b = 0; b < 256; ++b)\n {\n if (is_printable(b))\n {\n byte_decoder[b] = static_cast(b);\n }\n }\n\n // handle remaining characters\n // for these chars, key starts from 256 and increments\n int n = 0;\n for (int b = 0; b < 256; ++b)\n {\n if (!is_printable(b))\n {\n byte_decoder[256 + n] = static_cast(b);\n n++;\n }\n }\n }\n\n // convert utf-8 string to code points\n std::vector utf8_to_codepoints(const std::string& s) const\n {\n std::vector codepoints;\n for (size_t i = 0; i < s.length();)\n {\n uint32_t cp = 0;\n int len = 0;\n unsigned char c = s[i];\n\n if (c < 0x80) // 1-byte\n {\n cp = c;\n len = 1;\n }\n else if ((c & 0xE0) == 0xC0) // 2-byte\n {\n cp = ((s[i] & 0x1F) << 6) | (s[i + 1] & 0x3F);\n len = 2;\n }\n else if ((c & 0xF0) == 0xE0) // 3-byte\n {\n cp = ((s[i] & 0x0F) << 12) | ((s[i + 1] & 0x3F) << 6) | (s[i + 2] & 0x3F);\n len = 3;\n }\n else if ((c & 0xF8) == 0xF0) // 4-byte\n {\n cp = ((s[i] & 0x07) << 18) | ((s[i + 1] & 0x3F) << 12) | ((s[i + 2] & 0x3F) << 6) | (s[i + 3] & 0x3F);\n len = 4;\n }\n else\n {\n // invalid utf-8 start byte, skip\n i++;\n continue;\n }\n codepoints.push_back(cp);\n i += len;\n }\n return codepoints;\n }\n\n bool load(const char* vocab_path)\n {\n // generate decoder when loading\n generate_byte_decoder();\n\n {\n FILE* fp = fopen(vocab_path, \"rb\");\n if (!fp)\n {\n fprintf(stderr, \"fopen %s failed\\n\", vocab_path);\n return false;\n }\n\n char line[256];\n while (!feof(fp))\n {\n char* s = fgets(line, 255, fp);\n if (!s)\n break;\n\n int vocab_len = strlen(line);\n if (vocab_len > 1)\n {\n // drop the tail newline\n vocab_len -= 1;\n }\n\n reverse_vocab.push_back(std::string(line, vocab_len));\n }\n\n fclose(fp);\n }\n\n return true;\n }\n\n // decode token ids to text\n std::string decode(const std::vector& tokens) const\n {\n std::string outstring;\n bool in_timestamp = false;\n\n // step 1: concatenate token ids to a string with special unicode characters\n std::string text_buffer;\n for (int token_id : tokens)\n {\n if (token_id < token_endoftext)\n {\n text_buffer += reverse_vocab[token_id];\n continue;\n }\n\n // handle timestamp tokens\n if (token_id >= token_timestamp_first && token_id <= token_timestamp_last)\n {\n int timestamp = (token_id - token_timestamp_first) * 2;\n\n char tmp[256];\n sprintf(tmp, \" [%d.%02d] \", timestamp / 100, timestamp % 100);\n\n if (in_timestamp)\n {\n // step 2: translate the special string back to original byte stream\n std::vector codepoints = utf8_to_codepoints(text_buffer);\n\n std::vector byte_sequence;\n for (uint32_t cp : codepoints)\n {\n byte_sequence.push_back(byte_decoder[cp]);\n }\n\n std::string s(byte_sequence.begin(), byte_sequence.end());\n\n text_buffer.clear();\n\n outstring += s;\n outstring += tmp;\n outstring += \"\\n\";\n\n in_timestamp = false;\n }\n else\n {\n outstring += tmp;\n in_timestamp = true;\n }\n }\n\n // ignore functional/special tokens\n }\n\n if (!text_buffer.empty())\n {\n // step 2: translate the special string back to original byte stream\n std::vector codepoints = utf8_to_codepoints(text_buffer);\n\n std::vector byte_sequence;\n for (uint32_t cp : codepoints)\n {\n byte_sequence.push_back(byte_decoder[cp]);\n }\n\n std::string s(byte_sequence.begin(), byte_sequence.end());\n\n outstring += s;\n }\n\n return outstring;\n }\n};\n\n// result class for beam search\nclass Result\n{\npublic:\n std::vector ids;\n float score;\n\n std::vector kvcache;\n};\n\n// main whisper implementation class\nclass Whisper\n{\npublic:\n int load();\n\n int detect_lang(const std::vector& samples, std::string& lang) const;\n int transcribe(const std::vector& samples, const char* lang, std::string& text) const;\n\nprotected:\n int extract_fbank_feature(const std::vector& samples, ncnn::Mat& input_features) const;\n int run_encoder(const ncnn::Mat& input_features, ncnn::Mat& encoder_states) const;\n int run_decoder_prefill(const std::vector& tokens, const ncnn::Mat& encoder_states, ncnn::Mat& last_logits, std::vector& out_kvcache) const;\n int run_decoder_step(const std::vector& tokens, const ncnn::Mat& encoder_states, ncnn::Mat& last_logits, const std::vector& kvcache, std::vector& out_kvcache) const;\n\nprotected:\n ncnn::Net fbank;\n\n ncnn::Net encoder;\n\n ncnn::Net embed_token;\n ncnn::Net embed_position;\n ncnn::Net decoder;\n\n ncnn::Net proj_out;\n\n Tokenizer tokenizer;\n\nprotected:\n std::vector kv_cache_indexes;\n std::vector out_kv_cache_indexes;\n};\n\nint Whisper::load()\n{\n // whisper models could be found at\n // https://github.com/nihui/ncnn-android-whisper/releases\n // https://github.com/nihui/ncnn-android-whisper/tree/master/app/src/main/assets\n\n fbank.opt.use_vulkan_compute = true;\n fbank.opt.use_fp16_packed = false;\n fbank.opt.use_fp16_storage = false;\n fbank.opt.use_fp16_arithmetic = false;\n\n encoder.opt.use_vulkan_compute = true;\n encoder.opt.use_fp16_packed = false;\n encoder.opt.use_fp16_storage = false;\n encoder.opt.use_fp16_arithmetic = false;\n\n decoder.opt.use_vulkan_compute = true;\n decoder.opt.use_fp16_packed = false;\n decoder.opt.use_fp16_storage = false;\n decoder.opt.use_fp16_arithmetic = false;\n\n proj_out.opt.use_vulkan_compute = true;\n proj_out.opt.use_fp16_packed = false;\n proj_out.opt.use_fp16_storage = false;\n proj_out.opt.use_fp16_arithmetic = false;\n\n fbank.load_param(\"whisper_tiny_fbank.ncnn.param\");\n fbank.load_model(\"whisper_tiny_fbank.ncnn.bin\");\n\n encoder.load_param(\"whisper_tiny_encoder.ncnn.param\");\n encoder.load_model(\"whisper_tiny_encoder.ncnn.bin\");\n\n embed_token.load_param(\"whisper_tiny_embed_token.ncnn.param\");\n embed_token.load_model(\"whisper_tiny_embed_token.ncnn.bin\");\n\n embed_position.load_param(\"whisper_tiny_embed_position.ncnn.param\");\n embed_position.load_model(\"whisper_tiny_embed_position.ncnn.bin\");\n\n decoder.load_param(\"whisper_tiny_decoder.ncnn.param\");\n decoder.load_model(\"whisper_tiny_decoder.ncnn.bin\");\n\n proj_out.load_param(\"whisper_tiny_proj_out.ncnn.param\");\n proj_out.load_model(\"whisper_tiny_proj_out.ncnn.bin\");\n\n // fbank.load_param(\"whisper_large_v3_turbo_fbank.ncnn.param\");\n // fbank.load_model(\"whisper_large_v3_turbo_fbank.ncnn.bin\");\n //\n // encoder.load_param(\"whisper_large_v3_turbo_encoder.ncnn.param\");\n // encoder.load_model(\"whisper_large_v3_turbo_encoder.ncnn.bin\");\n //\n // embed_token.load_param(\"whisper_large_v3_turbo_embed_token.ncnn.param\");\n // embed_token.load_model(\"whisper_large_v3_turbo_embed_token.ncnn.bin\");\n //\n // embed_position.load_param(\"whisper_large_v3_turbo_embed_position.ncnn.param\");\n // embed_position.load_model(\"whisper_large_v3_turbo_embed_position.ncnn.bin\");\n //\n // decoder.load_param(\"whisper_large_v3_turbo_decoder.ncnn.param\");\n // decoder.load_model(\"whisper_large_v3_turbo_decoder.ncnn.bin\");\n //\n // proj_out.load_param(\"whisper_large_v3_turbo_proj_out.ncnn.param\");\n // proj_out.load_model(\"whisper_large_v3_turbo_proj_out.ncnn.bin\");\n\n tokenizer.load(\"whisper_vocab.txt\");\n\n // resolve kv cache blob indexes\n for (size_t i = 0; i < decoder.layers().size(); i++)\n {\n const ncnn::Layer* mha = decoder.layers()[i];\n if (mha->typeindex != ncnn::LayerType::MultiHeadAttention)\n continue;\n\n const size_t input_count = mha->bottoms.size();\n const size_t output_count = mha->tops.size();\n\n if (output_count == 3)\n {\n kv_cache_indexes.push_back(mha->bottoms[input_count - 2]);\n kv_cache_indexes.push_back(mha->bottoms[input_count - 1]);\n out_kv_cache_indexes.push_back(mha->tops[output_count - 2]);\n out_kv_cache_indexes.push_back(mha->tops[output_count - 1]);\n }\n }\n\n return 0;\n}\n\n// apply log_softmax in-place\nstatic void log_softmax_inplace(ncnn::Mat& m)\n{\n ncnn::Option opt;\n opt.use_packing_layout = false;\n opt.use_fp16_storage = false;\n\n {\n ncnn::Layer* softmax = ncnn::create_layer_cpu(\"Softmax\");\n ncnn::ParamDict pd;\n pd.set(0, 0); // axis\n softmax->load_param(pd);\n softmax->forward_inplace(m, opt);\n delete softmax;\n }\n\n {\n ncnn::Layer* log = ncnn::create_layer_cpu(\"UnaryOp\");\n ncnn::ParamDict pd;\n pd.set(0, 8); // log\n log->load_param(pd);\n log->forward_inplace(m, opt);\n delete log;\n }\n}\n\nint Whisper::detect_lang(const std::vector& samples, std::string& lang) const\n{\n std::vector ids(1);\n ids[0] = token_startoftranscript;\n\n ncnn::Mat input_features;\n extract_fbank_feature(samples, input_features);\n\n ncnn::Mat encoder_states;\n run_encoder(input_features, encoder_states);\n\n ncnn::Mat logits;\n std::vector out_kvcache;\n run_decoder_prefill(ids, encoder_states, logits, out_kvcache);\n\n // find the lang token with highest prob\n // we are only interested in lang part and no_speech\n int lang_id = token_lang_first;\n float max_prob = logits[token_lang_first];\n for (int i = token_lang_first; i <= token_lang_last; i++)\n {\n float prob = logits[i];\n if (prob > max_prob)\n {\n max_prob = prob;\n lang_id = i;\n }\n }\n\n lang = token_langs[lang_id - token_lang_first];\n\n return 0;\n}\n\nint Whisper::transcribe(const std::vector& samples, const char* lang, std::string& text) const\n{\n // find lang token id by lang string\n int token_lang = -1;\n for (int i = 0; i < token_lang_count; i++)\n {\n if (strcmp(token_langs[i], lang) == 0)\n {\n token_lang = token_lang_first + i;\n break;\n }\n }\n\n if (token_lang == -1)\n {\n fprintf(stderr, \"language %s not supported\\n\", lang);\n return -1;\n }\n\n // initialize with prompt tokens\n std::vector ids(4);\n ids[0] = token_startoftranscript;\n ids[1] = token_lang;\n ids[2] = token_transcribe;\n ids[3] = token_notimestamps;\n\n ncnn::Mat input_features;\n extract_fbank_feature(samples, input_features);\n\n ncnn::Mat encoder_states;\n run_encoder(input_features, encoder_states);\n\n const int beam_size = 5;\n const int max_candidates = 5;\n\n std::vector finished_beams;\n\n std::vector beams(1);\n beams[0].ids = ids;\n beams[0].score = 0.f;\n\n int step = 0;\n\n // beam search loop\n for (;;)\n {\n std::vector candidates;\n\n for (size_t i = 0; i < beams.size(); i++)\n {\n const Result& beam = beams[i];\n\n ncnn::Mat logits;\n std::vector out_kvcache;\n if (step == 0)\n {\n run_decoder_prefill(beam.ids, encoder_states, logits, out_kvcache);\n }\n else\n {\n run_decoder_step(beam.ids, encoder_states, logits, beam.kvcache, out_kvcache);\n }\n\n log_softmax_inplace(logits);\n\n // get topk candidates\n const int topk = 5;\n std::vector > vec(logits.w);\n for (int j = 0; j < logits.w; j++)\n {\n vec[j] = std::make_pair(logits[j], j);\n }\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(), std::greater >());\n\n for (int j = 0; j < topk; j++)\n {\n int next_id = vec[j].second;\n float next_id_score = vec[j].first;\n\n Result candidate;\n candidate.ids = beam.ids;\n candidate.ids.push_back(next_id);\n candidate.score = beam.score + next_id_score;\n candidate.kvcache = out_kvcache;\n\n candidates.push_back(candidate);\n }\n }\n\n // sort candidates by score\n std::sort(candidates.begin(), candidates.end(), [](const Result& a, const Result& b) {\n return a.score > b.score;\n });\n\n beams.clear();\n for (size_t i = 0; i < candidates.size(); i++)\n {\n const Result& candidate = candidates[i];\n\n if (candidate.ids.back() == token_endoftext)\n {\n finished_beams.push_back(candidate);\n }\n else\n {\n beams.push_back(candidate);\n }\n }\n\n if (beams.size() > beam_size)\n {\n beams.resize(beam_size);\n }\n\n step++;\n\n if (beams.empty())\n {\n break;\n }\n\n if (finished_beams.size() >= max_candidates)\n {\n break;\n }\n }\n\n if (finished_beams.empty())\n {\n // no results\n return 0;\n }\n\n // find the best result based on average score\n int max_avg_score_index = 0;\n float max_avg_score = -FLT_MAX;\n for (size_t i = 0; i < finished_beams.size(); i++)\n {\n const Result& result = finished_beams[i];\n float avg_score = result.score / result.ids.size();\n if (avg_score > max_avg_score)\n {\n max_avg_score_index = (int)i;\n max_avg_score = avg_score;\n }\n }\n\n const Result& best_result = finished_beams[max_avg_score_index];\n\n text = tokenizer.decode(best_result.ids);\n\n return 0;\n}\n\nint Whisper::extract_fbank_feature(const std::vector& samples, ncnn::Mat& input_features) const\n{\n const int samples_size = (int)samples.size();\n\n // pad to 480000, normalize samples to -1~1\n ncnn::Mat waveform(480000);\n waveform.fill(0.f);\n {\n for (int i = 0; i < samples_size; i++)\n {\n waveform[i] = samples[i] / 32768.0f;\n }\n }\n\n ncnn::Extractor ex = fbank.create_extractor();\n\n ex.input(\"in0\", waveform);\n\n ex.extract(\"out0\", input_features);\n\n // drop the last frame\n {\n ncnn::Mat input_features_3k(input_features.w - 1, input_features.h);\n for (int i = 0; i < input_features.h; i++)\n {\n memcpy(input_features_3k.row(i), input_features.row(i), (input_features.w - 1) * sizeof(float));\n }\n input_features = input_features_3k;\n }\n\n return 0;\n}\n\nint Whisper::run_encoder(const ncnn::Mat& input_features, ncnn::Mat& encoder_states) const\n{\n ncnn::Extractor ex = encoder.create_extractor();\n\n ex.input(\"in0\", input_features);\n\n ex.extract(\"out0\", encoder_states);\n\n return 0;\n}\n\nint Whisper::run_decoder_prefill(const std::vector& tokens, const ncnn::Mat& encoder_states, ncnn::Mat& last_logits, std::vector& out_kvcache) const\n{\n const int dst_seqlen = tokens.size();\n\n // token embedding\n ncnn::Mat token_embeds;\n {\n ncnn::Mat input_tokens(dst_seqlen);\n int* p = input_tokens;\n memcpy(p, tokens.data(), tokens.size() * sizeof(int));\n\n ncnn::Extractor ex = embed_token.create_extractor();\n ex.input(\"in0\", input_tokens);\n ex.extract(\"out0\", token_embeds);\n }\n\n // position embedding\n ncnn::Mat position_embeds;\n {\n ncnn::Mat input_positions(dst_seqlen);\n int* p = input_positions;\n for (int i = 0; i < dst_seqlen; i++)\n {\n p[i] = i;\n }\n\n ncnn::Extractor ex = embed_position.create_extractor();\n ex.input(\"in0\", input_positions);\n ex.extract(\"out0\", position_embeds);\n }\n\n // input embedding = token + position\n ncnn::Mat input_embeds;\n {\n input_embeds.create_like(token_embeds);\n for (int i = 0; i < input_embeds.total(); i++)\n {\n input_embeds[i] = token_embeds[i] + position_embeds[i];\n }\n }\n\n // create attention mask (causal mask)\n ncnn::Mat attention_mask(dst_seqlen, dst_seqlen);\n attention_mask.fill(0.f);\n for (int i = 0; i < dst_seqlen; i++)\n {\n for (int j = i + 1; j < dst_seqlen; j++)\n {\n attention_mask.row(i)[j] = -INFINITY;\n }\n }\n\n ncnn::Mat output_states;\n {\n ncnn::Extractor ex = decoder.create_extractor();\n ex.input(\"in0\", input_embeds);\n ex.input(\"in1\", encoder_states);\n ex.input(\"in2\", attention_mask);\n\n out_kvcache.resize(out_kv_cache_indexes.size());\n for (size_t i = 0; i < out_kv_cache_indexes.size(); i++)\n {\n ex.extract(out_kv_cache_indexes[i], out_kvcache[i], 1);\n }\n\n ex.extract(\"out0\", output_states);\n }\n\n // get last token's state for next token prediction\n ncnn::Mat last_state = output_states.row_range(dst_seqlen - 1, 1).clone();\n {\n ncnn::Extractor ex = proj_out.create_extractor();\n ex.input(\"in0\", last_state);\n ex.extract(\"out0\", last_logits);\n }\n\n last_logits = last_logits.reshape(last_logits.w);\n\n return 0;\n}\n\nint Whisper::run_decoder_step(const std::vector& tokens, const ncnn::Mat& encoder_states, ncnn::Mat& last_logits, const std::vector& kvcache, std::vector& out_kvcache) const\n{\n const int token_id = tokens.back();\n const int dst_seqlen = 1;\n\n // token embedding\n ncnn::Mat token_embeds;\n {\n ncnn::Mat input_tokens(dst_seqlen);\n ((int*)input_tokens)[0] = token_id;\n\n ncnn::Extractor ex = embed_token.create_extractor();\n ex.input(\"in0\", input_tokens);\n ex.extract(\"out0\", token_embeds);\n }\n\n // position embedding\n ncnn::Mat position_embeds;\n {\n ncnn::Mat input_positions(dst_seqlen);\n ((int*)input_positions)[0] = tokens.size() - 1;\n\n ncnn::Extractor ex = embed_position.create_extractor();\n ex.input(\"in0\", input_positions);\n ex.extract(\"out0\", position_embeds);\n }\n\n // input embedding = token + position\n ncnn::Mat input_embeds;\n {\n input_embeds.create_like(token_embeds);\n for (int i = 0; i < input_embeds.total(); i++)\n {\n input_embeds[i] = token_embeds[i] + position_embeds[i];\n }\n }\n\n // single token doesn't need attention mask\n ncnn::Mat attention_mask(dst_seqlen, dst_seqlen);\n attention_mask.fill(0.f);\n\n ncnn::Mat output_states;\n {\n ncnn::Extractor ex = decoder.create_extractor();\n ex.input(\"in0\", input_embeds);\n ex.input(\"in1\", encoder_states);\n ex.input(\"in2\", attention_mask);\n\n // pass in kv cache from previous steps\n for (size_t i = 0; i < kv_cache_indexes.size(); i++)\n {\n ex.input(kv_cache_indexes[i], kvcache[i]);\n }\n\n // extract updated kv cache\n out_kvcache.resize(out_kv_cache_indexes.size());\n for (size_t i = 0; i < out_kv_cache_indexes.size(); i++)\n {\n ex.extract(out_kv_cache_indexes[i], out_kvcache[i], 1);\n }\n\n ex.extract(\"out0\", output_states);\n }\n\n // get last token's state for prediction\n ncnn::Mat last_state = output_states.row_range(dst_seqlen - 1, 1).clone();\n {\n ncnn::Extractor ex = proj_out.create_extractor();\n ex.input(\"in0\", last_state);\n ex.extract(\"out0\", last_logits);\n }\n\n last_logits = last_logits.reshape(last_logits.w);\n\n return 0;\n}\n\nstatic int load_wav_samples(const char* wavpath, std::vector& samples)\n{\n FILE* fp = fopen(wavpath, \"rb\");\n if (!fp)\n {\n fprintf(stderr, \"open %s failed\\n\", wavpath);\n return -1;\n }\n\n// https://stackoverflow.com/questions/1537964/visual-c-equivalent-of-gccs-attribute-packed\n#ifdef _MSC_VER\n#define PACK(__Declaration__) __pragma(pack(push, 1)) __Declaration__ __pragma(pack(pop))\n#else\n#define PACK(__Declaration__) __Declaration__ __attribute__((__packed__))\n#endif\n\n PACK(struct wav_header {\n char riff[4];\n uint32_t chunk_size;\n char wave[4];\n char fmt[4];\n uint32_t subchunk1_size;\n uint16_t audio_format;\n uint16_t num_channels;\n uint32_t sample_rate;\n uint32_t byte_rate;\n uint16_t block_align;\n uint16_t bits_per_sample;\n char data[4];\n uint32_t data_size;\n });\n\n wav_header header;\n if (fread(&header, sizeof(wav_header), 1, fp) != 1)\n {\n fprintf(stderr, \"failed to read wav header from %s\\n\", wavpath);\n fclose(fp);\n return -1;\n }\n\n if (memcmp(header.riff, \"RIFF\", 4) != 0 || memcmp(header.wave, \"WAVE\", 4) != 0\n || memcmp(header.fmt, \"fmt \", 4) != 0 || memcmp(header.data, \"data\", 4) != 0)\n {\n fprintf(stderr, \"%s is not a valid wav file\\n\", wavpath);\n fclose(fp);\n return -1;\n }\n\n if (header.subchunk1_size != 16 || header.audio_format != 1 || header.num_channels != 1\n || header.sample_rate != 16000 || header.bits_per_sample != 16)\n {\n fprintf(stderr, \"%s is not pcm s16le 16k wav\\n\", wavpath);\n fprintf(stderr, \"ffmpeg -i input.xxx -vn -c:a pcm_s16le -ac 1 -ar 16000 -fflags bitexact output.wav\\n\");\n fclose(fp);\n return -1;\n }\n\n fseek(fp, 0, SEEK_END);\n long len = ftell(fp);\n\n samples.resize((len - sizeof(wav_header)) / sizeof(short));\n\n rewind(fp);\n\n fseek(fp, sizeof(wav_header), SEEK_SET);\n\n fread(samples.data(), 1, len - sizeof(wav_header), fp);\n\n fclose(fp);\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [wavpath]\\n\", argv[0]);\n return -1;\n }\n\n const char* wavpath = argv[1];\n\n std::vector samples;\n int ret = load_wav_samples(wavpath, samples);\n if (ret != 0)\n {\n fprintf(stderr, \"load wav failed\\n\");\n return -1;\n }\n\n if (samples.size() > 480000)\n {\n fprintf(stderr, \"audio duration too long, truncate to 30s\\n\");\n samples.resize(480000);\n }\n\n Whisper whisper;\n whisper.load();\n\n // detect language first\n std::string lang;\n whisper.detect_lang(samples, lang);\n fprintf(stderr, \"lang = %s\\n\", lang.c_str());\n\n // transcribe audio to text\n std::string text;\n whisper.transcribe(samples, lang.c_str(), text);\n fprintf(stderr, \"text = %s\\n\", text.c_str());\n\n return 0;\n}\n" }, { "path": "examples/yolact.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n std::vector maskdata;\n cv::Mat mask;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic int detect_yolact(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolact;\n\n yolact.opt.use_vulkan_compute = true;\n\n // original model converted from https://github.com/dbolya/yolact\n // yolact_resnet50_54_800000.pth\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (yolact.load_param(\"yolact.param\"))\n exit(-1);\n if (yolact.load_model(\"yolact.bin\"))\n exit(-1);\n\n const int target_size = 550;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, target_size, target_size);\n\n const float mean_vals[3] = {123.68f, 116.78f, 103.94f};\n const float norm_vals[3] = {1.0 / 58.40f, 1.0 / 57.12f, 1.0 / 57.38f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = yolact.create_extractor();\n\n ex.input(\"input.1\", in);\n\n ncnn::Mat maskmaps;\n ncnn::Mat location;\n ncnn::Mat mask;\n ncnn::Mat confidence;\n\n ex.extract(\"619\", maskmaps); // 138x138 x 32\n\n ex.extract(\"816\", location); // 4 x 19248\n ex.extract(\"818\", mask); // maskdim 32 x 19248\n ex.extract(\"820\", confidence); // 81 x 19248\n\n int num_class = confidence.w;\n int num_priors = confidence.h;\n\n // make priorbox\n ncnn::Mat priorbox(4, num_priors);\n {\n const int conv_ws[5] = {69, 35, 18, 9, 5};\n const int conv_hs[5] = {69, 35, 18, 9, 5};\n\n const float aspect_ratios[3] = {1.f, 0.5f, 2.f};\n const float scales[5] = {24.f, 48.f, 96.f, 192.f, 384.f};\n\n float* pb = priorbox;\n\n for (int p = 0; p < 5; p++)\n {\n int conv_w = conv_ws[p];\n int conv_h = conv_hs[p];\n\n float scale = scales[p];\n\n for (int i = 0; i < conv_h; i++)\n {\n for (int j = 0; j < conv_w; j++)\n {\n // +0.5 because priors are in center-size notation\n float cx = (j + 0.5f) / conv_w;\n float cy = (i + 0.5f) / conv_h;\n\n for (int k = 0; k < 3; k++)\n {\n float ar = aspect_ratios[k];\n\n ar = sqrt(ar);\n\n float w = scale * ar / 550;\n float h = scale / ar / 550;\n\n // This is for backward compatibility with a bug where I made everything square by accident\n // cfg.backbone.use_square_anchors:\n h = w;\n\n pb[0] = cx;\n pb[1] = cy;\n pb[2] = w;\n pb[3] = h;\n\n pb += 4;\n }\n }\n }\n }\n }\n\n const float confidence_thresh = 0.05f;\n const float nms_threshold = 0.5f;\n const int keep_top_k = 200;\n\n std::vector > class_candidates;\n class_candidates.resize(num_class);\n\n for (int i = 0; i < num_priors; i++)\n {\n const float* conf = confidence.row(i);\n const float* loc = location.row(i);\n const float* pb = priorbox.row(i);\n const float* maskdata = mask.row(i);\n\n // find class id with highest score\n // start from 1 to skip background\n int label = 0;\n float score = 0.f;\n for (int j = 1; j < num_class; j++)\n {\n float class_score = conf[j];\n if (class_score > score)\n {\n label = j;\n score = class_score;\n }\n }\n\n // ignore background or low score\n if (label == 0 || score <= confidence_thresh)\n continue;\n\n // CENTER_SIZE\n float var[4] = {0.1f, 0.1f, 0.2f, 0.2f};\n\n float pb_cx = pb[0];\n float pb_cy = pb[1];\n float pb_w = pb[2];\n float pb_h = pb[3];\n\n float bbox_cx = var[0] * loc[0] * pb_w + pb_cx;\n float bbox_cy = var[1] * loc[1] * pb_h + pb_cy;\n float bbox_w = (float)(exp(var[2] * loc[2]) * pb_w);\n float bbox_h = (float)(exp(var[3] * loc[3]) * pb_h);\n\n float obj_x1 = bbox_cx - bbox_w * 0.5f;\n float obj_y1 = bbox_cy - bbox_h * 0.5f;\n float obj_x2 = bbox_cx + bbox_w * 0.5f;\n float obj_y2 = bbox_cy + bbox_h * 0.5f;\n\n // clip\n obj_x1 = std::max(std::min(obj_x1 * bgr.cols, (float)(bgr.cols - 1)), 0.f);\n obj_y1 = std::max(std::min(obj_y1 * bgr.rows, (float)(bgr.rows - 1)), 0.f);\n obj_x2 = std::max(std::min(obj_x2 * bgr.cols, (float)(bgr.cols - 1)), 0.f);\n obj_y2 = std::max(std::min(obj_y2 * bgr.rows, (float)(bgr.rows - 1)), 0.f);\n\n // append object\n Object obj;\n obj.rect = cv::Rect_(obj_x1, obj_y1, obj_x2 - obj_x1 + 1, obj_y2 - obj_y1 + 1);\n obj.label = label;\n obj.prob = score;\n obj.maskdata = std::vector(maskdata, maskdata + mask.w);\n\n class_candidates[label].push_back(obj);\n }\n\n objects.clear();\n for (int i = 0; i < (int)class_candidates.size(); i++)\n {\n std::vector& candidates = class_candidates[i];\n\n qsort_descent_inplace(candidates);\n\n std::vector picked;\n nms_sorted_bboxes(candidates, picked, nms_threshold);\n\n for (int j = 0; j < (int)picked.size(); j++)\n {\n int z = picked[j];\n objects.push_back(candidates[z]);\n }\n }\n\n qsort_descent_inplace(objects);\n\n // keep_top_k\n if (keep_top_k < (int)objects.size())\n {\n objects.resize(keep_top_k);\n }\n\n // generate mask\n for (int i = 0; i < (int)objects.size(); i++)\n {\n Object& obj = objects[i];\n\n cv::Mat mask(maskmaps.h, maskmaps.w, CV_32FC1);\n {\n mask = cv::Scalar(0.f);\n\n for (int p = 0; p < maskmaps.c; p++)\n {\n const float* maskmap = maskmaps.channel(p);\n float coeff = obj.maskdata[p];\n float* mp = (float*)mask.data;\n\n // mask += m * coeff\n for (int j = 0; j < maskmaps.w * maskmaps.h; j++)\n {\n mp[j] += maskmap[j] * coeff;\n }\n }\n }\n\n cv::Mat mask2;\n cv::resize(mask, mask2, cv::Size(img_w, img_h));\n\n // crop obj box and binarize\n obj.mask = cv::Mat(img_h, img_w, CV_8UC1);\n {\n obj.mask = cv::Scalar(0);\n\n for (int y = 0; y < img_h; y++)\n {\n if (y < obj.rect.y || y > obj.rect.y + obj.rect.height)\n continue;\n\n const float* mp2 = mask2.ptr(y);\n uchar* bmp = obj.mask.ptr(y);\n\n for (int x = 0; x < img_w; x++)\n {\n if (x < obj.rect.x || x > obj.rect.x + obj.rect.width)\n continue;\n\n bmp[x] = mp2[x] > 0.5f ? 255 : 0;\n }\n }\n }\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\",\n \"train\", \"truck\", \"boat\", \"traffic light\", \"fire hydrant\",\n \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\",\n \"horse\", \"sheep\", \"cow\", \"elephant\", \"bear\", \"zebra\", \"giraffe\",\n \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\",\n \"baseball glove\", \"skateboard\", \"surfboard\", \"tennis racket\",\n \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\",\n \"banana\", \"apple\", \"sandwich\", \"orange\", \"broccoli\", \"carrot\",\n \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\",\n \"mouse\", \"remote\", \"keyboard\", \"cell phone\", \"microwave\", \"oven\",\n \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\",\n \"scissors\", \"teddy bear\", \"hair drier\", \"toothbrush\"\n };\n\n static const unsigned char colors[81][3] = {\n {56, 0, 255},\n {226, 255, 0},\n {0, 94, 255},\n {0, 37, 255},\n {0, 255, 94},\n {255, 226, 0},\n {0, 18, 255},\n {255, 151, 0},\n {170, 0, 255},\n {0, 255, 56},\n {255, 0, 75},\n {0, 75, 255},\n {0, 255, 169},\n {255, 0, 207},\n {75, 255, 0},\n {207, 0, 255},\n {37, 0, 255},\n {0, 207, 255},\n {94, 0, 255},\n {0, 255, 113},\n {255, 18, 0},\n {255, 0, 56},\n {18, 0, 255},\n {0, 255, 226},\n {170, 255, 0},\n {255, 0, 245},\n {151, 255, 0},\n {132, 255, 0},\n {75, 0, 255},\n {151, 0, 255},\n {0, 151, 255},\n {132, 0, 255},\n {0, 255, 245},\n {255, 132, 0},\n {226, 0, 255},\n {255, 37, 0},\n {207, 255, 0},\n {0, 255, 207},\n {94, 255, 0},\n {0, 226, 255},\n {56, 255, 0},\n {255, 94, 0},\n {255, 113, 0},\n {0, 132, 255},\n {255, 0, 132},\n {255, 170, 0},\n {255, 0, 188},\n {113, 255, 0},\n {245, 0, 255},\n {113, 0, 255},\n {255, 188, 0},\n {0, 113, 255},\n {255, 0, 0},\n {0, 56, 255},\n {255, 0, 113},\n {0, 255, 188},\n {255, 0, 94},\n {255, 0, 18},\n {18, 255, 0},\n {0, 255, 132},\n {0, 188, 255},\n {0, 245, 255},\n {0, 169, 255},\n {37, 255, 0},\n {255, 0, 151},\n {188, 0, 255},\n {0, 255, 37},\n {0, 255, 0},\n {255, 0, 170},\n {255, 0, 37},\n {255, 75, 0},\n {0, 0, 255},\n {255, 207, 0},\n {255, 0, 226},\n {255, 245, 0},\n {188, 255, 0},\n {0, 255, 18},\n {0, 255, 75},\n {0, 255, 151},\n {255, 56, 0},\n {245, 255, 0}\n };\n\n cv::Mat image = bgr.clone();\n\n int color_index = 0;\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n if (obj.prob < 0.15)\n continue;\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n const unsigned char* color = colors[color_index % 81];\n color_index++;\n\n cv::rectangle(image, obj.rect, cv::Scalar(color[0], color[1], color[2]));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n\n // draw mask\n for (int y = 0; y < image.rows; y++)\n {\n const uchar* mp = obj.mask.ptr(y);\n uchar* p = image.ptr(y);\n for (int x = 0; x < image.cols; x++)\n {\n if (mp[x] == 255)\n {\n p[0] = cv::saturate_cast(p[0] * 0.5 + color[0] * 0.5);\n p[1] = cv::saturate_cast(p[1] * 0.5 + color[1] * 0.5);\n p[2] = cv::saturate_cast(p[2] * 0.5 + color[2] * 0.5);\n }\n p += 3;\n }\n }\n }\n\n cv::imwrite(\"result.png\", image);\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolact(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolo11.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolo11 torchscript\n// yolo export model=yolo11n.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolo11n.torchscript\n// 4. modify yolo11n_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_235 = v_204.view(1, 144, 6400)\n// v_236 = v_219.view(1, 144, 1600)\n// v_237 = v_234.view(1, 144, 400)\n// v_238 = torch.cat((v_235, v_236, v_237), dim=2)\n// ...\n// after:\n// v_235 = v_204.view(1, 144, -1).transpose(1, 2)\n// v_236 = v_219.view(1, 144, -1).transpose(1, 2)\n// v_237 = v_234.view(1, 144, -1).transpose(1, 2)\n// v_238 = torch.cat((v_235, v_236, v_237), dim=1)\n// return v_238\n// D. modify area attention for dynamic shape inference\n// before:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, 400)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, 20, 20)\n// v_107 = v_99.reshape(1, 128, 20, 20)\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// after:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, -1)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, v_95.size(2), v_95.size(3))\n// v_107 = v_99.reshape(1, 128, v_95.size(2), v_95.size(3))\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// 5. re-export yolo11 torchscript\n// python3 -c 'import yolo11n_pnnx; yolo11n_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolo11n_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolo11n_pnnx.py.ncnn.param yolo11n.ncnn.param\n// mv yolo11n_pnnx.py.ncnn.bin yolo11n.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=144 h=8400\n//\n// | bbox-reg 16 x 4 | per-class scores(80) |\n// +-----+-----+-----+-----+----------------------+\n// | dx0 | dy0 | dx1 | dy1 |0.1 0.0 0.0 0.5 ......|\n// all /| | | | | . |\n// boxes | .. | .. | .. | .. |0.0 0.9 0.0 0.0 ......|\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+----------------------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 80 for COCO\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4);\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolo11(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolo11;\n\n yolo11.opt.use_vulkan_compute = true;\n // yolo11.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolo11/tree/master/app/src/main/assets\n yolo11.load_param(\"yolo11n.ncnn.param\");\n yolo11.load_model(\"yolo11n.ncnn.bin\");\n // yolo11.load_param(\"yolo11s.ncnn.param\");\n // yolo11.load_model(\"yolo11s.ncnn.bin\");\n // yolo11.load_param(\"yolo11m.ncnn.param\");\n // yolo11.load_model(\"yolo11m.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolo11.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolo11.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n std::vector proposals;\n generate_proposals(out, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolo11(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolo11_cls.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolo11-cls torchscript\n// yolo export model=yolo11n-cls.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolo11n-cls.torchscript\n// 4. now you get ncnn model files\n// yolo11n_cls.ncnn.param\n// yolo11n_cls.ncnn.bin\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n int label;\n float prob;\n};\n\nstatic void get_topk(const ncnn::Mat& cls_scores, int topk, std::vector& objects)\n{\n // partial sort topk with index\n int size = cls_scores.w;\n std::vector > vec;\n vec.resize(size);\n for (int i = 0; i < size; i++)\n {\n vec[i] = std::make_pair(cls_scores[i], i);\n }\n\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(),\n std::greater >());\n\n objects.resize(topk);\n for (int i = 0; i < topk; i++)\n {\n objects[i].label = vec[i].second;\n objects[i].prob = vec[i].first;\n }\n}\n\nstatic int detect_yolo11_cls(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolo11;\n\n yolo11.opt.use_vulkan_compute = true;\n // yolo11.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolo11/tree/master/app/src/main/assets\n yolo11.load_param(\"yolo11n_cls.ncnn.param\");\n yolo11.load_model(\"yolo11n_cls.ncnn.bin\");\n // yolo11.load_param(\"yolo11s_cls.ncnn.param\");\n // yolo11.load_model(\"yolo11s_cls.ncnn.bin\");\n // yolo11.load_param(\"yolo11m_cls.ncnn.param\");\n // yolo11.load_model(\"yolo11m_cls.ncnn.bin\");\n\n const int target_size = 224;\n const int topk = 5;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // letterbox pad\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = target_size - w;\n int hpad = target_size - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolo11.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n // return top-5\n get_topk(out, topk, objects);\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"tench\", \"goldfish\", \"great white shark\", \"tiger shark\", \"hammerhead\", \"electric ray\", \"stingray\", \"cock\",\n \"hen\", \"ostrich\", \"brambling\", \"goldfinch\", \"house finch\", \"junco\", \"indigo bunting\", \"robin\", \"bulbul\",\n \"jay\", \"magpie\", \"chickadee\", \"water ouzel\", \"kite\", \"bald eagle\", \"vulture\", \"great grey owl\",\n \"European fire salamander\", \"common newt\", \"eft\", \"spotted salamander\", \"axolotl\", \"bullfrog\", \"tree frog\",\n \"tailed frog\", \"loggerhead\", \"leatherback turtle\", \"mud turtle\", \"terrapin\", \"box turtle\", \"banded gecko\",\n \"common iguana\", \"American chameleon\", \"whiptail\", \"agama\", \"frilled lizard\", \"alligator lizard\",\n \"Gila monster\", \"green lizard\", \"African chameleon\", \"Komodo dragon\", \"African crocodile\",\n \"American alligator\", \"triceratops\", \"thunder snake\", \"ringneck snake\", \"hognose snake\", \"green snake\",\n \"king snake\", \"garter snake\", \"water snake\", \"vine snake\", \"night snake\", \"boa constrictor\", \"rock python\",\n \"Indian cobra\", \"green mamba\", \"sea snake\", \"horned viper\", \"diamondback\", \"sidewinder\", \"trilobite\",\n \"harvestman\", \"scorpion\", \"black and gold garden spider\", \"barn spider\", \"garden spider\", \"black widow\",\n \"tarantula\", \"wolf spider\", \"tick\", \"centipede\", \"black grouse\", \"ptarmigan\", \"ruffed grouse\",\n \"prairie chicken\", \"peacock\", \"quail\", \"partridge\", \"African grey\", \"macaw\", \"sulphur-crested cockatoo\",\n \"lorikeet\", \"coucal\", \"bee eater\", \"hornbill\", \"hummingbird\", \"jacamar\", \"toucan\", \"drake\",\n \"red-breasted merganser\", \"goose\", \"black swan\", \"tusker\", \"echidna\", \"platypus\", \"wallaby\", \"koala\",\n \"wombat\", \"jellyfish\", \"sea anemone\", \"brain coral\", \"flatworm\", \"nematode\", \"conch\", \"snail\", \"slug\",\n \"sea slug\", \"chiton\", \"chambered nautilus\", \"Dungeness crab\", \"rock crab\", \"fiddler crab\", \"king crab\",\n \"American lobster\", \"spiny lobster\", \"crayfish\", \"hermit crab\", \"isopod\", \"white stork\", \"black stork\",\n \"spoonbill\", \"flamingo\", \"little blue heron\", \"American egret\", \"bittern\", \"crane (bird)\", \"limpkin\",\n \"European gallinule\", \"American coot\", \"bustard\", \"ruddy turnstone\", \"red-backed sandpiper\", \"redshank\",\n \"dowitcher\", \"oystercatcher\", \"pelican\", \"king penguin\", \"albatross\", \"grey whale\", \"killer whale\",\n \"dugong\", \"sea lion\", \"Chihuahua\", \"Japanese spaniel\", \"Maltese dog\", \"Pekinese\", \"Shih-Tzu\",\n \"Blenheim spaniel\", \"papillon\", \"toy terrier\", \"Rhodesian ridgeback\", \"Afghan hound\", \"basset\", \"beagle\",\n \"bloodhound\", \"bluetick\", \"black-and-tan coonhound\", \"Walker hound\", \"English foxhound\", \"redbone\",\n \"borzoi\", \"Irish wolfhound\", \"Italian greyhound\", \"whippet\", \"Ibizan hound\", \"Norwegian elkhound\",\n \"otterhound\", \"Saluki\", \"Scottish deerhound\", \"Weimaraner\", \"Staffordshire bullterrier\",\n \"American Staffordshire terrier\", \"Bedlington terrier\", \"Border terrier\", \"Kerry blue terrier\",\n \"Irish terrier\", \"Norfolk terrier\", \"Norwich terrier\", \"Yorkshire terrier\", \"wire-haired fox terrier\",\n \"Lakeland terrier\", \"Sealyham terrier\", \"Airedale\", \"cairn\", \"Australian terrier\", \"Dandie Dinmont\",\n \"Boston bull\", \"miniature schnauzer\", \"giant schnauzer\", \"standard schnauzer\", \"Scotch terrier\",\n \"Tibetan terrier\", \"silky terrier\", \"soft-coated wheaten terrier\", \"West Highland white terrier\",\n \"Lhasa\", \"flat-coated retriever\", \"curly-coated retriever\", \"golden retriever\", \"Labrador retriever\",\n \"Chesapeake Bay retriever\", \"German short-haired pointer\", \"vizsla\", \"English setter\", \"Irish setter\",\n \"Gordon setter\", \"Brittany spaniel\", \"clumber\", \"English springer\", \"Welsh springer spaniel\",\n \"cocker spaniel\", \"Sussex spaniel\", \"Irish water spaniel\", \"kuvasz\", \"schipperke\", \"groenendael\",\n \"malinois\", \"briard\", \"kelpie\", \"komondor\", \"Old English sheepdog\", \"Shetland sheepdog\", \"collie\",\n \"Border collie\", \"Bouvier des Flandres\", \"Rottweiler\", \"German shepherd\", \"Doberman\",\n \"miniature pinscher\", \"Greater Swiss Mountain dog\", \"Bernese mountain dog\", \"Appenzeller\", \"EntleBucher\",\n \"boxer\", \"bull mastiff\", \"Tibetan mastiff\", \"French bulldog\", \"Great Dane\", \"Saint Bernard\",\n \"Eskimo dog\", \"malamute\", \"Siberian husky\", \"dalmatian\", \"affenpinscher\", \"basenji\", \"pug\", \"Leonberg\",\n \"Newfoundland\", \"Great Pyrenees\", \"Samoyed\", \"Pomeranian\", \"chow\", \"keeshond\", \"Brabancon griffon\",\n \"Pembroke\", \"Cardigan\", \"toy poodle\", \"miniature poodle\", \"standard poodle\", \"Mexican hairless\",\n \"timber wolf\", \"white wolf\", \"red wolf\", \"coyote\", \"dingo\", \"dhole\", \"African hunting dog\", \"hyena\",\n \"red fox\", \"kit fox\", \"Arctic fox\", \"grey fox\", \"tabby\", \"tiger cat\", \"Persian cat\", \"Siamese cat\",\n \"Egyptian cat\", \"cougar\", \"lynx\", \"leopard\", \"snow leopard\", \"jaguar\", \"lion\", \"tiger\", \"cheetah\",\n \"brown bear\", \"American black bear\", \"ice bear\", \"sloth bear\", \"mongoose\", \"meerkat\", \"tiger beetle\",\n \"ladybug\", \"ground beetle\", \"long-horned beetle\", \"leaf beetle\", \"dung beetle\", \"rhinoceros beetle\",\n \"weevil\", \"fly\", \"bee\", \"ant\", \"grasshopper\", \"cricket\", \"walking stick\", \"cockroach\", \"mantis\",\n \"cicada\", \"leafhopper\", \"lacewing\", \"dragonfly\", \"damselfly\", \"admiral\", \"ringlet\", \"monarch\",\n \"cabbage butterfly\", \"sulphur butterfly\", \"lycaenid\", \"starfish\", \"sea urchin\", \"sea cucumber\",\n \"wood rabbit\", \"hare\", \"Angora\", \"hamster\", \"porcupine\", \"fox squirrel\", \"marmot\", \"beaver\",\n \"guinea pig\", \"sorrel\", \"zebra\", \"hog\", \"wild boar\", \"warthog\", \"hippopotamus\", \"ox\", \"water buffalo\",\n \"bison\", \"ram\", \"bighorn\", \"ibex\", \"hartebeest\", \"impala\", \"gazelle\", \"Arabian camel\", \"llama\",\n \"weasel\", \"mink\", \"polecat\", \"black-footed ferret\", \"otter\", \"skunk\", \"badger\", \"armadillo\",\n \"three-toed sloth\", \"orangutan\", \"gorilla\", \"chimpanzee\", \"gibbon\", \"siamang\", \"guenon\", \"patas\",\n \"baboon\", \"macaque\", \"langur\", \"colobus\", \"proboscis monkey\", \"marmoset\", \"capuchin\", \"howler monkey\",\n \"titi\", \"spider monkey\", \"squirrel monkey\", \"Madagascar cat\", \"indri\", \"Indian elephant\",\n \"African elephant\", \"lesser panda\", \"giant panda\", \"barracouta\", \"eel\", \"coho\", \"rock beauty\",\n \"anemone fish\", \"sturgeon\", \"gar\", \"lionfish\", \"puffer\", \"abacus\", \"abaya\", \"academic gown\",\n \"accordion\", \"acoustic guitar\", \"aircraft carrier\", \"airliner\", \"airship\", \"altar\", \"ambulance\",\n \"amphibian\", \"analog clock\", \"apiary\", \"apron\", \"ashcan\", \"assault rifle\", \"backpack\", \"bakery\",\n \"balance beam\", \"balloon\", \"ballpoint\", \"Band Aid\", \"banjo\", \"bannister\", \"barbell\", \"barber chair\",\n \"barbershop\", \"barn\", \"barometer\", \"barrel\", \"barrow\", \"baseball\", \"basketball\", \"bassinet\", \"bassoon\",\n \"bathing cap\", \"bath towel\", \"bathtub\", \"beach wagon\", \"beacon\", \"beaker\", \"bearskin\", \"beer bottle\",\n \"beer glass\", \"bell cote\", \"bib\", \"bicycle-built-for-two\", \"bikini\", \"binder\", \"binoculars\",\n \"birdhouse\", \"boathouse\", \"bobsled\", \"bolo tie\", \"bonnet\", \"bookcase\", \"bookshop\", \"bottlecap\", \"bow\",\n \"bow tie\", \"brass\", \"brassiere\", \"breakwater\", \"breastplate\", \"broom\", \"bucket\", \"buckle\",\n \"bulletproof vest\", \"bullet train\", \"butcher shop\", \"cab\", \"caldron\", \"candle\", \"cannon\", \"canoe\",\n \"can opener\", \"cardigan\", \"car mirror\", \"carousel\", \"carpenter's kit\", \"carton\", \"car wheel\",\n \"cash machine\", \"cassette\", \"cassette player\", \"castle\", \"catamaran\", \"CD player\", \"cello\",\n \"cellular telephone\", \"chain\", \"chainlink fence\", \"chain mail\", \"chain saw\", \"chest\", \"chiffonier\",\n \"chime\", \"china cabinet\", \"Christmas stocking\", \"church\", \"cinema\", \"cleaver\", \"cliff dwelling\",\n \"cloak\", \"clog\", \"cocktail shaker\", \"coffee mug\", \"coffeepot\", \"coil\", \"combination lock\",\n \"computer keyboard\", \"confectionery\", \"container ship\", \"convertible\", \"corkscrew\", \"cornet\",\n \"cowboy boot\", \"cowboy hat\", \"cradle\", \"crane (machine)\", \"crash helmet\", \"crate\", \"crib\",\n \"Crock Pot\", \"croquet ball\", \"crutch\", \"cuirass\", \"dam\", \"desk\", \"desktop computer\", \"dial telephone\",\n \"diaper\", \"digital clock\", \"digital watch\", \"dining table\", \"dishrag\", \"dishwasher\", \"disk brake\",\n \"dock\", \"dogsled\", \"dome\", \"doormat\", \"drilling platform\", \"drum\", \"drumstick\", \"dumbbell\",\n \"Dutch oven\", \"electric fan\", \"electric guitar\", \"electric locomotive\", \"entertainment center\",\n \"envelope\", \"espresso maker\", \"face powder\", \"feather boa\", \"file\", \"fireboat\", \"fire engine\",\n \"fire screen\", \"flagpole\", \"flute\", \"folding chair\", \"football helmet\", \"forklift\", \"fountain\",\n \"fountain pen\", \"four-poster\", \"freight car\", \"French horn\", \"frying pan\", \"fur coat\", \"garbage truck\",\n \"gasmask\", \"gas pump\", \"goblet\", \"go-kart\", \"golf ball\", \"golfcart\", \"gondola\", \"gong\", \"gown\",\n \"grand piano\", \"greenhouse\", \"grille\", \"grocery store\", \"guillotine\", \"hair slide\", \"hair spray\",\n \"half track\", \"hammer\", \"hamper\", \"hand blower\", \"hand-held computer\", \"handkerchief\", \"hard disc\",\n \"harmonica\", \"harp\", \"harvester\", \"hatchet\", \"holster\", \"home theater\", \"honeycomb\", \"hook\",\n \"hoopskirt\", \"horizontal bar\", \"horse cart\", \"hourglass\", \"iPod\", \"iron\", \"jack-o'-lantern\", \"jean\",\n \"jeep\", \"jersey\", \"jigsaw puzzle\", \"jinrikisha\", \"joystick\", \"kimono\", \"knee pad\", \"knot\", \"lab coat\",\n \"ladle\", \"lampshade\", \"laptop\", \"lawn mower\", \"lens cap\", \"letter opener\", \"library\", \"lifeboat\",\n \"lighter\", \"limousine\", \"liner\", \"lipstick\", \"Loafer\", \"lotion\", \"loudspeaker\", \"loupe\", \"lumbermill\",\n \"magnetic compass\", \"mailbag\", \"mailbox\", \"maillot (tights)\", \"maillot (tank suit)\", \"manhole cover\",\n \"maraca\", \"marimba\", \"mask\", \"matchstick\", \"maypole\", \"maze\", \"measuring cup\", \"medicine chest\",\n \"megalith\", \"microphone\", \"microwave\", \"military uniform\", \"milk can\", \"minibus\", \"miniskirt\",\n \"minivan\", \"missile\", \"mitten\", \"mixing bowl\", \"mobile home\", \"Model T\", \"modem\", \"monastery\",\n \"monitor\", \"moped\", \"mortar\", \"mortarboard\", \"mosque\", \"mosquito net\", \"motor scooter\", \"mountain bike\",\n \"mountain tent\", \"mouse\", \"mousetrap\", \"moving van\", \"muzzle\", \"nail\", \"neck brace\", \"necklace\",\n \"nipple\", \"notebook\", \"obelisk\", \"oboe\", \"ocarina\", \"odometer\", \"oil filter\", \"organ\", \"oscilloscope\",\n \"overskirt\", \"oxcart\", \"oxygen mask\", \"packet\", \"paddle\", \"paddlewheel\", \"padlock\", \"paintbrush\",\n \"pajama\", \"palace\", \"panpipe\", \"paper towel\", \"parachute\", \"parallel bars\", \"park bench\",\n \"parking meter\", \"passenger car\", \"patio\", \"pay-phone\", \"pedestal\", \"pencil box\", \"pencil sharpener\",\n \"perfume\", \"Petri dish\", \"photocopier\", \"pick\", \"pickelhaube\", \"picket fence\", \"pickup\", \"pier\",\n \"piggy bank\", \"pill bottle\", \"pillow\", \"ping-pong ball\", \"pinwheel\", \"pirate\", \"pitcher\", \"plane\",\n \"planetarium\", \"plastic bag\", \"plate rack\", \"plow\", \"plunger\", \"Polaroid camera\", \"pole\",\n \"police van\", \"poncho\", \"pool table\", \"pop bottle\", \"pot\", \"potter's wheel\", \"power drill\",\n \"prayer rug\", \"printer\", \"prison\", \"projectile\", \"projector\", \"puck\", \"punching bag\", \"purse\",\n \"quill\", \"quilt\", \"racer\", \"racket\", \"radiator\", \"radio\", \"radio telescope\", \"rain barrel\",\n \"recreational vehicle\", \"reel\", \"reflex camera\", \"refrigerator\", \"remote control\", \"restaurant\",\n \"revolver\", \"rifle\", \"rocking chair\", \"rotisserie\", \"rubber eraser\", \"rugby ball\", \"rule\",\n \"running shoe\", \"safe\", \"safety pin\", \"saltshaker\", \"sandal\", \"sarong\", \"sax\", \"scabbard\", \"scale\",\n \"school bus\", \"schooner\", \"scoreboard\", \"screen\", \"screw\", \"screwdriver\", \"seat belt\", \"sewing machine\",\n \"shield\", \"shoe shop\", \"shoji\", \"shopping basket\", \"shopping cart\", \"shovel\", \"shower cap\",\n \"shower curtain\", \"ski\", \"ski mask\", \"sleeping bag\", \"slide rule\", \"sliding door\", \"slot\", \"snorkel\",\n \"snowmobile\", \"snowplow\", \"soap dispenser\", \"soccer ball\", \"sock\", \"solar dish\", \"sombrero\",\n \"soup bowl\", \"space bar\", \"space heater\", \"space shuttle\", \"spatula\", \"speedboat\", \"spider web\",\n \"spindle\", \"sports car\", \"spotlight\", \"stage\", \"steam locomotive\", \"steel arch bridge\", \"steel drum\",\n \"stethoscope\", \"stole\", \"stone wall\", \"stopwatch\", \"stove\", \"strainer\", \"streetcar\", \"stretcher\",\n \"studio couch\", \"stupa\", \"submarine\", \"suit\", \"sundial\", \"sunglass\", \"sunglasses\", \"sunscreen\",\n \"suspension bridge\", \"swab\", \"sweatshirt\", \"swimming trunks\", \"swing\", \"switch\", \"syringe\",\n \"table lamp\", \"tank\", \"tape player\", \"teapot\", \"teddy\", \"television\", \"tennis ball\", \"thatch\",\n \"theater curtain\", \"thimble\", \"thresher\", \"throne\", \"tile roof\", \"toaster\", \"tobacco shop\",\n \"toilet seat\", \"torch\", \"totem pole\", \"tow truck\", \"toyshop\", \"tractor\", \"trailer truck\", \"tray\",\n \"trench coat\", \"tricycle\", \"trimaran\", \"tripod\", \"triumphal arch\", \"trolleybus\", \"trombone\", \"tub\",\n \"turnstile\", \"typewriter keyboard\", \"umbrella\", \"unicycle\", \"upright\", \"vacuum\", \"vase\", \"vault\",\n \"velvet\", \"vending machine\", \"vestment\", \"viaduct\", \"violin\", \"volleyball\", \"waffle iron\", \"wall clock\",\n \"wallet\", \"wardrobe\", \"warplane\", \"washbasin\", \"washer\", \"water bottle\", \"water jug\", \"water tower\",\n \"whiskey jug\", \"whistle\", \"wig\", \"window screen\", \"window shade\", \"Windsor tie\", \"wine bottle\", \"wing\",\n \"wok\", \"wooden spoon\", \"wool\", \"worm fence\", \"wreck\", \"yawl\", \"yurt\", \"web site\", \"comic book\",\n \"crossword puzzle\", \"street sign\", \"traffic light\", \"book jacket\", \"menu\", \"plate\", \"guacamole\",\n \"consomme\", \"hot pot\", \"trifle\", \"ice cream\", \"ice lolly\", \"French loaf\", \"bagel\", \"pretzel\",\n \"cheeseburger\", \"hotdog\", \"mashed potato\", \"head cabbage\", \"broccoli\", \"cauliflower\", \"zucchini\",\n \"spaghetti squash\", \"acorn squash\", \"butternut squash\", \"cucumber\", \"artichoke\", \"bell pepper\",\n \"cardoon\", \"mushroom\", \"Granny Smith\", \"strawberry\", \"orange\", \"lemon\", \"fig\", \"pineapple\", \"banana\",\n \"jackfruit\", \"custard apple\", \"pomegranate\", \"hay\", \"carbonara\", \"chocolate sauce\", \"dough\",\n \"meat loaf\", \"pizza\", \"potpie\", \"burrito\", \"red wine\", \"espresso\", \"cup\", \"eggnog\", \"alp\", \"bubble\",\n \"cliff\", \"coral reef\", \"geyser\", \"lakeside\", \"promontory\", \"sandbar\", \"seashore\", \"valley\", \"volcano\",\n \"ballplayer\", \"groom\", \"scuba diver\", \"rapeseed\", \"daisy\", \"yellow lady's slipper\", \"corn\", \"acorn\",\n \"hip\", \"buckeye\", \"coral fungus\", \"agaric\", \"gyromitra\", \"stinkhorn\", \"earthstar\", \"hen-of-the-woods\",\n \"bolete\", \"ear\", \"toilet tissue\"\n };\n\n cv::Mat image = bgr.clone();\n\n int y_offset = 0;\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f\\n\", obj.label, obj.prob);\n\n char text[256];\n sprintf(text, \"%4.1f%% %s\", obj.prob * 100, class_names[obj.label]);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = 0;\n int y = y_offset;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n\n y_offset += label_size.height;\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolo11_cls(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolo11_obb.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolo11-obb torchscript\n// yolo export model=yolo11n-obb.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolo11n-obb.torchscript\n// 4. modify yolo11n_obb_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_195 = v_194.view(1, 1, 16384)\n// v_201 = v_200.view(1, 1, 4096)\n// v_207 = v_206.view(1, 1, 1024)\n// v_208 = torch.cat((v_195, v_201, v_207), dim=2)\n// ...\n// v_256 = v_225.view(1, 79, 16384)\n// v_257 = v_240.view(1, 79, 4096)\n// v_258 = v_255.view(1, 79, 1024)\n// v_259 = torch.cat((v_256, v_257, v_258), dim=2)\n// ...\n// after:\n// v_195 = v_194.view(1, 1, -1).transpose(1, 2)\n// v_201 = v_200.view(1, 1, -1).transpose(1, 2)\n// v_207 = v_206.view(1, 1, -1).transpose(1, 2)\n// v_208 = torch.cat((v_195, v_201, v_207), dim=1)\n// ...\n// v_256 = v_225.view(1, 79, -1).transpose(1, 2)\n// v_257 = v_240.view(1, 79, -1).transpose(1, 2)\n// v_258 = v_255.view(1, 79, -1).transpose(1, 2)\n// v_259 = torch.cat((v_256, v_257, v_258), dim=1)\n// return v_259, v_208\n// D. modify area attention for dynamic shape inference\n// before:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, 1024)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, 32, 32)\n// v_107 = v_99.reshape(1, 128, 32, 32)\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// after:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, -1)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, v_95.size(2), v_95.size(3))\n// v_107 = v_99.reshape(1, 128, v_95.size(2), v_95.size(3))\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// 5. re-export yolo11-obb torchscript\n// python3 -c 'import yolo11n_obb_pnnx; yolo11n_obb_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolo11n_obb_pnnx.py.pt inputshape=[1,3,1024,1024] inputshape2=[1,3,512,512]\n// 7. now you get ncnn model files\n// mv yolo11n_obb_pnnx.py.ncnn.param yolo11n_obb.ncnn.param\n// mv yolo11n_obb_pnnx.py.ncnn.bin yolo11n_obb.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=79 h=21504\n//\n// | bbox-reg 16 x 4 |score(15)|\n// +-----+-----+-----+-----+---------+\n// | dx0 | dy0 | dx1 | dy1 | 0.1 ... |\n// all /| | | | | ... |\n// boxes | .. | .. | .. | .. | 0.0 ... |\n// (21504)| | | | | . ... |\n// \\| | | | | . ... |\n// +-----+-----+-----+-----+---------+\n//\n\n// the out blob would be a 2-dim tensor with w=1 h=21504\n//\n// | degree(1)|\n// +----------+\n// | 0.1 |\n// all /| |\n// boxes | 0.0 |\n// (21504)| . |\n// \\| . |\n// +----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n\nstruct Object\n{\n cv::RotatedRect rrect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n std::vector intersection;\n cv::rotatedRectangleIntersection(a.rrect, b.rrect, intersection);\n if (intersection.empty())\n return 0.f;\n\n return cv::contourArea(intersection);\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rrect.size.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area;\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_angle, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 15 for DOTAv1\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n const float angle = sigmoid(pred_angle.row(y * num_grid_x + x)[0]) - 0.25f;\n\n const float angle_rad = angle * 3.14159265358979323846f;\n const float angle_degree = angle * 180.f;\n\n float cos = cosf(angle_rad);\n float sin = sinf(angle_rad);\n\n float xx = (pred_ltrb[2] - pred_ltrb[0]) * 0.5f;\n float yy = (pred_ltrb[3] - pred_ltrb[1]) * 0.5f;\n float xr = xx * cos - yy * sin;\n float yr = xx * sin + yy * cos;\n const float cx = pb_cx + xr;\n const float cy = pb_cy + yr;\n const float ww = pred_ltrb[2] + pred_ltrb[0];\n const float hh = pred_ltrb[3] + pred_ltrb[1];\n\n Object obj;\n obj.rrect = cv::RotatedRect(cv::Point2f(cx, cy), cv::Size_(ww, hh), angle_degree);\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_angle, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), pred_angle.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolo11_obb(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolo11;\n\n yolo11.opt.use_vulkan_compute = true;\n // yolo11.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolo11/tree/master/app/src/main/assets\n yolo11.load_param(\"yolo11n_obb.ncnn.param\");\n yolo11.load_model(\"yolo11n_obb.ncnn.bin\");\n // yolo11.load_param(\"yolo11s_obb.ncnn.param\");\n // yolo11.load_model(\"yolo11s_obb.ncnn.bin\");\n // yolo11.load_param(\"yolo11m_obb.ncnn.param\");\n // yolo11.load_model(\"yolo11m_obb.ncnn.bin\");\n\n const int target_size = 1024;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolo11.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolo11.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat out_angle;\n ex.extract(\"out1\", out_angle);\n\n std::vector proposals;\n generate_proposals(out, out_angle, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n Object obj = proposals[picked[i]];\n\n // adjust offset to original unpadded\n obj.rrect.center.x = (obj.rrect.center.x - (wpad / 2)) / scale;\n obj.rrect.center.y = (obj.rrect.center.y - (hpad / 2)) / scale;\n obj.rrect.size.width = (obj.rrect.size.width) / scale;\n obj.rrect.size.height = (obj.rrect.size.height) / scale;\n\n objects[i] = obj;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"plane\", \"ship\", \"storage tank\", \"baseball diamond\", \"tennis court\",\n \"basketball court\", \"ground track field\", \"harbor\", \"bridge\", \"large vehicle\",\n \"small vehicle\", \"helicopter\", \"roundabout\", \"soccer ball field\", \"swimming pool\"\n };\n\n static const cv::Scalar colors[] = {\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[obj.label];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f @ %.2f\\n\", obj.label, obj.prob,\n obj.rrect.center.x, obj.rrect.center.y, obj.rrect.size.width, obj.rrect.size.height, obj.rrect.angle);\n\n cv::Point2f corners[4];\n obj.rrect.points(corners);\n cv::line(image, corners[0], corners[1], color);\n cv::line(image, corners[1], corners[2], color);\n cv::line(image, corners[2], corners[3], color);\n cv::line(image, corners[3], corners[0], color);\n }\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[obj.label];\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rrect.center.x - label_size.width / 2;\n int y = obj.rrect.center.y - label_size.height / 2 - baseLine;\n if (y < 0)\n y = 0;\n if (y + label_size.height > image.rows)\n y = image.rows - label_size.height;\n if (x < 0)\n x = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolo11_obb(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolo11_pose.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolo11-pose torchscript\n// yolo export model=yolo11n-pose.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolo11n-pose.torchscript\n// 4. modify yolo11n_pose_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_195 = v_194.view(1, 51, 6400)\n// v_201 = v_200.view(1, 51, 1600)\n// v_207 = v_206.view(1, 51, 400)\n// v_208 = torch.cat((v_195, v_201, v_207), dim=-1)\n// ...\n// v_254 = v_223.view(1, 65, 6400)\n// v_255 = v_238.view(1, 65, 1600)\n// v_256 = v_253.view(1, 65, 400)\n// v_257 = torch.cat((v_254, v_255, v_256), dim=2)\n// ...\n// after:\n// v_195 = v_194.view(1, 51, -1).transpose(1, 2)\n// v_201 = v_200.view(1, 51, -1).transpose(1, 2)\n// v_207 = v_206.view(1, 51, -1).transpose(1, 2)\n// v_208 = torch.cat((v_195, v_201, v_207), dim=1)\n// ...\n// v_254 = v_223.view(1, 65, -1).transpose(1, 2)\n// v_255 = v_238.view(1, 65, -1).transpose(1, 2)\n// v_256 = v_253.view(1, 65, -1).transpose(1, 2)\n// v_257 = torch.cat((v_254, v_255, v_256), dim=1)\n// return v_257, v_208\n// D. modify area attention for dynamic shape inference\n// before:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, 400)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, 20, 20)\n// v_107 = v_99.reshape(1, 128, 20, 20)\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// after:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, -1)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, v_95.size(2), v_95.size(3))\n// v_107 = v_99.reshape(1, 128, v_95.size(2), v_95.size(3))\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// 5. re-export yolo11-pose torchscript\n// python3 -c 'import yolo11n_pose_pnnx; yolo11n_pose_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolo11n_pose_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolo11n_pose_pnnx.py.ncnn.param yolo11n_pose.ncnn.param\n// mv yolo11n_pose_pnnx.py.ncnn.bin yolo11n_pose.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=65 h=8400\n//\n// | bbox-reg 16 x 4 |score(1)|\n// +-----+-----+-----+-----+--------+\n// | dx0 | dy0 | dx1 | dy1 | 0.1 |\n// all /| | | | | |\n// boxes | .. | .. | .. | .. | 0.0 |\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+--------+\n//\n\n//\n// | pose (51) |\n// +-----------+\n// |0.1........|\n// all /| |\n// boxes |0.0........|\n// (8400)| . |\n// \\| . |\n// +-----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct KeyPoint\n{\n cv::Point2f p;\n float prob;\n};\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n std::vector keypoints;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_points, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_points = pred_points.w / 3;\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n const ncnn::Mat pred_points_grid = pred_points.row_range(y * num_grid_x + x, 1).reshape(3, num_points);\n\n // find label with max score\n int label = 0;\n float score = sigmoid(pred_grid[reg_max_1 * 4]);\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n std::vector keypoints;\n for (int k = 0; k < num_points; k++)\n {\n KeyPoint keypoint;\n keypoint.p.x = (x + pred_points_grid.row(k)[0] * 2) * stride;\n keypoint.p.y = (y + pred_points_grid.row(k)[1] * 2) * stride;\n keypoint.prob = sigmoid(pred_points_grid.row(k)[2]);\n keypoints.push_back(keypoint);\n }\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n obj.keypoints = keypoints;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_points, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), pred_points.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolo11_pose(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolo11;\n\n yolo11.opt.use_vulkan_compute = true;\n // yolo11.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolo11/tree/master/app/src/main/assets\n yolo11.load_param(\"yolo11n_pose.ncnn.param\");\n yolo11.load_model(\"yolo11n_pose.ncnn.bin\");\n // yolo11.load_param(\"yolo11s_pose.ncnn.param\");\n // yolo11.load_model(\"yolo11s_pose.ncnn.bin\");\n // yolo11.load_param(\"yolo11m_pose.ncnn.param\");\n // yolo11.load_model(\"yolo11m_pose.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n const float mask_threshold = 0.5f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolo11.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolo11.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat out_points;\n ex.extract(\"out1\", out_points);\n\n std::vector proposals;\n generate_proposals(out, out_points, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n const int num_points = out_points.w / 3;\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n for (int j = 0; j < num_points; j++)\n {\n objects[i].keypoints[j].p.x = (objects[i].keypoints[j].p.x - (wpad / 2)) / scale;\n objects[i].keypoints[j].p.y = (objects[i].keypoints[j].p.y - (hpad / 2)) / scale;\n }\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"person\"};\n\n static const cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n // draw bone\n static const int joint_pairs[16][2] = {\n {0, 1}, {1, 3}, {0, 2}, {2, 4}, {5, 6}, {5, 7}, {7, 9}, {6, 8}, {8, 10}, {5, 11}, {6, 12}, {11, 12}, {11, 13}, {12, 14}, {13, 15}, {14, 16}\n };\n static const cv::Scalar bone_colors[] = {\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 51, 255),\n cv::Scalar(255, 51, 255),\n cv::Scalar(255, 51, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n };\n\n for (int j = 0; j < 16; j++)\n {\n const KeyPoint& p1 = obj.keypoints[joint_pairs[j][0]];\n const KeyPoint& p2 = obj.keypoints[joint_pairs[j][1]];\n\n if (p1.prob < 0.2f || p2.prob < 0.2f)\n continue;\n\n cv::line(image, p1.p, p2.p, bone_colors[j], 2);\n }\n\n // draw joint\n for (size_t j = 0; j < obj.keypoints.size(); j++)\n {\n const KeyPoint& keypoint = obj.keypoints[j];\n\n fprintf(stderr, \"%.2f %.2f = %.5f\\n\", keypoint.p.x, keypoint.p.y, keypoint.prob);\n\n if (keypoint.prob < 0.2f)\n continue;\n\n cv::circle(image, keypoint.p, 3, color, -1);\n }\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolo11_pose(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolo11_seg.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolo11-seg torchscript\n// yolo export model=yolo11n-seg.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolo11n-seg.torchscript\n// 4. modify yolo11n_seg_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_202 = v_201.view(1, 32, 6400)\n// v_208 = v_207.view(1, 32, 1600)\n// v_214 = v_213.view(1, 32, 400)\n// v_215 = torch.cat((v_202, v_208, v_214), dim=2)\n// ...\n// v_261 = v_230.view(1, 144, 6400)\n// v_262 = v_245.view(1, 144, 1600)\n// v_263 = v_260.view(1, 144, 400)\n// v_264 = torch.cat((v_261, v_262, v_263), dim=2)\n// ...\n// v_285 = (v_284, v_196, )\n// return v_285\n// after:\n// v_202 = v_201.view(1, 32, -1).transpose(1, 2)\n// v_208 = v_207.view(1, 32, -1).transpose(1, 2)\n// v_214 = v_213.view(1, 32, -1).transpose(1, 2)\n// v_215 = torch.cat((v_202, v_208, v_214), dim=1)\n// ...\n// v_261 = v_230.view(1, 144, -1).transpose(1, 2)\n// v_262 = v_245.view(1, 144, -1).transpose(1, 2)\n// v_263 = v_260.view(1, 144, -1).transpose(1, 2)\n// v_264 = torch.cat((v_261, v_262, v_263), dim=1)\n// return v_264, v_215, v_196\n// D. modify area attention for dynamic shape inference\n// before:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, 400)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, 20, 20)\n// v_107 = v_99.reshape(1, 128, 20, 20)\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// after:\n// v_95 = self.model_10_m_0_attn_qkv_conv(v_94)\n// v_96 = v_95.view(1, 2, 128, -1)\n// v_97, v_98, v_99 = torch.split(tensor=v_96, dim=2, split_size_or_sections=(32,32,64))\n// v_100 = torch.transpose(input=v_97, dim0=-2, dim1=-1)\n// v_101 = torch.matmul(input=v_100, other=v_98)\n// v_102 = (v_101 * 0.176777)\n// v_103 = F.softmax(input=v_102, dim=-1)\n// v_104 = torch.transpose(input=v_103, dim0=-2, dim1=-1)\n// v_105 = torch.matmul(input=v_99, other=v_104)\n// v_106 = v_105.view(1, 128, v_95.size(2), v_95.size(3))\n// v_107 = v_99.reshape(1, 128, v_95.size(2), v_95.size(3))\n// v_108 = self.model_10_m_0_attn_pe_conv(v_107)\n// v_109 = (v_106 + v_108)\n// v_110 = self.model_10_m_0_attn_proj_conv(v_109)\n// 5. re-export yolo11-seg torchscript\n// python3 -c 'import yolo11n_seg_pnnx; yolo11n_seg_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolo11n_seg_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolo11n_seg_pnnx.py.ncnn.param yolo11n_seg.ncnn.param\n// mv yolo11n_seg_pnnx.py.ncnn.bin yolo11n_seg.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=176 h=8400\n//\n// | bbox-reg 16 x 4 | per-class scores(80) |\n// +-----+-----+-----+-----+----------------------+\n// | dx0 | dy0 | dx1 | dy1 |0.1 0.0 0.0 0.5 ......|\n// all /| | | | | . |\n// boxes | .. | .. | .. | .. |0.0 0.9 0.0 0.0 ......|\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+----------------------+\n//\n\n//\n// | mask (32) |\n// +-----------+\n// |0.1........|\n// all /| |\n// boxes |0.0........|\n// (8400)| . |\n// \\| . |\n// +-----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n int gindex;\n cv::Mat mask;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 80 for COCO\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n obj.gindex = y * num_grid_x + x;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n std::vector objects_stride;\n generate_proposals(pred.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects_stride);\n\n for (size_t j = 0; j < objects_stride.size(); j++)\n {\n Object obj = objects_stride[j];\n obj.gindex += pred_row_offset;\n objects.push_back(obj);\n }\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolo11_seg(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolo11;\n\n yolo11.opt.use_vulkan_compute = true;\n // yolo11.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolo11/tree/master/app/src/main/assets\n yolo11.load_param(\"yolo11n_seg.ncnn.param\");\n yolo11.load_model(\"yolo11n_seg.ncnn.bin\");\n // yolo11.load_param(\"yolo11s_seg.ncnn.param\");\n // yolo11.load_model(\"yolo11s_seg.ncnn.bin\");\n // yolo11.load_param(\"yolo11m_seg.ncnn.param\");\n // yolo11.load_model(\"yolo11m_seg.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n const float mask_threshold = 0.5f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolo11.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolo11.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n std::vector proposals;\n generate_proposals(out, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n ncnn::Mat mask_feat;\n ex.extract(\"out1\", mask_feat);\n\n ncnn::Mat mask_protos;\n ex.extract(\"out2\", mask_protos);\n\n ncnn::Mat objects_mask_feat(mask_feat.w, 1, count);\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n\n // pick mask feat\n memcpy(objects_mask_feat.channel(i), mask_feat.row(objects[i].gindex), mask_feat.w * sizeof(float));\n }\n\n // process mask\n ncnn::Mat objects_mask;\n {\n ncnn::Layer* gemm = ncnn::create_layer(\"Gemm\");\n\n ncnn::ParamDict pd;\n pd.set(6, 1); // constantC\n pd.set(7, count); // constantM\n pd.set(8, mask_protos.w * mask_protos.h); // constantN\n pd.set(9, mask_feat.w); // constantK\n pd.set(10, -1); // constant_broadcast_type_C\n pd.set(11, 1); // output_N1M\n gemm->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n gemm->create_pipeline(opt);\n\n std::vector gemm_inputs(2);\n gemm_inputs[0] = objects_mask_feat;\n gemm_inputs[1] = mask_protos.reshape(mask_protos.w * mask_protos.h, 1, mask_protos.c);\n std::vector gemm_outputs(1);\n gemm->forward(gemm_inputs, gemm_outputs, opt);\n objects_mask = gemm_outputs[0].reshape(mask_protos.w, mask_protos.h, count);\n\n gemm->destroy_pipeline(opt);\n\n delete gemm;\n }\n {\n ncnn::Layer* sigmoid = ncnn::create_layer(\"Sigmoid\");\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n sigmoid->create_pipeline(opt);\n\n sigmoid->forward_inplace(objects_mask, opt);\n\n sigmoid->destroy_pipeline(opt);\n\n delete sigmoid;\n }\n\n // resize mask map\n {\n ncnn::Mat objects_mask_resized;\n ncnn::resize_bilinear(objects_mask, objects_mask_resized, in_pad.w / scale, in_pad.h / scale);\n objects_mask = objects_mask_resized;\n }\n\n // create per-object mask\n for (int i = 0; i < count; i++)\n {\n Object& obj = objects[i];\n\n const ncnn::Mat mm = objects_mask.channel(i);\n\n obj.mask = cv::Mat((int)obj.rect.height, (int)obj.rect.width, CV_8UC1);\n\n // adjust offset to original unpadded and clip inside object box\n for (int y = 0; y < (int)obj.rect.height; y++)\n {\n const float* pmm = mm.row((int)(hpad / 2 / scale + obj.rect.y + y)) + (int)(wpad / 2 / scale + obj.rect.x);\n uchar* pmask = obj.mask.ptr(y);\n for (int x = 0; x < (int)obj.rect.width; x++)\n {\n pmask[x] = pmm[x] > mask_threshold ? 1 : 0;\n }\n }\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n for (int y = 0; y < (int)obj.rect.height; y++)\n {\n const uchar* maskptr = obj.mask.ptr(y);\n uchar* bgrptr = image.ptr((int)obj.rect.y + y) + (int)obj.rect.x * 3;\n for (int x = 0; x < (int)obj.rect.width; x++)\n {\n if (maskptr[x])\n {\n bgrptr[0] = bgrptr[0] * 0.5 + color[0] * 0.5;\n bgrptr[1] = bgrptr[1] * 0.5 + color[1] * 0.5;\n bgrptr[2] = bgrptr[2] * 0.5 + color[2] * 0.5;\n }\n bgrptr += 3;\n }\n }\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolo11_seg(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov2.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_yolov2(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov2;\n\n yolov2.opt.use_vulkan_compute = true;\n\n // original pretrained model from https://github.com/eric612/MobileNet-YOLO\n // https://github.com/eric612/MobileNet-YOLO/blob/master/models/yolov2/mobilenet_yolo_deploy.prototxt\n // https://github.com/eric612/MobileNet-YOLO/blob/master/models/yolov2/mobilenet_yolo_deploy_iter_80000.caffemodel\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (yolov2.load_param(\"mobilenet_yolo.param\"))\n exit(-1);\n if (yolov2.load_model(\"mobilenet_yolo.bin\"))\n exit(-1);\n\n const int target_size = 416;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n // the Caffe-YOLOv2-Windows style\n // X' = X * scale - mean\n const float mean_vals[3] = {1.0f, 1.0f, 1.0f};\n const float norm_vals[3] = {0.007843f, 0.007843f, 0.007843f};\n in.substract_mean_normalize(0, norm_vals);\n in.substract_mean_normalize(mean_vals, 0);\n\n ncnn::Extractor ex = yolov2.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov2(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov3.cpp", "content": "// Copyright 2018 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int detect_yolov3(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov3;\n\n yolov3.opt.use_vulkan_compute = true;\n\n // original pretrained model from https://github.com/eric612/MobileNet-YOLO\n // param : https://drive.google.com/open?id=1V9oKHP6G6XvXZqhZbzNKL6FI_clRWdC-\n // bin : https://drive.google.com/open?id=1DBcuFCr-856z3FRQznWL_S5h-Aj3RawA\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (yolov3.load_param(\"mobilenetv2_yolov3.param\"))\n exit(-1);\n if (yolov3.load_model(\"mobilenetv2_yolov3.bin\"))\n exit(-1);\n\n const int target_size = 352;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {127.5f, 127.5f, 127.5f};\n const float norm_vals[3] = {0.007843f, 0.007843f, 0.007843f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = yolov3.create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"detection_out\", out);\n\n // printf(\"%d %d %d\\n\", out.w, out.h, out.c);\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"background\",\n \"aeroplane\", \"bicycle\", \"bird\", \"boat\",\n \"bottle\", \"bus\", \"car\", \"cat\", \"chair\",\n \"cow\", \"diningtable\", \"dog\", \"horse\",\n \"motorbike\", \"person\", \"pottedplant\",\n \"sheep\", \"sofa\", \"train\", \"tvmonitor\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov3(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov4.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"net.h\"\n\n#include \n#include \n#include \n\n#if CV_MAJOR_VERSION >= 3\n#include \n#endif\n\n#include \n\n#include \n\n#define NCNN_PROFILING\n#define YOLOV4_TINY //Using yolov4_tiny, if undef, using original yolov4\n\n#ifdef NCNN_PROFILING\n#include \"benchmark.h\"\n#endif\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic int init_yolov4(ncnn::Net* yolov4, int* target_size)\n{\n /* --> Set the params you need for the ncnn inference <-- */\n\n yolov4->opt.num_threads = 4; //You need to compile with libgomp for multi thread support\n\n yolov4->opt.use_vulkan_compute = true; //You need to compile with libvulkan for gpu support\n\n /* --> End of setting params <-- */\n int ret = 0;\n\n // original pretrained model from https://github.com/AlexeyAB/darknet\n // the ncnn model https://drive.google.com/drive/folders/1YzILvh0SKQPS_lrb33dmGNq7aVTKPWS0?usp=sharing\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n#ifdef YOLOV4_TINY\n const char* yolov4_param = \"yolov4-tiny-opt.param\";\n const char* yolov4_model = \"yolov4-tiny-opt.bin\";\n *target_size = 416;\n#else\n const char* yolov4_param = \"yolov4-opt.param\";\n const char* yolov4_model = \"yolov4-opt.bin\";\n *target_size = 608;\n#endif\n\n if (yolov4->load_param(yolov4_param))\n exit(-1);\n if (yolov4->load_model(yolov4_model))\n exit(-1);\n\n return 0;\n}\n\nstatic int detect_yolov4(const cv::Mat& bgr, std::vector& objects, int target_size, ncnn::Net* yolov4)\n{\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, bgr.cols, bgr.rows, target_size, target_size);\n\n const float mean_vals[3] = {0, 0, 0};\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in.substract_mean_normalize(mean_vals, norm_vals);\n\n ncnn::Extractor ex = yolov4->create_extractor();\n\n ex.input(\"data\", in);\n\n ncnn::Mat out;\n ex.extract(\"output\", out);\n\n objects.clear();\n for (int i = 0; i < out.h; i++)\n {\n const float* values = out.row(i);\n\n Object object;\n object.label = values[0];\n object.prob = values[1];\n object.rect.x = values[2] * img_w;\n object.rect.y = values[3] * img_h;\n object.rect.width = values[4] * img_w - object.rect.x;\n object.rect.height = values[5] * img_h - object.rect.y;\n\n objects.push_back(object);\n }\n\n return 0;\n}\n\nstatic int draw_objects(const cv::Mat& bgr, const std::vector& objects, int is_streaming)\n{\n static const char* class_names[] = {\"background\", \"person\", \"bicycle\",\n \"car\", \"motorbike\", \"aeroplane\", \"bus\", \"train\", \"truck\",\n \"boat\", \"traffic light\", \"fire hydrant\", \"stop sign\",\n \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\",\n \"sheep\", \"cow\", \"elephant\", \"bear\", \"zebra\", \"giraffe\",\n \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\",\n \"frisbee\", \"skis\", \"snowboard\", \"sports ball\", \"kite\",\n \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\",\n \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\", \"sandwich\",\n \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\",\n \"cake\", \"chair\", \"sofa\", \"pottedplant\", \"bed\", \"diningtable\",\n \"toilet\", \"tvmonitor\", \"laptop\", \"mouse\", \"remote\", \"keyboard\",\n \"cell phone\", \"microwave\", \"oven\", \"toaster\", \"sink\",\n \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\",\n \"teddy bear\", \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n\n if (is_streaming)\n {\n cv::waitKey(1);\n }\n else\n {\n cv::waitKey(0);\n }\n\n return 0;\n}\n\nint main(int argc, char** argv)\n{\n cv::Mat frame;\n std::vector objects;\n\n cv::VideoCapture cap;\n\n ncnn::Net yolov4;\n\n const char* devicepath;\n\n int target_size = 0;\n int is_streaming = 0;\n\n if (argc < 2)\n {\n fprintf(stderr, \"Usage: %s [v4l input device or image]\\n\", argv[0]);\n return -1;\n }\n\n devicepath = argv[1];\n\n#ifdef NCNN_PROFILING\n double t_load_start = ncnn::get_current_time();\n#endif\n\n int ret = init_yolov4(&yolov4, &target_size); //We load model and param first!\n if (ret != 0)\n {\n fprintf(stderr, \"Failed to load model or param, error %d\", ret);\n return -1;\n }\n\n#ifdef NCNN_PROFILING\n double t_load_end = ncnn::get_current_time();\n fprintf(stdout, \"NCNN Init time %.02lfms\\n\", t_load_end - t_load_start);\n#endif\n\n if (strstr(devicepath, \"/dev/video\") == NULL)\n {\n frame = cv::imread(argv[1], 1);\n if (frame.empty())\n {\n fprintf(stderr, \"Failed to read image %s.\\n\", argv[1]);\n return -1;\n }\n }\n else\n {\n cap.open(devicepath);\n\n if (!cap.isOpened())\n {\n fprintf(stderr, \"Failed to open %s\", devicepath);\n return -1;\n }\n\n cap >> frame;\n\n if (frame.empty())\n {\n fprintf(stderr, \"Failed to read from device %s.\\n\", devicepath);\n return -1;\n }\n\n is_streaming = 1;\n }\n\n while (1)\n {\n if (is_streaming)\n {\n#ifdef NCNN_PROFILING\n double t_capture_start = ncnn::get_current_time();\n#endif\n\n cap >> frame;\n\n#ifdef NCNN_PROFILING\n double t_capture_end = ncnn::get_current_time();\n fprintf(stdout, \"NCNN OpenCV capture time %.02lfms\\n\", t_capture_end - t_capture_start);\n#endif\n if (frame.empty())\n {\n fprintf(stderr, \"OpenCV Failed to Capture from device %s\\n\", devicepath);\n return -1;\n }\n }\n\n#ifdef NCNN_PROFILING\n double t_detect_start = ncnn::get_current_time();\n#endif\n\n detect_yolov4(frame, objects, target_size, &yolov4); //Create an extractor and run detection\n\n#ifdef NCNN_PROFILING\n double t_detect_end = ncnn::get_current_time();\n fprintf(stdout, \"NCNN detection time %.02lfms\\n\", t_detect_end - t_detect_start);\n#endif\n\n#ifdef NCNN_PROFILING\n double t_draw_start = ncnn::get_current_time();\n#endif\n\n draw_objects(frame, objects, is_streaming); //Draw detection results on opencv image\n\n#ifdef NCNN_PROFILING\n double t_draw_end = ncnn::get_current_time();\n fprintf(stdout, \"NCNN OpenCV draw result time %.02lfms\\n\", t_draw_end - t_draw_start);\n#endif\n\n if (!is_streaming)\n { //If it is a still image, exit!\n return 0;\n }\n }\n\n return 0;\n}\n" }, { "path": "examples/yolov5.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\n//#define YOLOV5_V60 1 //YOLOv5 v6.0\n#define YOLOV5_V62 1 //YOLOv5 v6.2 export onnx model method https://github.com/shaoshengsong/yolov5_62_export_ncnn\n\n#if YOLOV5_V60 || YOLOV5_V62\n#define MAX_STRIDE 64\n#else\n#define MAX_STRIDE 32\nclass YoloV5Focus : public ncnn::Layer\n{\npublic:\n YoloV5Focus()\n {\n one_blob_only = true;\n }\n\n virtual int forward(const ncnn::Mat& bottom_blob, ncnn::Mat& top_blob, const ncnn::Option& opt) const\n {\n int w = bottom_blob.w;\n int h = bottom_blob.h;\n int channels = bottom_blob.c;\n\n int outw = w / 2;\n int outh = h / 2;\n int outc = channels * 4;\n\n top_blob.create(outw, outh, outc, 4u, 1, opt.blob_allocator);\n if (top_blob.empty())\n return -100;\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int p = 0; p < outc; p++)\n {\n const float* ptr = bottom_blob.channel(p % channels).row((p / channels) % 2) + ((p / channels) / 2);\n float* outptr = top_blob.channel(p);\n\n for (int i = 0; i < outh; i++)\n {\n for (int j = 0; j < outw; j++)\n {\n *outptr = *ptr;\n\n outptr += 1;\n ptr += 2;\n }\n\n ptr += w;\n }\n }\n\n return 0;\n }\n};\n\nDEFINE_LAYER_CREATOR(YoloV5Focus)\n#endif //YOLOV5_V60 YOLOV5_V62\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return static_cast(1.f / (1.f + exp(-x)));\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int stride, const ncnn::Mat& in_pad, const ncnn::Mat& feat_blob, float prob_threshold, std::vector& objects)\n{\n const int num_grid = feat_blob.h;\n\n int num_grid_x;\n int num_grid_y;\n if (in_pad.w > in_pad.h)\n {\n num_grid_x = in_pad.w / stride;\n num_grid_y = num_grid / num_grid_x;\n }\n else\n {\n num_grid_y = in_pad.h / stride;\n num_grid_x = num_grid / num_grid_y;\n }\n\n const int num_class = feat_blob.w - 5;\n\n const int num_anchors = anchors.w / 2;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float anchor_w = anchors[q * 2];\n const float anchor_h = anchors[q * 2 + 1];\n\n const ncnn::Mat feat = feat_blob.channel(q);\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n const float* featptr = feat.row(i * num_grid_x + j);\n float box_confidence = sigmoid(featptr[4]);\n if (box_confidence >= prob_threshold)\n {\n // find class index with max class score\n int class_index = 0;\n float class_score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n float score = featptr[5 + k];\n if (score > class_score)\n {\n class_index = k;\n class_score = score;\n }\n }\n float confidence = box_confidence * sigmoid(class_score);\n if (confidence >= prob_threshold)\n {\n // yolov5/models/yolo.py Detect forward\n // y = x[i].sigmoid()\n // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy\n // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh\n\n float dx = sigmoid(featptr[0]);\n float dy = sigmoid(featptr[1]);\n float dw = sigmoid(featptr[2]);\n float dh = sigmoid(featptr[3]);\n\n float pb_cx = (dx * 2.f - 0.5f + j) * stride;\n float pb_cy = (dy * 2.f - 0.5f + i) * stride;\n\n float pb_w = pow(dw * 2.f, 2) * anchor_w;\n float pb_h = pow(dh * 2.f, 2) * anchor_h;\n\n float x0 = pb_cx - pb_w * 0.5f;\n float y0 = pb_cy - pb_h * 0.5f;\n float x1 = pb_cx + pb_w * 0.5f;\n float y1 = pb_cy + pb_h * 0.5f;\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = class_index;\n obj.prob = confidence;\n\n objects.push_back(obj);\n }\n }\n }\n }\n }\n}\n\nstatic int detect_yolov5(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov5;\n\n yolov5.opt.use_vulkan_compute = true;\n // yolov5.opt.use_bf16_storage = true;\n\n // original pretrained model from https://github.com/ultralytics/yolov5\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n#if YOLOV5_V62\n if (yolov5.load_param(\"yolov5s_6.2.param\"))\n exit(-1);\n if (yolov5.load_model(\"yolov5s_6.2.bin\"))\n exit(-1);\n#elif YOLOV5_V60\n if (yolov5.load_param(\"yolov5s_6.0.param\"))\n exit(-1);\n if (yolov5.load_model(\"yolov5s_6.0.bin\"))\n exit(-1);\n#else\n yolov5.register_custom_layer(\"YoloV5Focus\", YoloV5Focus_layer_creator);\n\n if (yolov5.load_param(\"yolov5s.param\"))\n exit(-1);\n if (yolov5.load_model(\"yolov5s.bin\"))\n exit(-1);\n#endif\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // letterbox pad to multiple of MAX_STRIDE\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // pad to target_size rectangle\n // yolov5/utils/datasets.py letterbox\n int wpad = (w + MAX_STRIDE - 1) / MAX_STRIDE * MAX_STRIDE - w;\n int hpad = (h + MAX_STRIDE - 1) / MAX_STRIDE * MAX_STRIDE - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov5.create_extractor();\n\n ex.input(\"images\", in_pad);\n\n std::vector proposals;\n\n // anchor setting from yolov5/models/yolov5s.yaml\n\n // stride 8\n {\n ncnn::Mat out;\n ex.extract(\"output\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 10.f;\n anchors[1] = 13.f;\n anchors[2] = 16.f;\n anchors[3] = 30.f;\n anchors[4] = 33.f;\n anchors[5] = 23.f;\n\n std::vector objects8;\n generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat out;\n\n#if YOLOV5_V62\n ex.extract(\"353\", out);\n#elif YOLOV5_V60\n ex.extract(\"376\", out);\n#else\n ex.extract(\"781\", out);\n#endif\n\n ncnn::Mat anchors(6);\n anchors[0] = 30.f;\n anchors[1] = 61.f;\n anchors[2] = 62.f;\n anchors[3] = 45.f;\n anchors[4] = 59.f;\n anchors[5] = 119.f;\n\n std::vector objects16;\n generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat out;\n#if YOLOV5_V62\n ex.extract(\"367\", out);\n#elif YOLOV5_V60\n ex.extract(\"401\", out);\n#else\n ex.extract(\"801\", out);\n#endif\n ncnn::Mat anchors(6);\n anchors[0] = 116.f;\n anchors[1] = 90.f;\n anchors[2] = 156.f;\n anchors[3] = 198.f;\n anchors[4] = 373.f;\n anchors[5] = 326.f;\n\n std::vector objects32;\n generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov5(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov5_pnnx.cpp", "content": "// Copyright 2022 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return static_cast(1.f / (1.f + exp(-x)));\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int stride, const ncnn::Mat& in_pad, const ncnn::Mat& feat_blob, float prob_threshold, std::vector& objects)\n{\n const int num_grid_x = feat_blob.w;\n const int num_grid_y = feat_blob.h;\n\n const int num_anchors = anchors.w / 2;\n\n const int num_class = feat_blob.c / num_anchors - 5;\n\n const int feat_offset = num_class + 5;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float anchor_w = anchors[q * 2];\n const float anchor_h = anchors[q * 2 + 1];\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n // find class index with max class score\n int class_index = 0;\n float class_score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n float score = feat_blob.channel(q * feat_offset + 5 + k).row(i)[j];\n if (score > class_score)\n {\n class_index = k;\n class_score = score;\n }\n }\n\n float box_score = feat_blob.channel(q * feat_offset + 4).row(i)[j];\n\n float confidence = sigmoid(box_score) * sigmoid(class_score);\n\n if (confidence >= prob_threshold)\n {\n // yolov5/models/yolo.py Detect forward\n // y = x[i].sigmoid()\n // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy\n // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh\n\n float dx = sigmoid(feat_blob.channel(q * feat_offset + 0).row(i)[j]);\n float dy = sigmoid(feat_blob.channel(q * feat_offset + 1).row(i)[j]);\n float dw = sigmoid(feat_blob.channel(q * feat_offset + 2).row(i)[j]);\n float dh = sigmoid(feat_blob.channel(q * feat_offset + 3).row(i)[j]);\n\n float pb_cx = (dx * 2.f - 0.5f + j) * stride;\n float pb_cy = (dy * 2.f - 0.5f + i) * stride;\n\n float pb_w = pow(dw * 2.f, 2) * anchor_w;\n float pb_h = pow(dh * 2.f, 2) * anchor_h;\n\n float x0 = pb_cx - pb_w * 0.5f;\n float y0 = pb_cy - pb_h * 0.5f;\n float x1 = pb_cx + pb_w * 0.5f;\n float y1 = pb_cy + pb_h * 0.5f;\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = class_index;\n obj.prob = confidence;\n\n objects.push_back(obj);\n }\n }\n }\n }\n}\n\nstatic int detect_yolov5(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov5;\n\n yolov5.opt.use_vulkan_compute = true;\n // yolov5.opt.use_bf16_storage = true;\n\n // original pretrained model from https://github.com/ultralytics/yolov5\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if (yolov5.load_param(\"yolov5s.ncnn.param\"))\n exit(-1);\n if (yolov5.load_model(\"yolov5s.ncnn.bin\"))\n exit(-1);\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // yolov5/models/common.py DetectMultiBackend\n const int max_stride = 64;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // pad to target_size rectangle\n // yolov5/utils/datasets.py letterbox\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov5.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n std::vector proposals;\n\n // anchor setting from yolov5/models/yolov5s.yaml\n\n // stride 8\n {\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 10.f;\n anchors[1] = 13.f;\n anchors[2] = 16.f;\n anchors[3] = 30.f;\n anchors[4] = 33.f;\n anchors[5] = 23.f;\n\n std::vector objects8;\n generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat out;\n ex.extract(\"out1\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 30.f;\n anchors[1] = 61.f;\n anchors[2] = 62.f;\n anchors[3] = 45.f;\n anchors[4] = 59.f;\n anchors[5] = 119.f;\n\n std::vector objects16;\n generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat out;\n ex.extract(\"out2\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 116.f;\n anchors[1] = 90.f;\n anchors[2] = 156.f;\n anchors[3] = 198.f;\n anchors[4] = 373.f;\n anchors[5] = 326.f;\n\n std::vector objects32;\n generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov5(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov7.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\n#define MAX_STRIDE 32\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return static_cast(1.f / (1.f + exp(-x)));\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int stride, const ncnn::Mat& in_pad, const ncnn::Mat& feat_blob, float prob_threshold, std::vector& objects)\n{\n const int num_grid = feat_blob.h;\n\n int num_grid_x;\n int num_grid_y;\n if (in_pad.w > in_pad.h)\n {\n num_grid_x = in_pad.w / stride;\n num_grid_y = num_grid / num_grid_x;\n }\n else\n {\n num_grid_y = in_pad.h / stride;\n num_grid_x = num_grid / num_grid_y;\n }\n\n const int num_class = feat_blob.w - 5;\n\n const int num_anchors = anchors.w / 2;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float anchor_w = anchors[q * 2];\n const float anchor_h = anchors[q * 2 + 1];\n\n const ncnn::Mat feat = feat_blob.channel(q);\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n const float* featptr = feat.row(i * num_grid_x + j);\n float box_confidence = sigmoid(featptr[4]);\n if (box_confidence >= prob_threshold)\n {\n // find class index with max class score\n int class_index = 0;\n float class_score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n float score = featptr[5 + k];\n if (score > class_score)\n {\n class_index = k;\n class_score = score;\n }\n }\n float confidence = box_confidence * sigmoid(class_score);\n if (confidence >= prob_threshold)\n {\n float dx = sigmoid(featptr[0]);\n float dy = sigmoid(featptr[1]);\n float dw = sigmoid(featptr[2]);\n float dh = sigmoid(featptr[3]);\n\n float pb_cx = (dx * 2.f - 0.5f + j) * stride;\n float pb_cy = (dy * 2.f - 0.5f + i) * stride;\n\n float pb_w = pow(dw * 2.f, 2) * anchor_w;\n float pb_h = pow(dh * 2.f, 2) * anchor_h;\n\n float x0 = pb_cx - pb_w * 0.5f;\n float y0 = pb_cy - pb_h * 0.5f;\n float x1 = pb_cx + pb_w * 0.5f;\n float y1 = pb_cy + pb_h * 0.5f;\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = class_index;\n obj.prob = confidence;\n\n objects.push_back(obj);\n }\n }\n }\n }\n }\n}\n\nstatic int detect_yolov7(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov7;\n\n yolov7.opt.use_vulkan_compute = true;\n // yolov7.opt.use_bf16_storage = true;\n\n // original pretrained model from https://github.com/WongKinYiu/yolov7\n // the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n yolov7.load_param(\"yolov7-tiny.param\");\n yolov7.load_model(\"yolov7-tiny.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // letterbox pad to multiple of MAX_STRIDE\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n int wpad = (w + MAX_STRIDE - 1) / MAX_STRIDE * MAX_STRIDE - w;\n int hpad = (h + MAX_STRIDE - 1) / MAX_STRIDE * MAX_STRIDE - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov7.create_extractor();\n\n ex.input(\"images\", in_pad);\n\n std::vector proposals;\n\n // stride 8\n {\n ncnn::Mat out;\n ex.extract(\"output\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 12.f;\n anchors[1] = 16.f;\n anchors[2] = 19.f;\n anchors[3] = 36.f;\n anchors[4] = 40.f;\n anchors[5] = 28.f;\n\n std::vector objects8;\n generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat out;\n\n ex.extract(\"288\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 36.f;\n anchors[1] = 75.f;\n anchors[2] = 76.f;\n anchors[3] = 55.f;\n anchors[4] = 72.f;\n anchors[5] = 146.f;\n\n std::vector objects16;\n generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat out;\n\n ex.extract(\"302\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 142.f;\n anchors[1] = 110.f;\n anchors[2] = 192.f;\n anchors[3] = 243.f;\n anchors[4] = 459.f;\n anchors[5] = 401.f;\n\n std::vector objects32;\n generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static const unsigned char colors[19][3] = {\n {54, 67, 244},\n {99, 30, 233},\n {176, 39, 156},\n {183, 58, 103},\n {181, 81, 63},\n {243, 150, 33},\n {244, 169, 3},\n {212, 188, 0},\n {136, 150, 0},\n {80, 175, 76},\n {74, 195, 139},\n {57, 220, 205},\n {59, 235, 255},\n {7, 193, 255},\n {0, 152, 255},\n {34, 87, 255},\n {72, 85, 121},\n {158, 158, 158},\n {139, 125, 96}\n };\n\n int color_index = 0;\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const unsigned char* color = colors[color_index % 19];\n color_index++;\n\n cv::Scalar cc(color[0], color[1], color[2]);\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cc, 2);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cc, -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(255, 255, 255));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov7(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov7_pnnx.cpp", "content": "// Copyright 2022 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects)\n{\n if (faceobjects.empty())\n return;\n\n qsort_descent_inplace(faceobjects, 0, faceobjects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return static_cast(1.f / (1.f + exp(-x)));\n}\n\nstatic void generate_proposals(const ncnn::Mat& anchors, int stride, const ncnn::Mat& in_pad, const ncnn::Mat& feat_blob, float prob_threshold, std::vector& objects)\n{\n const int num_grid_x = feat_blob.w;\n const int num_grid_y = feat_blob.h;\n\n const int num_anchors = anchors.w / 2;\n\n const int num_class = 80;\n\n for (int q = 0; q < num_anchors; q++)\n {\n const float anchor_w = anchors[q * 2];\n const float anchor_h = anchors[q * 2 + 1];\n\n for (int i = 0; i < num_grid_y; i++)\n {\n for (int j = 0; j < num_grid_x; j++)\n {\n // find class index with max class score\n int class_index = 0;\n float class_score = -FLT_MAX;\n for (int k = 0; k < num_class; k++)\n {\n float score = feat_blob.channel(q * 85 + 5 + k).row(i)[j];\n if (score > class_score)\n {\n class_index = k;\n class_score = score;\n }\n }\n\n float box_score = feat_blob.channel(q * 85 + 4).row(i)[j];\n\n float confidence = sigmoid(box_score) * sigmoid(class_score);\n\n if (confidence >= prob_threshold)\n {\n // yolov5/models/yolo.py Detect forward\n // y = x[i].sigmoid()\n // y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i] # xy\n // y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh\n\n float dx = sigmoid(feat_blob.channel(q * 85 + 0).row(i)[j]);\n float dy = sigmoid(feat_blob.channel(q * 85 + 1).row(i)[j]);\n float dw = sigmoid(feat_blob.channel(q * 85 + 2).row(i)[j]);\n float dh = sigmoid(feat_blob.channel(q * 85 + 3).row(i)[j]);\n\n float pb_cx = (dx * 2.f - 0.5f + j) * stride;\n float pb_cy = (dy * 2.f - 0.5f + i) * stride;\n\n float pb_w = pow(dw * 2.f, 2) * anchor_w;\n float pb_h = pow(dh * 2.f, 2) * anchor_h;\n\n float x0 = pb_cx - pb_w * 0.5f;\n float y0 = pb_cy - pb_h * 0.5f;\n float x1 = pb_cx + pb_w * 0.5f;\n float y1 = pb_cy + pb_h * 0.5f;\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = class_index;\n obj.prob = confidence;\n\n objects.push_back(obj);\n }\n }\n }\n }\n}\n\nstatic int detect_yolov7(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov7;\n\n yolov7.opt.use_vulkan_compute = true;\n // yolov7.opt.use_bf16_storage = true;\n\n // git clone https://github.com/WongKinYiu/yolov7\n // cd yolov7\n // wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt\n // python models/export.py --weights yolov7.pt\n // pnnx yolov7.torchscript.pt inputshape=[1,3,640,640] inputshape=[1,3,320,320]\n yolov7.load_param(\"yolov7.param\");\n yolov7.load_model(\"yolov7.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // yolov5/models/common.py DetectMultiBackend\n const int max_stride = 64;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // pad to target_size rectangle\n // yolov5/utils/datasets.py letterbox\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov7.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n std::vector proposals;\n\n // anchor setting from yolov5/models/yolov5s.yaml\n\n // stride 8\n {\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 12.f;\n anchors[1] = 16.f;\n anchors[2] = 19.f;\n anchors[3] = 36.f;\n anchors[4] = 40.f;\n anchors[5] = 28.f;\n\n std::vector objects8;\n generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects8);\n\n proposals.insert(proposals.end(), objects8.begin(), objects8.end());\n }\n\n // stride 16\n {\n ncnn::Mat out;\n ex.extract(\"out1\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 36.f;\n anchors[1] = 75.f;\n anchors[2] = 76.f;\n anchors[3] = 55.f;\n anchors[4] = 72.f;\n anchors[5] = 146.f;\n\n std::vector objects16;\n generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects16);\n\n proposals.insert(proposals.end(), objects16.begin(), objects16.end());\n }\n\n // stride 32\n {\n ncnn::Mat out;\n ex.extract(\"out2\", out);\n\n ncnn::Mat anchors(6);\n anchors[0] = 142.f;\n anchors[1] = 110.f;\n anchors[2] = 192.f;\n anchors[3] = 243.f;\n anchors[4] = 459.f;\n anchors[5] = 401.f;\n\n std::vector objects32;\n generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects32);\n\n proposals.insert(proposals.end(), objects32.begin(), objects32.end());\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov7(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov8.cpp", "content": "// Copyright 2024 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolov8 torchscript\n// yolo export model=yolov8n.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8n.torchscript\n// 4. modify yolov8n_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_165 = v_142.view(1, 144, 6400)\n// v_166 = v_153.view(1, 144, 1600)\n// v_167 = v_164.view(1, 144, 400)\n// v_168 = torch.cat((v_165, v_166, v_167), dim=2)\n// ...\n// after:\n// v_165 = v_142.view(1, 144, -1).transpose(1, 2)\n// v_166 = v_153.view(1, 144, -1).transpose(1, 2)\n// v_167 = v_164.view(1, 144, -1).transpose(1, 2)\n// v_168 = torch.cat((v_165, v_166, v_167), dim=1)\n// return v_168\n// 5. re-export yolov8 torchscript\n// python3 -c 'import yolov8n_pnnx; yolov8n_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolov8n_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolov8n_pnnx.py.ncnn.param yolov8n.ncnn.param\n// mv yolov8n_pnnx.py.ncnn.bin yolov8n.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=144 h=8400\n//\n// | bbox-reg 16 x 4 | per-class scores(80) |\n// +-----+-----+-----+-----+----------------------+\n// | dx0 | dy0 | dx1 | dy1 |0.1 0.0 0.0 0.5 ......|\n// all /| | | | | . |\n// boxes | .. | .. | .. | .. |0.0 0.9 0.0 0.0 ......|\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+----------------------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 80 for COCO\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4);\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolov8(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov8;\n\n yolov8.opt.use_vulkan_compute = true;\n // yolov8.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolov8/tree/master/app/src/main/assets\n yolov8.load_param(\"yolov8n.ncnn.param\");\n yolov8.load_model(\"yolov8n.ncnn.bin\");\n // yolov8.load_param(\"yolov8s.ncnn.param\");\n // yolov8.load_model(\"yolov8s.ncnn.bin\");\n // yolov8.load_param(\"yolov8m.ncnn.param\");\n // yolov8.load_model(\"yolov8m.ncnn.bin\");\n\n // if you use oiv7 models, you shall call draw_objects_oiv() instead\n // yolov8.load_param(\"yolov8n_oiv7.ncnn.param\");\n // yolov8.load_model(\"yolov8n_oiv7.ncnn.bin\");\n // yolov8.load_param(\"yolov8s_oiv7.ncnn.param\");\n // yolov8.load_model(\"yolov8s_oiv7.ncnn.bin\");\n // yolov8.load_param(\"yolov8m_oiv7.ncnn.param\");\n // yolov8.load_model(\"yolov8m_oiv7.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolov8.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov8.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n std::vector proposals;\n generate_proposals(out, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects_coco(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nstatic void draw_objects_oiv(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"Accordion\", \"Adhesive tape\", \"Aircraft\", \"Airplane\", \"Alarm clock\", \"Alpaca\", \"Ambulance\", \"Animal\",\n \"Ant\", \"Antelope\", \"Apple\", \"Armadillo\", \"Artichoke\", \"Auto part\", \"Axe\", \"Backpack\", \"Bagel\",\n \"Baked goods\", \"Balance beam\", \"Ball\", \"Balloon\", \"Banana\", \"Band-aid\", \"Banjo\", \"Barge\", \"Barrel\",\n \"Baseball bat\", \"Baseball glove\", \"Bat (Animal)\", \"Bathroom accessory\", \"Bathroom cabinet\", \"Bathtub\",\n \"Beaker\", \"Bear\", \"Bed\", \"Bee\", \"Beehive\", \"Beer\", \"Beetle\", \"Bell pepper\", \"Belt\", \"Bench\", \"Bicycle\",\n \"Bicycle helmet\", \"Bicycle wheel\", \"Bidet\", \"Billboard\", \"Billiard table\", \"Binoculars\", \"Bird\",\n \"Blender\", \"Blue jay\", \"Boat\", \"Bomb\", \"Book\", \"Bookcase\", \"Boot\", \"Bottle\", \"Bottle opener\",\n \"Bow and arrow\", \"Bowl\", \"Bowling equipment\", \"Box\", \"Boy\", \"Brassiere\", \"Bread\", \"Briefcase\",\n \"Broccoli\", \"Bronze sculpture\", \"Brown bear\", \"Building\", \"Bull\", \"Burrito\", \"Bus\", \"Bust\", \"Butterfly\",\n \"Cabbage\", \"Cabinetry\", \"Cake\", \"Cake stand\", \"Calculator\", \"Camel\", \"Camera\", \"Can opener\", \"Canary\",\n \"Candle\", \"Candy\", \"Cannon\", \"Canoe\", \"Cantaloupe\", \"Car\", \"Carnivore\", \"Carrot\", \"Cart\", \"Cassette deck\",\n \"Castle\", \"Cat\", \"Cat furniture\", \"Caterpillar\", \"Cattle\", \"Ceiling fan\", \"Cello\", \"Centipede\",\n \"Chainsaw\", \"Chair\", \"Cheese\", \"Cheetah\", \"Chest of drawers\", \"Chicken\", \"Chime\", \"Chisel\", \"Chopsticks\",\n \"Christmas tree\", \"Clock\", \"Closet\", \"Clothing\", \"Coat\", \"Cocktail\", \"Cocktail shaker\", \"Coconut\",\n \"Coffee\", \"Coffee cup\", \"Coffee table\", \"Coffeemaker\", \"Coin\", \"Common fig\", \"Common sunflower\",\n \"Computer keyboard\", \"Computer monitor\", \"Computer mouse\", \"Container\", \"Convenience store\", \"Cookie\",\n \"Cooking spray\", \"Corded phone\", \"Cosmetics\", \"Couch\", \"Countertop\", \"Cowboy hat\", \"Crab\", \"Cream\",\n \"Cricket ball\", \"Crocodile\", \"Croissant\", \"Crown\", \"Crutch\", \"Cucumber\", \"Cupboard\", \"Curtain\",\n \"Cutting board\", \"Dagger\", \"Dairy Product\", \"Deer\", \"Desk\", \"Dessert\", \"Diaper\", \"Dice\", \"Digital clock\",\n \"Dinosaur\", \"Dishwasher\", \"Dog\", \"Dog bed\", \"Doll\", \"Dolphin\", \"Door\", \"Door handle\", \"Doughnut\",\n \"Dragonfly\", \"Drawer\", \"Dress\", \"Drill (Tool)\", \"Drink\", \"Drinking straw\", \"Drum\", \"Duck\", \"Dumbbell\",\n \"Eagle\", \"Earrings\", \"Egg (Food)\", \"Elephant\", \"Envelope\", \"Eraser\", \"Face powder\", \"Facial tissue holder\",\n \"Falcon\", \"Fashion accessory\", \"Fast food\", \"Fax\", \"Fedora\", \"Filing cabinet\", \"Fire hydrant\",\n \"Fireplace\", \"Fish\", \"Flag\", \"Flashlight\", \"Flower\", \"Flowerpot\", \"Flute\", \"Flying disc\", \"Food\",\n \"Food processor\", \"Football\", \"Football helmet\", \"Footwear\", \"Fork\", \"Fountain\", \"Fox\", \"French fries\",\n \"French horn\", \"Frog\", \"Fruit\", \"Frying pan\", \"Furniture\", \"Garden Asparagus\", \"Gas stove\", \"Giraffe\",\n \"Girl\", \"Glasses\", \"Glove\", \"Goat\", \"Goggles\", \"Goldfish\", \"Golf ball\", \"Golf cart\", \"Gondola\",\n \"Goose\", \"Grape\", \"Grapefruit\", \"Grinder\", \"Guacamole\", \"Guitar\", \"Hair dryer\", \"Hair spray\", \"Hamburger\",\n \"Hammer\", \"Hamster\", \"Hand dryer\", \"Handbag\", \"Handgun\", \"Harbor seal\", \"Harmonica\", \"Harp\",\n \"Harpsichord\", \"Hat\", \"Headphones\", \"Heater\", \"Hedgehog\", \"Helicopter\", \"Helmet\", \"High heels\",\n \"Hiking equipment\", \"Hippopotamus\", \"Home appliance\", \"Honeycomb\", \"Horizontal bar\", \"Horse\", \"Hot dog\",\n \"House\", \"Houseplant\", \"Human arm\", \"Human beard\", \"Human body\", \"Human ear\", \"Human eye\", \"Human face\",\n \"Human foot\", \"Human hair\", \"Human hand\", \"Human head\", \"Human leg\", \"Human mouth\", \"Human nose\",\n \"Humidifier\", \"Ice cream\", \"Indoor rower\", \"Infant bed\", \"Insect\", \"Invertebrate\", \"Ipod\", \"Isopod\",\n \"Jacket\", \"Jacuzzi\", \"Jaguar (Animal)\", \"Jeans\", \"Jellyfish\", \"Jet ski\", \"Jug\", \"Juice\", \"Kangaroo\",\n \"Kettle\", \"Kitchen & dining room table\", \"Kitchen appliance\", \"Kitchen knife\", \"Kitchen utensil\",\n \"Kitchenware\", \"Kite\", \"Knife\", \"Koala\", \"Ladder\", \"Ladle\", \"Ladybug\", \"Lamp\", \"Land vehicle\",\n \"Lantern\", \"Laptop\", \"Lavender (Plant)\", \"Lemon\", \"Leopard\", \"Light bulb\", \"Light switch\", \"Lighthouse\",\n \"Lily\", \"Limousine\", \"Lion\", \"Lipstick\", \"Lizard\", \"Lobster\", \"Loveseat\", \"Luggage and bags\", \"Lynx\",\n \"Magpie\", \"Mammal\", \"Man\", \"Mango\", \"Maple\", \"Maracas\", \"Marine invertebrates\", \"Marine mammal\",\n \"Measuring cup\", \"Mechanical fan\", \"Medical equipment\", \"Microphone\", \"Microwave oven\", \"Milk\",\n \"Miniskirt\", \"Mirror\", \"Missile\", \"Mixer\", \"Mixing bowl\", \"Mobile phone\", \"Monkey\", \"Moths and butterflies\",\n \"Motorcycle\", \"Mouse\", \"Muffin\", \"Mug\", \"Mule\", \"Mushroom\", \"Musical instrument\", \"Musical keyboard\",\n \"Nail (Construction)\", \"Necklace\", \"Nightstand\", \"Oboe\", \"Office building\", \"Office supplies\", \"Orange\",\n \"Organ (Musical Instrument)\", \"Ostrich\", \"Otter\", \"Oven\", \"Owl\", \"Oyster\", \"Paddle\", \"Palm tree\",\n \"Pancake\", \"Panda\", \"Paper cutter\", \"Paper towel\", \"Parachute\", \"Parking meter\", \"Parrot\", \"Pasta\",\n \"Pastry\", \"Peach\", \"Pear\", \"Pen\", \"Pencil case\", \"Pencil sharpener\", \"Penguin\", \"Perfume\", \"Person\",\n \"Personal care\", \"Personal flotation device\", \"Piano\", \"Picnic basket\", \"Picture frame\", \"Pig\",\n \"Pillow\", \"Pineapple\", \"Pitcher (Container)\", \"Pizza\", \"Pizza cutter\", \"Plant\", \"Plastic bag\", \"Plate\",\n \"Platter\", \"Plumbing fixture\", \"Polar bear\", \"Pomegranate\", \"Popcorn\", \"Porch\", \"Porcupine\", \"Poster\",\n \"Potato\", \"Power plugs and sockets\", \"Pressure cooker\", \"Pretzel\", \"Printer\", \"Pumpkin\", \"Punching bag\",\n \"Rabbit\", \"Raccoon\", \"Racket\", \"Radish\", \"Ratchet (Device)\", \"Raven\", \"Rays and skates\", \"Red panda\",\n \"Refrigerator\", \"Remote control\", \"Reptile\", \"Rhinoceros\", \"Rifle\", \"Ring binder\", \"Rocket\",\n \"Roller skates\", \"Rose\", \"Rugby ball\", \"Ruler\", \"Salad\", \"Salt and pepper shakers\", \"Sandal\",\n \"Sandwich\", \"Saucer\", \"Saxophone\", \"Scale\", \"Scarf\", \"Scissors\", \"Scoreboard\", \"Scorpion\",\n \"Screwdriver\", \"Sculpture\", \"Sea lion\", \"Sea turtle\", \"Seafood\", \"Seahorse\", \"Seat belt\", \"Segway\",\n \"Serving tray\", \"Sewing machine\", \"Shark\", \"Sheep\", \"Shelf\", \"Shellfish\", \"Shirt\", \"Shorts\",\n \"Shotgun\", \"Shower\", \"Shrimp\", \"Sink\", \"Skateboard\", \"Ski\", \"Skirt\", \"Skull\", \"Skunk\", \"Skyscraper\",\n \"Slow cooker\", \"Snack\", \"Snail\", \"Snake\", \"Snowboard\", \"Snowman\", \"Snowmobile\", \"Snowplow\",\n \"Soap dispenser\", \"Sock\", \"Sofa bed\", \"Sombrero\", \"Sparrow\", \"Spatula\", \"Spice rack\", \"Spider\",\n \"Spoon\", \"Sports equipment\", \"Sports uniform\", \"Squash (Plant)\", \"Squid\", \"Squirrel\", \"Stairs\",\n \"Stapler\", \"Starfish\", \"Stationary bicycle\", \"Stethoscope\", \"Stool\", \"Stop sign\", \"Strawberry\",\n \"Street light\", \"Stretcher\", \"Studio couch\", \"Submarine\", \"Submarine sandwich\", \"Suit\", \"Suitcase\",\n \"Sun hat\", \"Sunglasses\", \"Surfboard\", \"Sushi\", \"Swan\", \"Swim cap\", \"Swimming pool\", \"Swimwear\",\n \"Sword\", \"Syringe\", \"Table\", \"Table tennis racket\", \"Tablet computer\", \"Tableware\", \"Taco\", \"Tank\",\n \"Tap\", \"Tart\", \"Taxi\", \"Tea\", \"Teapot\", \"Teddy bear\", \"Telephone\", \"Television\", \"Tennis ball\",\n \"Tennis racket\", \"Tent\", \"Tiara\", \"Tick\", \"Tie\", \"Tiger\", \"Tin can\", \"Tire\", \"Toaster\", \"Toilet\",\n \"Toilet paper\", \"Tomato\", \"Tool\", \"Toothbrush\", \"Torch\", \"Tortoise\", \"Towel\", \"Tower\", \"Toy\",\n \"Traffic light\", \"Traffic sign\", \"Train\", \"Training bench\", \"Treadmill\", \"Tree\", \"Tree house\",\n \"Tripod\", \"Trombone\", \"Trousers\", \"Truck\", \"Trumpet\", \"Turkey\", \"Turtle\", \"Umbrella\", \"Unicycle\",\n \"Van\", \"Vase\", \"Vegetable\", \"Vehicle\", \"Vehicle registration plate\", \"Violin\", \"Volleyball (Ball)\",\n \"Waffle\", \"Waffle iron\", \"Wall clock\", \"Wardrobe\", \"Washing machine\", \"Waste container\", \"Watch\",\n \"Watercraft\", \"Watermelon\", \"Weapon\", \"Whale\", \"Wheel\", \"Wheelchair\", \"Whisk\", \"Whiteboard\", \"Willow\",\n \"Window\", \"Window blind\", \"Wine\", \"Wine glass\", \"Wine rack\", \"Winter melon\", \"Wok\", \"Woman\",\n \"Wood-burning stove\", \"Woodpecker\", \"Worm\", \"Wrench\", \"Zebra\", \"Zucchini\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov8(m, objects);\n\n draw_objects_coco(m, objects);\n // draw_objects_oiv(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov8_cls.cpp", "content": "// Copyright 2024 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolov8-cls torchscript\n// yolo export model=yolov8n-cls.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8n-cls.torchscript\n// 4. now you get ncnn model files\n// yolov8n_cls.ncnn.param\n// yolov8n_cls.ncnn.bin\n\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n int label;\n float prob;\n};\n\nstatic void get_topk(const ncnn::Mat& cls_scores, int topk, std::vector& objects)\n{\n // partial sort topk with index\n int size = cls_scores.w;\n std::vector > vec;\n vec.resize(size);\n for (int i = 0; i < size; i++)\n {\n vec[i] = std::make_pair(cls_scores[i], i);\n }\n\n std::partial_sort(vec.begin(), vec.begin() + topk, vec.end(),\n std::greater >());\n\n objects.resize(topk);\n for (int i = 0; i < topk; i++)\n {\n objects[i].label = vec[i].second;\n objects[i].prob = vec[i].first;\n }\n}\n\nstatic int detect_yolov8_cls(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov8;\n\n yolov8.opt.use_vulkan_compute = true;\n // yolov8.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolov8/tree/master/app/src/main/assets\n yolov8.load_param(\"yolov8n_cls.ncnn.param\");\n yolov8.load_model(\"yolov8n_cls.ncnn.bin\");\n // yolov8.load_param(\"yolov8s_cls.ncnn.param\");\n // yolov8.load_model(\"yolov8s_cls.ncnn.bin\");\n // yolov8.load_param(\"yolov8m_cls.ncnn.param\");\n // yolov8.load_model(\"yolov8m_cls.ncnn.bin\");\n\n const int target_size = 224;\n const int topk = 5;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // letterbox pad\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = target_size - w;\n int hpad = target_size - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov8.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n // return top-5\n get_topk(out, topk, objects);\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"tench\", \"goldfish\", \"great white shark\", \"tiger shark\", \"hammerhead\", \"electric ray\", \"stingray\", \"cock\",\n \"hen\", \"ostrich\", \"brambling\", \"goldfinch\", \"house finch\", \"junco\", \"indigo bunting\", \"robin\", \"bulbul\",\n \"jay\", \"magpie\", \"chickadee\", \"water ouzel\", \"kite\", \"bald eagle\", \"vulture\", \"great grey owl\",\n \"European fire salamander\", \"common newt\", \"eft\", \"spotted salamander\", \"axolotl\", \"bullfrog\", \"tree frog\",\n \"tailed frog\", \"loggerhead\", \"leatherback turtle\", \"mud turtle\", \"terrapin\", \"box turtle\", \"banded gecko\",\n \"common iguana\", \"American chameleon\", \"whiptail\", \"agama\", \"frilled lizard\", \"alligator lizard\",\n \"Gila monster\", \"green lizard\", \"African chameleon\", \"Komodo dragon\", \"African crocodile\",\n \"American alligator\", \"triceratops\", \"thunder snake\", \"ringneck snake\", \"hognose snake\", \"green snake\",\n \"king snake\", \"garter snake\", \"water snake\", \"vine snake\", \"night snake\", \"boa constrictor\", \"rock python\",\n \"Indian cobra\", \"green mamba\", \"sea snake\", \"horned viper\", \"diamondback\", \"sidewinder\", \"trilobite\",\n \"harvestman\", \"scorpion\", \"black and gold garden spider\", \"barn spider\", \"garden spider\", \"black widow\",\n \"tarantula\", \"wolf spider\", \"tick\", \"centipede\", \"black grouse\", \"ptarmigan\", \"ruffed grouse\",\n \"prairie chicken\", \"peacock\", \"quail\", \"partridge\", \"African grey\", \"macaw\", \"sulphur-crested cockatoo\",\n \"lorikeet\", \"coucal\", \"bee eater\", \"hornbill\", \"hummingbird\", \"jacamar\", \"toucan\", \"drake\",\n \"red-breasted merganser\", \"goose\", \"black swan\", \"tusker\", \"echidna\", \"platypus\", \"wallaby\", \"koala\",\n \"wombat\", \"jellyfish\", \"sea anemone\", \"brain coral\", \"flatworm\", \"nematode\", \"conch\", \"snail\", \"slug\",\n \"sea slug\", \"chiton\", \"chambered nautilus\", \"Dungeness crab\", \"rock crab\", \"fiddler crab\", \"king crab\",\n \"American lobster\", \"spiny lobster\", \"crayfish\", \"hermit crab\", \"isopod\", \"white stork\", \"black stork\",\n \"spoonbill\", \"flamingo\", \"little blue heron\", \"American egret\", \"bittern\", \"crane (bird)\", \"limpkin\",\n \"European gallinule\", \"American coot\", \"bustard\", \"ruddy turnstone\", \"red-backed sandpiper\", \"redshank\",\n \"dowitcher\", \"oystercatcher\", \"pelican\", \"king penguin\", \"albatross\", \"grey whale\", \"killer whale\",\n \"dugong\", \"sea lion\", \"Chihuahua\", \"Japanese spaniel\", \"Maltese dog\", \"Pekinese\", \"Shih-Tzu\",\n \"Blenheim spaniel\", \"papillon\", \"toy terrier\", \"Rhodesian ridgeback\", \"Afghan hound\", \"basset\", \"beagle\",\n \"bloodhound\", \"bluetick\", \"black-and-tan coonhound\", \"Walker hound\", \"English foxhound\", \"redbone\",\n \"borzoi\", \"Irish wolfhound\", \"Italian greyhound\", \"whippet\", \"Ibizan hound\", \"Norwegian elkhound\",\n \"otterhound\", \"Saluki\", \"Scottish deerhound\", \"Weimaraner\", \"Staffordshire bullterrier\",\n \"American Staffordshire terrier\", \"Bedlington terrier\", \"Border terrier\", \"Kerry blue terrier\",\n \"Irish terrier\", \"Norfolk terrier\", \"Norwich terrier\", \"Yorkshire terrier\", \"wire-haired fox terrier\",\n \"Lakeland terrier\", \"Sealyham terrier\", \"Airedale\", \"cairn\", \"Australian terrier\", \"Dandie Dinmont\",\n \"Boston bull\", \"miniature schnauzer\", \"giant schnauzer\", \"standard schnauzer\", \"Scotch terrier\",\n \"Tibetan terrier\", \"silky terrier\", \"soft-coated wheaten terrier\", \"West Highland white terrier\",\n \"Lhasa\", \"flat-coated retriever\", \"curly-coated retriever\", \"golden retriever\", \"Labrador retriever\",\n \"Chesapeake Bay retriever\", \"German short-haired pointer\", \"vizsla\", \"English setter\", \"Irish setter\",\n \"Gordon setter\", \"Brittany spaniel\", \"clumber\", \"English springer\", \"Welsh springer spaniel\",\n \"cocker spaniel\", \"Sussex spaniel\", \"Irish water spaniel\", \"kuvasz\", \"schipperke\", \"groenendael\",\n \"malinois\", \"briard\", \"kelpie\", \"komondor\", \"Old English sheepdog\", \"Shetland sheepdog\", \"collie\",\n \"Border collie\", \"Bouvier des Flandres\", \"Rottweiler\", \"German shepherd\", \"Doberman\",\n \"miniature pinscher\", \"Greater Swiss Mountain dog\", \"Bernese mountain dog\", \"Appenzeller\", \"EntleBucher\",\n \"boxer\", \"bull mastiff\", \"Tibetan mastiff\", \"French bulldog\", \"Great Dane\", \"Saint Bernard\",\n \"Eskimo dog\", \"malamute\", \"Siberian husky\", \"dalmatian\", \"affenpinscher\", \"basenji\", \"pug\", \"Leonberg\",\n \"Newfoundland\", \"Great Pyrenees\", \"Samoyed\", \"Pomeranian\", \"chow\", \"keeshond\", \"Brabancon griffon\",\n \"Pembroke\", \"Cardigan\", \"toy poodle\", \"miniature poodle\", \"standard poodle\", \"Mexican hairless\",\n \"timber wolf\", \"white wolf\", \"red wolf\", \"coyote\", \"dingo\", \"dhole\", \"African hunting dog\", \"hyena\",\n \"red fox\", \"kit fox\", \"Arctic fox\", \"grey fox\", \"tabby\", \"tiger cat\", \"Persian cat\", \"Siamese cat\",\n \"Egyptian cat\", \"cougar\", \"lynx\", \"leopard\", \"snow leopard\", \"jaguar\", \"lion\", \"tiger\", \"cheetah\",\n \"brown bear\", \"American black bear\", \"ice bear\", \"sloth bear\", \"mongoose\", \"meerkat\", \"tiger beetle\",\n \"ladybug\", \"ground beetle\", \"long-horned beetle\", \"leaf beetle\", \"dung beetle\", \"rhinoceros beetle\",\n \"weevil\", \"fly\", \"bee\", \"ant\", \"grasshopper\", \"cricket\", \"walking stick\", \"cockroach\", \"mantis\",\n \"cicada\", \"leafhopper\", \"lacewing\", \"dragonfly\", \"damselfly\", \"admiral\", \"ringlet\", \"monarch\",\n \"cabbage butterfly\", \"sulphur butterfly\", \"lycaenid\", \"starfish\", \"sea urchin\", \"sea cucumber\",\n \"wood rabbit\", \"hare\", \"Angora\", \"hamster\", \"porcupine\", \"fox squirrel\", \"marmot\", \"beaver\",\n \"guinea pig\", \"sorrel\", \"zebra\", \"hog\", \"wild boar\", \"warthog\", \"hippopotamus\", \"ox\", \"water buffalo\",\n \"bison\", \"ram\", \"bighorn\", \"ibex\", \"hartebeest\", \"impala\", \"gazelle\", \"Arabian camel\", \"llama\",\n \"weasel\", \"mink\", \"polecat\", \"black-footed ferret\", \"otter\", \"skunk\", \"badger\", \"armadillo\",\n \"three-toed sloth\", \"orangutan\", \"gorilla\", \"chimpanzee\", \"gibbon\", \"siamang\", \"guenon\", \"patas\",\n \"baboon\", \"macaque\", \"langur\", \"colobus\", \"proboscis monkey\", \"marmoset\", \"capuchin\", \"howler monkey\",\n \"titi\", \"spider monkey\", \"squirrel monkey\", \"Madagascar cat\", \"indri\", \"Indian elephant\",\n \"African elephant\", \"lesser panda\", \"giant panda\", \"barracouta\", \"eel\", \"coho\", \"rock beauty\",\n \"anemone fish\", \"sturgeon\", \"gar\", \"lionfish\", \"puffer\", \"abacus\", \"abaya\", \"academic gown\",\n \"accordion\", \"acoustic guitar\", \"aircraft carrier\", \"airliner\", \"airship\", \"altar\", \"ambulance\",\n \"amphibian\", \"analog clock\", \"apiary\", \"apron\", \"ashcan\", \"assault rifle\", \"backpack\", \"bakery\",\n \"balance beam\", \"balloon\", \"ballpoint\", \"Band Aid\", \"banjo\", \"bannister\", \"barbell\", \"barber chair\",\n \"barbershop\", \"barn\", \"barometer\", \"barrel\", \"barrow\", \"baseball\", \"basketball\", \"bassinet\", \"bassoon\",\n \"bathing cap\", \"bath towel\", \"bathtub\", \"beach wagon\", \"beacon\", \"beaker\", \"bearskin\", \"beer bottle\",\n \"beer glass\", \"bell cote\", \"bib\", \"bicycle-built-for-two\", \"bikini\", \"binder\", \"binoculars\",\n \"birdhouse\", \"boathouse\", \"bobsled\", \"bolo tie\", \"bonnet\", \"bookcase\", \"bookshop\", \"bottlecap\", \"bow\",\n \"bow tie\", \"brass\", \"brassiere\", \"breakwater\", \"breastplate\", \"broom\", \"bucket\", \"buckle\",\n \"bulletproof vest\", \"bullet train\", \"butcher shop\", \"cab\", \"caldron\", \"candle\", \"cannon\", \"canoe\",\n \"can opener\", \"cardigan\", \"car mirror\", \"carousel\", \"carpenter's kit\", \"carton\", \"car wheel\",\n \"cash machine\", \"cassette\", \"cassette player\", \"castle\", \"catamaran\", \"CD player\", \"cello\",\n \"cellular telephone\", \"chain\", \"chainlink fence\", \"chain mail\", \"chain saw\", \"chest\", \"chiffonier\",\n \"chime\", \"china cabinet\", \"Christmas stocking\", \"church\", \"cinema\", \"cleaver\", \"cliff dwelling\",\n \"cloak\", \"clog\", \"cocktail shaker\", \"coffee mug\", \"coffeepot\", \"coil\", \"combination lock\",\n \"computer keyboard\", \"confectionery\", \"container ship\", \"convertible\", \"corkscrew\", \"cornet\",\n \"cowboy boot\", \"cowboy hat\", \"cradle\", \"crane (machine)\", \"crash helmet\", \"crate\", \"crib\",\n \"Crock Pot\", \"croquet ball\", \"crutch\", \"cuirass\", \"dam\", \"desk\", \"desktop computer\", \"dial telephone\",\n \"diaper\", \"digital clock\", \"digital watch\", \"dining table\", \"dishrag\", \"dishwasher\", \"disk brake\",\n \"dock\", \"dogsled\", \"dome\", \"doormat\", \"drilling platform\", \"drum\", \"drumstick\", \"dumbbell\",\n \"Dutch oven\", \"electric fan\", \"electric guitar\", \"electric locomotive\", \"entertainment center\",\n \"envelope\", \"espresso maker\", \"face powder\", \"feather boa\", \"file\", \"fireboat\", \"fire engine\",\n \"fire screen\", \"flagpole\", \"flute\", \"folding chair\", \"football helmet\", \"forklift\", \"fountain\",\n \"fountain pen\", \"four-poster\", \"freight car\", \"French horn\", \"frying pan\", \"fur coat\", \"garbage truck\",\n \"gasmask\", \"gas pump\", \"goblet\", \"go-kart\", \"golf ball\", \"golfcart\", \"gondola\", \"gong\", \"gown\",\n \"grand piano\", \"greenhouse\", \"grille\", \"grocery store\", \"guillotine\", \"hair slide\", \"hair spray\",\n \"half track\", \"hammer\", \"hamper\", \"hand blower\", \"hand-held computer\", \"handkerchief\", \"hard disc\",\n \"harmonica\", \"harp\", \"harvester\", \"hatchet\", \"holster\", \"home theater\", \"honeycomb\", \"hook\",\n \"hoopskirt\", \"horizontal bar\", \"horse cart\", \"hourglass\", \"iPod\", \"iron\", \"jack-o'-lantern\", \"jean\",\n \"jeep\", \"jersey\", \"jigsaw puzzle\", \"jinrikisha\", \"joystick\", \"kimono\", \"knee pad\", \"knot\", \"lab coat\",\n \"ladle\", \"lampshade\", \"laptop\", \"lawn mower\", \"lens cap\", \"letter opener\", \"library\", \"lifeboat\",\n \"lighter\", \"limousine\", \"liner\", \"lipstick\", \"Loafer\", \"lotion\", \"loudspeaker\", \"loupe\", \"lumbermill\",\n \"magnetic compass\", \"mailbag\", \"mailbox\", \"maillot (tights)\", \"maillot (tank suit)\", \"manhole cover\",\n \"maraca\", \"marimba\", \"mask\", \"matchstick\", \"maypole\", \"maze\", \"measuring cup\", \"medicine chest\",\n \"megalith\", \"microphone\", \"microwave\", \"military uniform\", \"milk can\", \"minibus\", \"miniskirt\",\n \"minivan\", \"missile\", \"mitten\", \"mixing bowl\", \"mobile home\", \"Model T\", \"modem\", \"monastery\",\n \"monitor\", \"moped\", \"mortar\", \"mortarboard\", \"mosque\", \"mosquito net\", \"motor scooter\", \"mountain bike\",\n \"mountain tent\", \"mouse\", \"mousetrap\", \"moving van\", \"muzzle\", \"nail\", \"neck brace\", \"necklace\",\n \"nipple\", \"notebook\", \"obelisk\", \"oboe\", \"ocarina\", \"odometer\", \"oil filter\", \"organ\", \"oscilloscope\",\n \"overskirt\", \"oxcart\", \"oxygen mask\", \"packet\", \"paddle\", \"paddlewheel\", \"padlock\", \"paintbrush\",\n \"pajama\", \"palace\", \"panpipe\", \"paper towel\", \"parachute\", \"parallel bars\", \"park bench\",\n \"parking meter\", \"passenger car\", \"patio\", \"pay-phone\", \"pedestal\", \"pencil box\", \"pencil sharpener\",\n \"perfume\", \"Petri dish\", \"photocopier\", \"pick\", \"pickelhaube\", \"picket fence\", \"pickup\", \"pier\",\n \"piggy bank\", \"pill bottle\", \"pillow\", \"ping-pong ball\", \"pinwheel\", \"pirate\", \"pitcher\", \"plane\",\n \"planetarium\", \"plastic bag\", \"plate rack\", \"plow\", \"plunger\", \"Polaroid camera\", \"pole\",\n \"police van\", \"poncho\", \"pool table\", \"pop bottle\", \"pot\", \"potter's wheel\", \"power drill\",\n \"prayer rug\", \"printer\", \"prison\", \"projectile\", \"projector\", \"puck\", \"punching bag\", \"purse\",\n \"quill\", \"quilt\", \"racer\", \"racket\", \"radiator\", \"radio\", \"radio telescope\", \"rain barrel\",\n \"recreational vehicle\", \"reel\", \"reflex camera\", \"refrigerator\", \"remote control\", \"restaurant\",\n \"revolver\", \"rifle\", \"rocking chair\", \"rotisserie\", \"rubber eraser\", \"rugby ball\", \"rule\",\n \"running shoe\", \"safe\", \"safety pin\", \"saltshaker\", \"sandal\", \"sarong\", \"sax\", \"scabbard\", \"scale\",\n \"school bus\", \"schooner\", \"scoreboard\", \"screen\", \"screw\", \"screwdriver\", \"seat belt\", \"sewing machine\",\n \"shield\", \"shoe shop\", \"shoji\", \"shopping basket\", \"shopping cart\", \"shovel\", \"shower cap\",\n \"shower curtain\", \"ski\", \"ski mask\", \"sleeping bag\", \"slide rule\", \"sliding door\", \"slot\", \"snorkel\",\n \"snowmobile\", \"snowplow\", \"soap dispenser\", \"soccer ball\", \"sock\", \"solar dish\", \"sombrero\",\n \"soup bowl\", \"space bar\", \"space heater\", \"space shuttle\", \"spatula\", \"speedboat\", \"spider web\",\n \"spindle\", \"sports car\", \"spotlight\", \"stage\", \"steam locomotive\", \"steel arch bridge\", \"steel drum\",\n \"stethoscope\", \"stole\", \"stone wall\", \"stopwatch\", \"stove\", \"strainer\", \"streetcar\", \"stretcher\",\n \"studio couch\", \"stupa\", \"submarine\", \"suit\", \"sundial\", \"sunglass\", \"sunglasses\", \"sunscreen\",\n \"suspension bridge\", \"swab\", \"sweatshirt\", \"swimming trunks\", \"swing\", \"switch\", \"syringe\",\n \"table lamp\", \"tank\", \"tape player\", \"teapot\", \"teddy\", \"television\", \"tennis ball\", \"thatch\",\n \"theater curtain\", \"thimble\", \"thresher\", \"throne\", \"tile roof\", \"toaster\", \"tobacco shop\",\n \"toilet seat\", \"torch\", \"totem pole\", \"tow truck\", \"toyshop\", \"tractor\", \"trailer truck\", \"tray\",\n \"trench coat\", \"tricycle\", \"trimaran\", \"tripod\", \"triumphal arch\", \"trolleybus\", \"trombone\", \"tub\",\n \"turnstile\", \"typewriter keyboard\", \"umbrella\", \"unicycle\", \"upright\", \"vacuum\", \"vase\", \"vault\",\n \"velvet\", \"vending machine\", \"vestment\", \"viaduct\", \"violin\", \"volleyball\", \"waffle iron\", \"wall clock\",\n \"wallet\", \"wardrobe\", \"warplane\", \"washbasin\", \"washer\", \"water bottle\", \"water jug\", \"water tower\",\n \"whiskey jug\", \"whistle\", \"wig\", \"window screen\", \"window shade\", \"Windsor tie\", \"wine bottle\", \"wing\",\n \"wok\", \"wooden spoon\", \"wool\", \"worm fence\", \"wreck\", \"yawl\", \"yurt\", \"web site\", \"comic book\",\n \"crossword puzzle\", \"street sign\", \"traffic light\", \"book jacket\", \"menu\", \"plate\", \"guacamole\",\n \"consomme\", \"hot pot\", \"trifle\", \"ice cream\", \"ice lolly\", \"French loaf\", \"bagel\", \"pretzel\",\n \"cheeseburger\", \"hotdog\", \"mashed potato\", \"head cabbage\", \"broccoli\", \"cauliflower\", \"zucchini\",\n \"spaghetti squash\", \"acorn squash\", \"butternut squash\", \"cucumber\", \"artichoke\", \"bell pepper\",\n \"cardoon\", \"mushroom\", \"Granny Smith\", \"strawberry\", \"orange\", \"lemon\", \"fig\", \"pineapple\", \"banana\",\n \"jackfruit\", \"custard apple\", \"pomegranate\", \"hay\", \"carbonara\", \"chocolate sauce\", \"dough\",\n \"meat loaf\", \"pizza\", \"potpie\", \"burrito\", \"red wine\", \"espresso\", \"cup\", \"eggnog\", \"alp\", \"bubble\",\n \"cliff\", \"coral reef\", \"geyser\", \"lakeside\", \"promontory\", \"sandbar\", \"seashore\", \"valley\", \"volcano\",\n \"ballplayer\", \"groom\", \"scuba diver\", \"rapeseed\", \"daisy\", \"yellow lady's slipper\", \"corn\", \"acorn\",\n \"hip\", \"buckeye\", \"coral fungus\", \"agaric\", \"gyromitra\", \"stinkhorn\", \"earthstar\", \"hen-of-the-woods\",\n \"bolete\", \"ear\", \"toilet tissue\"\n };\n\n cv::Mat image = bgr.clone();\n\n int y_offset = 0;\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f\\n\", obj.label, obj.prob);\n\n char text[256];\n sprintf(text, \"%4.1f%% %s\", obj.prob * 100, class_names[obj.label]);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = 0;\n int y = y_offset;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n\n y_offset += label_size.height;\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov8_cls(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov8_obb.cpp", "content": "// Copyright 2024 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolov8-obb torchscript\n// yolo export model=yolov8n-obb.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8n-obb.torchscript\n// 4. modify yolov8n_obb_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_137 = v_136.view(1, 1, 16384)\n// v_143 = v_142.view(1, 1, 4096)\n// v_149 = v_148.view(1, 1, 1024)\n// v_150 = torch.cat((v_137, v_143, v_149), dim=2)\n// ...\n// v_186 = v_163.view(1, 79, 16384)\n// v_187 = v_174.view(1, 79, 4096)\n// v_188 = v_185.view(1, 79, 1024)\n// v_189 = torch.cat((v_186, v_187, v_188), dim=2)\n// ...\n// after:\n// v_137 = v_136.view(1, 1, -1).transpose(1, 2)\n// v_143 = v_142.view(1, 1, -1).transpose(1, 2)\n// v_149 = v_148.view(1, 1, -1).transpose(1, 2)\n// v_150 = torch.cat((v_137, v_143, v_149), dim=1)\n// ...\n// v_186 = v_163.view(1, 79, -1).transpose(1, 2)\n// v_187 = v_174.view(1, 79, -1).transpose(1, 2)\n// v_188 = v_185.view(1, 79, -1).transpose(1, 2)\n// v_189 = torch.cat((v_186, v_187, v_188), dim=1)\n// return v_189, v_150\n// 5. re-export yolov8-obb torchscript\n// python3 -c 'import yolov8n_obb_pnnx; yolov8n_obb_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolov8n_obb_pnnx.py.pt inputshape=[1,3,1024,1024] inputshape2=[1,3,512,512]\n// 7. now you get ncnn model files\n// mv yolov8n_obb_pnnx.py.ncnn.param yolov8n_obb.ncnn.param\n// mv yolov8n_obb_pnnx.py.ncnn.bin yolov8n_obb.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=79 h=21504\n//\n// | bbox-reg 16 x 4 |score(15)|\n// +-----+-----+-----+-----+---------+\n// | dx0 | dy0 | dx1 | dy1 | 0.1 ... |\n// all /| | | | | ... |\n// boxes | .. | .. | .. | .. | 0.0 ... |\n// (21504)| | | | | . ... |\n// \\| | | | | . ... |\n// +-----+-----+-----+-----+---------+\n//\n\n// the out blob would be a 2-dim tensor with w=1 h=21504\n//\n// | degree(1)|\n// +----------+\n// | 0.1 |\n// all /| |\n// boxes | 0.0 |\n// (21504)| . |\n// \\| . |\n// +----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n\nstruct Object\n{\n cv::RotatedRect rrect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n std::vector intersection;\n cv::rotatedRectangleIntersection(a.rrect, b.rrect, intersection);\n if (intersection.empty())\n return 0.f;\n\n return cv::contourArea(intersection);\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rrect.size.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area;\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_angle, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 15 for DOTAv1\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n const float angle = sigmoid(pred_angle.row(y * num_grid_x + x)[0]) - 0.25f;\n\n const float angle_rad = angle * 3.14159265358979323846f;\n const float angle_degree = angle * 180.f;\n\n float cos = cosf(angle_rad);\n float sin = sinf(angle_rad);\n\n float xx = (pred_ltrb[2] - pred_ltrb[0]) * 0.5f;\n float yy = (pred_ltrb[3] - pred_ltrb[1]) * 0.5f;\n float xr = xx * cos - yy * sin;\n float yr = xx * sin + yy * cos;\n const float cx = pb_cx + xr;\n const float cy = pb_cy + yr;\n const float ww = pred_ltrb[2] + pred_ltrb[0];\n const float hh = pred_ltrb[3] + pred_ltrb[1];\n\n Object obj;\n obj.rrect = cv::RotatedRect(cv::Point2f(cx, cy), cv::Size_(ww, hh), angle_degree);\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_angle, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), pred_angle.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolov8_obb(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov8;\n\n yolov8.opt.use_vulkan_compute = true;\n // yolov8.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolov8/tree/master/app/src/main/assets\n yolov8.load_param(\"yolov8n_obb.ncnn.param\");\n yolov8.load_model(\"yolov8n_obb.ncnn.bin\");\n // yolov8.load_param(\"yolov8s_obb.ncnn.param\");\n // yolov8.load_model(\"yolov8s_obb.ncnn.bin\");\n // yolov8.load_param(\"yolov8m_obb.ncnn.param\");\n // yolov8.load_model(\"yolov8m_obb.ncnn.bin\");\n\n const int target_size = 1024;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolov8.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov8.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat out_angle;\n ex.extract(\"out1\", out_angle);\n\n std::vector proposals;\n generate_proposals(out, out_angle, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n Object obj = proposals[picked[i]];\n\n // adjust offset to original unpadded\n obj.rrect.center.x = (obj.rrect.center.x - (wpad / 2)) / scale;\n obj.rrect.center.y = (obj.rrect.center.y - (hpad / 2)) / scale;\n obj.rrect.size.width = (obj.rrect.size.width) / scale;\n obj.rrect.size.height = (obj.rrect.size.height) / scale;\n\n objects[i] = obj;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"plane\", \"ship\", \"storage tank\", \"baseball diamond\", \"tennis court\",\n \"basketball court\", \"ground track field\", \"harbor\", \"bridge\", \"large vehicle\",\n \"small vehicle\", \"helicopter\", \"roundabout\", \"soccer ball field\", \"swimming pool\"\n };\n\n static const cv::Scalar colors[] = {\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[obj.label];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f @ %.2f\\n\", obj.label, obj.prob,\n obj.rrect.center.x, obj.rrect.center.y, obj.rrect.size.width, obj.rrect.size.height, obj.rrect.angle);\n\n cv::Point2f corners[4];\n obj.rrect.points(corners);\n cv::line(image, corners[0], corners[1], color);\n cv::line(image, corners[1], corners[2], color);\n cv::line(image, corners[2], corners[3], color);\n cv::line(image, corners[3], corners[0], color);\n }\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[obj.label];\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rrect.center.x - label_size.width / 2;\n int y = obj.rrect.center.y - label_size.height / 2 - baseLine;\n if (y < 0)\n y = 0;\n if (y + label_size.height > image.rows)\n y = image.rows - label_size.height;\n if (x < 0)\n x = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov8_obb(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov8_pose.cpp", "content": "// Copyright 2024 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolov8-pose torchscript\n// yolo export model=yolov8n-pose.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8n-pose.torchscript\n// 4. modify yolov8n_pose_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_137 = v_136.view(1, 51, 6400)\n// v_143 = v_142.view(1, 51, 1600)\n// v_149 = v_148.view(1, 51, 400)\n// v_150 = torch.cat((v_137, v_143, v_149), dim=-1)\n// ...\n// v_184 = v_161.view(1, 65, 6400)\n// v_185 = v_172.view(1, 65, 1600)\n// v_186 = v_183.view(1, 65, 400)\n// v_187 = torch.cat((v_184, v_185, v_186), dim=2)\n// ...\n// after:\n// v_137 = v_136.view(1, 51, -1).transpose(1, 2)\n// v_143 = v_142.view(1, 51, -1).transpose(1, 2)\n// v_149 = v_148.view(1, 51, -1).transpose(1, 2)\n// v_150 = torch.cat((v_137, v_143, v_149), dim=1)\n// ...\n// v_184 = v_161.view(1, 65, -1).transpose(1, 2)\n// v_185 = v_172.view(1, 65, -1).transpose(1, 2)\n// v_186 = v_183.view(1, 65, -1).transpose(1, 2)\n// v_187 = torch.cat((v_184, v_185, v_186), dim=1)\n// return v_187, v_150\n// 5. re-export yolov8-pose torchscript\n// python3 -c 'import yolov8n_pose_pnnx; yolov8n_pose_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolov8n_pose_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolov8n_pose_pnnx.py.ncnn.param yolov8n_pose.ncnn.param\n// mv yolov8n_pose_pnnx.py.ncnn.bin yolov8n_pose.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=65 h=8400\n//\n// | bbox-reg 16 x 4 |score(1)|\n// +-----+-----+-----+-----+--------+\n// | dx0 | dy0 | dx1 | dy1 | 0.1 |\n// all /| | | | | |\n// boxes | .. | .. | .. | .. | 0.0 |\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+--------+\n//\n\n//\n// | pose (51) |\n// +-----------+\n// |0.1........|\n// all /| |\n// boxes |0.0........|\n// (8400)| . |\n// \\| . |\n// +-----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct KeyPoint\n{\n cv::Point2f p;\n float prob;\n};\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n std::vector keypoints;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_points, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_points = pred_points.w / 3;\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n const ncnn::Mat pred_points_grid = pred_points.row_range(y * num_grid_x + x, 1).reshape(3, num_points);\n\n // find label with max score\n int label = 0;\n float score = sigmoid(pred_grid[reg_max_1 * 4]);\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n std::vector keypoints;\n for (int k = 0; k < num_points; k++)\n {\n KeyPoint keypoint;\n keypoint.p.x = (x + pred_points_grid.row(k)[0] * 2) * stride;\n keypoint.p.y = (y + pred_points_grid.row(k)[1] * 2) * stride;\n keypoint.prob = sigmoid(pred_points_grid.row(k)[2]);\n keypoints.push_back(keypoint);\n }\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n obj.keypoints = keypoints;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const ncnn::Mat& pred_points, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n generate_proposals(pred.row_range(pred_row_offset, num_grid), pred_points.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects);\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolov8_pose(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov8;\n\n yolov8.opt.use_vulkan_compute = true;\n // yolov8.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolov8/tree/master/app/src/main/assets\n yolov8.load_param(\"yolov8n_pose.ncnn.param\");\n yolov8.load_model(\"yolov8n_pose.ncnn.bin\");\n // yolov8.load_param(\"yolov8s_pose.ncnn.param\");\n // yolov8.load_model(\"yolov8s_pose.ncnn.bin\");\n // yolov8.load_param(\"yolov8m_pose.ncnn.param\");\n // yolov8.load_model(\"yolov8m_pose.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n const float mask_threshold = 0.5f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolov8.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov8.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n ncnn::Mat out_points;\n ex.extract(\"out1\", out_points);\n\n std::vector proposals;\n generate_proposals(out, out_points, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n const int num_points = out_points.w / 3;\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n for (int j = 0; j < num_points; j++)\n {\n objects[i].keypoints[j].p.x = (objects[i].keypoints[j].p.x - (wpad / 2)) / scale;\n objects[i].keypoints[j].p.y = (objects[i].keypoints[j].p.y - (hpad / 2)) / scale;\n }\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\"person\"};\n\n static const cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n // draw bone\n static const int joint_pairs[16][2] = {\n {0, 1}, {1, 3}, {0, 2}, {2, 4}, {5, 6}, {5, 7}, {7, 9}, {6, 8}, {8, 10}, {5, 11}, {6, 12}, {11, 12}, {11, 13}, {12, 14}, {13, 15}, {14, 16}\n };\n static const cv::Scalar bone_colors[] = {\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(0, 255, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 128, 0),\n cv::Scalar(255, 51, 255),\n cv::Scalar(255, 51, 255),\n cv::Scalar(255, 51, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n cv::Scalar(51, 153, 255),\n };\n\n for (int j = 0; j < 16; j++)\n {\n const KeyPoint& p1 = obj.keypoints[joint_pairs[j][0]];\n const KeyPoint& p2 = obj.keypoints[joint_pairs[j][1]];\n\n if (p1.prob < 0.2f || p2.prob < 0.2f)\n continue;\n\n cv::line(image, p1.p, p2.p, bone_colors[j], 2);\n }\n\n // draw joint\n for (size_t j = 0; j < obj.keypoints.size(); j++)\n {\n const KeyPoint& keypoint = obj.keypoints[j];\n\n fprintf(stderr, \"%.2f %.2f = %.5f\\n\", keypoint.p.x, keypoint.p.y, keypoint.prob);\n\n if (keypoint.prob < 0.2f)\n continue;\n\n cv::circle(image, keypoint.p, 3, color, -1);\n }\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov8_pose(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolov8_seg.cpp", "content": "// Copyright 2024 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yolov8-seg torchscript\n// yolo export model=yolov8n-seg.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8n-seg.torchscript\n// 4. modify yolov8n_seg_pnnx.py for dynamic shape inference\n// A. modify reshape to support dynamic image sizes\n// B. permute tensor before concat and adjust concat axis\n// C. drop post-process part\n// before:\n// v_144 = v_143.view(1, 32, 6400)\n// v_150 = v_149.view(1, 32, 1600)\n// v_156 = v_155.view(1, 32, 400)\n// v_157 = torch.cat((v_144, v_150, v_156), dim=2)\n// ...\n// v_191 = v_168.view(1, 144, 6400)\n// v_192 = v_179.view(1, 144, 1600)\n// v_193 = v_190.view(1, 144, 400)\n// v_194 = torch.cat((v_191, v_192, v_193), dim=2)\n// ...\n// v_215 = (v_214, v_138, )\n// return v_215\n// after:\n// v_144 = v_143.view(1, 32, -1).transpose(1, 2)\n// v_150 = v_149.view(1, 32, -1).transpose(1, 2)\n// v_156 = v_155.view(1, 32, -1).transpose(1, 2)\n// v_157 = torch.cat((v_144, v_150, v_156), dim=1)\n// ...\n// v_191 = v_168.view(1, 144, -1).transpose(1, 2)\n// v_192 = v_179.view(1, 144, -1).transpose(1, 2)\n// v_193 = v_190.view(1, 144, -1).transpose(1, 2)\n// v_194 = torch.cat((v_191, v_192, v_193), dim=1)\n// return v_194, v_157, v_138\n// 5. re-export yolov8-seg torchscript\n// python3 -c 'import yolov8n_seg_pnnx; yolov8n_seg_pnnx.export_torchscript()'\n// 6. convert new torchscript with dynamic shape\n// pnnx yolov8n_seg_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]\n// 7. now you get ncnn model files\n// mv yolov8n_seg_pnnx.py.ncnn.param yolov8n_seg.ncnn.param\n// mv yolov8n_seg_pnnx.py.ncnn.bin yolov8n_seg.ncnn.bin\n\n// the out blob would be a 2-dim tensor with w=176 h=8400\n//\n// | bbox-reg 16 x 4 | per-class scores(80) |\n// +-----+-----+-----+-----+----------------------+\n// | dx0 | dy0 | dx1 | dy1 |0.1 0.0 0.0 0.5 ......|\n// all /| | | | | . |\n// boxes | .. | .. | .. | .. |0.0 0.9 0.0 0.0 ......|\n// (8400)| | | | | . |\n// \\| | | | | . |\n// +-----+-----+-----+-----+----------------------+\n//\n\n//\n// | mask (32) |\n// +-----------+\n// |0.1........|\n// all /| |\n// boxes |0.0........|\n// (8400)| . |\n// \\| . |\n// +-----------+\n//\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n int gindex;\n cv::Mat mask;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic inline float sigmoid(float x)\n{\n return 1.0f / (1.0f + expf(-x));\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, int stride, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n\n const int reg_max_1 = 16;\n const int num_class = pred.w - reg_max_1 * 4; // number of classes. 80 for COCO\n\n for (int y = 0; y < num_grid_y; y++)\n {\n for (int x = 0; x < num_grid_x; x++)\n {\n const ncnn::Mat pred_grid = pred.row_range(y * num_grid_x + x, 1);\n\n // find label with max score\n int label = -1;\n float score = -FLT_MAX;\n {\n const ncnn::Mat pred_score = pred_grid.range(reg_max_1 * 4, num_class);\n\n for (int k = 0; k < num_class; k++)\n {\n float s = pred_score[k];\n if (s > score)\n {\n label = k;\n score = s;\n }\n }\n\n score = sigmoid(score);\n }\n\n if (score >= prob_threshold)\n {\n ncnn::Mat pred_bbox = pred_grid.range(0, reg_max_1 * 4).reshape(reg_max_1, 4).clone();\n\n {\n ncnn::Layer* softmax = ncnn::create_layer(\"Softmax\");\n\n ncnn::ParamDict pd;\n pd.set(0, 1); // axis\n pd.set(1, 1);\n softmax->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n softmax->create_pipeline(opt);\n\n softmax->forward_inplace(pred_bbox, opt);\n\n softmax->destroy_pipeline(opt);\n\n delete softmax;\n }\n\n float pred_ltrb[4];\n for (int k = 0; k < 4; k++)\n {\n float dis = 0.f;\n const float* dis_after_sm = pred_bbox.row(k);\n for (int l = 0; l < reg_max_1; l++)\n {\n dis += l * dis_after_sm[l];\n }\n\n pred_ltrb[k] = dis * stride;\n }\n\n float pb_cx = (x + 0.5f) * stride;\n float pb_cy = (y + 0.5f) * stride;\n\n float x0 = pb_cx - pred_ltrb[0];\n float y0 = pb_cy - pred_ltrb[1];\n float x1 = pb_cx + pred_ltrb[2];\n float y1 = pb_cy + pred_ltrb[3];\n\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = x1 - x0;\n obj.rect.height = y1 - y0;\n obj.label = label;\n obj.prob = score;\n obj.gindex = y * num_grid_x + x;\n\n objects.push_back(obj);\n }\n }\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, const std::vector& strides, const ncnn::Mat& in_pad, float prob_threshold, std::vector& objects)\n{\n const int w = in_pad.w;\n const int h = in_pad.h;\n\n int pred_row_offset = 0;\n for (size_t i = 0; i < strides.size(); i++)\n {\n const int stride = strides[i];\n\n const int num_grid_x = w / stride;\n const int num_grid_y = h / stride;\n const int num_grid = num_grid_x * num_grid_y;\n\n std::vector objects_stride;\n generate_proposals(pred.row_range(pred_row_offset, num_grid), stride, in_pad, prob_threshold, objects_stride);\n\n for (size_t j = 0; j < objects_stride.size(); j++)\n {\n Object obj = objects_stride[j];\n obj.gindex += pred_row_offset;\n objects.push_back(obj);\n }\n\n pred_row_offset += num_grid;\n }\n}\n\nstatic int detect_yolov8_seg(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolov8;\n\n yolov8.opt.use_vulkan_compute = true;\n // yolov8.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-android-yolov8/tree/master/app/src/main/assets\n yolov8.load_param(\"yolov8n_seg.ncnn.param\");\n yolov8.load_model(\"yolov8n_seg.ncnn.bin\");\n // yolov8.load_param(\"yolov8s_seg.ncnn.param\");\n // yolov8.load_model(\"yolov8s_seg.ncnn.bin\");\n // yolov8.load_param(\"yolov8m_seg.ncnn.param\");\n // yolov8.load_model(\"yolov8m_seg.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n const float mask_threshold = 0.5f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // ultralytics/cfg/models/v8/yolov8.yaml\n std::vector strides(3);\n strides[0] = 8;\n strides[1] = 16;\n strides[2] = 32;\n const int max_stride = 32;\n\n // letterbox pad to multiple of max_stride\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = (w + max_stride - 1) / max_stride * max_stride - w;\n int hpad = (h + max_stride - 1) / max_stride * max_stride - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yolov8.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n std::vector proposals;\n generate_proposals(out, strides, in_pad, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n if (count == 0)\n return 0;\n\n ncnn::Mat mask_feat;\n ex.extract(\"out1\", mask_feat);\n\n ncnn::Mat mask_protos;\n ex.extract(\"out2\", mask_protos);\n\n ncnn::Mat objects_mask_feat(mask_feat.w, 1, count);\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n\n // pick mask feat\n memcpy(objects_mask_feat.channel(i), mask_feat.row(objects[i].gindex), mask_feat.w * sizeof(float));\n }\n\n // process mask\n ncnn::Mat objects_mask;\n {\n ncnn::Layer* gemm = ncnn::create_layer(\"Gemm\");\n\n ncnn::ParamDict pd;\n pd.set(6, 1); // constantC\n pd.set(7, count); // constantM\n pd.set(8, mask_protos.w * mask_protos.h); // constantN\n pd.set(9, mask_feat.w); // constantK\n pd.set(10, -1); // constant_broadcast_type_C\n pd.set(11, 1); // output_N1M\n gemm->load_param(pd);\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n gemm->create_pipeline(opt);\n\n std::vector gemm_inputs(2);\n gemm_inputs[0] = objects_mask_feat;\n gemm_inputs[1] = mask_protos.reshape(mask_protos.w * mask_protos.h, 1, mask_protos.c);\n std::vector gemm_outputs(1);\n gemm->forward(gemm_inputs, gemm_outputs, opt);\n objects_mask = gemm_outputs[0].reshape(mask_protos.w, mask_protos.h, count);\n\n gemm->destroy_pipeline(opt);\n\n delete gemm;\n }\n {\n ncnn::Layer* sigmoid = ncnn::create_layer(\"Sigmoid\");\n\n ncnn::Option opt;\n opt.num_threads = 1;\n opt.use_packing_layout = false;\n\n sigmoid->create_pipeline(opt);\n\n sigmoid->forward_inplace(objects_mask, opt);\n\n sigmoid->destroy_pipeline(opt);\n\n delete sigmoid;\n }\n\n // resize mask map\n {\n ncnn::Mat objects_mask_resized;\n ncnn::resize_bilinear(objects_mask, objects_mask_resized, in_pad.w / scale, in_pad.h / scale);\n objects_mask = objects_mask_resized;\n }\n\n // create per-object mask\n for (int i = 0; i < count; i++)\n {\n Object& obj = objects[i];\n\n const ncnn::Mat mm = objects_mask.channel(i);\n\n obj.mask = cv::Mat((int)obj.rect.height, (int)obj.rect.width, CV_8UC1);\n\n // adjust offset to original unpadded and clip inside object box\n for (int y = 0; y < (int)obj.rect.height; y++)\n {\n const float* pmm = mm.row((int)(hpad / 2 / scale + obj.rect.y + y)) + (int)(wpad / 2 / scale + obj.rect.x);\n uchar* pmask = obj.mask.ptr(y);\n for (int x = 0; x < (int)obj.rect.width; x++)\n {\n pmask[x] = pmm[x] > mask_threshold ? 1 : 0;\n }\n }\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n for (int y = 0; y < (int)obj.rect.height; y++)\n {\n const uchar* maskptr = obj.mask.ptr(y);\n uchar* bgrptr = image.ptr((int)obj.rect.y + y) + (int)obj.rect.x * 3;\n for (int x = 0; x < (int)obj.rect.width; x++)\n {\n if (maskptr[x])\n {\n bgrptr[0] = bgrptr[0] * 0.5 + color[0] * 0.5;\n bgrptr[1] = bgrptr[1] * 0.5 + color[1] * 0.5;\n bgrptr[2] = bgrptr[2] * 0.5 + color[2] * 0.5;\n }\n bgrptr += 3;\n }\n }\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolov8_seg(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yoloworld.cpp", "content": "// Copyright 2025 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n// 1. install\n// pip3 install -U ultralytics pnnx ncnn\n// 2. export yoloworld torchscript\n// yolo export model=yolov8s-world.pt format=torchscript\n// yolo export model=yolov8m-world.pt format=torchscript\n// yolo export model=yolov8l-world.pt format=torchscript\n// yolo export model=yolov8x-world.pt format=torchscript\n// yolo export model=yolov8s-worldv2.pt format=torchscript\n// yolo export model=yolov8m-worldv2.pt format=torchscript\n// yolo export model=yolov8l-worldv2.pt format=torchscript\n// yolo export model=yolov8x-worldv2.pt format=torchscript\n// 3. convert torchscript with static shape\n// pnnx yolov8s-world.torchscript\n// pnnx yolov8m-world.torchscript\n// pnnx yolov8l-world.torchscript\n// pnnx yolov8x-world.torchscript\n// pnnx yolov8s-worldv2.torchscript\n// pnnx yolov8m-worldv2.torchscript\n// pnnx yolov8l-worldv2.torchscript\n// pnnx yolov8x-worldv2.torchscript\n\n// the out blob would be a 2-dim tensor with w=8400 h=84\n//\n// | all boxes (8400) |\n// +-------------------------+\n// | center-x . |\n// bbox | center-y . |\n// | w . |\n// | h . |\n// +-------------------------+\n// | 0.1 . |\n// per | 0.0 . |\n// class | 0.5 . |\n// scores | . . |\n// (80) | . . |\n// +-------------------------+\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& objects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = objects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (objects[i].prob > p)\n i++;\n\n while (objects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(objects[i], objects[j]);\n\n i++;\n j--;\n }\n }\n\n // #pragma omp parallel sections\n {\n // #pragma omp section\n {\n if (left < j) qsort_descent_inplace(objects, left, j);\n }\n // #pragma omp section\n {\n if (i < right) qsort_descent_inplace(objects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& objects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = objects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = objects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = objects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = objects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic void generate_proposals(const ncnn::Mat& pred, float prob_threshold, std::vector& objects)\n{\n const int num_boxes = pred.w;\n const int num_class = pred.h - 4;\n\n const ncnn::Mat pred_bbox = pred.row_range(0, 4);\n const ncnn::Mat pred_score = pred.row_range(4, num_class);\n\n for (int i = 0; i < num_boxes; i++)\n {\n int label = 0;\n float score = -9999.f;\n for (int j = 0; j < num_class; j++)\n {\n const float prob = pred_score.row(j)[i];\n if (prob > score)\n {\n score = prob;\n label = j;\n }\n }\n\n if (score >= prob_threshold)\n {\n const float cx = pred_bbox.row(0)[i];\n const float cy = pred_bbox.row(1)[i];\n const float w = pred_bbox.row(2)[i];\n const float h = pred_bbox.row(3)[i];\n\n Object obj;\n obj.rect.x = cx - w / 2;\n obj.rect.y = cy - h / 2;\n obj.rect.width = w;\n obj.rect.height = h;\n obj.label = label;\n obj.prob = score;\n\n objects.push_back(obj);\n }\n }\n}\n\nstatic int detect_yoloworld(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yoloworld;\n\n yoloworld.opt.use_vulkan_compute = true;\n // yoloworld.opt.use_bf16_storage = true;\n\n // https://github.com/nihui/ncnn-assets/tree/master/models\n // yoloworld.load_param(\"yolov8s_world.ncnn.param\");\n // yoloworld.load_model(\"yolov8s_world.ncnn.bin\");\n // yoloworld.load_param(\"yolov8m_world.ncnn.param\");\n // yoloworld.load_model(\"yolov8m_world.ncnn.bin\");\n // yoloworld.load_param(\"yolov8l_world.ncnn.param\");\n // yoloworld.load_model(\"yolov8l_world.ncnn.bin\");\n // yoloworld.load_param(\"yolov8x_world.ncnn.param\");\n // yoloworld.load_model(\"yolov8x_world.ncnn.bin\");\n yoloworld.load_param(\"yolov8s_worldv2.ncnn.param\");\n yoloworld.load_model(\"yolov8s_worldv2.ncnn.bin\");\n // yoloworld.load_param(\"yolov8m_worldv2.ncnn.param\");\n // yoloworld.load_model(\"yolov8m_worldv2.ncnn.bin\");\n // yoloworld.load_param(\"yolov8l_worldv2.ncnn.param\");\n // yoloworld.load_model(\"yolov8l_worldv2.ncnn.bin\");\n // yoloworld.load_param(\"yolov8x_worldv2.ncnn.param\");\n // yoloworld.load_model(\"yolov8x_worldv2.ncnn.bin\");\n\n const int target_size = 640;\n const float prob_threshold = 0.25f;\n const float nms_threshold = 0.45f;\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n // letterbox pad\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)target_size / w;\n w = target_size;\n h = h * scale;\n }\n else\n {\n scale = (float)target_size / h;\n h = target_size;\n w = w * scale;\n }\n\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR2RGB, img_w, img_h, w, h);\n\n // letterbox pad to target_size rectangle\n int wpad = target_size - w;\n int hpad = target_size - h;\n ncnn::Mat in_pad;\n ncnn::copy_make_border(in, in_pad, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, ncnn::BORDER_CONSTANT, 114.f);\n\n const float norm_vals[3] = {1 / 255.f, 1 / 255.f, 1 / 255.f};\n in_pad.substract_mean_normalize(0, norm_vals);\n\n ncnn::Extractor ex = yoloworld.create_extractor();\n\n ex.input(\"in0\", in_pad);\n\n ncnn::Mat out;\n ex.extract(\"out0\", out);\n\n std::vector proposals;\n generate_proposals(out, prob_threshold, proposals);\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, nms_threshold);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x - (wpad / 2)) / scale;\n float y0 = (objects[i].rect.y - (hpad / 2)) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width - (wpad / 2)) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height - (hpad / 2)) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n static cv::Scalar colors[] = {\n cv::Scalar(244, 67, 54),\n cv::Scalar(233, 30, 99),\n cv::Scalar(156, 39, 176),\n cv::Scalar(103, 58, 183),\n cv::Scalar(63, 81, 181),\n cv::Scalar(33, 150, 243),\n cv::Scalar(3, 169, 244),\n cv::Scalar(0, 188, 212),\n cv::Scalar(0, 150, 136),\n cv::Scalar(76, 175, 80),\n cv::Scalar(139, 195, 74),\n cv::Scalar(205, 220, 57),\n cv::Scalar(255, 235, 59),\n cv::Scalar(255, 193, 7),\n cv::Scalar(255, 152, 0),\n cv::Scalar(255, 87, 34),\n cv::Scalar(121, 85, 72),\n cv::Scalar(158, 158, 158),\n cv::Scalar(96, 125, 139)\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n const cv::Scalar& color = colors[i % 19];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, color);\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yoloworld(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "examples/yolox.cpp", "content": "// Copyright 2020 Tencent\n// Copyright 2020-2021 Megvii Inc.\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \"layer.h\"\n#include \"net.h\"\n\n#if defined(USE_NCNN_SIMPLEOCV)\n#include \"simpleocv.h\"\n#else\n#include \n#include \n#include \n#endif\n#include \n#include \n#include \n\n#define YOLOX_NMS_THRESH 0.45 // nms threshold\n#define YOLOX_CONF_THRESH 0.25 // threshold of bounding box prob\n#define YOLOX_TARGET_SIZE 640 // target image size after resize, might use 416 for small model\n\n// YOLOX use the same focus in yolov5\nclass YoloV5Focus : public ncnn::Layer\n{\npublic:\n YoloV5Focus()\n {\n one_blob_only = true;\n }\n\n virtual int forward(const ncnn::Mat& bottom_blob, ncnn::Mat& top_blob, const ncnn::Option& opt) const\n {\n int w = bottom_blob.w;\n int h = bottom_blob.h;\n int channels = bottom_blob.c;\n\n int outw = w / 2;\n int outh = h / 2;\n int outc = channels * 4;\n\n top_blob.create(outw, outh, outc, 4u, 1, opt.blob_allocator);\n if (top_blob.empty())\n return -100;\n\n #pragma omp parallel for num_threads(opt.num_threads)\n for (int p = 0; p < outc; p++)\n {\n const float* ptr = bottom_blob.channel(p % channels).row((p / channels) % 2) + ((p / channels) / 2);\n float* outptr = top_blob.channel(p);\n\n for (int i = 0; i < outh; i++)\n {\n for (int j = 0; j < outw; j++)\n {\n *outptr = *ptr;\n\n outptr += 1;\n ptr += 2;\n }\n\n ptr += w;\n }\n }\n\n return 0;\n }\n};\n\nDEFINE_LAYER_CREATOR(YoloV5Focus)\n\nstruct Object\n{\n cv::Rect_ rect;\n int label;\n float prob;\n};\n\nstruct GridAndStride\n{\n int grid0;\n int grid1;\n int stride;\n};\n\nstatic inline float intersection_area(const Object& a, const Object& b)\n{\n cv::Rect_ inter = a.rect & b.rect;\n return inter.area();\n}\n\nstatic void qsort_descent_inplace(std::vector& faceobjects, int left, int right)\n{\n int i = left;\n int j = right;\n float p = faceobjects[(left + right) / 2].prob;\n\n while (i <= j)\n {\n while (faceobjects[i].prob > p)\n i++;\n\n while (faceobjects[j].prob < p)\n j--;\n\n if (i <= j)\n {\n // swap\n std::swap(faceobjects[i], faceobjects[j]);\n\n i++;\n j--;\n }\n }\n\n #pragma omp parallel sections\n {\n #pragma omp section\n {\n if (left < j) qsort_descent_inplace(faceobjects, left, j);\n }\n #pragma omp section\n {\n if (i < right) qsort_descent_inplace(faceobjects, i, right);\n }\n }\n}\n\nstatic void qsort_descent_inplace(std::vector& objects)\n{\n if (objects.empty())\n return;\n\n qsort_descent_inplace(objects, 0, objects.size() - 1);\n}\n\nstatic void nms_sorted_bboxes(const std::vector& faceobjects, std::vector& picked, float nms_threshold, bool agnostic = false)\n{\n picked.clear();\n\n const int n = faceobjects.size();\n\n std::vector areas(n);\n for (int i = 0; i < n; i++)\n {\n areas[i] = faceobjects[i].rect.area();\n }\n\n for (int i = 0; i < n; i++)\n {\n const Object& a = faceobjects[i];\n\n int keep = 1;\n for (int j = 0; j < (int)picked.size(); j++)\n {\n const Object& b = faceobjects[picked[j]];\n\n if (!agnostic && a.label != b.label)\n continue;\n\n // intersection over union\n float inter_area = intersection_area(a, b);\n float union_area = areas[i] + areas[picked[j]] - inter_area;\n // float IoU = inter_area / union_area\n if (inter_area / union_area > nms_threshold)\n keep = 0;\n }\n\n if (keep)\n picked.push_back(i);\n }\n}\n\nstatic void generate_grids_and_stride(const int target_w, const int target_h, std::vector& strides, std::vector& grid_strides)\n{\n for (int i = 0; i < (int)strides.size(); i++)\n {\n int stride = strides[i];\n int num_grid_w = target_w / stride;\n int num_grid_h = target_h / stride;\n for (int g1 = 0; g1 < num_grid_h; g1++)\n {\n for (int g0 = 0; g0 < num_grid_w; g0++)\n {\n GridAndStride gs;\n gs.grid0 = g0;\n gs.grid1 = g1;\n gs.stride = stride;\n grid_strides.push_back(gs);\n }\n }\n }\n}\n\nstatic void generate_yolox_proposals(std::vector grid_strides, const ncnn::Mat& feat_blob, float prob_threshold, std::vector& objects)\n{\n const int num_grid = feat_blob.h;\n const int num_class = feat_blob.w - 5;\n const int num_anchors = grid_strides.size();\n\n const float* feat_ptr = feat_blob.channel(0);\n for (int anchor_idx = 0; anchor_idx < num_anchors; anchor_idx++)\n {\n const int grid0 = grid_strides[anchor_idx].grid0;\n const int grid1 = grid_strides[anchor_idx].grid1;\n const int stride = grid_strides[anchor_idx].stride;\n\n // yolox/models/yolo_head.py decode logic\n // outputs[..., :2] = (outputs[..., :2] + grids) * strides\n // outputs[..., 2:4] = torch.exp(outputs[..., 2:4]) * strides\n float x_center = (feat_ptr[0] + grid0) * stride;\n float y_center = (feat_ptr[1] + grid1) * stride;\n float w = exp(feat_ptr[2]) * stride;\n float h = exp(feat_ptr[3]) * stride;\n float x0 = x_center - w * 0.5f;\n float y0 = y_center - h * 0.5f;\n\n float box_objectness = feat_ptr[4];\n for (int class_idx = 0; class_idx < num_class; class_idx++)\n {\n float box_cls_score = feat_ptr[5 + class_idx];\n float box_prob = box_objectness * box_cls_score;\n if (box_prob > prob_threshold)\n {\n Object obj;\n obj.rect.x = x0;\n obj.rect.y = y0;\n obj.rect.width = w;\n obj.rect.height = h;\n obj.label = class_idx;\n obj.prob = box_prob;\n\n objects.push_back(obj);\n }\n\n } // class loop\n feat_ptr += feat_blob.w;\n\n } // point anchor loop\n}\n\nstatic int detect_yolox(const cv::Mat& bgr, std::vector& objects)\n{\n ncnn::Net yolox;\n\n yolox.opt.use_vulkan_compute = true;\n // yolox.opt.use_bf16_storage = true;\n\n // Focus in yolov5\n yolox.register_custom_layer(\"YoloV5Focus\", YoloV5Focus_layer_creator);\n\n // original pretrained model from https://github.com/Megvii-BaseDetection/YOLOX\n // ncnn model param: https://github.com/Megvii-BaseDetection/YOLOX/releases/download/0.1.1rc0/yolox_s_ncnn.tar.gz\n // NOTE that newest version YOLOX remove normalization of model (minus mean and then div by std),\n // which might cause your model outputs becoming a total mess, plz check carefully.\n if (yolox.load_param(\"yolox.param\"))\n exit(-1);\n if (yolox.load_model(\"yolox.bin\"))\n exit(-1);\n\n int img_w = bgr.cols;\n int img_h = bgr.rows;\n\n int w = img_w;\n int h = img_h;\n float scale = 1.f;\n if (w > h)\n {\n scale = (float)YOLOX_TARGET_SIZE / w;\n w = YOLOX_TARGET_SIZE;\n h = h * scale;\n }\n else\n {\n scale = (float)YOLOX_TARGET_SIZE / h;\n h = YOLOX_TARGET_SIZE;\n w = w * scale;\n }\n ncnn::Mat in = ncnn::Mat::from_pixels_resize(bgr.data, ncnn::Mat::PIXEL_BGR, img_w, img_h, w, h);\n\n // pad to YOLOX_TARGET_SIZE rectangle\n int wpad = (w + 31) / 32 * 32 - w;\n int hpad = (h + 31) / 32 * 32 - h;\n ncnn::Mat in_pad;\n // different from yolov5, yolox only pad on bottom and right side,\n // which means users don't need to extra padding info to decode boxes coordinate.\n ncnn::copy_make_border(in, in_pad, 0, hpad, 0, wpad, ncnn::BORDER_CONSTANT, 114.f);\n\n ncnn::Extractor ex = yolox.create_extractor();\n\n ex.input(\"images\", in_pad);\n\n std::vector proposals;\n\n {\n ncnn::Mat out;\n ex.extract(\"output\", out);\n\n static const int stride_arr[] = {8, 16, 32}; // might have stride=64 in YOLOX\n std::vector strides(stride_arr, stride_arr + sizeof(stride_arr) / sizeof(stride_arr[0]));\n std::vector grid_strides;\n generate_grids_and_stride(in_pad.w, in_pad.h, strides, grid_strides);\n generate_yolox_proposals(grid_strides, out, YOLOX_CONF_THRESH, proposals);\n }\n\n // sort all proposals by score from highest to lowest\n qsort_descent_inplace(proposals);\n\n // apply nms with nms_threshold\n std::vector picked;\n nms_sorted_bboxes(proposals, picked, YOLOX_NMS_THRESH);\n\n int count = picked.size();\n\n objects.resize(count);\n for (int i = 0; i < count; i++)\n {\n objects[i] = proposals[picked[i]];\n\n // adjust offset to original unpadded\n float x0 = (objects[i].rect.x) / scale;\n float y0 = (objects[i].rect.y) / scale;\n float x1 = (objects[i].rect.x + objects[i].rect.width) / scale;\n float y1 = (objects[i].rect.y + objects[i].rect.height) / scale;\n\n // clip\n x0 = std::max(std::min(x0, (float)(img_w - 1)), 0.f);\n y0 = std::max(std::min(y0, (float)(img_h - 1)), 0.f);\n x1 = std::max(std::min(x1, (float)(img_w - 1)), 0.f);\n y1 = std::max(std::min(y1, (float)(img_h - 1)), 0.f);\n\n objects[i].rect.x = x0;\n objects[i].rect.y = y0;\n objects[i].rect.width = x1 - x0;\n objects[i].rect.height = y1 - y0;\n }\n\n return 0;\n}\n\nstatic void draw_objects(const cv::Mat& bgr, const std::vector& objects)\n{\n static const char* class_names[] = {\n \"person\", \"bicycle\", \"car\", \"motorcycle\", \"airplane\", \"bus\", \"train\", \"truck\", \"boat\", \"traffic light\",\n \"fire hydrant\", \"stop sign\", \"parking meter\", \"bench\", \"bird\", \"cat\", \"dog\", \"horse\", \"sheep\", \"cow\",\n \"elephant\", \"bear\", \"zebra\", \"giraffe\", \"backpack\", \"umbrella\", \"handbag\", \"tie\", \"suitcase\", \"frisbee\",\n \"skis\", \"snowboard\", \"sports ball\", \"kite\", \"baseball bat\", \"baseball glove\", \"skateboard\", \"surfboard\",\n \"tennis racket\", \"bottle\", \"wine glass\", \"cup\", \"fork\", \"knife\", \"spoon\", \"bowl\", \"banana\", \"apple\",\n \"sandwich\", \"orange\", \"broccoli\", \"carrot\", \"hot dog\", \"pizza\", \"donut\", \"cake\", \"chair\", \"couch\",\n \"potted plant\", \"bed\", \"dining table\", \"toilet\", \"tv\", \"laptop\", \"mouse\", \"remote\", \"keyboard\", \"cell phone\",\n \"microwave\", \"oven\", \"toaster\", \"sink\", \"refrigerator\", \"book\", \"clock\", \"vase\", \"scissors\", \"teddy bear\",\n \"hair drier\", \"toothbrush\"\n };\n\n cv::Mat image = bgr.clone();\n\n for (size_t i = 0; i < objects.size(); i++)\n {\n const Object& obj = objects[i];\n\n fprintf(stderr, \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\", obj.label, obj.prob,\n obj.rect.x, obj.rect.y, obj.rect.width, obj.rect.height);\n\n cv::rectangle(image, obj.rect, cv::Scalar(255, 0, 0));\n\n char text[256];\n sprintf(text, \"%s %.1f%%\", class_names[obj.label], obj.prob * 100);\n\n int baseLine = 0;\n cv::Size label_size = cv::getTextSize(text, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);\n\n int x = obj.rect.x;\n int y = obj.rect.y - label_size.height - baseLine;\n if (y < 0)\n y = 0;\n if (x + label_size.width > image.cols)\n x = image.cols - label_size.width;\n\n cv::rectangle(image, cv::Rect(cv::Point(x, y), cv::Size(label_size.width, label_size.height + baseLine)),\n cv::Scalar(255, 255, 255), -1);\n\n cv::putText(image, text, cv::Point(x, y + label_size.height),\n cv::FONT_HERSHEY_SIMPLEX, 0.5, cv::Scalar(0, 0, 0));\n }\n\n cv::imshow(\"image\", image);\n cv::waitKey(0);\n}\n\nint main(int argc, char** argv)\n{\n if (argc != 2)\n {\n fprintf(stderr, \"Usage: %s [imagepath]\\n\", argv[0]);\n return -1;\n }\n\n const char* imagepath = argv[1];\n\n cv::Mat m = cv::imread(imagepath, 1);\n if (m.empty())\n {\n fprintf(stderr, \"cv::imread %s failed\\n\", imagepath);\n return -1;\n }\n\n std::vector objects;\n detect_yolox(m, objects);\n\n draw_objects(m, objects);\n\n return 0;\n}\n" }, { "path": "package.sh", "content": "#!/usr/bin/bash\n\nNAME=ncnn\n\n##### package android lib\nANDROIDPKGNAME=${NAME}-android-lib\nrm -rf $ANDROIDPKGNAME\nmkdir -p $ANDROIDPKGNAME\nmkdir -p $ANDROIDPKGNAME/armeabi-v7a\nmkdir -p $ANDROIDPKGNAME/arm64-v8a\nmkdir -p $ANDROIDPKGNAME/x86\nmkdir -p $ANDROIDPKGNAME/x86_64\nmkdir -p $ANDROIDPKGNAME/include\ncp build-android-armv7/install/lib/lib*.a $ANDROIDPKGNAME/armeabi-v7a/\ncp build-android-aarch64/install/lib/lib*.a $ANDROIDPKGNAME/arm64-v8a/\ncp build-android-x86/install/lib/lib*.a $ANDROIDPKGNAME/x86/\ncp build-android-x86_64/install/lib/lib*.a $ANDROIDPKGNAME/x86_64/\ncp -r build-android-aarch64/install/include/* $ANDROIDPKGNAME/include/\nrm -f $ANDROIDPKGNAME.zip\nzip -9 -r $ANDROIDPKGNAME.zip $ANDROIDPKGNAME\n\n##### package ios framework\nIOSPKGNAME=${NAME}.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/${NAME} $IOSPKGNAME/${NAME}\nlipo -create \\\n build-ios/install/lib/lib${NAME}.a \\\n build-ios-sim/install/lib/lib${NAME}.a \\\n -o $IOSPKGNAME/Versions/A/${NAME}\ncp -r build-ios/install/include/* $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME.zip\nzip -9 -y -r $IOSPKGNAME.zip $IOSPKGNAME\n\n##### package ios framework bitcode\nIOSPKGNAME=${NAME}.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/${NAME} $IOSPKGNAME/${NAME}\nlipo -create \\\n build-ios-bitcode/install/lib/lib${NAME}.a \\\n build-ios-sim-bitcode/install/lib/lib${NAME}.a \\\n -o $IOSPKGNAME/Versions/A/${NAME}\ncp -r build-ios-bitcode/install/include/ncnn $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME-bitcode.zip\nzip -9 -y -r $IOSPKGNAME-bitcode.zip $IOSPKGNAME\n\n\n##### package android lib vulkan\nANDROIDPKGNAME=${NAME}-android-vulkan-lib\nrm -rf $ANDROIDPKGNAME\nmkdir -p $ANDROIDPKGNAME\nmkdir -p $ANDROIDPKGNAME/armeabi-v7a\nmkdir -p $ANDROIDPKGNAME/arm64-v8a\nmkdir -p $ANDROIDPKGNAME/x86\nmkdir -p $ANDROIDPKGNAME/x86_64\nmkdir -p $ANDROIDPKGNAME/include\ncp build-android-armv7-vulkan/install/lib/lib*.a $ANDROIDPKGNAME/armeabi-v7a/\ncp build-android-aarch64-vulkan/install/lib/lib*.a $ANDROIDPKGNAME/arm64-v8a/\ncp build-android-x86-vulkan/install/lib/lib*.a $ANDROIDPKGNAME/x86/\ncp build-android-x86_64-vulkan/install/lib/lib*.a $ANDROIDPKGNAME/x86_64/\ncp -r build-android-aarch64-vulkan/install/include/* $ANDROIDPKGNAME/include/\nrm -f $ANDROIDPKGNAME.zip\nzip -9 -r $ANDROIDPKGNAME.zip $ANDROIDPKGNAME\n\n##### package ios framework vulkan\nIOSPKGNAME=${NAME}.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/${NAME} $IOSPKGNAME/${NAME}\nlipo -create \\\n build-ios-vulkan/install/lib/lib${NAME}.a \\\n build-ios-sim-vulkan/install/lib/lib${NAME}.a \\\n -o $IOSPKGNAME/Versions/A/${NAME}\ncp -r build-ios-vulkan/install/include/ncnn $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME-vulkan.zip\nzip -9 -y -r $IOSPKGNAME-vulkan.zip $IOSPKGNAME\n\n##### package ios framework vulkan bitcode\nIOSPKGNAME=${NAME}.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/${NAME} $IOSPKGNAME/${NAME}\nlipo -create \\\n build-ios-vulkan-bitcode/install/lib/lib${NAME}.a \\\n build-ios-sim-vulkan-bitcode/install/lib/lib${NAME}.a \\\n -o $IOSPKGNAME/Versions/A/${NAME}\ncp -r build-ios-vulkan-bitcode/install/include/ncnn $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME-vulkan-bitcode.zip\nzip -9 -y -r $IOSPKGNAME-vulkan-bitcode.zip $IOSPKGNAME\n\n\n##### package ios framework glslang\nIOSPKGNAME=glslang.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/glslang $IOSPKGNAME/glslang\nlibtool -static \\\n build-ios-vulkan/install/lib/libglslang.a \\\n build-ios-vulkan/install/lib/libSPIRV.a \\\n build-ios-vulkan/install/lib/libOGLCompiler.a \\\n build-ios-vulkan/install/lib/libOSDependent.a \\\n -o build-ios-vulkan/install/lib/libglslang_combined.a\nlibtool -static \\\n build-ios-sim-vulkan/install/lib/libglslang.a \\\n build-ios-sim-vulkan/install/lib/libSPIRV.a \\\n build-ios-sim-vulkan/install/lib/libOGLCompiler.a \\\n build-ios-sim-vulkan/install/lib/libOSDependent.a \\\n -o build-ios-sim-vulkan/install/lib/libglslang_combined.a\nlipo -create \\\n build-ios-vulkan/install/lib/libglslang_combined.a \\\n build-ios-sim-vulkan/install/lib/libglslang_combined.a \\\n -o $IOSPKGNAME/Versions/A/glslang\ncp -r build-ios-vulkan/install/include/glslang $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME.zip\nzip -9 -y -r $IOSPKGNAME.zip $IOSPKGNAME\n\n##### package ios framework glslang bitcode\nIOSPKGNAME=glslang.framework\nrm -rf $IOSPKGNAME\nmkdir -p $IOSPKGNAME/Versions/A/Headers\nmkdir -p $IOSPKGNAME/Versions/A/Resources\nln -s A $IOSPKGNAME/Versions/Current\nln -s Versions/Current/Headers $IOSPKGNAME/Headers\nln -s Versions/Current/Resources $IOSPKGNAME/Resources\nln -s Versions/Current/glslang $IOSPKGNAME/glslang\nlibtool -static \\\n build-ios-vulkan-bitcode/install/lib/libglslang.a \\\n build-ios-vulkan-bitcode/install/lib/libSPIRV.a \\\n build-ios-vulkan-bitcode/install/lib/libOGLCompiler.a \\\n build-ios-vulkan-bitcode/install/lib/libOSDependent.a \\\n -o build-ios-vulkan-bitcode/install/lib/libglslang_combined.a\nlibtool -static \\\n build-ios-sim-vulkan-bitcode/install/lib/libglslang.a \\\n build-ios-sim-vulkan-bitcode/install/lib/libSPIRV.a \\\n build-ios-sim-vulkan-bitcode/install/lib/libOGLCompiler.a \\\n build-ios-sim-vulkan-bitcode/install/lib/libOSDependent.a \\\n -o build-ios-sim-vulkan-bitcode/install/lib/libglslang_combined.a\nlipo -create \\\n build-ios-vulkan-bitcode/install/lib/libglslang_combined.a \\\n build-ios-sim-vulkan-bitcode/install/lib/libglslang_combined.a \\\n -o $IOSPKGNAME/Versions/A/glslang\ncp -r build-ios-vulkan-bitcode/install/include/glslang $IOSPKGNAME/Versions/A/Headers/\ncp Info.plist ${IOSPKGNAME}/Versions/A/Resources/\nrm -f $IOSPKGNAME-bitcode.zip\nzip -9 -y -r $IOSPKGNAME-bitcode.zip $IOSPKGNAME\n" }, { "path": "pyproject.toml", "content": "[build-system]\nrequires = [\n \"setuptools>=42\",\n \"wheel\",\n \"importlib-metadata\",\n]\nbuild-backend = \"setuptools.build_meta\"\n" }, { "path": "python/CMakeLists.txt", "content": "cmake_minimum_required(VERSION 3.4...3.10)\n\nproject(pyncnn)\n\nset(PACKAGE_VERSION ${NCNN_VERSION_STRING})\nadd_definitions(-DVERSION_INFO=\"${PACKAGE_VERSION}\")\n\nset( CMAKE_CXX_STANDARD 11 )\nset( CMAKE_CXX_STANDARD_REQUIRED ON )\n\noption(NCNN_SYSTEM_PYBIND11 \"use system pybind11\" OFF)\n\nif(CMAKE_CXX_COMPILER_ARCHITECTURE_ID MATCHES \"ARM64\")\n option(PYBIND11_PYTHONLIBS_OVERWRITE \"\" OFF)\n\n set(PYTHON_PREFIX \"$ENV{LOCALAPPDATA}/pypa/cibuildwheel/Cache/nuget-cpython/pythonarm64.$ENV{PYTHON_VERSION}/tools\")\n if(NOT DEFINED $ENV{CIBUILDWHEEL})\n message(WARNING\n \" This is hack for cibuildwheel on github action\\n\"\n \" Use the right way to cross-compile python module for windows arm64 like follows\\n\"\n \" set(PYTHON_PREFIX \\\"\\\")\\n\"\n )\n endif()\nendif()\n\nif(NCNN_SYSTEM_PYBIND11)\n find_package(pybind11)\n if(NOT pybind11_FOUND)\n message(WARNING \"pybind11 package not found! NCNN_SYSTEM_PYBIND11 will be turned off.\")\n set(NCNN_SYSTEM_PYBIND11 OFF)\n endif()\nendif()\n\nif(NOT NCNN_SYSTEM_PYBIND11)\n if(NOT EXISTS \"${CMAKE_CURRENT_SOURCE_DIR}/pybind11/CMakeLists.txt\")\n message(FATAL_ERROR \"The submodules were not downloaded! Please update submodules with \\\"git submodule update --init\\\" and try again.\")\n else()\n add_subdirectory(pybind11)\n endif()\nendif()\n\nif(\"${CMAKE_LIBRARY_OUTPUT_DIRECTORY}\" STREQUAL \"\")\n if(MSVC OR CMAKE_GENERATOR STREQUAL \"Xcode\")\n set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_DEBUG ${CMAKE_CURRENT_BINARY_DIR}/ncnn/)\n set(CMAKE_LIBRARY_OUTPUT_DIRECTORY_RELEASE ${CMAKE_CURRENT_BINARY_DIR}/ncnn/)\n endif(MSVC OR CMAKE_GENERATOR STREQUAL \"Xcode\")\nendif(\"${CMAKE_LIBRARY_OUTPUT_DIRECTORY}\" STREQUAL \"\")\n\n# enable global link time optimization\ncmake_policy(SET CMP0069 NEW)\nset(CMAKE_POLICY_DEFAULT_CMP0069 NEW)\ninclude(CheckIPOSupported)\ncheck_ipo_supported(RESULT ipo_supported OUTPUT ipo_supported_output)\nif(ipo_supported)\n set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)\nendif()\n\ninclude_directories(${pybind11_INCLUDE_DIR} ${PYTHON_INCLUDE_DIRS})\npybind11_add_module(pyncnn src/main.cpp)\nset_target_properties(pyncnn PROPERTIES OUTPUT_NAME \"ncnn\")\ntarget_link_libraries(pyncnn PUBLIC ncnn)\nset_target_properties(pyncnn PROPERTIES PREFIX \"\" LIBRARY_OUTPUT_DIRECTORY \"${CMAKE_CURRENT_BINARY_DIR}/ncnn\")\nset_property(TARGET pyncnn PROPERTY FOLDER \"python\")\nif(\"${CMAKE_LIBRARY_OUTPUT_DIRECTORY}\" STREQUAL \"\")\n add_custom_command(TARGET pyncnn POST_BUILD \n COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_BINARY_DIR}/ncnn/ncnn${PYTHON_MODULE_PREFIX}${PYTHON_MODULE_EXTENSION} \n ${PROJECT_SOURCE_DIR}/ncnn/ncnn${PYTHON_MODULE_PREFIX}${PYTHON_MODULE_EXTENSION})\nendif(\"${CMAKE_LIBRARY_OUTPUT_DIRECTORY}\" STREQUAL \"\")\n\nconfigure_file(setup.py.i ${PROJECT_SOURCE_DIR}/setup.py)\n" }, { "path": "python/README.md", "content": "# ncnn\npython wrapper of ncnn with [pybind11](https://github.com/pybind/pybind11), only support python3.x now.\n\n\nInstall from pip\n==================\n\nncnn is available as wheel packages for macOS, Windows and Linux distributions, you can install with pip:\n\n```\npython -m pip install -U pip\npython -m pip install -U ncnn\n```\n\n# Build from source\n\nIf you want to build ncnn with some options not as default, or just like to build everything yourself, it is not difficult to build ncnn from source.\n\n## Prerequisites\n\n**On Unix (Linux, OS X)**\n\n* A compiler with C++11 support\n* CMake >= 3.4\n\n**On Mac**\n\n* A compiler with C++11 support\n* CMake >= 3.4\n\n**On Windows**\n\n* Visual Studio 2015 or higher\n* CMake >= 3.4\n\n## Build & Install\n\n1. clone ncnn and init submodule.\n\n```bash\ncd /pathto/ncnn\ngit submodule init && git submodule update\n```\n\n2. build and install.\n\n```\npython setup.py install\n```\n\nIf you want to use a custom toolchain, you can install with the `CMAKE_TOOLCHAIN_FILE` environment variable, like this:\n\n```\nCMAKE_TOOLCHAIN_FILE=\"../../toolchains/power9le-linux-gnu-vsx.clang.toolchain.cmake\" python setup.py install\n```\n\nif you want to enable the usage of vulkan, you can install as following:\n\n```\npython setup.py install --vulkan=on\n```\n\n> **Attention:**\n>\n> To enable Vulkan support, you must first install the Vulkan SDK.\n>\n> **For Windows or Linux Users:**\n>\n> Ensure that the `VULKAN_SDK` environment variable is set to the path of the Vulkan SDK.\n>\n> **For MacOS Users:**\n>\n> On MacOS, you will need to specify additional environment variables. For guidance on setting these variables, please refer to lines 279-286 in the following file: [ncnn/.github/workflows/release-python.yml at master · Tencent/ncnn](https://github.com/Tencent/ncnn/blob/master/.github/workflows/release-python.yml).\n\n## Custom-build & Install\n\n1. clone ncnn and init submodule.\n```bash\ncd /pathto/ncnn\ngit submodule init && git submodule update\n```\n2. build.\n```bash\nmkdir build\ncd build\ncmake -DNCNN_PYTHON=ON ..\nmake\n```\n\nTo use the pybind11 package provided by your system, set the CMake variable `NCNN_SYSTEM_PYBIND11` to `ON` during the build process, like this:\n\n```bash\nmkdir build\ncd build\ncmake -DNCNN_PYTHON=ON -DNCNN_SYSTEM_PYBIND11=ON ..\nmake\n```\n\n3. install\n\n```bash\ncd /pathto/ncnn\npip install .\n```\n\nif you use conda or miniconda, you can also install as following:\n```bash\ncd /pathto/ncnn\npython3 setup.py install\n```\n\n## Tests\n\n**test**\n```bash\ncd /pathto/ncnn/python\npython3 tests/test.py\n```\n\n**benchmark**\n\n```bash\ncd /pathto/ncnn/python\npython3 tests/benchmark.py\n```\n\n## Numpy\n**ncnn.Mat->numpy.array, with no memory copy**\n\n```bash\nmat = ncnn.Mat(...)\nmat_np = np.array(mat)\n```\n\n**numpy.array->ncnn.Mat, with no memory copy**\n```bash\nmat_np = np.array(...)\nmat = ncnn.Mat(mat_np)\n```\n\n# Model Zoo\ninstall requirements\n```bash\npip install -r requirements.txt\n```\nthen you can import ncnn.model_zoo and get model list as follow:\n```bash\nimport ncnn\nimport ncnn.model_zoo as model_zoo\n\nprint(model_zoo.get_model_list())\n```\nmodels now in model zoo are as list below:\n```bash\nmobilenet_yolov2\nmobilenetv2_yolov3\nyolov4_tiny\nyolov4\nyolov5s\nyolact\nmobilenet_ssd\nsqueezenet_ssd\nmobilenetv2_ssdlite\nmobilenetv3_ssdlite\nsqueezenet\nfaster_rcnn\npeleenet_ssd\nretinaface\nrfcn\nshufflenetv2\nsimplepose\nnanodet\n```\nall model in model zoo has example in ncnn/python/examples folder\n\n# Custom Layer\n\ncustom layer demo is in ncnn/python/ncnn/model_zoo/yolov5.py:23\n" }, { "path": "python/examples/fasterrcnn.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"faster_rcnn\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/mobilenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"mobilenet_ssd\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/mobilenetv2ssdlite.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"mobilenetv2_ssdlite\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/mobilenetv3ssdlite.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"mobilenetv3_ssdlite\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects, 0.6)\n" }, { "path": "python/examples/model_zoo.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom ncnn.model_zoo import get_model_list\n\nif __name__ == \"__main__\":\n print(get_model_list())\n" }, { "path": "python/examples/nanodet.py", "content": "# Copyright 2021 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\n \"nanodet\",\n target_size=320,\n prob_threshold=0.4,\n nms_threshold=0.5,\n num_threads=4,\n use_gpu=True,\n )\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/peleenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nimport numpy as np\nfrom ncnn.model_zoo import get_model\n\n\ndef draw_detection_objects_seg(image, class_names, objects, mat_map):\n color = [128, 255, 128, 244, 35, 232]\n color_count = len(color)\n\n for obj in objects:\n print(\n \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\"\n % (obj.label, obj.prob, obj.rect.x, obj.rect.y, obj.rect.w, obj.rect.h)\n )\n\n cv2.rectangle(\n image,\n (int(obj.rect.x), int(obj.rect.y)),\n (int(obj.rect.x + obj.rect.w), int(obj.rect.y + obj.rect.h)),\n (255, 0, 0),\n )\n\n text = \"%s %.1f%%\" % (class_names[int(obj.label)], obj.prob * 100)\n\n label_size, baseLine = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)\n\n x = obj.rect.x\n y = obj.rect.y - label_size[1] - baseLine\n if y < 0:\n y = 0\n if x + label_size[0] > image.shape[1]:\n x = image.shape[1] - label_size[0]\n\n cv2.rectangle(\n image,\n (int(x), int(y)),\n (int(x + label_size[0]), int(y + label_size[1] + baseLine)),\n (255, 255, 255),\n -1,\n )\n\n cv2.putText(\n image,\n text,\n (int(x), int(y + label_size[1])),\n cv2.FONT_HERSHEY_SIMPLEX,\n 0.5,\n (0, 0, 0),\n )\n\n width = mat_map.w\n height = mat_map.h\n size = mat_map.c\n img_index2 = 0\n threshold = 0.45\n ptr2 = np.array(mat_map)\n for i in range(height):\n ptr1 = image[i].flatten()\n img_index1 = 0\n for j in range(width):\n maxima = threshold\n index = -1\n for c in range(size):\n # const float* ptr3 = ptr2 + c*width*height\n ptr3 = ptr2[c].flatten()\n if ptr3[img_index2] > maxima:\n maxima = ptr3[img_index2]\n index = c\n\n if index > -1:\n color_index = (index) * 3\n if color_index < color_count:\n b = color[color_index]\n g = color[color_index + 1]\n r = color[color_index + 2]\n ptr1[img_index1] = b / 2 + ptr1[img_index1] / 2\n ptr1[img_index1 + 1] = g / 2 + ptr1[img_index1 + 1] / 2\n ptr1[img_index1 + 2] = r / 2 + ptr1[img_index1 + 2] / 2\n\n img_index1 += 3\n img_index2 += 1\n\n image[i] = ptr1.reshape(image[i].shape)\n\n cv2.imshow(\"image\", image)\n cv2.waitKey(0)\n\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"peleenet_ssd\", num_threads=4, use_gpu=True)\n\n objects, seg_out = net(m)\n\n draw_detection_objects_seg(m, net.class_names, objects, seg_out)\n" }, { "path": "python/examples/retinaface.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_faceobjects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"retinaface\", num_threads=4, use_gpu=True)\n\n faceobjects = net(m)\n\n draw_faceobjects(m, faceobjects)\n" }, { "path": "python/examples/rfcn.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"rfcn\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/shufflenetv2.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import print_topk\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"shufflenetv2\", num_threads=4, use_gpu=True)\n\n cls_scores = net(m)\n\n print_topk(cls_scores, 3)\n" }, { "path": "python/examples/simplepose.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_pose\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"simplepose\", num_threads=4, use_gpu=True)\n\n keypoints = net(m)\n\n draw_pose(m, keypoints)\n" }, { "path": "python/examples/squeezenet.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import print_topk\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"squeezenet\", num_threads=4, use_gpu=True)\n\n cls_scores = net(m)\n\n print_topk(cls_scores, 5)\n" }, { "path": "python/examples/squeezenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"squeezenet_ssd\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/yolact.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nimport numpy as np\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\n\ndef draw_result(image, class_names, boxes, masks, classes, scores):\n colors = [\n [56, 0, 255],\n [226, 255, 0],\n [0, 94, 255],\n [0, 37, 255],\n [0, 255, 94],\n [255, 226, 0],\n [0, 18, 255],\n [255, 151, 0],\n [170, 0, 255],\n [0, 255, 56],\n [255, 0, 75],\n [0, 75, 255],\n [0, 255, 169],\n [255, 0, 207],\n [75, 255, 0],\n [207, 0, 255],\n [37, 0, 255],\n [0, 207, 255],\n [94, 0, 255],\n [0, 255, 113],\n [255, 18, 0],\n [255, 0, 56],\n [18, 0, 255],\n [0, 255, 226],\n [170, 255, 0],\n [255, 0, 245],\n [151, 255, 0],\n [132, 255, 0],\n [75, 0, 255],\n [151, 0, 255],\n [0, 151, 255],\n [132, 0, 255],\n [0, 255, 245],\n [255, 132, 0],\n [226, 0, 255],\n [255, 37, 0],\n [207, 255, 0],\n [0, 255, 207],\n [94, 255, 0],\n [0, 226, 255],\n [56, 255, 0],\n [255, 94, 0],\n [255, 113, 0],\n [0, 132, 255],\n [255, 0, 132],\n [255, 170, 0],\n [255, 0, 188],\n [113, 255, 0],\n [245, 0, 255],\n [113, 0, 255],\n [255, 188, 0],\n [0, 113, 255],\n [255, 0, 0],\n [0, 56, 255],\n [255, 0, 113],\n [0, 255, 188],\n [255, 0, 94],\n [255, 0, 18],\n [18, 255, 0],\n [0, 255, 132],\n [0, 188, 255],\n [0, 245, 255],\n [0, 169, 255],\n [37, 255, 0],\n [255, 0, 151],\n [188, 0, 255],\n [0, 255, 37],\n [0, 255, 0],\n [255, 0, 170],\n [255, 0, 37],\n [255, 75, 0],\n [0, 0, 255],\n [255, 207, 0],\n [255, 0, 226],\n [255, 245, 0],\n [188, 255, 0],\n [0, 255, 18],\n [0, 255, 75],\n [0, 255, 151],\n [255, 56, 0],\n [245, 255, 0],\n ]\n\n color_index = 0\n\n for box, mask, label, score in zip(boxes, masks, classes, scores):\n if score < 0.15:\n continue\n\n print(\n \"%s = %.5f at %.2f %.2f %.2f x %.2f\\n\"\n % (label, score, box[0], box[1], box[2], box[3])\n )\n\n cv2.rectangle(\n image,\n (int(box[0]), int(box[1])),\n (int(box[0] + box[2]), int(int(box[1] + box[3]))),\n (255, 0, 0),\n )\n\n text = \"%s %.1f%%\" % (class_names[int(label)], score * 100)\n\n label_size, baseLine = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)\n\n x = box[0]\n y = box[1] - label_size[1] - baseLine\n if y < 0:\n y = 0\n if x + label_size[0] > image.shape[1]:\n x = image.shape[1] - label_size[0]\n\n cv2.rectangle(\n image,\n (int(x), int(y)),\n (int(x + label_size[0]), int(y + label_size[1] + baseLine)),\n (255, 255, 255),\n -1,\n )\n\n cv2.putText(\n image,\n text,\n (int(x), int(y + label_size[1])),\n cv2.FONT_HERSHEY_SIMPLEX,\n 0.5,\n (0, 0, 0),\n )\n\n image[mask] = image[mask] * 0.5 + np.array(colors[color_index]) * 0.5\n color_index += 1\n\n cv2.imshow(\"image\", image)\n cv2.waitKey(0)\n\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\n \"yolact\",\n target_size=550,\n confidence_threshold=0.05,\n nms_threshold=0.5,\n keep_top_k=200,\n num_threads=4,\n use_gpu=True,\n )\n\n boxes, masks, classes, scores = net(m)\n\n draw_result(m, net.class_names, boxes, masks, classes, scores)\n" }, { "path": "python/examples/yolov2.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"mobilenet_yolov2\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/yolov3.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\"mobilenetv2_yolov3\", num_threads=4, use_gpu=True)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/yolov4.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [v4l input device or image]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n devicepath = sys.argv[1]\n\n net = get_model(\"yolov4_tiny\", num_threads=4, use_gpu=True)\n # net = get_model(\"yolov4\", num_threads=4, use_gpu=True)\n\n if devicepath.find(\"/dev/video\") == -1:\n m = cv2.imread(devicepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (devicepath))\n sys.exit(0)\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n else:\n cap = cv2.VideoCapture(devicepath)\n\n if cap.isOpened() == False:\n print(\"Failed to open %s\" % (devicepath))\n sys.exit(0)\n\n while True:\n ret, frame = cap.read()\n\n objects = net(frame)\n\n draw_detection_objects(frame, net.class_names, objects)\n" }, { "path": "python/examples/yolov5.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\n \"yolov5s\",\n target_size=640,\n prob_threshold=0.25,\n nms_threshold=0.45,\n num_threads=4,\n use_gpu=True,\n )\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/examples/yolov8.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport sys\nimport cv2\nfrom ncnn.model_zoo import get_model\nfrom ncnn.utils import draw_detection_objects\n\nif __name__ == \"__main__\":\n if len(sys.argv) != 2:\n print(\"Usage: %s [imagepath]\\n\" % (sys.argv[0]))\n sys.exit(0)\n\n imagepath = sys.argv[1]\n\n m = cv2.imread(imagepath)\n if m is None:\n print(\"cv2.imread %s failed\\n\" % (imagepath))\n sys.exit(0)\n\n net = get_model(\n \"yolov8s\",\n target_size=640,\n prob_threshold=0.25,\n nms_threshold=0.45,\n num_threads=4,\n use_gpu=True,\n )\n\n objects = net(m)\n\n draw_detection_objects(m, net.class_names, objects)\n" }, { "path": "python/ncnn/__init__.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom .ncnn import *\n\n__version__ = ncnn.__version__\n" }, { "path": "python/ncnn/model_zoo/__init__.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\n# coding: utf-8\n\"\"\"Predefined and pretrained models.\"\"\"\n\nfrom . import model_store\n\nfrom .model_zoo import get_model, get_model_list\n" }, { "path": "python/ncnn/model_zoo/fasterrcnn.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass Faster_RCNN:\n def __init__(\n self,\n img_width=600,\n img_height=600,\n num_threads=1,\n use_gpu=False,\n max_per_image=100,\n confidence_thresh=0.05,\n nms_threshold=0.3,\n ):\n self.img_width = img_width\n self.img_height = img_height\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [102.9801, 115.9465, 122.7717]\n self.norm_vals = []\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # original pretrained model from https://github.com/rbgirshick/py-faster-rcnn\n # py-faster-rcnn/models/pascal_voc/ZF/faster_rcnn_alt_opt/faster_rcnn_test.pt\n # https://dl.dropboxusercontent.com/s/o6ii098bu51d139/faster_rcnn_models.tgz?dl=0\n # ZF_faster_rcnn_final.caffemodel\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"ZF_faster_rcnn_final.param\"))\n self.net.load_model(get_model_file(\"ZF_faster_rcnn_final.bin\"))\n\n self.max_per_image = max_per_image\n self.confidence_thresh = confidence_thresh\n self.nms_threshold = nms_threshold\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n # scale to target detect size\n h = img.shape[0]\n w = img.shape[1]\n scale = 1.0\n if w < h:\n scale = float(self.img_width) / w\n w = self.img_width\n h = int(h * scale)\n else:\n scale = float(self.img_height) / h\n h = self.img_height\n w = int(w * scale)\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img, ncnn.Mat.PixelType.PIXEL_BGR, img.shape[1], img.shape[0], w, h\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n # method 1 use numpy to Mat interface\n # im_info = ncnn.Mat(np.array([h, w, scale], dtype=np.float32))\n\n # method 2 use ncnn.Mat interface\n im_info = ncnn.Mat(3)\n im_info[0] = h\n im_info[1] = w\n im_info[2] = scale\n\n ex1 = self.net.create_extractor()\n\n ex1.input(\"data\", mat_in)\n ex1.input(\"im_info\", im_info)\n\n ret1, conv5_relu5 = ex1.extract(\"conv5_relu5\")\n ret2, rois = ex1.extract(\"rois\")\n\n class_candidates = []\n for i in range(rois.c):\n ex2 = self.net.create_extractor()\n\n roi = rois.channel(i) # get single roi\n ex2.input(\"conv5_relu5\", conv5_relu5)\n ex2.input(\"rois\", roi)\n\n ret1, bbox_pred = ex2.extract(\"bbox_pred\")\n ret2, cls_prob = ex2.extract(\"cls_prob\")\n\n num_class = cls_prob.w\n while len(class_candidates) < num_class:\n class_candidates.append([])\n\n # find class id with highest score\n label = 0\n score = 0.0\n for j in range(num_class):\n class_score = cls_prob[j]\n if class_score > score:\n label = j\n score = class_score\n\n # ignore background or low score\n if label == 0 or score <= self.confidence_thresh:\n continue\n\n # fprintf(stderr, \"%d = %f\\n\", label, score);\n\n # unscale to image size\n x1 = roi[0] / scale\n y1 = roi[1] / scale\n x2 = roi[2] / scale\n y2 = roi[3] / scale\n\n pb_w = x2 - x1 + 1\n pb_h = y2 - y1 + 1\n\n # apply bbox regression\n dx = bbox_pred[label * 4]\n dy = bbox_pred[label * 4 + 1]\n dw = bbox_pred[label * 4 + 2]\n dh = bbox_pred[label * 4 + 3]\n\n cx = x1 + pb_w * 0.5\n cy = y1 + pb_h * 0.5\n\n obj_cx = cx + pb_w * dx\n obj_cy = cy + pb_h * dy\n\n obj_w = pb_w * np.exp(dw)\n obj_h = pb_h * np.exp(dh)\n\n obj_x1 = obj_cx - obj_w * 0.5\n obj_y1 = obj_cy - obj_h * 0.5\n obj_x2 = obj_cx + obj_w * 0.5\n obj_y2 = obj_cy + obj_h * 0.5\n\n # clip\n obj_x1 = np.maximum(np.minimum(obj_x1, float(img.shape[1] - 1)), 0.0)\n obj_y1 = np.maximum(np.minimum(obj_y1, float(img.shape[0] - 1)), 0.0)\n obj_x2 = np.maximum(np.minimum(obj_x2, float(img.shape[1] - 1)), 0.0)\n obj_y2 = np.maximum(np.minimum(obj_y2, float(img.shape[0] - 1)), 0.0)\n\n # append object\n obj = Detect_Object()\n obj.rect.x = obj_x1\n obj.rect.y = obj_y1\n obj.rect.w = obj_x2 - obj_x1 + 1\n obj.rect.h = obj_y2 - obj_y1 + 1\n obj.label = label\n obj.prob = score\n\n class_candidates[label].append(obj)\n\n # post process\n objects = []\n for candidates in class_candidates:\n if len(candidates) == 0:\n continue\n\n candidates.sort(key=lambda obj: obj.prob, reverse=True)\n\n picked = self.nms_sorted_bboxes(candidates, self.nms_threshold)\n\n for j in range(len(picked)):\n z = picked[j]\n objects.append(candidates[z])\n\n objects.sort(key=lambda obj: obj.prob, reverse=True)\n\n objects = objects[: self.max_per_image]\n\n return objects\n\n def nms_sorted_bboxes(self, objects, nms_threshold):\n picked = []\n\n n = len(objects)\n\n areas = np.zeros((n,), dtype=np.float32)\n for i in range(n):\n areas[i] = objects[i].rect.area()\n\n for i in range(n):\n a = objects[i]\n\n keep = True\n for j in range(len(picked)):\n b = objects[picked[j]]\n\n # intersection over union\n inter_area = a.rect.intersection_area(b.rect)\n union_area = areas[i] + areas[picked[j]] - inter_area\n # float IoU = inter_area / union_area\n if inter_area / union_area > nms_threshold:\n keep = False\n\n if keep:\n picked.append(i)\n\n return picked\n" }, { "path": "python/ncnn/model_zoo/mobilenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass MobileNet_SSD:\n def __init__(self, target_size=300, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [127.5, 127.5, 127.5]\n self.norm_vals = [0.007843, 0.007843, 0.007843]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # model is converted from https://github.com/chuanqi305/MobileNet-SSD\n # and can be downloaded from https://drive.google.com/open?id=0ByaKLD9QaPtucWk0Y0dha1VVY0U\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mobilenet_ssd_voc_ncnn.param\"))\n self.net.load_model(get_model_file(\"mobilenet_ssd_voc_ncnn.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/mobilenetv2ssdlite.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass Noop(ncnn.Layer):\n pass\n\n\ndef Noop_layer_creator():\n return Noop()\n\n\nclass MobileNetV2_SSDLite:\n def __init__(self, target_size=300, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [127.5, 127.5, 127.5]\n self.norm_vals = [0.007843, 0.007843, 0.007843]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n # self.net.register_custom_layer(\"Silence\", Noop_layer_creator)\n\n # original pretrained model from https://github.com/chuanqi305/MobileNetv2-SSDLite\n # https://github.com/chuanqi305/MobileNetv2-SSDLite/blob/master/ssdlite/voc/deploy.prototxt\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mobilenetv2_ssdlite_voc.param\"))\n self.net.load_model(get_model_file(\"mobilenetv2_ssdlite_voc.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img_w,\n img_h,\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/mobilenetv3ssdlite.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\ndef clamp(v, lo, hi):\n if v < lo:\n return lo\n elif hi < v:\n return hi\n else:\n return v\n\n\nclass MobileNetV3_SSDLite:\n def __init__(self, target_size=300, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [123.675, 116.28, 103.53]\n self.norm_vals = [1.0, 1.0, 1.0]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # converted ncnn model from https://github.com/ujsyehao/mobilenetv3-ssd\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mobilenetv3_ssdlite_voc.param\"))\n self.net.load_model(get_model_file(\"mobilenetv3_ssdlite_voc.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR2RGB,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize([], self.norm_vals)\n mat_in.substract_mean_normalize(self.mean_vals, [])\n\n ex = self.net.create_extractor()\n\n ex.input(\"input\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n\n x1 = (\n clamp(values[2] * self.target_size, 0.0, float(self.target_size - 1))\n / self.target_size\n * img_w\n )\n y1 = (\n clamp(values[3] * self.target_size, 0.0, float(self.target_size - 1))\n / self.target_size\n * img_h\n )\n x2 = (\n clamp(values[4] * self.target_size, 0.0, float(self.target_size - 1))\n / self.target_size\n * img_w\n )\n y2 = (\n clamp(values[5] * self.target_size, 0.0, float(self.target_size - 1))\n / self.target_size\n * img_h\n )\n\n if np.isnan(x1) or np.isnan(y1) or np.isnan(x2) or np.isnan(y2):\n continue\n\n obj.rect.x = x1\n obj.rect.y = y1\n obj.rect.w = x2 - x1\n obj.rect.h = y2 - y1\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n\n x1 = clamp(values[2] * self.img_width, 0.0, float(self.img_width - 1)) / self.img_width * img_w\n y1 = clamp(values[3] * self.img_height, 0.0, float(self.img_height - 1)) / self.img_height * img_h\n x2 = clamp(values[4] * self.img_width, 0.0, float(self.img_width - 1)) / self.img_width * img_w\n y2 = clamp(values[5] * self.img_height, 0.0, float(self.img_height - 1)) / self.img_height * img_h\n\n obj.rect.x = x1\n obj.rect.y = y1\n obj.rect.w = x2 - x1\n obj.rect.h = y2 - y1\n\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/model_store.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\n\"\"\"Model store which provides pretrained models.\"\"\"\nfrom __future__ import print_function\n\n__all__ = [\"get_model_file\", \"purge\"]\n\nimport os\nimport zipfile\nimport logging\nimport portalocker\n\nfrom ..utils import download, check_sha1\n\n_model_sha1 = {\n name: checksum\n for checksum, name in [\n (\"4ff279e78cdb0b8bbc9363181df6f094ad46dc36\", \"mobilenet_yolo.param\"),\n (\"1528cf08b9823fc01aaebfc932ec8c8d4a3b1613\", \"mobilenet_yolo.bin\"),\n (\"3f5b78b0c982f8bdf3a2c3a27e6136d4d2680e96\", \"mobilenetv2_yolov3.param\"),\n (\"0705b0f8fe5a77718561b9b7d6ed4f33fcd3d455\", \"mobilenetv2_yolov3.bin\"),\n (\"de59186323ebad5650631e12a6cc66b526ec7df4\", \"yolov4-tiny-opt.param\"),\n (\"1765c3b251c041dd6ac59d2ec3ddf7b983fe9ee9\", \"yolov4-tiny-opt.bin\"),\n (\"e92d3a3a8ac5e6a6c08c433aa2252b0680124328\", \"yolov4-opt.param\"),\n (\"69d128b42b70fb790e9d3ccabcf1b6e8cc2859fe\", \"yolov4-opt.bin\"),\n (\"6fa8ccc8cabc0f5633ab3c6ffa268e6042b8888f\", \"yolov5s.param\"),\n (\"0cbab3664deb090480ea748c1305f6fe850b9ac4\", \"yolov5s.bin\"),\n (\"35ab0c1ce2864e0759d5794aa818df2de3013ab3\", \"yolov7-tiny.param\"),\n (\"c0454f072b41997aa230c3fe1c1d504566574b6c\", \"yolov7-tiny.bin\"),\n (\"e9de3c929d1c93f7dc94ed0f125795ac16ecc120\", \"yolov8s.param\"),\n (\"90f4eb9e90086e2ec3af4c7837f00757e710b9c6\", \"yolov8s.bin\"),\n (\"e65bae7052d9e9b9d45e1214a8d1b5fe6f64e8af\", \"yolact.param\"),\n (\"9bda99f50b1c14c98c5c6bbc08d4f782eed66548\", \"yolact.bin\"),\n (\"3723ce3e312db6a102cff1a5c39dae80e1de658e\", \"mobilenet_ssd_voc_ncnn.param\"),\n (\"8e2d2139550dcbee1ce5e200b7697b25aab29656\", \"mobilenet_ssd_voc_ncnn.bin\"),\n (\"52c669821dc32ef5b7ab30749fa71a3bc27786b8\", \"squeezenet_ssd_voc.param\"),\n (\"347e31d1cbe469259fa8305860a7c24a95039202\", \"squeezenet_ssd_voc.bin\"),\n (\"52dab628ecac8137e61ce3aea1a912f9c5a0a638\", \"mobilenetv2_ssdlite_voc.param\"),\n (\"9fea06f74f7c60d753cf703ea992f92e50a986d4\", \"mobilenetv2_ssdlite_voc.bin\"),\n (\"f36661eff1eda1e36185e7f2f28fc722ad8b66bb\", \"mobilenetv3_ssdlite_voc.param\"),\n (\"908f63ca9bff0061a499512664b9c533a0b7f485\", \"mobilenetv3_ssdlite_voc.bin\"),\n (\"a63d779a1f789af976bc4e2eae86fdd9b0bb6c2c\", \"squeezenet_v1.1.param\"),\n (\"262f0e33e37aeac69021b5a3556664be65fc0aeb\", \"squeezenet_v1.1.bin\"),\n (\"3ba57cccd1d4a583f6eb76eae25a2dbda7ce7f74\", \"ZF_faster_rcnn_final.param\"),\n (\"1095fbb5f846a1f311b40941add5fef691acaf8d\", \"ZF_faster_rcnn_final.bin\"),\n (\"3586ec3d663b1cc8ec8c662768caa9c7fbcf4fdc\", \"pelee.param\"),\n (\"2442ad483dc546940271591b86db0d9c8b1c7118\", \"pelee.bin\"),\n (\"6cfeda08d5494a1274199089fda77c421be1ecac\", \"mnet.25-opt.param\"),\n (\"3ff9a51dc81cdf506a87543dbf752071ffc50b8d\", \"mnet.25-opt.bin\"),\n (\"50acebff393c91468a73a7b7c604ef231429d068\", \"rfcn_end2end.param\"),\n (\"9a68cd937959b4dda9c5bf9c99181cb0e40f266b\", \"rfcn_end2end.bin\"),\n (\"d6b289cda068e9a9d8a171fb909352a05a39a494\", \"shufflenet_v2_x0.5.param\"),\n (\"2ccd631d04a1b7e05483cd8a8def76bca7d330a8\", \"shufflenet_v2_x0.5.bin\"),\n (\"7c8f8d72c60aab6802985423686b36c61be2f68c\", \"pose.param\"),\n (\"7f691540972715298c611a3e595b20c59c2147ce\", \"pose.bin\"),\n (\"979d09942881cf1207a93cbfa9853005a434469b\", \"nanodet_m.param\"),\n (\"51d868905361e4ba9c45bd12e8a5608e7aadd1bd\", \"nanodet_m.bin\"),\n ]\n}\n\n\n_split_model_bins = {\n \"ZF_faster_rcnn_final.bin\": 3,\n \"rfcn_end2end.bin\": 2,\n \"yolov4-opt.bin\": 7,\n}\n\n\ngithub_repo_url = \"https://github.com/nihui/ncnn-assets/raw/master/models/\"\n_url_format = \"{repo_url}{file_name}\"\n\n\ndef merge_file(root, files_in, file_out, remove=True):\n with open(file_out, \"wb\") as fd_out:\n for file_in in files_in:\n file = os.path.join(root, file_in)\n with open(file, \"rb\") as fd_in:\n fd_out.write(fd_in.read())\n if remove == True:\n os.remove(file)\n\n\ndef short_hash(name):\n if name not in _model_sha1:\n raise ValueError(\n \"Pretrained model for {name} is not available.\".format(name=name)\n )\n return _model_sha1[name][:8]\n\n\ndef get_model_file(name, tag=None, root=os.path.join(\"~\", \".ncnn\", \"models\")):\n r\"\"\"Return location for the pretrained on local file system.\n\n This function will download from online model zoo when model cannot be found or has mismatch.\n The root directory will be created if it doesn't exist.\n\n Parameters\n ----------\n name : str\n Name of the model.\n root : str, default '~/.ncnn/models'\n Location for keeping the model parameters.\n\n Returns\n -------\n file_path\n Path to the requested pretrained model file.\n \"\"\"\n if \"NCNN_HOME\" in os.environ:\n root = os.path.join(os.environ[\"NCNN_HOME\"], \"models\")\n\n use_tag = isinstance(tag, str)\n if use_tag:\n file_name = \"{name}-{short_hash}\".format(name=name, short_hash=tag)\n else:\n file_name = \"{name}\".format(name=name)\n\n root = os.path.expanduser(root)\n params_path = os.path.join(root, file_name)\n lockfile = os.path.join(root, file_name + \".lock\")\n if use_tag:\n sha1_hash = tag\n else:\n sha1_hash = _model_sha1[name]\n\n if not os.path.exists(root):\n os.makedirs(root)\n\n with portalocker.Lock(\n lockfile, timeout=int(os.environ.get(\"NCNN_MODEL_LOCK_TIMEOUT\", 300))\n ):\n if os.path.exists(params_path):\n if check_sha1(params_path, sha1_hash):\n return params_path\n else:\n logging.warning(\n \"Hash mismatch in the content of model file '%s' detected. \"\n \"Downloading again.\",\n params_path,\n )\n else:\n logging.info(\"Model file not found. Downloading.\")\n\n zip_file_path = os.path.join(root, file_name)\n if file_name in _split_model_bins:\n file_name_parts = [\n \"%s.part%02d\" % (file_name, i + 1)\n for i in range(_split_model_bins[file_name])\n ]\n for file_name_part in file_name_parts:\n file_path = os.path.join(root, file_name_part)\n repo_url = os.environ.get(\"NCNN_REPO\", github_repo_url)\n if repo_url[-1] != \"/\":\n repo_url = repo_url + \"/\"\n download(\n _url_format.format(repo_url=repo_url, file_name=file_name_part),\n path=file_path,\n overwrite=True,\n )\n\n merge_file(root, file_name_parts, zip_file_path)\n else:\n repo_url = os.environ.get(\"NCNN_REPO\", github_repo_url)\n if repo_url[-1] != \"/\":\n repo_url = repo_url + \"/\"\n download(\n _url_format.format(repo_url=repo_url, file_name=file_name),\n path=zip_file_path,\n overwrite=True,\n )\n if zip_file_path.endswith(\".zip\"):\n with zipfile.ZipFile(zip_file_path) as zf:\n zf.extractall(root)\n os.remove(zip_file_path)\n # Make sure we write the model file on networked filesystems\n try:\n os.sync()\n except AttributeError:\n pass\n if check_sha1(params_path, sha1_hash):\n return params_path\n else:\n raise ValueError(\"Downloaded file has different hash. Please try again.\")\n\n\ndef purge(root=os.path.join(\"~\", \".ncnn\", \"models\")):\n r\"\"\"Purge all pretrained model files in local file store.\n\n Parameters\n ----------\n root : str, default '~/.ncnn/models'\n Location for keeping the model parameters.\n \"\"\"\n root = os.path.expanduser(root)\n files = os.listdir(root)\n for f in files:\n if f.endswith(\".params\"):\n os.remove(os.path.join(root, f))\n" }, { "path": "python/ncnn/model_zoo/model_zoo.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom .yolov2 import MobileNet_YoloV2\nfrom .yolov3 import MobileNetV2_YoloV3\nfrom .yolov4 import YoloV4_Tiny, YoloV4\nfrom .yolov5 import YoloV5s\nfrom .yolov7 import YoloV7_Tiny\nfrom .yolov8 import YoloV8s\nfrom .yolact import Yolact\nfrom .mobilenetssd import MobileNet_SSD\nfrom .squeezenetssd import SqueezeNet_SSD\nfrom .mobilenetv2ssdlite import MobileNetV2_SSDLite\nfrom .mobilenetv3ssdlite import MobileNetV3_SSDLite\nfrom .squeezenet import SqueezeNet\nfrom .fasterrcnn import Faster_RCNN\nfrom .peleenetssd import PeleeNet_SSD\nfrom .retinaface import RetinaFace\nfrom .rfcn import RFCN\nfrom .shufflenetv2 import ShuffleNetV2\nfrom .simplepose import SimplePose\nfrom .nanodet import NanoDet\n\n__all__ = [\"get_model\", \"get_model_list\"]\n\n_models = {\n \"mobilenet_yolov2\": MobileNet_YoloV2,\n \"mobilenetv2_yolov3\": MobileNetV2_YoloV3,\n \"yolov4_tiny\": YoloV4_Tiny,\n \"yolov4\": YoloV4,\n \"yolov5s\": YoloV5s,\n \"yolov7_tiny\": YoloV7_Tiny,\n \"yolov8s\": YoloV8s,\n \"yolact\": Yolact,\n \"mobilenet_ssd\": MobileNet_SSD,\n \"squeezenet_ssd\": SqueezeNet_SSD,\n \"mobilenetv2_ssdlite\": MobileNetV2_SSDLite,\n \"mobilenetv3_ssdlite\": MobileNetV3_SSDLite,\n \"squeezenet\": SqueezeNet,\n \"faster_rcnn\": Faster_RCNN,\n \"peleenet_ssd\": PeleeNet_SSD,\n \"retinaface\": RetinaFace,\n \"rfcn\": RFCN,\n \"shufflenetv2\": ShuffleNetV2,\n \"simplepose\": SimplePose,\n \"nanodet\": NanoDet,\n}\n\n\ndef get_model(name, **kwargs):\n name = name.lower()\n if name not in _models:\n err_str = '\"%s\" is not among the following model list:\\n\\t' % (name)\n err_str += \"%s\" % (\"\\n\\t\".join(sorted(_models.keys())))\n raise ValueError(err_str)\n net = _models[name](**kwargs)\n return net\n\n\ndef get_model_list():\n return list(_models.keys())\n" }, { "path": "python/ncnn/model_zoo/nanodet.py", "content": "# Copyright 2021 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\nfrom ..utils.functional import *\n\n\nclass NanoDet:\n def __init__(\n self,\n target_size=320,\n prob_threshold=0.4,\n nms_threshold=0.3,\n num_threads=1,\n use_gpu=False,\n ):\n self.target_size = target_size\n self.prob_threshold = prob_threshold\n self.nms_threshold = nms_threshold\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [103.53, 116.28, 123.675]\n self.norm_vals = [0.017429, 0.017507, 0.017125]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n # original pretrained model from https://github.com/RangiLyu/nanodet\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"nanodet_m.param\"))\n self.net.load_model(get_model_file(\"nanodet_m.bin\"))\n\n self.reg_max = 7\n self.strides = [8, 16, 32]\n self.num_candidate = 1000\n self.top_k = -1\n\n self.class_names = [\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorcycle\",\n \"airplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"couch\",\n \"potted plant\",\n \"bed\",\n \"dining table\",\n \"toilet\",\n \"tv\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_w = img.shape[1]\n img_h = img.shape[0]\n\n w = img_w\n h = img_h\n scale = 1.0\n if w > h:\n scale = float(self.target_size) / w\n w = self.target_size\n h = int(h * scale)\n else:\n scale = float(self.target_size) / h\n h = self.target_size\n w = int(w * scale)\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img, ncnn.Mat.PixelType.PIXEL_BGR, img_w, img_h, w, h\n )\n\n # pad to target_size rectangle\n wpad = (w + 31) // 32 * 32 - w\n hpad = (h + 31) // 32 * 32 - h\n mat_in_pad = ncnn.copy_make_border(\n mat_in,\n hpad // 2,\n hpad - hpad // 2,\n wpad // 2,\n wpad - wpad // 2,\n ncnn.BorderType.BORDER_CONSTANT,\n 0,\n )\n\n mat_in_pad.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"input.1\", mat_in_pad)\n\n score_out_name = [\"792\", \"814\", \"836\"]\n scores = [ex.extract(x)[1] for x in score_out_name]\n scores = [np.reshape(x, (-1, 80)) for x in scores]\n\n boxes_out_name = [\"795\", \"817\", \"839\"]\n raw_boxes = [ex.extract(x)[1] for x in boxes_out_name]\n raw_boxes = [np.reshape(x, (-1, 32)) for x in raw_boxes]\n\n # generate centers\n decode_boxes = []\n select_scores = []\n for stride, box_distribute, score in zip(self.strides, raw_boxes, scores):\n # centers\n if mat_in_pad.w > mat_in_pad.h:\n fm_w = mat_in_pad.w // stride\n fm_h = score.shape[0] // fm_w\n else:\n fm_h = mat_in_pad.h // stride\n fm_w = score.shape[1] // fm_h\n h_range = np.arange(fm_h)\n w_range = np.arange(fm_w)\n ww, hh = np.meshgrid(w_range, h_range)\n ct_row = (hh.flatten() + 0.5) * stride\n ct_col = (ww.flatten() + 0.5) * stride\n center = np.stack((ct_col, ct_row, ct_col, ct_row), axis=1)\n\n # box distribution to distance\n reg_range = np.arange(self.reg_max + 1)\n box_distance = box_distribute.reshape((-1, self.reg_max + 1))\n box_distance = softmax(box_distance)\n box_distance = box_distance * np.expand_dims(reg_range, axis=0)\n box_distance = np.sum(box_distance, axis=1).reshape((-1, 4))\n box_distance = box_distance * stride\n\n # top K candidate\n topk_idx = np.argsort(score.max(axis=1))[::-1]\n topk_idx = topk_idx[: self.num_candidate]\n center = center[topk_idx]\n score = score[topk_idx]\n box_distance = box_distance[topk_idx]\n\n # decode box\n decode_box = center + [-1, -1, 1, 1] * box_distance\n\n select_scores.append(score)\n decode_boxes.append(decode_box)\n\n # nms\n bboxes = np.concatenate(decode_boxes, axis=0)\n confidences = np.concatenate(select_scores, axis=0)\n picked_box = []\n picked_probs = []\n picked_labels = []\n for class_index in range(0, confidences.shape[1]):\n probs = confidences[:, class_index]\n mask = probs > self.prob_threshold\n probs = probs[mask]\n if probs.shape[0] == 0:\n continue\n subset_boxes = bboxes[mask, :]\n picked = nms(\n subset_boxes,\n probs,\n iou_threshold=self.nms_threshold,\n top_k=self.top_k,\n )\n picked_box.append(subset_boxes[picked])\n picked_probs.append(probs[picked])\n picked_labels.extend([class_index] * len(picked))\n\n if not picked_box:\n return []\n\n picked_box = np.concatenate(picked_box)\n picked_probs = np.concatenate(picked_probs)\n\n # result with clip\n objects = [\n Detect_Object(\n label,\n score,\n (bbox[0] - wpad / 2) / scale if bbox[0] > 0 else 0,\n (bbox[1] - hpad / 2) / scale if bbox[1] > 0 else 0,\n (bbox[2] - bbox[0]) / scale\n if bbox[2] < mat_in_pad.w\n else (mat_in_pad.w - bbox[0]) / scale,\n (bbox[3] - bbox[1]) / scale\n if bbox[3] < mat_in_pad.h\n else (mat_in_pad.h - bbox[1]) / scale,\n )\n for label, score, bbox in zip(picked_labels, picked_probs, picked_box)\n ]\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/peleenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass PeleeNet_SSD:\n def __init__(self, target_size=304, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [103.9, 116.7, 123.6]\n self.norm_vals = [0.017, 0.017, 0.017]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # model is converted from https://github.com/eric612/MobileNet-YOLO\n # and can be downloaded from https://drive.google.com/open?id=1Wt6jKv13sBRMHgrGAJYlOlRF-o80pC0g\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"pelee.param\"))\n self.net.load_model(get_model_file(\"pelee.bin\"))\n\n self.class_names = [\n \"background\",\n \"person\",\n \"rider\",\n \"car\",\n \"bus\",\n \"truck\",\n \"bike\",\n \"motor\",\n \"traffic light\",\n \"traffic sign\",\n \"train\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n objects.append(obj)\n \"\"\"\n\n ret, seg_out = ex.extract(\"sigmoid\")\n\n resized = ncnn.Mat()\n ncnn.resize_bilinear(seg_out, resized, img_w, img_h)\n\n return objects, resized\n" }, { "path": "python/ncnn/model_zoo/retinaface.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Point, Face_Object\n\n\nclass RetinaFace:\n def __init__(\n self, prob_threshold=0.8, nms_threshold=0.4, num_threads=1, use_gpu=False\n ):\n self.prob_threshold = prob_threshold\n self.nms_threshold = nms_threshold\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # model is converted from\n # https://github.com/deepinsight/insightface/tree/master/RetinaFace#retinaface-pretrained-models\n # https://github.com/deepinsight/insightface/issues/669\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mnet.25-opt.param\"))\n self.net.load_model(get_model_file(\"mnet.25-opt.bin\"))\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels(\n img, ncnn.Mat.PixelType.PIXEL_BGR2RGB, img_w, img_h\n )\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n faceobjects32 = self.detect_stride32(ex)\n faceobjects16 = self.detect_stride16(ex)\n faceobjects8 = self.detect_stride8(ex)\n\n faceproposals = [*faceobjects32, *faceobjects16, *faceobjects8]\n\n # sort all proposals by score from highest to lowest\n faceproposals.sort(key=lambda obj: obj.prob, reverse=True)\n\n # apply nms with nms_threshold\n picked = self.nms_sorted_bboxes(faceproposals, self.nms_threshold)\n\n face_count = len(picked)\n\n faceobjects = []\n for i in range(face_count):\n faceobjects.append(faceproposals[picked[i]])\n\n # clip to image size\n x0 = faceobjects[i].rect.x\n y0 = faceobjects[i].rect.y\n x1 = x0 + faceobjects[i].rect.w\n y1 = y0 + faceobjects[i].rect.h\n\n x0 = np.maximum(np.minimum(x0, float(img_w) - 1), 0.0)\n y0 = np.maximum(np.minimum(y0, float(img_h) - 1), 0.0)\n x1 = np.maximum(np.minimum(x1, float(img_w) - 1), 0.0)\n y1 = np.maximum(np.minimum(y1, float(img_h) - 1), 0.0)\n\n faceobjects[i].rect.x = x0\n faceobjects[i].rect.y = y0\n faceobjects[i].rect.w = x1 - x0\n faceobjects[i].rect.h = y1 - y0\n\n return faceobjects\n\n def detect_stride32(self, ex):\n ret1, score_blob = ex.extract(\"face_rpn_cls_prob_reshape_stride32\")\n ret2, bbox_blob = ex.extract(\"face_rpn_bbox_pred_stride32\")\n ret3, landmark_blob = ex.extract(\"face_rpn_landmark_pred_stride32\")\n\n base_size = 16\n feat_stride = 32\n ratios = ncnn.Mat(1)\n ratios[0] = 1.0\n scales = ncnn.Mat(2)\n scales[0] = 32.0\n scales[1] = 16.0\n anchors = self.generate_anchors(base_size, ratios, scales)\n\n faceobjects32 = self.generate_proposals(\n anchors,\n feat_stride,\n score_blob,\n bbox_blob,\n landmark_blob,\n self.prob_threshold,\n )\n\n return faceobjects32\n\n def detect_stride16(self, ex):\n ret1, score_blob = ex.extract(\"face_rpn_cls_prob_reshape_stride16\")\n ret2, bbox_blob = ex.extract(\"face_rpn_bbox_pred_stride16\")\n ret3, landmark_blob = ex.extract(\"face_rpn_landmark_pred_stride16\")\n\n base_size = 16\n feat_stride = 16\n ratios = ncnn.Mat(1)\n ratios[0] = 1.0\n scales = ncnn.Mat(2)\n scales[0] = 8.0\n scales[1] = 4.0\n anchors = self.generate_anchors(base_size, ratios, scales)\n\n faceobjects16 = self.generate_proposals(\n anchors,\n feat_stride,\n score_blob,\n bbox_blob,\n landmark_blob,\n self.prob_threshold,\n )\n\n return faceobjects16\n\n def detect_stride8(self, ex):\n ret1, score_blob = ex.extract(\"face_rpn_cls_prob_reshape_stride8\")\n ret2, bbox_blob = ex.extract(\"face_rpn_bbox_pred_stride8\")\n ret3, landmark_blob = ex.extract(\"face_rpn_landmark_pred_stride8\")\n\n base_size = 16\n feat_stride = 8\n ratios = ncnn.Mat(1)\n ratios[0] = 1.0\n scales = ncnn.Mat(2)\n scales[0] = 2.0\n scales[1] = 1.0\n anchors = self.generate_anchors(base_size, ratios, scales)\n\n faceobjects8 = self.generate_proposals(\n anchors,\n feat_stride,\n score_blob,\n bbox_blob,\n landmark_blob,\n self.prob_threshold,\n )\n\n return faceobjects8\n\n def generate_anchors(self, base_size, ratios, scales):\n num_ratio = ratios.w\n num_scale = scales.w\n\n # anchors = ncnn.Mat()\n # anchors.create(w=4, h=num_ratio * num_scale)\n\n anchors_np = np.zeros((2, 4), dtype=np.float32)\n\n cx = base_size * 0.5\n cy = base_size * 0.5\n\n for i in range(num_ratio):\n ar = ratios[i]\n\n r_w = np.round(base_size / np.sqrt(ar))\n r_h = np.round(r_w * ar) # round(base_size * np.sqrt(ar))\n\n for j in range(num_scale):\n scale = scales[j]\n\n rs_w = r_w * scale\n rs_h = r_h * scale\n\n anchor = anchors_np[i * num_scale + j]\n\n anchor[0] = cx - rs_w * 0.5\n anchor[1] = cy - rs_h * 0.5\n anchor[2] = cx + rs_w * 0.5\n anchor[3] = cy + rs_h * 0.5\n\n anchors = ncnn.Mat(anchors_np)\n return anchors\n\n def generate_proposals(\n self, anchors, feat_stride, score_blob, bbox_blob, landmark_blob, prob_threshold\n ):\n faceobjects = []\n\n w = score_blob.w\n h = score_blob.h\n\n # generate face proposal from bbox deltas and shifted anchors\n num_anchors = anchors.h\n\n for q in range(num_anchors):\n anchor = anchors.row(q)\n\n score = score_blob.channel(q + num_anchors)\n bbox = bbox_blob.channel_range(q * 4, 4)\n landmark = landmark_blob.channel_range(q * 10, 10)\n\n # shifted anchor\n anchor_y = anchor[1]\n\n anchor_w = anchor[2] - anchor[0]\n anchor_h = anchor[3] - anchor[1]\n\n for i in range(h):\n anchor_x = anchor[0]\n\n for j in range(w):\n index = i * w + j\n\n prob = score[index]\n\n if prob >= prob_threshold:\n # apply center size\n dx = bbox.channel(0)[index]\n dy = bbox.channel(1)[index]\n dw = bbox.channel(2)[index]\n dh = bbox.channel(3)[index]\n\n cx = anchor_x + anchor_w * 0.5\n cy = anchor_y + anchor_h * 0.5\n\n pb_cx = cx + anchor_w * dx\n pb_cy = cy + anchor_h * dy\n\n pb_w = anchor_w * np.exp(dw)\n pb_h = anchor_h * np.exp(dh)\n\n x0 = pb_cx - pb_w * 0.5\n y0 = pb_cy - pb_h * 0.5\n x1 = pb_cx + pb_w * 0.5\n y1 = pb_cy + pb_h * 0.5\n\n obj = Face_Object()\n obj.rect.x = x0\n obj.rect.y = y0\n obj.rect.w = x1 - x0 + 1\n obj.rect.h = y1 - y0 + 1\n obj.landmark = [Point(), Point(), Point(), Point(), Point()]\n obj.landmark[0].x = (\n cx + (anchor_w + 1) * landmark.channel(0)[index]\n )\n obj.landmark[0].y = (\n cy + (anchor_h + 1) * landmark.channel(1)[index]\n )\n obj.landmark[1].x = (\n cx + (anchor_w + 1) * landmark.channel(2)[index]\n )\n obj.landmark[1].y = (\n cy + (anchor_h + 1) * landmark.channel(3)[index]\n )\n obj.landmark[2].x = (\n cx + (anchor_w + 1) * landmark.channel(4)[index]\n )\n obj.landmark[2].y = (\n cy + (anchor_h + 1) * landmark.channel(5)[index]\n )\n obj.landmark[3].x = (\n cx + (anchor_w + 1) * landmark.channel(6)[index]\n )\n obj.landmark[3].y = (\n cy + (anchor_h + 1) * landmark.channel(7)[index]\n )\n obj.landmark[4].x = (\n cx + (anchor_w + 1) * landmark.channel(8)[index]\n )\n obj.landmark[4].y = (\n cy + (anchor_h + 1) * landmark.channel(9)[index]\n )\n obj.prob = prob\n\n faceobjects.append(obj)\n\n anchor_x += feat_stride\n\n anchor_y += feat_stride\n\n return faceobjects\n\n def nms_sorted_bboxes(self, faceobjects, nms_threshold):\n picked = []\n\n n = len(faceobjects)\n\n areas = []\n for i in range(n):\n areas.append(faceobjects[i].rect.area())\n\n for i in range(n):\n a = faceobjects[i]\n\n keep = True\n for j in range(len(picked)):\n b = faceobjects[picked[j]]\n\n # intersection over union\n inter_area = a.rect.intersection_area(b.rect)\n union_area = areas[i] + areas[picked[j]] - inter_area\n # float IoU = inter_area / union_area\n if inter_area / union_area > nms_threshold:\n keep = False\n\n if keep:\n picked.append(i)\n\n return picked\n" }, { "path": "python/ncnn/model_zoo/rfcn.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass RFCN:\n def __init__(\n self,\n target_size=224,\n max_per_image=100,\n confidence_thresh=0.6,\n nms_threshold=0.3,\n num_threads=1,\n use_gpu=False,\n ):\n self.target_size = target_size\n self.max_per_image = max_per_image\n self.confidence_thresh = confidence_thresh\n self.nms_threshold = nms_threshold\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [102.9801, 115.9465, 122.7717]\n self.norm_vals = []\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # original pretrained model from https://github.com/YuwenXiong/py-R-FCN\n # https://github.com/YuwenXiong/py-R-FCN/blob/master/models/pascal_voc/ResNet-50/rfcn_end2end/test_agnostic.prototxt\n # https://1drv.ms/u/s!AoN7vygOjLIQqUWHpY67oaC7mopf\n # resnet50_rfcn_final.caffemodel\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"rfcn_end2end.param\"))\n self.net.load_model(get_model_file(\"rfcn_end2end.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n h = img.shape[0]\n w = img.shape[1]\n\n scale = 1.0\n if w < h:\n scale = float(self.target_size) / w\n w = self.target_size\n h = h * scale\n else:\n scale = float(self.target_size) / h\n h = self.target_size\n w = w * scale\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n int(w),\n int(h),\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n im_info = ncnn.Mat(3)\n im_info[0] = h\n im_info[1] = w\n im_info[2] = scale\n\n # step1, extract feature and all rois\n ex1 = self.net.create_extractor()\n\n ex1.input(\"data\", mat_in)\n ex1.input(\"im_info\", im_info)\n\n ret1, rfcn_cls = ex1.extract(\"rfcn_cls\")\n ret2, rfcn_bbox = ex1.extract(\"rfcn_bbox\")\n ret3, rois = ex1.extract(\"rois\") # all rois\n\n # step2, extract bbox and score for each roi\n class_candidates = []\n for i in range(rois.c):\n ex2 = self.net.create_extractor()\n\n roi = rois.channel(i) # get single roi\n ex2.input(\"rfcn_cls\", rfcn_cls)\n ex2.input(\"rfcn_bbox\", rfcn_bbox)\n ex2.input(\"rois\", roi)\n\n ret1, bbox_pred = ex2.extract(\"bbox_pred\")\n ret2, cls_prob = ex2.extract(\"cls_prob\")\n\n num_class = cls_prob.w\n while len(class_candidates) < num_class:\n class_candidates.append([])\n\n # find class id with highest score\n label = 0\n score = 0.0\n for j in range(num_class):\n class_score = cls_prob[j]\n if class_score > score:\n label = j\n score = class_score\n\n # ignore background or low score\n if label == 0 or score <= self.confidence_thresh:\n continue\n\n # fprintf(stderr, \"%d = %f\\n\", label, score)\n\n # unscale to image size\n x1 = roi[0] / scale\n y1 = roi[1] / scale\n x2 = roi[2] / scale\n y2 = roi[3] / scale\n\n pb_w = x2 - x1 + 1\n pb_h = y2 - y1 + 1\n\n # apply bbox regression\n dx = bbox_pred[4]\n dy = bbox_pred[4 + 1]\n dw = bbox_pred[4 + 2]\n dh = bbox_pred[4 + 3]\n\n cx = x1 + pb_w * 0.5\n cy = y1 + pb_h * 0.5\n\n obj_cx = cx + pb_w * dx\n obj_cy = cy + pb_h * dy\n\n obj_w = pb_w * np.exp(dw)\n obj_h = pb_h * np.exp(dh)\n\n obj_x1 = obj_cx - obj_w * 0.5\n obj_y1 = obj_cy - obj_h * 0.5\n obj_x2 = obj_cx + obj_w * 0.5\n obj_y2 = obj_cy + obj_h * 0.5\n\n # clip\n obj_x1 = np.maximum(np.minimum(obj_x1, float(img.shape[1] - 1)), 0.0)\n obj_y1 = np.maximum(np.minimum(obj_y1, float(img.shape[0] - 1)), 0.0)\n obj_x2 = np.maximum(np.minimum(obj_x2, float(img.shape[1] - 1)), 0.0)\n obj_y2 = np.maximum(np.minimum(obj_y2, float(img.shape[0] - 1)), 0.0)\n\n # append object\n obj = Detect_Object()\n obj.rect.x = obj_x1\n obj.rect.y = obj_y1\n obj.rect.w = obj_x2 - obj_x1 + 1\n obj.rect.h = obj_y2 - obj_y1 + 1\n obj.label = label\n obj.prob = score\n\n class_candidates[label].append(obj)\n\n # post process\n objects = []\n for candidates in class_candidates:\n if len(candidates) == 0:\n continue\n\n candidates.sort(key=lambda obj: obj.prob, reverse=True)\n\n picked = self.nms_sorted_bboxes(candidates, self.nms_threshold)\n\n for j in range(len(picked)):\n z = picked[j]\n objects.append(candidates[z])\n\n objects.sort(key=lambda obj: obj.prob, reverse=True)\n\n objects = objects[: self.max_per_image]\n\n return objects\n\n def nms_sorted_bboxes(self, objects, nms_threshold):\n picked = []\n\n n = len(objects)\n\n areas = np.zeros((n,), dtype=np.float32)\n for i in range(n):\n areas[i] = objects[i].rect.area()\n\n for i in range(n):\n a = objects[i]\n\n keep = True\n for j in range(len(picked)):\n b = objects[picked[j]]\n\n # intersection over union\n inter_area = a.rect.intersection_area(b.rect)\n union_area = areas[i] + areas[picked[j]] - inter_area\n # float IoU = inter_area / union_area\n if inter_area / union_area > nms_threshold:\n keep = False\n\n if keep:\n picked.append(i)\n\n return picked\n" }, { "path": "python/ncnn/model_zoo/shufflenetv2.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\n\n\nclass ShuffleNetV2:\n def __init__(self, target_size=224, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = []\n self.norm_vals = [1 / 255.0, 1 / 255.0, 1 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe\n # models can be downloaded from https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe/releases\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"shufflenet_v2_x0.5.param\"))\n self.net.load_model(get_model_file(\"shufflenet_v2_x0.5.bin\"))\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"fc\")\n\n # manually call softmax on the fc output\n # convert result into probability\n # skip if your model already has softmax operation\n softmax = ncnn.create_layer(\"Softmax\")\n\n pd = ncnn.ParamDict()\n softmax.load_param(pd)\n\n softmax.forward_inplace(mat_out, self.net.opt)\n\n mat_out = mat_out.reshape(mat_out.w * mat_out.h * mat_out.c)\n\n cls_scores = np.array(mat_out)\n return cls_scores\n" }, { "path": "python/ncnn/model_zoo/simplepose.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import KeyPoint\n\n\nclass SimplePose:\n def __init__(\n self, target_width=192, target_height=256, num_threads=1, use_gpu=False\n ):\n self.target_width = target_width\n self.target_height = target_height\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [0.485 * 255.0, 0.456 * 255.0, 0.406 * 255.0]\n self.norm_vals = [1 / 0.229 / 255.0, 1 / 0.224 / 255.0, 1 / 0.225 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # the simple baseline human pose estimation from gluon-cv\n # https://gluon-cv.mxnet.io/build/examples_pose/demo_simple_pose.html\n # mxnet model exported via\n # pose_net.hybridize()\n # pose_net.export('pose')\n # then mxnet2ncnn\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"pose.param\"))\n self.net.load_model(get_model_file(\"pose.bin\"))\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n h = img.shape[0]\n w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR2RGB,\n img.shape[1],\n img.shape[0],\n self.target_width,\n self.target_height,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"conv3_fwd\")\n\n keypoints = []\n\n for p in range(mat_out.c):\n m = mat_out.channel(p)\n\n max_prob = 0.0\n max_x = 0\n max_y = 0\n for y in range(mat_out.h):\n ptr = m.row(y)\n for x in range(mat_out.w):\n prob = ptr[x]\n if prob > max_prob:\n max_prob = prob\n max_x = x\n max_y = y\n\n keypoint = KeyPoint()\n keypoint.p.x = max_x * w / float(mat_out.w)\n keypoint.p.y = max_y * h / float(mat_out.h)\n keypoint.prob = max_prob\n\n keypoints.append(keypoint)\n\n return keypoints\n" }, { "path": "python/ncnn/model_zoo/squeezenet.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\n\n\nclass SqueezeNet:\n def __init__(self, target_size=227, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [104.0, 117.0, 123.0]\n self.norm_vals = []\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"squeezenet_v1.1.param\"))\n self.net.load_model(get_model_file(\"squeezenet_v1.1.bin\"))\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"prob\")\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n out = np.array(mat_out)\n return out\n" }, { "path": "python/ncnn/model_zoo/squeezenetssd.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass SqueezeNet_SSD:\n def __init__(self, target_size=300, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [104.0, 117.0, 123.0]\n self.norm_vals = []\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # original pretrained model from https://github.com/chuanqi305/SqueezeNet-SSD\n # squeezenet_ssd_voc_deploy.prototxt\n # https://drive.google.com/open?id=0B3gersZ2cHIxdGpyZlZnbEQ5Snc\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"squeezenet_ssd_voc.param\"))\n self.net.load_model(get_model_file(\"squeezenet_ssd_voc.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/yolact.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom math import sqrt\nimport numpy as np\nimport cv2\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.functional import sigmoid, nms\n\n\nclass Yolact:\n def __init__(\n self,\n target_size=550,\n confidence_threshold=0.05,\n nms_threshold=0.5,\n keep_top_k=200,\n num_threads=1,\n use_gpu=False,\n ):\n self.target_size = target_size\n self.confidence_threshold = confidence_threshold\n self.nms_threshold = nms_threshold\n self.keep_top_k = keep_top_k\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [123.68, 116.78, 103.94]\n self.norm_vals = [1.0 / 58.40, 1.0 / 57.12, 1.0 / 57.38]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n # original model converted from https://github.com/dbolya/yolact\n # yolact_resnet50_54_800000.pth\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"yolact.param\"))\n self.net.load_model(get_model_file(\"yolact.bin\"))\n\n self.conv_ws = [69, 35, 18, 9, 5]\n self.conv_hs = [69, 35, 18, 9, 5]\n self.aspect_ratios = [1, 0.5, 2]\n self.scales = [24, 48, 96, 192, 384]\n\n self.priors = None\n self.last_img_size = None\n\n self.make_priors()\n\n self.class_names = [\n \"background\",\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorcycle\",\n \"airplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"couch\",\n \"potted plant\",\n \"bed\",\n \"dining table\",\n \"toilet\",\n \"tv\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR2RGB,\n img_w,\n img_h,\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"input.1\", mat_in)\n\n ret1, proto_data = ex.extract(\"619\") # 138x138 x 32\n ret2, loc_data = ex.extract(\"816\") # 4 x 19248\n ret3, mask_data = ex.extract(\"818\") # maskdim 32 x 19248\n ret4, conf_data = ex.extract(\"820\") # 81 x 19248\n\n proto_data = np.array(proto_data)\n loc_data = np.array(loc_data)\n mask_data = np.array(mask_data)\n conf_data = np.array(conf_data)\n prior_data = self.make_priors()\n\n # decoded_boxes = self.decode(loc_data, prior_data)\n boxes, masks, classes, scores = self.detect(\n conf_data, loc_data, prior_data, mask_data, img_w, img_h\n )\n\n # generate mask\n masks = proto_data.transpose(1, 2, 0) @ masks.T\n masks = sigmoid(masks)\n\n # Scale masks up to the full image\n masks = cv2.resize(masks, (img_w, img_h), interpolation=cv2.INTER_LINEAR)\n\n # transpose into the correct output shape [num_dets, proto_h, proto_w]\n masks = masks.transpose(2, 0, 1)\n\n masks = masks > 0.5\n\n return boxes, masks, classes, scores\n\n def make_priors(self):\n \"\"\" Note that priors are [x,y,width,height] where (x,y) is the center of the box. \"\"\"\n if self.last_img_size != (self.target_size, self.target_size):\n prior_data = []\n\n for conv_w, conv_h, scale in zip(self.conv_ws, self.conv_hs, self.scales):\n for i in range(conv_h):\n for j in range(conv_w):\n # +0.5 because priors are in center-size notation\n cx = (j + 0.5) / conv_w\n cy = (i + 0.5) / conv_h\n\n for ar in self.aspect_ratios:\n ar = sqrt(ar)\n\n w = scale * ar / self.target_size\n h = scale / ar / self.target_size\n\n # This is for backward compatibility with a bug where I made everything square by accident\n h = w\n\n prior_data += [cx, cy, w, h]\n\n self.priors = np.array(prior_data).reshape(-1, 4)\n self.last_img_size = (self.target_size, self.target_size)\n\n return self.priors\n\n def decode(self, loc, priors, img_w, img_h):\n \"\"\"\n Decode predicted bbox coordinates using the same scheme\n employed by Yolov2: https://arxiv.org/pdf/1612.08242.pdf\n\n b_x = (sigmoid(pred_x) - .5) / conv_w + prior_x\n b_y = (sigmoid(pred_y) - .5) / conv_h + prior_y\n b_w = prior_w * exp(loc_w)\n b_h = prior_h * exp(loc_h)\n\n Note that loc is inputed as [(s(x)-.5)/conv_w, (s(y)-.5)/conv_h, w, h]\n while priors are inputed as [x, y, w, h] where each coordinate\n is relative to size of the image (even sigmoid(x)). We do this\n in the network by dividing by the 'cell size', which is just\n the size of the convouts.\n\n Also note that prior_x and prior_y are center coordinates which\n is why we have to subtract .5 from sigmoid(pred_x and pred_y).\n\n Args:\n - loc: The predicted bounding boxes of size [num_priors, 4]\n - priors: The priorbox coords with size [num_priors, 4]\n\n Returns: A tensor of decoded relative coordinates in point form\n form with size [num_priors, 4(x, y, w, h)]\n \"\"\"\n\n variances = [0.1, 0.2]\n\n boxes = np.concatenate(\n (\n priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:],\n priors[:, 2:] * np.exp(loc[:, 2:] * variances[1]),\n ),\n 1,\n )\n boxes[:, :2] -= boxes[:, 2:] / 2\n # boxes[:, 2:] += boxes[:, :2]\n\n # crop\n np.where(boxes[:, 0] < 0, 0, boxes[:, 0])\n np.where(boxes[:, 1] < 0, 0, boxes[:, 1])\n np.where(boxes[:, 2] > 1, 1, boxes[:, 2])\n np.where(boxes[:, 3] > 1, 1, boxes[:, 3])\n\n # decode to img size\n boxes[:, 0] *= img_w\n boxes[:, 1] *= img_h\n boxes[:, 2] = boxes[:, 2] * img_w + 1\n boxes[:, 3] = boxes[:, 3] * img_h + 1\n\n return boxes\n\n def detect(self, conf_preds, loc_data, prior_data, mask_data, img_w, img_h):\n \"\"\" Perform nms for only the max scoring class that isn't background (class 0) \"\"\"\n cur_scores = conf_preds[:, 1:]\n num_class = cur_scores.shape[1]\n\n classes = np.argmax(cur_scores, axis=1)\n conf_scores = cur_scores[range(cur_scores.shape[0]), classes]\n\n # filte by confidence_threshold\n keep = conf_scores > self.confidence_threshold\n conf_scores = conf_scores[keep]\n classes = classes[keep]\n loc_data = loc_data[keep, :]\n prior_data = prior_data[keep, :]\n masks = mask_data[keep, :]\n\n # decode x, y, w, h\n boxes = self.decode(loc_data, prior_data, img_w, img_h)\n\n # nms for every class\n boxes_result = []\n masks_result = []\n classes_result = []\n conf_scores_result = []\n for i in range(num_class):\n where = np.where(classes == i)\n if len(where) == 0:\n continue\n\n boxes_tmp = boxes[where]\n masks_tmp = masks[where]\n classes_tmp = classes[where]\n conf_scores_tmp = conf_scores[where]\n\n score_mask = conf_scores_tmp > self.confidence_threshold\n boxes_tmp = boxes_tmp[score_mask]\n masks_tmp = masks_tmp[score_mask]\n classes_tmp = classes_tmp[score_mask]\n conf_scores_tmp = conf_scores_tmp[score_mask]\n\n indexes = nms(\n boxes_tmp,\n conf_scores_tmp,\n iou_threshold=self.nms_threshold,\n top_k=self.keep_top_k,\n )\n\n for index in indexes:\n boxes_result.append(boxes_tmp[index])\n masks_result.append(masks_tmp[index])\n classes_result.append(classes_tmp[index] + 1)\n conf_scores_result.append(conf_scores_tmp[index])\n\n # keep top k\n if len(conf_scores_result) > self.keep_top_k:\n indexes = np.argsort(conf_scores_result)\n indexes = indexes[: self.keep_top_k]\n\n boxes_result = boxes_result[indexes]\n masks_result = masks_result[indexes]\n classes_result = classes_result[indexes]\n conf_scores_result = conf_scores_result[indexes]\n\n return (\n np.array(boxes_result),\n np.array(masks_result),\n np.array(classes_result),\n np.array(conf_scores_result),\n )\n" }, { "path": "python/ncnn/model_zoo/yolov2.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass MobileNet_YoloV2:\n def __init__(self, target_size=416, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [1.0, 1.0, 1.0]\n self.norm_vals = [0.007843, 0.007843, 0.007843]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # original pretrained model from https://github.com/eric612/MobileNet-YOLO\n # https://github.com/eric612/MobileNet-YOLO/blob/master/models/yolov2/mobilenet_yolo_deploy.prototxt\n # https://github.com/eric612/MobileNet-YOLO/blob/master/models/yolov2/mobilenet_yolo_deploy_iter_80000.caffemodel\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mobilenet_yolo.param\"))\n self.net.load_model(get_model_file(\"mobilenet_yolo.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize([], self.norm_vals)\n mat_in.substract_mean_normalize(self.mean_vals, [])\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/yolov3.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass MobileNetV2_YoloV3:\n def __init__(self, target_size=352, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = [127.5, 127.5, 127.5]\n self.norm_vals = [0.007843, 0.007843, 0.007843]\n\n self.net = ncnn.Net()\n self.net.opt.num_threads = self.num_threads\n self.net.opt.use_vulkan_compute = self.use_gpu\n\n # original pretrained model from https://github.com/eric612/MobileNet-YOLO\n # param : https://drive.google.com/open?id=1V9oKHP6G6XvXZqhZbzNKL6FI_clRWdC-\n # bin : https://drive.google.com/open?id=1DBcuFCr-856z3FRQznWL_S5h-Aj3RawA\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"mobilenetv2_yolov3.param\"))\n self.net.load_model(get_model_file(\"mobilenetv2_yolov3.bin\"))\n\n self.class_names = [\n \"background\",\n \"aeroplane\",\n \"bicycle\",\n \"bird\",\n \"boat\",\n \"bottle\",\n \"bus\",\n \"car\",\n \"cat\",\n \"chair\",\n \"cow\",\n \"diningtable\",\n \"dog\",\n \"horse\",\n \"motorbike\",\n \"person\",\n \"pottedplant\",\n \"sheep\",\n \"sofa\",\n \"train\",\n \"tvmonitor\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"detection_out\")\n\n objects = []\n\n # printf(\"%d %d %d\\n\", mat_out.w, mat_out.h, mat_out.c)\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.x = values[2] * img_w\n obj.y = values[3] * img_h\n obj.w = values[4] * img_w - obj.x\n obj.h = values[5] * img_h - obj.y\n objects.append(obj)\n \"\"\"\n\n return objects\n" }, { "path": "python/ncnn/model_zoo/yolov4.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\n\n\nclass YoloV4_Base:\n def __init__(self, tiny, target_size, num_threads=1, use_gpu=False):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = []\n self.norm_vals = [1 / 255.0, 1 / 255.0, 1 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n # original pretrained model from https://github.com/AlexeyAB/darknet\n # the ncnn model https://drive.google.com/drive/folders/1YzILvh0SKQPS_lrb33dmGNq7aVTKPWS0?usp=sharing\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n if tiny == True:\n self.net.load_param(get_model_file(\"yolov4-tiny-opt.param\"))\n self.net.load_model(get_model_file(\"yolov4-tiny-opt.bin\"))\n else:\n self.net.load_param(get_model_file(\"yolov4-opt.param\"))\n self.net.load_model(get_model_file(\"yolov4-opt.bin\"))\n\n self.class_names = [\n \"background\",\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorbike\",\n \"aeroplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"sofa\",\n \"pottedplant\",\n \"bed\",\n \"diningtable\",\n \"toilet\",\n \"tvmonitor\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR2RGB,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"data\", mat_in)\n\n ret, mat_out = ex.extract(\"output\")\n\n objects = []\n\n # method 1, use ncnn.Mat.row to get the result, no memory copy\n for i in range(mat_out.h):\n values = mat_out.row(i)\n\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.rect.x = values[2] * img_w\n obj.rect.y = values[3] * img_h\n obj.rect.w = values[4] * img_w - obj.rect.x\n obj.rect.h = values[5] * img_h - obj.rect.y\n\n objects.append(obj)\n\n \"\"\"\n #method 2, use ncnn.Mat->numpy.array to get the result, no memory copy too\n out = np.array(mat_out)\n for i in range(len(out)):\n values = out[i]\n obj = Detect_Object()\n obj.label = values[0]\n obj.prob = values[1]\n obj.x = values[2] * img_w\n obj.y = values[3] * img_h\n obj.w = values[4] * img_w - obj.x\n obj.h = values[5] * img_h - obj.y\n objects.append(obj)\n \"\"\"\n\n return objects\n\n\nclass YoloV4_Tiny(YoloV4_Base):\n def __init__(self, **kwargs):\n super(YoloV4_Tiny, self).__init__(True, 416, **kwargs)\n\n\nclass YoloV4(YoloV4_Base):\n def __init__(self, **kwargs):\n super(YoloV4, self).__init__(False, 608, **kwargs)\n" }, { "path": "python/ncnn/model_zoo/yolov5.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport time\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\nfrom ..utils.functional import *\n\n\nclass YoloV5Focus(ncnn.Layer):\n yolov5FocusLayers = []\n\n def __init__(self):\n ncnn.Layer.__init__(self)\n self.one_blob_only = True\n\n self.yolov5FocusLayers.append(self)\n\n def forward(self, bottom_blob, top_blob, opt):\n x = np.array(bottom_blob)\n x = np.concatenate(\n [\n x[..., ::2, ::2],\n x[..., 1::2, ::2],\n x[..., ::2, 1::2],\n x[..., 1::2, 1::2],\n ]\n )\n\n top_blob.clone_from(ncnn.Mat(x), opt.blob_allocator)\n if top_blob.empty():\n return -100\n\n return 0\n\n\ndef YoloV5Focus_layer_creator():\n return YoloV5Focus()\n\n\ndef YoloV5Focus_layer_destroyer(layer):\n for i in range(len(YoloV5Focus.yolov5FocusLayers)):\n if YoloV5Focus.yolov5FocusLayers[i] == layer:\n del YoloV5Focus.yolov5FocusLayers[i]\n break\n\n\nclass YoloV5s:\n def __init__(\n self,\n target_size=640,\n prob_threshold=0.25,\n nms_threshold=0.45,\n num_threads=1,\n use_gpu=False,\n ):\n self.target_size = target_size\n self.prob_threshold = prob_threshold\n self.nms_threshold = nms_threshold\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.mean_vals = []\n self.norm_vals = [1 / 255.0, 1 / 255.0, 1 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n self.net.register_custom_layer(\n \"YoloV5Focus\", YoloV5Focus_layer_creator, YoloV5Focus_layer_destroyer\n )\n\n # original pretrained model from https://github.com/ultralytics/yolov5\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"yolov5s.param\"))\n self.net.load_model(get_model_file(\"yolov5s.bin\"))\n\n self.grid = [make_grid(10, 6), make_grid(20, 12), make_grid(40, 24)]\n self.stride = np.array([32, 16, 8])\n self.anchor_grid = np.array(\n [\n [116, 90, 156, 198, 373, 326],\n [30, 61, 62, 45, 59, 119],\n [10, 13, 16, 30, 33, 23],\n ]\n ).reshape((3, 1, 3, 1, 1, 2))\n\n self.class_names = [\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorcycle\",\n \"airplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"couch\",\n \"potted plant\",\n \"bed\",\n \"dining table\",\n \"toilet\",\n \"tv\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_w = img.shape[1]\n img_h = img.shape[0]\n\n w = img_w\n h = img_h\n scale = 1.0\n if w > h:\n scale = float(self.target_size) / w\n w = self.target_size\n h = int(h * scale)\n else:\n scale = float(self.target_size) / h\n h = self.target_size\n w = int(w * scale)\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img, ncnn.Mat.PixelType.PIXEL_BGR2RGB, img_w, img_h, w, h\n )\n # pad to target_size rectangle\n # yolov5/utils/datasets.py letterbox\n wpad = (w + 31) // 32 * 32 - w\n hpad = (h + 31) // 32 * 32 - h\n mat_in_pad = ncnn.copy_make_border(\n mat_in,\n hpad // 2,\n hpad - hpad // 2,\n wpad // 2,\n wpad - wpad // 2,\n ncnn.BorderType.BORDER_CONSTANT,\n 114.0,\n )\n\n mat_in_pad.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"images\", mat_in_pad)\n\n # anchor setting from yolov5/models/yolov5s.yaml\n ret1, mat_out1 = ex.extract(\"output\") # stride 8\n ret2, mat_out2 = ex.extract(\"781\") # stride 16\n ret3, mat_out3 = ex.extract(\"801\") # stride 32\n\n pred = [np.array(mat_out3), np.array(mat_out2), np.array(mat_out1)]\n z = []\n for i in range(len(pred)):\n num_grid = pred[i].shape[1]\n if mat_in_pad.w > mat_in_pad.h:\n num_grid_x = mat_in_pad.w // self.stride[i]\n num_grid_y = num_grid // num_grid_x\n else:\n num_grid_y = mat_in_pad.h // self.stride[i]\n num_grid_x = num_grid // num_grid_y\n if (\n self.grid[i].shape[0] != num_grid_x\n or self.grid[i].shape[1] != num_grid_y\n ):\n self.grid[i] = make_grid(num_grid_x, num_grid_y)\n\n y = sigmoid(pred[i])\n y = y.reshape(pred[i].shape[0], num_grid_y, num_grid_x, pred[i].shape[2])\n y[..., 0:2] = (y[..., 0:2] * 2.0 - 0.5 + self.grid[i]) * self.stride[\n i\n ] # xy\n y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh\n z.append(y.reshape(1, -1, y.shape[-1]))\n pred = np.concatenate(z, 1)\n\n result = self.non_max_suppression(\n pred, self.prob_threshold, self.nms_threshold\n )[0]\n\n objects = [\n Detect_Object(\n obj[5],\n obj[4],\n (obj[0] - (wpad / 2)) / scale,\n (obj[1] - (hpad / 2)) / scale,\n (obj[2] - obj[0]) / scale,\n (obj[3] - obj[1]) / scale,\n )\n for obj in result\n ]\n\n return objects\n\n def non_max_suppression(\n self,\n prediction,\n conf_thres=0.1,\n iou_thres=0.6,\n merge=False,\n classes=None,\n agnostic=False,\n ):\n \"\"\"Performs Non-Maximum Suppression (NMS) on inference results\n\n Returns:\n detections with shape: nx6 (x1, y1, x2, y2, conf, cls)\n \"\"\"\n nc = prediction[0].shape[1] - 5 # number of classes\n xc = prediction[..., 4] > conf_thres # candidates\n\n # Settings\n min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height\n max_det = 300 # maximum number of detections per image\n time_limit = 10.0 # seconds to quit after\n redundant = True # require redundant detections\n multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img)\n\n t = time.time()\n output = [None] * prediction.shape[0]\n for xi, x in enumerate(prediction): # image index, image inference\n # Apply constraints\n # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height\n x = x[xc[xi]] # confidence\n\n # If none remain process next image\n if not x.shape[0]:\n continue\n\n # Compute conf\n x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf\n\n # Box (center x, center y, width, height) to (x1, y1, x2, y2)\n box = xywh2xyxy(x[:, :4])\n\n # Detections matrix nx6 (xyxy, conf, cls)\n if multi_label:\n i, j = (x[:, 5:] > conf_thres).nonzero()\n x = np.concatenate(\n (box[i], x[i, j + 5, None], j[:, None].astype(np.float32)), axis=1\n )\n else: # best class only\n conf, j = x[:, 5:].max(1, keepdim=True)\n x = np.concatenate((box, conf, j.float()), axis=1)[\n conf.view(-1) > conf_thres\n ]\n\n # Filter by class\n if classes:\n x = x[(x[:, 5:6] == np.array(classes)).any(1)]\n\n # Apply finite constraint\n # if not torch.isfinite(x).all():\n # x = x[torch.isfinite(x).all(1)]\n\n # If none remain process next image\n n = x.shape[0] # number of boxes\n if not n:\n continue\n\n # Sort by confidence\n # x = x[x[:, 4].argsort(descending=True)]\n\n # Batched NMS\n c = x[:, 5:6] * (0 if agnostic else max_wh) # classes\n boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores\n i = nms(boxes, scores, iou_threshold=iou_thres)\n if len(i) > max_det: # limit detections\n i = i[:max_det]\n if merge and (1 < n < 3e3): # Merge NMS (boxes merged using weighted mean)\n try: # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)\n iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix\n weights = iou * scores[None] # box weights\n x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(\n 1, keepdim=True\n ) # merged boxes\n if redundant:\n i = i[iou.sum(1) > 1] # require redundancy\n except: # possible CUDA error https://github.com/ultralytics/yolov3/issues/1139\n print(x, i, x.shape, i.shape)\n pass\n\n output[xi] = x[i]\n if (time.time() - t) > time_limit:\n break # time limit exceeded\n\n return output\n" }, { "path": "python/ncnn/model_zoo/yolov7.py", "content": "# Copyright 2020 Tencent\n# Copyright 2023 Kenny Bradley\n# SPDX-License-Identifier: BSD-3-Clause\n\n# Ported yolov7-tiny to python based on:\n# - https://github.com/Qengineering/YoloV7-ncnn-Raspberry-Pi-4/blob/main/yolo.cpp\n# Format based on the ncnn yolov4 implementation\n\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\nimport numpy as np\n\n\n#def sigmoid_binned(val)\n# this could use a much faster binned lookup table instead of np.exp and floating division\n\ndef sigmoid(val):\n return 1.0 / (1.0 + np.exp(-val))\n\n#IOU functions:\n#find the overlap width given ([x1,x2], [x3,x4]) or ([y1,y2], [y3,y4])\ndef calcOverlap(r1, r2):\n #r1 contains r2\n if r1[0] <= r2[0] and r1[1] >= r2[1]:\n return r2[1] - r2[0]\n #r2 contains r1\n elif r1[0] >= r2[0] and r1[1] <= r2[1]:\n return r1[1] - r1[0]\n #r1.1 is between r2.0 and r2.1\n elif r1[0] <= r2[0] and r1[1] >= r2[0]: # r1[1] <= r2[1] is true since the first if failed\n return r1[1] - r2[0]\n #r1.0 is between r2.0 and r2.1\n elif r1[0] >= r2[0] and r1[0] <= r2[1]: # r1[1] >= r2[1] is true since the second if failed\n return r2[1] - r1[0]\n else:\n return 0\n\n#find X and Y overlaps and return intersection area\ndef calcIntersection(r1 : Detect_Object, r2 : Detect_Object):\n xOverlap = calcOverlap([r1.rect.x, r1.rect.x+r1.rect.w], [r2.rect.x, r2.rect.x+r2.rect.w])\n yOverlap = calcOverlap([r1.rect.y, r1.rect.y+r1.rect.h], [r2.rect.y, r2.rect.y+r2.rect.h])\n return xOverlap*yOverlap\n\n\n#with r = [X1,X2,Y1,Y2] as the format return the IOU\ndef IOU(r1 : Detect_Object, r2 : Detect_Object):\n intersection = calcIntersection(r1,r2)\n #union = r1 area + r2 area - duplicate area\n union = (r1.rect.w*r1.rect.h) + (r2.rect.w*r2.rect.h) - intersection\n if union == 0:\n return 0\n else:\n return intersection/union\n\n#NMS\n#detections are pre-sorted in ascending confidence order\n#detections are a list of Detect_Objects with : label, prob, rect\ndef NMS(detections, iou_thresh=0.45):\n cleanDetections = []\n detByClasses = {}\n #group by class\n for det in detections:\n #det.label is the class\n if det.label not in detByClasses.keys():\n detByClasses[det.label] = []\n detByClasses[det.label].append(det)\n\n #for each class find the values to keep\n for key, dets in detByClasses.items():\n for i in range(len(dets)):\n keep = 1\n #keep unless a higher priority det has IOU > thresh\n for j in range(i+1,len(dets)):\n iou = IOU(dets[i], dets[j])\n if iou > iou_thresh:\n keep = 0\n break\n if keep:\n cleanDetections.append(dets[i])\n\n #return cleaner list of Detect_Object values\n return cleanDetections\n \nclass YoloV7_Base:\n def __init__(self, target_size, num_threads=1, use_gpu=False, use_strides=[8,16,32]):\n self.target_size = target_size\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n self.use_strides = use_strides\n\n self.mean_vals = []\n self.norm_vals = [1 / 255.0, 1 / 255.0, 1 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n # original pretrained model from https://github.com/AlexeyAB/darknet\n # the ncnn model https://drive.google.com/drive/folders/1YzILvh0SKQPS_lrb33dmGNq7aVTKPWS0?usp=sharing\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"yolov7-tiny.param\"))\n self.net.load_model(get_model_file(\"yolov7-tiny.bin\"))\n\n self.class_names = [\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorbike\",\n \"aeroplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"sofa\",\n \"pottedplant\",\n \"bed\",\n \"diningtable\",\n \"toilet\",\n \"tvmonitor\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\"\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n\n img_h = img.shape[0]\n img_w = img.shape[1]\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img,\n ncnn.Mat.PixelType.PIXEL_BGR2RGB,\n img.shape[1],\n img.shape[0],\n self.target_size,\n self.target_size,\n )\n mat_in.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"images\", mat_in)\n\n outValues = []\n if 8 in self.use_strides:\n ret8, out8 = ex.extract(\"output\");\n outValues.append(out8)\n else:\n outValues.append(None)\n\n if 16 in self.use_strides:\n ret16, out16 = ex.extract(\"288\");\n outValues.append(out16)\n else:\n outValues.append(None)\n\n if 32 in self.use_strides:\n ret32, out32 = ex.extract(\"302\");\n outValues.append(out32)\n else:\n outValues.append(None)\n\n # P3/8, P4/16, P5/32\n anchors = [[12,16, 19,36, 40,28], [36,75, 76,55, 72,146], [142,110, 192,243, 459,401]]\n strides = [8,16,32]\n\n objects = []\n #this threshold is the value for which sigmoid gives 0.25 which is the threshold\n threshNonSigmoid = -1.098612\n for strideCount, mat_out in enumerate(outValues):\n if mat_out is None:\n continue\n\n stride = strides[strideCount]\n for c in range(3):\n mat = mat_out.channel(c)\n\n #yolo should always be square, it is expected to be 52x52\n # but sqrt() guarantees the correct size for side\n side = int(np.sqrt(mat.h))\n\n anchorW = anchors[strideCount][c*2]\n anchorH = anchors[strideCount][c*2+1]\n index = 0\n for i in range(side):\n for j in range(side):\n\n #values 5-84 are class data\n classData=mat.row(index)[5:]\n maxLabel = max(classData)\n \n #optimization\n #if either the objectness or max class score resolve to < 0.25 we can skip this\n # but the values are pre-sigmoid so compare to threshNonSigmoid.\n # 1 / (1+e^(-1.098612)) = 0.25 so just compare to the -1.098612 threshold\n if mat.row(index)[4] < threshNonSigmoid or maxLabel < threshNonSigmoid:\n index += 1\n continue\n\n #values 0-3 are coordinate data\n locData = mat.row(index)[0:4]\n #value 4 is the box confidence score\n box_score = sigmoid(mat.row(index)[4])\n #get the highest scoring class for this detection to multiply by the box_score\n label = np.argmax(classData)\n class_score = sigmoid(mat.row(index)[label+5])\n\n conf = box_score * class_score\n if conf > 0.25:\n obj = Detect_Object()\n obj.label = self.class_names[label]\n obj.prob = conf\n #convert from raw yolo output to W,H and X,Y\n obj.rect.w = ((sigmoid(locData[2]) *2) ** 2) * anchorW\n obj.rect.h = ((sigmoid(locData[3]) *2) ** 2) * anchorH\n obj.rect.x = ((sigmoid(locData[0]) * 2) - 0.5 + j) * stride - (obj.rect.w/2)\n obj.rect.y = ((sigmoid(locData[1]) * 2) - 0.5 + i) * stride - (obj.rect.h/2)\n objects.append(obj)\n\n index +=1\n\n #sort based on probability in ascending order\n objects.sort(key = lambda x: x.prob)\n filtered_objects = NMS(objects)\n\n #rescale to input image size\n XscaleAdj = img_w / self.target_size\n YscaleAdj = img_h / self.target_size\n for count in range(len(filtered_objects)):\n filtered_objects[count].rect.x *= XscaleAdj\n filtered_objects[count].rect.w *= XscaleAdj\n filtered_objects[count].rect.y *= YscaleAdj\n filtered_objects[count].rect.h *= YscaleAdj\n\n return filtered_objects\n\n\n\nclass YoloV7_Tiny(YoloV7_Base):\n def __init__(self, **kwargs):\n super(YoloV7_Tiny, self).__init__(416, **kwargs)\n" }, { "path": "python/ncnn/model_zoo/yolov8.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport time\nimport numpy as np\nimport ncnn\nfrom .model_store import get_model_file\nfrom ..utils.objects import Detect_Object\nfrom ..utils.functional import *\nfrom typing import Iterable\n\nclass YoloV8s:\n def __init__(\n self,\n target_size=640,\n prob_threshold=0.25,\n nms_threshold=0.45,\n num_threads=1,\n use_gpu=False,\n ):\n self.target_size = target_size\n self.prob_threshold = prob_threshold\n self.nms_threshold = nms_threshold\n self.num_threads = num_threads\n self.use_gpu = use_gpu\n\n self.reg_max = 16\n self.mean_vals = []\n self.norm_vals = [1 / 255.0, 1 / 255.0, 1 / 255.0]\n\n self.net = ncnn.Net()\n self.net.opt.use_vulkan_compute = self.use_gpu\n self.net.opt.num_threads = self.num_threads\n\n # original pretrained model from https://github.com/ultralytics/ultralytics\n # the ncnn model https://github.com/nihui/ncnn-assets/tree/master/models\n self.net.load_param(get_model_file(\"yolov8s.param\"))\n self.net.load_model(get_model_file(\"yolov8s.bin\"))\n\n self.grid = [make_grid(20, 20), make_grid(40, 40), make_grid(80, 80)]\n self.stride = np.array([32, 16, 8])\n\n self.class_names = [\n \"person\",\n \"bicycle\",\n \"car\",\n \"motorcycle\",\n \"airplane\",\n \"bus\",\n \"train\",\n \"truck\",\n \"boat\",\n \"traffic light\",\n \"fire hydrant\",\n \"stop sign\",\n \"parking meter\",\n \"bench\",\n \"bird\",\n \"cat\",\n \"dog\",\n \"horse\",\n \"sheep\",\n \"cow\",\n \"elephant\",\n \"bear\",\n \"zebra\",\n \"giraffe\",\n \"backpack\",\n \"umbrella\",\n \"handbag\",\n \"tie\",\n \"suitcase\",\n \"frisbee\",\n \"skis\",\n \"snowboard\",\n \"sports ball\",\n \"kite\",\n \"baseball bat\",\n \"baseball glove\",\n \"skateboard\",\n \"surfboard\",\n \"tennis racket\",\n \"bottle\",\n \"wine glass\",\n \"cup\",\n \"fork\",\n \"knife\",\n \"spoon\",\n \"bowl\",\n \"banana\",\n \"apple\",\n \"sandwich\",\n \"orange\",\n \"broccoli\",\n \"carrot\",\n \"hot dog\",\n \"pizza\",\n \"donut\",\n \"cake\",\n \"chair\",\n \"couch\",\n \"potted plant\",\n \"bed\",\n \"dining table\",\n \"toilet\",\n \"tv\",\n \"laptop\",\n \"mouse\",\n \"remote\",\n \"keyboard\",\n \"cell phone\",\n \"microwave\",\n \"oven\",\n \"toaster\",\n \"sink\",\n \"refrigerator\",\n \"book\",\n \"clock\",\n \"vase\",\n \"scissors\",\n \"teddy bear\",\n \"hair drier\",\n \"toothbrush\",\n ]\n\n def __del__(self):\n self.net = None\n\n def __call__(self, img):\n img_w = img.shape[1]\n img_h = img.shape[0]\n\n w = img_w\n h = img_h\n scale = 1.0\n if w > h:\n scale = float(self.target_size) / w\n w = self.target_size\n h = int(h * scale)\n else:\n scale = float(self.target_size) / h\n h = self.target_size\n w = int(w * scale)\n\n mat_in = ncnn.Mat.from_pixels_resize(\n img, ncnn.Mat.PixelType.PIXEL_BGR2RGB, img_w, img_h, w, h\n )\n # pad to target_size rectangle\n # yolov5/utils/datasets.py letterbox\n wpad = (w + 31) // 32 * 32 - w\n hpad = (h + 31) // 32 * 32 - h\n mat_in_pad = ncnn.copy_make_border(\n mat_in,\n hpad // 2,\n hpad - hpad // 2,\n wpad // 2,\n wpad - wpad // 2,\n ncnn.BorderType.BORDER_CONSTANT,\n 114.0,\n )\n\n mat_in_pad.substract_mean_normalize(self.mean_vals, self.norm_vals)\n\n ex = self.net.create_extractor()\n ex.input(\"in0\", mat_in_pad)\n\n ret1, mat_out1 = ex.extract(\"out0\") # stride 8\n ret2, mat_out2 = ex.extract(\"out1\") # stride 16\n ret3, mat_out3 = ex.extract(\"out2\") # stride 32\n\n pred = [np.array(mat_out3), np.array(mat_out2), np.array(mat_out1)]\n z = []\n for i in range(len(pred)):\n num_grid_x = mat_in_pad.w // self.stride[i]\n num_grid_y = mat_in_pad.h // self.stride[i]\n if (\n self.grid[i].shape[1] != num_grid_y\n or self.grid[i].shape[2] != num_grid_x\n ):\n self.grid[i] = make_grid(num_grid_x, num_grid_y)\n cls, box = np.split(pred[i].transpose((1, 2, 0)), [len(self.class_names), ], -1)\n box = softmax(box.reshape(-1, self.reg_max))\n box = box.reshape(num_grid_y, num_grid_x, 4, self.reg_max)\n box = box @ np.arange(0, self.reg_max, dtype=np.float32)\n cls = sigmoid(cls)\n conf = cls.max(-1, keepdims=True)\n x1y1 = (self.grid[i][0] + 0.5 - box[..., :2]) * self.stride[i]\n x2y2 = (self.grid[i][0] + 0.5 + box[..., 2:]) * self.stride[i]\n res = np.concatenate([x1y1, x2y2, conf, cls], -1)\n z.append(res.reshape((1, -1, len(self.class_names) + 5)))\n pred = np.concatenate(z, 1)\n\n result = self.non_max_suppression(\n pred, self.prob_threshold, self.nms_threshold\n )[0]\n\n if isinstance(result, Iterable):\n objects = [\n Detect_Object(\n obj[5],\n obj[4],\n (obj[0] - (wpad / 2)) / scale,\n (obj[1] - (hpad / 2)) / scale,\n (obj[2] - obj[0]) / scale,\n (obj[3] - obj[1]) / scale,\n )\n for obj in result\n ]\n else:\n objects = []\n\n return objects\n\n def non_max_suppression(\n self,\n prediction,\n conf_thres=0.1,\n iou_thres=0.6,\n merge=False,\n classes=None,\n agnostic=False,\n ):\n \"\"\"Performs Non-Maximum Suppression (NMS) on inference results\n\n Returns:\n detections with shape: nx6 (x1, y1, x2, y2, conf, cls)\n \"\"\"\n nc = prediction[0].shape[1] - 5 # number of classes\n xc = prediction[..., 4] > conf_thres # candidates\n\n # Settings\n min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height\n max_det = 300 # maximum number of detections per image\n time_limit = 10.0 # seconds to quit after\n redundant = True # require redundant detections\n multi_label = nc > 1 # multiple labels per box (adds 0.5ms/img)\n\n t = time.time()\n output = [None] * prediction.shape[0]\n for xi, x in enumerate(prediction): # image index, image inference\n # Apply constraints\n # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height\n x = x[xc[xi]] # confidence\n\n # If none remain process next image\n if not x.shape[0]:\n continue\n\n box = x[:, :4]\n\n # Detections matrix nx6 (xyxy, conf, cls)\n if multi_label:\n i, j = (x[:, 5:] > conf_thres).nonzero()\n x = np.concatenate(\n (box[i], x[i, j + 5, None], j[:, None].astype(np.float32)), axis=1\n )\n else: # best class only\n conf, j = x[:, 5:].max(1, keepdim=True)\n x = np.concatenate((box, conf, j.float()), axis=1)[\n conf.view(-1) > conf_thres\n ]\n\n # Filter by class\n if classes:\n x = x[(x[:, 5:6] == np.array(classes)).any(1)]\n\n # Apply finite constraint\n # if not torch.isfinite(x).all():\n # x = x[torch.isfinite(x).all(1)]\n\n # If none remain process next image\n n = x.shape[0] # number of boxes\n if not n:\n continue\n\n # Sort by confidence\n # x = x[x[:, 4].argsort(descending=True)]\n\n # Batched NMS\n c = x[:, 5:6] * (0 if agnostic else max_wh) # classes\n boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores\n i = nms(boxes, scores, iou_threshold=iou_thres)\n if len(i) > max_det: # limit detections\n i = i[:max_det]\n if merge and (1 < n < 3e3): # Merge NMS (boxes merged using weighted mean)\n try: # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)\n iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix\n weights = iou * scores[None] # box weights\n x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(\n 1, keepdim=True\n ) # merged boxes\n if redundant:\n i = i[iou.sum(1) > 1] # require redundancy\n except: # possible CUDA error https://github.com/ultralytics/yolov3/issues/1139\n print(x, i, x.shape, i.shape)\n pass\n\n output[xi] = x[i]\n if (time.time() - t) > time_limit:\n break # time limit exceeded\n\n return output\n" }, { "path": "python/ncnn/utils/__init__.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom .download import download, check_sha1\nfrom .visual import *\nfrom .objects import *\n" }, { "path": "python/ncnn/utils/download.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\n\"\"\"Download files with progress bar.\"\"\"\n\nimport os\nimport hashlib\nimport requests\nfrom tqdm import tqdm\n\n\ndef check_sha1(filename, sha1_hash):\n \"\"\"Check whether the sha1 hash of the file content matches the expected hash.\n Parameters\n ----------\n filename : str\n Path to the file.\n sha1_hash : str\n Expected sha1 hash in hexadecimal digits.\n Returns\n -------\n bool\n Whether the file content matches the expected hash.\n \"\"\"\n sha1 = hashlib.sha1()\n with open(filename, \"rb\") as f:\n while True:\n data = f.read(1048576)\n if not data:\n break\n sha1.update(data)\n\n sha1_file = sha1.hexdigest()\n l = min(len(sha1_file), len(sha1_hash))\n return sha1.hexdigest()[0:l] == sha1_hash[0:l]\n\n\ndef download(url, path=None, overwrite=False, sha1_hash=None):\n \"\"\"Download an given URL\n Parameters\n ----------\n url : str\n URL to download\n path : str, optional\n Destination path to store downloaded file. By default stores to the\n current directory with same name as in url.\n overwrite : bool, optional\n Whether to overwrite destination file if already exists.\n sha1_hash : str, optional\n Expected sha1 hash in hexadecimal digits. Will ignore existing file when hash is specified\n but doesn't match.\n Returns\n -------\n str\n The file path of the downloaded file.\n \"\"\"\n if path is None:\n fname = url.split(\"/\")[-1]\n else:\n path = os.path.expanduser(path)\n if os.path.isdir(path):\n fname = os.path.join(path, url.split(\"/\")[-1])\n else:\n fname = path\n\n if (\n overwrite\n or not os.path.exists(fname)\n or (sha1_hash and not check_sha1(fname, sha1_hash))\n ):\n dirname = os.path.dirname(os.path.abspath(os.path.expanduser(fname)))\n if not os.path.exists(dirname):\n os.makedirs(dirname)\n\n print(\"Downloading %s from %s...\" % (fname, url))\n r = requests.get(url, stream=True)\n if r.status_code != 200:\n raise RuntimeError(\"Failed downloading url %s\" % url)\n total_length = r.headers.get(\"content-length\")\n with open(fname, \"wb\") as f:\n if total_length is None: # no content length header\n for chunk in r.iter_content(chunk_size=1024):\n if chunk: # filter out keep-alive new chunks\n f.write(chunk)\n else:\n total_length = int(total_length)\n for chunk in tqdm(\n r.iter_content(chunk_size=1024),\n total=int(total_length / 1024.0 + 0.5),\n unit=\"KB\",\n unit_scale=False,\n dynamic_ncols=True,\n ):\n f.write(chunk)\n\n if sha1_hash and not check_sha1(fname, sha1_hash):\n raise UserWarning(\n \"File {} is downloaded but the content hash does not match. \"\n \"The repo may be outdated or download may be incomplete. \"\n 'If the \"repo_url\" is overridden, consider switching to '\n \"the default repo.\".format(fname)\n )\n\n return fname\n" }, { "path": "python/ncnn/utils/functional.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\n\n\ndef xywh2xyxy(x):\n # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right\n y = np.zeros_like(x)\n y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x\n y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y\n y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x\n y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y\n return y\n\n\ndef xyxy2xywh(x):\n # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right\n y = np.zeros_like(x)\n y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center\n y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center\n y[:, 2] = x[:, 2] - x[:, 0] # width\n y[:, 3] = x[:, 3] - x[:, 1] # height\n return y\n\n\ndef make_grid(nx=20, ny=20):\n xv1, yv1 = np.meshgrid(np.arange(nx), np.arange(ny))\n z1 = np.stack((xv1, yv1), 2).reshape((1, ny, nx, 2)).astype(np.float32)\n return z1\n\n\ndef sigmoid(x):\n return 1 / (1 + np.exp(-x))\n\n\ndef softmax(x):\n max_value = np.max(x, axis=-1)\n x -= max_value.reshape((x.shape[0], 1))\n x = np.exp(x)\n sum_value = np.sum(x, axis=-1)\n x /= sum_value.reshape((x.shape[0], 1))\n return x\n\n\ndef iou_of(boxes0, boxes1, eps=1e-5):\n \"\"\"Return intersection-over-union (Jaccard index) of boxes.\n\n Args:\n boxes0 (N, 4): ground truth boxes.\n boxes1 (N or 1, 4): predicted boxes.\n eps: a small number to avoid 0 as denominator.\n Returns:\n iou (N): IoU values.\n \"\"\"\n overlap_left_top = np.maximum(boxes0[..., :2], boxes1[..., :2])\n overlap_right_bottom = np.minimum(boxes0[..., 2:], boxes1[..., 2:])\n\n overlap_area = area_of(overlap_left_top, overlap_right_bottom)\n area0 = area_of(boxes0[..., :2], boxes0[..., 2:])\n area1 = area_of(boxes1[..., :2], boxes1[..., 2:])\n return overlap_area / (area0 + area1 - overlap_area + eps)\n\n\ndef area_of(left_top, right_bottom):\n \"\"\"Compute the areas of rectangles given two corners.\n\n Args:\n left_top (N, 2): left top corner.\n right_bottom (N, 2): right bottom corner.\n\n Returns:\n area (N): return the area.\n \"\"\"\n hw = np.clip(right_bottom - left_top, 0.0, None)\n return hw[..., 0] * hw[..., 1]\n\n\ndef nms(boxes, scores, iou_threshold, top_k=-1, candidate_size=200):\n \"\"\"\n\n Args:\n box_scores (N, 5): boxes in corner-form(x1, y1, x2, y2) and probabilities.\n iou_threshold: intersection over union threshold.\n top_k: keep top_k results. If k <= 0, keep all the results.\n candidate_size: only consider the candidates with the highest scores.\n Returns:\n picked: a list of indexes of the kept boxes\n \"\"\"\n\n picked = []\n indexes = np.argsort(scores)\n indexes = indexes[-candidate_size:]\n while len(indexes) > 0:\n current = indexes[-1]\n picked.append(current)\n if 0 < top_k == len(picked) or len(indexes) == 1:\n break\n\n current_box = boxes[current, :]\n indexes = indexes[:-1]\n rest_boxes = boxes[indexes, :]\n iou = iou_of(\n rest_boxes,\n np.expand_dims(current_box, axis=0),\n )\n indexes = indexes[iou <= iou_threshold]\n\n return picked\n" }, { "path": "python/ncnn/utils/objects.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\n\n\nclass Point(object):\n def __init__(self):\n self.x = 0.0\n self.y = 0.0\n\n\nclass Rect(object):\n def __init__(self, x=0, y=0, w=0, h=0):\n self.x = x\n self.y = y\n self.w = w\n self.h = h\n\n def area(self):\n return self.w * self.h\n\n def intersection_area(self, b):\n x1 = np.maximum(self.x, b.x)\n y1 = np.maximum(self.y, b.y)\n x2 = np.minimum(self.x + self.w, b.x + b.w)\n y2 = np.minimum(self.y + self.h, b.y + b.h)\n return np.abs(x1 - x2) * np.abs(y1 - y2)\n\n\nclass Detect_Object(object):\n def __init__(self, label=0, prob=0, x=0, y=0, w=0, h=0):\n self.label = label\n self.prob = prob\n self.rect = Rect(x, y, w, h)\n\n\nclass Face_Object(object):\n def __init__(self):\n self.prob = 0.0\n self.rect = Rect()\n self.landmark = []\n\n\nclass KeyPoint(object):\n def __init__(self):\n self.p = Point()\n self.prob = 0.0\n" }, { "path": "python/ncnn/utils/visual.py", "content": "# Copyright 2020 Tencent\n# SPDX-License-Identifier: BSD-3-Clause\n\nimport numpy as np\nimport cv2\n\n\ndef draw_detection_objects(image, class_names, objects, min_prob=0.0):\n for obj in objects:\n if obj.prob < min_prob:\n continue\n\n print(\n \"%d = %.5f at %.2f %.2f %.2f x %.2f\\n\"\n % (obj.label, obj.prob, obj.rect.x, obj.rect.y, obj.rect.w, obj.rect.h)\n )\n\n cv2.rectangle(\n image,\n (int(obj.rect.x), int(obj.rect.y)),\n (int(obj.rect.x + obj.rect.w), int(obj.rect.y + obj.rect.h)),\n (255, 0, 0),\n )\n\n text = \"%s %.1f%%\" % (class_names[int(obj.label)], obj.prob * 100)\n\n label_size, baseLine = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)\n\n x = obj.rect.x\n y = obj.rect.y - label_size[1] - baseLine\n if y < 0:\n y = 0\n if x + label_size[0] > image.shape[1]:\n x = image.shape[1] - label_size[0]\n\n cv2.rectangle(\n image,\n (int(x), int(y)),\n (int(x + label_size[0]), int(y + label_size[1] + baseLine)),\n (255, 255, 255),\n -1,\n )\n\n cv2.putText(\n image,\n text,\n (int(x), int(y + label_size[1])),\n cv2.FONT_HERSHEY_SIMPLEX,\n 0.5,\n (0, 0, 0),\n )\n\n cv2.imshow(\"image\", image)\n cv2.waitKey(0)\n\n\ndef print_topk(cls_scores, topk):\n indexes = np.argsort(cls_scores)[::-1][0:topk]\n scores = cls_scores[indexes]\n\n for index, score in zip(indexes, scores):\n print(\"%d=%f\" % (index, score))\n\n\ndef draw_faceobjects(image, faceobjects):\n for obj in faceobjects:\n print(\n \"%.5f at %.2f %.2f %.2f x %.2f\"\n % (obj.prob, obj.rect.x, obj.rect.y, obj.rect.w, obj.rect.h)\n )\n\n cv2.rectangle(\n image,\n (int(obj.rect.x), int(obj.rect.y)),\n (int(obj.rect.x + obj.rect.w), int(obj.rect.y + obj.rect.h)),\n (255, 0, 0),\n )\n\n cv2.circle(\n image,\n (int(obj.landmark[0].x), int(obj.landmark[0].y)),\n 2,\n (0, 255, 255),\n -1,\n )\n cv2.circle(\n image,\n (int(obj.landmark[1].x), int(obj.landmark[1].y)),\n 2,\n (0, 255, 255),\n -1,\n )\n cv2.circle(\n image,\n (int(obj.landmark[2].x), int(obj.landmark[2].y)),\n 2,\n (0, 255, 255),\n -1,\n )\n cv2.circle(\n image,\n (int(obj.landmark[3].x), int(obj.landmark[3].y)),\n 2,\n (0, 255, 255),\n -1,\n )\n cv2.circle(\n image,\n (int(obj.landmark[4].x), int(obj.landmark[4].y)),\n 2,\n (0, 255, 255),\n -1,\n )\n\n text = \"%.1f%%\" % (obj.prob * 100)\n\n label_size, baseLine = cv2.getTextSize(text, cv2.FONT_HERSHEY_SIMPLEX, 0.5, 1)\n\n x = obj.rect.x\n y = obj.rect.y - label_size[1] - baseLine\n if y < 0:\n y = 0\n if x + label_size[0] > image.shape[1]:\n x = image.shape[1] - label_size[0]\n\n cv2.rectangle(\n image,\n (int(x), int(y)),\n (int(x + label_size[0]), int(y + label_size[1] + baseLine)),\n (255, 255, 255),\n -1,\n )\n\n cv2.putText(\n image,\n text,\n (int(x), int(y + label_size[1])),\n cv2.FONT_HERSHEY_SIMPLEX,\n 0.5,\n (0, 0, 0),\n )\n\n cv2.imshow(\"image\", image)\n cv2.waitKey(0)\n\n\ndef draw_pose(image, keypoints):\n # draw bone\n joint_pairs = [\n (0, 1),\n (1, 3),\n (0, 2),\n (2, 4),\n (5, 6),\n (5, 7),\n (7, 9),\n (6, 8),\n (8, 10),\n (5, 11),\n (6, 12),\n (11, 12),\n (11, 13),\n (12, 14),\n (13, 15),\n (14, 16),\n ]\n\n for i in range(16):\n p1 = keypoints[joint_pairs[i][0]]\n p2 = keypoints[joint_pairs[i][1]]\n\n if p1.prob < 0.2 or p2.prob < 0.2:\n continue\n\n cv2.line(\n image,\n (int(p1.p.x), int(p1.p.y)),\n (int(p2.p.x), int(p2.p.y)),\n (255, 0, 0),\n 2,\n )\n\n # draw joint\n for keypoint in keypoints:\n print(\"%.2f %.2f = %.5f\" % (keypoint.p.x, keypoint.p.y, keypoint.prob))\n\n if keypoint.prob < 0.2:\n continue\n\n cv2.circle(image, (int(keypoint.p.x), int(keypoint.p.y)), 3, (0, 255, 0), -1)\n\n cv2.imshow(\"image\", image)\n cv2.waitKey(0)\n" }, { "path": "python/requirements.txt", "content": "numpy\ntqdm\nrequests\nportalocker\nopencv-python" }, { "path": "python/setup.py.i", "content": "import sys\nfrom setuptools import setup, find_packages\n\ntry:\n from wheel.bdist_wheel import bdist_wheel as _bdist_wheel\n\n class bdist_wheel(_bdist_wheel):\n def finalize_options(self):\n _bdist_wheel.finalize_options(self)\n self.root_is_pure = False\n\n\nexcept ImportError:\n bdist_wheel = None\n\nif sys.version_info < (3, 0):\n sys.exit(\"Sorry, Python < 3.0 is not supported\")\n\nrequirements = [\"numpy\", \"tqdm\", \"requests\", \"portalocker\", \"opencv-python\"]\n\nsetup(\n name=\"ncnn\",\n version=\"${PACKAGE_VERSION}\",\n author=\"nihui\",\n author_email=\"nihuini@tencent.com\",\n maintainer=\"caishanli\",\n maintainer_email=\"caishanli25@gmail.com\",\n description=\"ncnn is a high-performance neural network inference framework optimized for the mobile platform\",\n url=\"https://github.com/Tencent/ncnn\",\n classifiers=[\n \"Programming Language :: C++\",\n \"Programming Language :: Python :: 3\",\n \"Programming Language :: Python :: 3.6\",\n \"Programming Language :: Python :: 3.7\",\n \"Programming Language :: Python :: 3.8\",\n \"Programming Language :: Python :: 3.9\",\n \"Programming Language :: Python :: 3.10\",\n \"Programming Language :: Python :: 3.11\",\n \"Programming Language :: Python :: 3.12\",\n \"Programming Language :: Python :: 3.13\",\n \"Programming Language :: Python :: 3.14\",\n \"License :: OSI Approved :: BSD License\",\n \"Operating System :: OS Independent\",\n \"Topic :: Scientific/Engineering :: Artificial Intelligence\",\n ],\n license=\"BSD-3\",\n python_requires=\">=3.5\",\n packages=find_packages(),\n package_dir={\"\": \".\"},\n package_data={\"ncnn\": [\"ncnn${PYTHON_MODULE_PREFIX}${PYTHON_MODULE_EXTENSION}\"]},\n install_requires=requirements,\n cmdclass={\"bdist_wheel\": bdist_wheel},\n)\n" }, { "path": "python/src/main.cpp", "content": "// Copyright 2020 Tencent\n// SPDX-License-Identifier: BSD-3-Clause\n\n#include \n#include \n#include \n#include \n\n#include \n#include \n#include \n#include \n#include \n#include \n\n#include \"pybind11_mat.h\"\n#include \"pybind11_datareader.h\"\n#include \"pybind11_allocator.h\"\n#include \"pybind11_modelbin.h\"\n#include \"pybind11_layer.h\"\nusing namespace ncnn;\n\nnamespace py = pybind11;\n\nclass DataReaderFromMemoryCopy : public DataReaderFromMemory\n{\npublic:\n explicit DataReaderFromMemoryCopy(const unsigned char*& mem)\n : DataReaderFromMemory(mem)\n {\n }\n\n virtual size_t reference(size_t size, const void** buf) const\n {\n return 0;\n }\n};\n\nstruct LayerFactory\n{\n std::string name;\n int index;\n std::function creator;\n std::function destroyer;\n layer_creator_func creator_func;\n layer_destroyer_func destroyer_func;\n};\n\n#define LayerFactoryDeclear(n) \\\n static ncnn::Layer* LayerCreator##n(void*); \\\n static void LayerDestroyer##n(ncnn::Layer*, void*);\n\nLayerFactoryDeclear(0);\nLayerFactoryDeclear(1);\nLayerFactoryDeclear(2);\nLayerFactoryDeclear(3);\nLayerFactoryDeclear(4);\nLayerFactoryDeclear(5);\nLayerFactoryDeclear(6);\nLayerFactoryDeclear(7);\nLayerFactoryDeclear(8);\nLayerFactoryDeclear(9);\n\nstd::vector g_layer_factroys = {\n {\"\", -1, nullptr, nullptr, LayerCreator0, LayerDestroyer0},\n {\"\", -1, nullptr, nullptr, LayerCreator1, LayerDestroyer1},\n {\"\", -1, nullptr, nullptr, LayerCreator2, LayerDestroyer2},\n {\"\", -1, nullptr, nullptr, LayerCreator3, LayerDestroyer3},\n {\"\", -1, nullptr, nullptr, LayerCreator4, LayerDestroyer4},\n {\"\", -1, nullptr, nullptr, LayerCreator5, LayerDestroyer5},\n {\"\", -1, nullptr, nullptr, LayerCreator6, LayerDestroyer6},\n {\"\", -1, nullptr, nullptr, LayerCreator7, LayerDestroyer7},\n {\"\", -1, nullptr, nullptr, LayerCreator8, LayerDestroyer8},\n {\"\", -1, nullptr, nullptr, LayerCreator9, LayerDestroyer9},\n};\nint g_layer_factroy_index = 0;\n\n#define LayerFactoryDefine(n) \\\n static ncnn::Layer* LayerCreator##n(void* p) \\\n { \\\n if (g_layer_factroys[n].creator != nullptr) \\\n { \\\n return g_layer_factroys[n].creator(); \\\n } \\\n return nullptr; \\\n } \\\n static void LayerDestroyer##n(ncnn::Layer* layer, void* p) \\\n { \\\n if (g_layer_factroys[n].destroyer) \\\n { \\\n g_layer_factroys[n].destroyer(layer); \\\n } \\\n }\n\nLayerFactoryDefine(0);\nLayerFactoryDefine(1);\nLayerFactoryDefine(2);\nLayerFactoryDefine(3);\nLayerFactoryDefine(4);\nLayerFactoryDefine(5);\nLayerFactoryDefine(6);\nLayerFactoryDefine(7);\nLayerFactoryDefine(8);\nLayerFactoryDefine(9);\n\nPYBIND11_MODULE(ncnn, m)\n{\n auto atexit = py::module_::import(\"atexit\");\n atexit.attr(\"register\")(py::cpp_function([]() {\n for (int i = 0; i < g_layer_factroys.size(); i++)\n {\n g_layer_factroys[i].creator = nullptr;\n g_layer_factroys[i].destroyer = nullptr;\n }\n }));\n\n py::class_ >(m, \"Allocator\");\n py::class_ >(m, \"PoolAllocator\")\n .def(py::init<>())\n .def(\"set_size_compare_ratio\", &PoolAllocator::set_size_compare_ratio, py::arg(\"src\"))\n .def(\"clear\", &PoolAllocator::clear)\n .def(\"fastMalloc\", &PoolAllocator::fastMalloc, py::arg(\"size\"))\n .def(\"fastFree\", &PoolAllocator::fastFree, py::arg(\"ptr\"));\n py::class_ >(m, \"UnlockedPoolAllocator\")\n .def(py::init<>())\n .def(\"set_size_compare_ratio\", &UnlockedPoolAllocator::set_size_compare_ratio, py::arg(\"src\"))\n .def(\"clear\", &UnlockedPoolAllocator::clear)\n .def(\"fastMalloc\", &UnlockedPoolAllocator::fastMalloc, py::arg(\"size\"))\n .def(\"fastFree\", &UnlockedPoolAllocator::fastFree, py::arg(\"ptr\"));\n\n py::class_ >(m, \"DataReader\")\n .def(py::init<>())\n#if NCNN_STRING\n .def(\"scan\", &DataReader::scan, py::arg(\"format\"), py::arg(\"p\"))\n#endif // NCNN_STRING\n .def(\"read\", &DataReader::read, py::arg(\"buf\"), py::arg(\"size\"));\n py::class_ >(m, \"DataReaderFromEmpty\")\n .def(py::init<>())\n#if NCNN_STRING\n .def(\"scan\", &DataReaderFromEmpty::scan, py::arg(\"format\"), py::arg(\"p\"))\n#endif // NCNN_STRING\n .def(\"read\", &DataReaderFromEmpty::read, py::arg(\"buf\"), py::arg(\"size\"));\n\n py::class_(m, \"Blob\")\n .def(py::init<>())\n#if NCNN_STRING\n .def_readwrite(\"name\", &Blob::name)\n#endif // NCNN_STRING\n .def_readwrite(\"producer\", &Blob::producer)\n .def_readwrite(\"consumer\", &Blob::consumer)\n .def_readwrite(\"shape\", &Blob::shape);\n\n py::class_ >(m, \"ModelBin\")\n .def(py::init<>())\n .def(\"load\", (Mat(ModelBin::*)(int, int) const) & ModelBin::load, py::arg(\"w\"), py::arg(\"type\"))\n .def(\"load\", (Mat(ModelBin::*)(int, int, int) const) & ModelBin::load, py::arg(\"w\"), py::arg(\"h\"), py::arg(\"type\"))\n .def(\"load\", (Mat(ModelBin::*)(int, int, int, int) const) & ModelBin::load, py::arg(\"w\"), py::arg(\"h\"), py::arg(\"c\"), py::arg(\"type\"))\n .def(\"load\", (Mat(ModelBin::*)(int, int, int, int, int) const) & ModelBin::load, py::arg(\"w\"), py::arg(\"h\"), py::arg(\"d\"), py::arg(\"c\"), py::arg(\"type\"));\n py::class_ >(m, \"ModelBinFromDataReader\")\n .def(py::init(), py::arg(\"dr\"))\n .def(\"load\", &ModelBinFromDataReader::load, py::arg(\"w\"), py::arg(\"type\"));\n py::class_ >(m, \"ModelBinFromMatArray\")\n .def(py::init(), py::arg(\"weights\"))\n .def(\"load\", &ModelBinFromMatArray::load, py::arg(\"w\"), py::arg(\"type\"));\n\n py::class_(m, \"ParamDict\")\n .def(py::init<>())\n .def(\"type\", &ParamDict::type, py::arg(\"id\"))\n .def(\"get\", (int (ParamDict::*)(int, int) const) & ParamDict::get, py::arg(\"id\"), py::arg(\"def\"))\n .def(\"get\", (float (ParamDict::*)(int, float) const) & ParamDict::get, py::arg(\"id\"), py::arg(\"def\"))\n .def(\"get\", (Mat(ParamDict::*)(int, const Mat&) const) & ParamDict::get, py::arg(\"id\"), py::arg(\"def\"))\n .def(\"set\", (void (ParamDict::*)(int, int)) & ParamDict::set, py::arg(\"id\"), py::arg(\"i\"))\n .def(\"set\", (void (ParamDict::*)(int, float)) & ParamDict::set, py::arg(\"id\"), py::arg(\"f\"))\n .def(\"set\", (void (ParamDict::*)(int, const Mat&)) & ParamDict::set, py::arg(\"id\"), py::arg(\"v\"));\n\n py::class_